Version:  2.0.40 2.2.26 2.4.37 3.9 3.10 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

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 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         sa->util_avg = scale_load_down(SCHED_LOAD_SCALE);
686         sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
687         /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
688 }
689 
690 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq);
691 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq);
692 #else
693 void init_entity_runnable_average(struct sched_entity *se)
694 {
695 }
696 #endif
697 
698 /*
699  * Update the current task's runtime statistics.
700  */
701 static void update_curr(struct cfs_rq *cfs_rq)
702 {
703         struct sched_entity *curr = cfs_rq->curr;
704         u64 now = rq_clock_task(rq_of(cfs_rq));
705         u64 delta_exec;
706 
707         if (unlikely(!curr))
708                 return;
709 
710         delta_exec = now - curr->exec_start;
711         if (unlikely((s64)delta_exec <= 0))
712                 return;
713 
714         curr->exec_start = now;
715 
716         schedstat_set(curr->statistics.exec_max,
717                       max(delta_exec, curr->statistics.exec_max));
718 
719         curr->sum_exec_runtime += delta_exec;
720         schedstat_add(cfs_rq, exec_clock, delta_exec);
721 
722         curr->vruntime += calc_delta_fair(delta_exec, curr);
723         update_min_vruntime(cfs_rq);
724 
725         if (entity_is_task(curr)) {
726                 struct task_struct *curtask = task_of(curr);
727 
728                 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
729                 cpuacct_charge(curtask, delta_exec);
730                 account_group_exec_runtime(curtask, delta_exec);
731         }
732 
733         account_cfs_rq_runtime(cfs_rq, delta_exec);
734 }
735 
736 static void update_curr_fair(struct rq *rq)
737 {
738         update_curr(cfs_rq_of(&rq->curr->se));
739 }
740 
741 #ifdef CONFIG_SCHEDSTATS
742 static inline void
743 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
744 {
745         u64 wait_start = rq_clock(rq_of(cfs_rq));
746 
747         if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
748             likely(wait_start > se->statistics.wait_start))
749                 wait_start -= se->statistics.wait_start;
750 
751         se->statistics.wait_start = wait_start;
752 }
753 
754 static void
755 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
756 {
757         struct task_struct *p;
758         u64 delta;
759 
760         delta = rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start;
761 
762         if (entity_is_task(se)) {
763                 p = task_of(se);
764                 if (task_on_rq_migrating(p)) {
765                         /*
766                          * Preserve migrating task's wait time so wait_start
767                          * time stamp can be adjusted to accumulate wait time
768                          * prior to migration.
769                          */
770                         se->statistics.wait_start = delta;
771                         return;
772                 }
773                 trace_sched_stat_wait(p, delta);
774         }
775 
776         se->statistics.wait_max = max(se->statistics.wait_max, delta);
777         se->statistics.wait_count++;
778         se->statistics.wait_sum += delta;
779         se->statistics.wait_start = 0;
780 }
781 
782 /*
783  * Task is being enqueued - update stats:
784  */
785 static inline void
786 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
787 {
788         /*
789          * Are we enqueueing a waiting task? (for current tasks
790          * a dequeue/enqueue event is a NOP)
791          */
792         if (se != cfs_rq->curr)
793                 update_stats_wait_start(cfs_rq, se);
794 }
795 
796 static inline void
797 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
798 {
799         /*
800          * Mark the end of the wait period if dequeueing a
801          * waiting task:
802          */
803         if (se != cfs_rq->curr)
804                 update_stats_wait_end(cfs_rq, se);
805 
806         if (flags & DEQUEUE_SLEEP) {
807                 if (entity_is_task(se)) {
808                         struct task_struct *tsk = task_of(se);
809 
810                         if (tsk->state & TASK_INTERRUPTIBLE)
811                                 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
812                         if (tsk->state & TASK_UNINTERRUPTIBLE)
813                                 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
814                 }
815         }
816 
817 }
818 #else
819 static inline void
820 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
821 {
822 }
823 
824 static inline void
825 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
826 {
827 }
828 
829 static inline void
830 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
831 {
832 }
833 
834 static inline void
835 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
836 {
837 }
838 #endif
839 
840 /*
841  * We are picking a new current task - update its stats:
842  */
843 static inline void
844 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
845 {
846         /*
847          * We are starting a new run period:
848          */
849         se->exec_start = rq_clock_task(rq_of(cfs_rq));
850 }
851 
852 /**************************************************
853  * Scheduling class queueing methods:
854  */
855 
856 #ifdef CONFIG_NUMA_BALANCING
857 /*
858  * Approximate time to scan a full NUMA task in ms. The task scan period is
859  * calculated based on the tasks virtual memory size and
860  * numa_balancing_scan_size.
861  */
862 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
863 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
864 
865 /* Portion of address space to scan in MB */
866 unsigned int sysctl_numa_balancing_scan_size = 256;
867 
868 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
869 unsigned int sysctl_numa_balancing_scan_delay = 1000;
870 
871 static unsigned int task_nr_scan_windows(struct task_struct *p)
872 {
873         unsigned long rss = 0;
874         unsigned long nr_scan_pages;
875 
876         /*
877          * Calculations based on RSS as non-present and empty pages are skipped
878          * by the PTE scanner and NUMA hinting faults should be trapped based
879          * on resident pages
880          */
881         nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
882         rss = get_mm_rss(p->mm);
883         if (!rss)
884                 rss = nr_scan_pages;
885 
886         rss = round_up(rss, nr_scan_pages);
887         return rss / nr_scan_pages;
888 }
889 
890 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
891 #define MAX_SCAN_WINDOW 2560
892 
893 static unsigned int task_scan_min(struct task_struct *p)
894 {
895         unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
896         unsigned int scan, floor;
897         unsigned int windows = 1;
898 
899         if (scan_size < MAX_SCAN_WINDOW)
900                 windows = MAX_SCAN_WINDOW / scan_size;
901         floor = 1000 / windows;
902 
903         scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
904         return max_t(unsigned int, floor, scan);
905 }
906 
907 static unsigned int task_scan_max(struct task_struct *p)
908 {
909         unsigned int smin = task_scan_min(p);
910         unsigned int smax;
911 
912         /* Watch for min being lower than max due to floor calculations */
913         smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
914         return max(smin, smax);
915 }
916 
917 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
918 {
919         rq->nr_numa_running += (p->numa_preferred_nid != -1);
920         rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
921 }
922 
923 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
924 {
925         rq->nr_numa_running -= (p->numa_preferred_nid != -1);
926         rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
927 }
928 
929 struct numa_group {
930         atomic_t refcount;
931 
932         spinlock_t lock; /* nr_tasks, tasks */
933         int nr_tasks;
934         pid_t gid;
935         int active_nodes;
936 
937         struct rcu_head rcu;
938         unsigned long total_faults;
939         unsigned long max_faults_cpu;
940         /*
941          * Faults_cpu is used to decide whether memory should move
942          * towards the CPU. As a consequence, these stats are weighted
943          * more by CPU use than by memory faults.
944          */
945         unsigned long *faults_cpu;
946         unsigned long faults[0];
947 };
948 
949 /* Shared or private faults. */
950 #define NR_NUMA_HINT_FAULT_TYPES 2
951 
952 /* Memory and CPU locality */
953 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
954 
955 /* Averaged statistics, and temporary buffers. */
956 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
957 
958 pid_t task_numa_group_id(struct task_struct *p)
959 {
960         return p->numa_group ? p->numa_group->gid : 0;
961 }
962 
963 /*
964  * The averaged statistics, shared & private, memory & cpu,
965  * occupy the first half of the array. The second half of the
966  * array is for current counters, which are averaged into the
967  * first set by task_numa_placement.
968  */
969 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
970 {
971         return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
972 }
973 
974 static inline unsigned long task_faults(struct task_struct *p, int nid)
975 {
976         if (!p->numa_faults)
977                 return 0;
978 
979         return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
980                 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
981 }
982 
983 static inline unsigned long group_faults(struct task_struct *p, int nid)
984 {
985         if (!p->numa_group)
986                 return 0;
987 
988         return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
989                 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
990 }
991 
992 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
993 {
994         return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
995                 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
996 }
997 
998 /*
999  * A node triggering more than 1/3 as many NUMA faults as the maximum is
1000  * considered part of a numa group's pseudo-interleaving set. Migrations
1001  * between these nodes are slowed down, to allow things to settle down.
1002  */
1003 #define ACTIVE_NODE_FRACTION 3
1004 
1005 static bool numa_is_active_node(int nid, struct numa_group *ng)
1006 {
1007         return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1008 }
1009 
1010 /* Handle placement on systems where not all nodes are directly connected. */
1011 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1012                                         int maxdist, bool task)
1013 {
1014         unsigned long score = 0;
1015         int node;
1016 
1017         /*
1018          * All nodes are directly connected, and the same distance
1019          * from each other. No need for fancy placement algorithms.
1020          */
1021         if (sched_numa_topology_type == NUMA_DIRECT)
1022                 return 0;
1023 
1024         /*
1025          * This code is called for each node, introducing N^2 complexity,
1026          * which should be ok given the number of nodes rarely exceeds 8.
1027          */
1028         for_each_online_node(node) {
1029                 unsigned long faults;
1030                 int dist = node_distance(nid, node);
1031 
1032                 /*
1033                  * The furthest away nodes in the system are not interesting
1034                  * for placement; nid was already counted.
1035                  */
1036                 if (dist == sched_max_numa_distance || node == nid)
1037                         continue;
1038 
1039                 /*
1040                  * On systems with a backplane NUMA topology, compare groups
1041                  * of nodes, and move tasks towards the group with the most
1042                  * memory accesses. When comparing two nodes at distance
1043                  * "hoplimit", only nodes closer by than "hoplimit" are part
1044                  * of each group. Skip other nodes.
1045                  */
1046                 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1047                                         dist > maxdist)
1048                         continue;
1049 
1050                 /* Add up the faults from nearby nodes. */
1051                 if (task)
1052                         faults = task_faults(p, node);
1053                 else
1054                         faults = group_faults(p, node);
1055 
1056                 /*
1057                  * On systems with a glueless mesh NUMA topology, there are
1058                  * no fixed "groups of nodes". Instead, nodes that are not
1059                  * directly connected bounce traffic through intermediate
1060                  * nodes; a numa_group can occupy any set of nodes.
1061                  * The further away a node is, the less the faults count.
1062                  * This seems to result in good task placement.
1063                  */
1064                 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1065                         faults *= (sched_max_numa_distance - dist);
1066                         faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1067                 }
1068 
1069                 score += faults;
1070         }
1071 
1072         return score;
1073 }
1074 
1075 /*
1076  * These return the fraction of accesses done by a particular task, or
1077  * task group, on a particular numa node.  The group weight is given a
1078  * larger multiplier, in order to group tasks together that are almost
1079  * evenly spread out between numa nodes.
1080  */
1081 static inline unsigned long task_weight(struct task_struct *p, int nid,
1082                                         int dist)
1083 {
1084         unsigned long faults, total_faults;
1085 
1086         if (!p->numa_faults)
1087                 return 0;
1088 
1089         total_faults = p->total_numa_faults;
1090 
1091         if (!total_faults)
1092                 return 0;
1093 
1094         faults = task_faults(p, nid);
1095         faults += score_nearby_nodes(p, nid, dist, true);
1096 
1097         return 1000 * faults / total_faults;
1098 }
1099 
1100 static inline unsigned long group_weight(struct task_struct *p, int nid,
1101                                          int dist)
1102 {
1103         unsigned long faults, total_faults;
1104 
1105         if (!p->numa_group)
1106                 return 0;
1107 
1108         total_faults = p->numa_group->total_faults;
1109 
1110         if (!total_faults)
1111                 return 0;
1112 
1113         faults = group_faults(p, nid);
1114         faults += score_nearby_nodes(p, nid, dist, false);
1115 
1116         return 1000 * faults / total_faults;
1117 }
1118 
1119 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1120                                 int src_nid, int dst_cpu)
1121 {
1122         struct numa_group *ng = p->numa_group;
1123         int dst_nid = cpu_to_node(dst_cpu);
1124         int last_cpupid, this_cpupid;
1125 
1126         this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1127 
1128         /*
1129          * Multi-stage node selection is used in conjunction with a periodic
1130          * migration fault to build a temporal task<->page relation. By using
1131          * a two-stage filter we remove short/unlikely relations.
1132          *
1133          * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1134          * a task's usage of a particular page (n_p) per total usage of this
1135          * page (n_t) (in a given time-span) to a probability.
1136          *
1137          * Our periodic faults will sample this probability and getting the
1138          * same result twice in a row, given these samples are fully
1139          * independent, is then given by P(n)^2, provided our sample period
1140          * is sufficiently short compared to the usage pattern.
1141          *
1142          * This quadric squishes small probabilities, making it less likely we
1143          * act on an unlikely task<->page relation.
1144          */
1145         last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1146         if (!cpupid_pid_unset(last_cpupid) &&
1147                                 cpupid_to_nid(last_cpupid) != dst_nid)
1148                 return false;
1149 
1150         /* Always allow migrate on private faults */
1151         if (cpupid_match_pid(p, last_cpupid))
1152                 return true;
1153 
1154         /* A shared fault, but p->numa_group has not been set up yet. */
1155         if (!ng)
1156                 return true;
1157 
1158         /*
1159          * Destination node is much more heavily used than the source
1160          * node? Allow migration.
1161          */
1162         if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1163                                         ACTIVE_NODE_FRACTION)
1164                 return true;
1165 
1166         /*
1167          * Distribute memory according to CPU & memory use on each node,
1168          * with 3/4 hysteresis to avoid unnecessary memory migrations:
1169          *
1170          * faults_cpu(dst)   3   faults_cpu(src)
1171          * --------------- * - > ---------------
1172          * faults_mem(dst)   4   faults_mem(src)
1173          */
1174         return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1175                group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1176 }
1177 
1178 static unsigned long weighted_cpuload(const int cpu);
1179 static unsigned long source_load(int cpu, int type);
1180 static unsigned long target_load(int cpu, int type);
1181 static unsigned long capacity_of(int cpu);
1182 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1183 
1184 /* Cached statistics for all CPUs within a node */
1185 struct numa_stats {
1186         unsigned long nr_running;
1187         unsigned long load;
1188 
1189         /* Total compute capacity of CPUs on a node */
1190         unsigned long compute_capacity;
1191 
1192         /* Approximate capacity in terms of runnable tasks on a node */
1193         unsigned long task_capacity;
1194         int has_free_capacity;
1195 };
1196 
1197 /*
1198  * XXX borrowed from update_sg_lb_stats
1199  */
1200 static void update_numa_stats(struct numa_stats *ns, int nid)
1201 {
1202         int smt, cpu, cpus = 0;
1203         unsigned long capacity;
1204 
1205         memset(ns, 0, sizeof(*ns));
1206         for_each_cpu(cpu, cpumask_of_node(nid)) {
1207                 struct rq *rq = cpu_rq(cpu);
1208 
1209                 ns->nr_running += rq->nr_running;
1210                 ns->load += weighted_cpuload(cpu);
1211                 ns->compute_capacity += capacity_of(cpu);
1212 
1213                 cpus++;
1214         }
1215 
1216         /*
1217          * If we raced with hotplug and there are no CPUs left in our mask
1218          * the @ns structure is NULL'ed and task_numa_compare() will
1219          * not find this node attractive.
1220          *
1221          * We'll either bail at !has_free_capacity, or we'll detect a huge
1222          * imbalance and bail there.
1223          */
1224         if (!cpus)
1225                 return;
1226 
1227         /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1228         smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1229         capacity = cpus / smt; /* cores */
1230 
1231         ns->task_capacity = min_t(unsigned, capacity,
1232                 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1233         ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1234 }
1235 
1236 struct task_numa_env {
1237         struct task_struct *p;
1238 
1239         int src_cpu, src_nid;
1240         int dst_cpu, dst_nid;
1241 
1242         struct numa_stats src_stats, dst_stats;
1243 
1244         int imbalance_pct;
1245         int dist;
1246 
1247         struct task_struct *best_task;
1248         long best_imp;
1249         int best_cpu;
1250 };
1251 
1252 static void task_numa_assign(struct task_numa_env *env,
1253                              struct task_struct *p, long imp)
1254 {
1255         if (env->best_task)
1256                 put_task_struct(env->best_task);
1257 
1258         env->best_task = p;
1259         env->best_imp = imp;
1260         env->best_cpu = env->dst_cpu;
1261 }
1262 
1263 static bool load_too_imbalanced(long src_load, long dst_load,
1264                                 struct task_numa_env *env)
1265 {
1266         long imb, old_imb;
1267         long orig_src_load, orig_dst_load;
1268         long src_capacity, dst_capacity;
1269 
1270         /*
1271          * The load is corrected for the CPU capacity available on each node.
1272          *
1273          * src_load        dst_load
1274          * ------------ vs ---------
1275          * src_capacity    dst_capacity
1276          */
1277         src_capacity = env->src_stats.compute_capacity;
1278         dst_capacity = env->dst_stats.compute_capacity;
1279 
1280         /* We care about the slope of the imbalance, not the direction. */
1281         if (dst_load < src_load)
1282                 swap(dst_load, src_load);
1283 
1284         /* Is the difference below the threshold? */
1285         imb = dst_load * src_capacity * 100 -
1286               src_load * dst_capacity * env->imbalance_pct;
1287         if (imb <= 0)
1288                 return false;
1289 
1290         /*
1291          * The imbalance is above the allowed threshold.
1292          * Compare it with the old imbalance.
1293          */
1294         orig_src_load = env->src_stats.load;
1295         orig_dst_load = env->dst_stats.load;
1296 
1297         if (orig_dst_load < orig_src_load)
1298                 swap(orig_dst_load, orig_src_load);
1299 
1300         old_imb = orig_dst_load * src_capacity * 100 -
1301                   orig_src_load * dst_capacity * env->imbalance_pct;
1302 
1303         /* Would this change make things worse? */
1304         return (imb > old_imb);
1305 }
1306 
1307 /*
1308  * This checks if the overall compute and NUMA accesses of the system would
1309  * be improved if the source tasks was migrated to the target dst_cpu taking
1310  * into account that it might be best if task running on the dst_cpu should
1311  * be exchanged with the source task
1312  */
1313 static void task_numa_compare(struct task_numa_env *env,
1314                               long taskimp, long groupimp)
1315 {
1316         struct rq *src_rq = cpu_rq(env->src_cpu);
1317         struct rq *dst_rq = cpu_rq(env->dst_cpu);
1318         struct task_struct *cur;
1319         long src_load, dst_load;
1320         long load;
1321         long imp = env->p->numa_group ? groupimp : taskimp;
1322         long moveimp = imp;
1323         int dist = env->dist;
1324         bool assigned = false;
1325 
1326         rcu_read_lock();
1327 
1328         raw_spin_lock_irq(&dst_rq->lock);
1329         cur = dst_rq->curr;
1330         /*
1331          * No need to move the exiting task or idle task.
1332          */
1333         if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1334                 cur = NULL;
1335         else {
1336                 /*
1337                  * The task_struct must be protected here to protect the
1338                  * p->numa_faults access in the task_weight since the
1339                  * numa_faults could already be freed in the following path:
1340                  * finish_task_switch()
1341                  *     --> put_task_struct()
1342                  *         --> __put_task_struct()
1343                  *             --> task_numa_free()
1344                  */
1345                 get_task_struct(cur);
1346         }
1347 
1348         raw_spin_unlock_irq(&dst_rq->lock);
1349 
1350         /*
1351          * Because we have preemption enabled we can get migrated around and
1352          * end try selecting ourselves (current == env->p) as a swap candidate.
1353          */
1354         if (cur == env->p)
1355                 goto unlock;
1356 
1357         /*
1358          * "imp" is the fault differential for the source task between the
1359          * source and destination node. Calculate the total differential for
1360          * the source task and potential destination task. The more negative
1361          * the value is, the more rmeote accesses that would be expected to
1362          * be incurred if the tasks were swapped.
1363          */
1364         if (cur) {
1365                 /* Skip this swap candidate if cannot move to the source cpu */
1366                 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1367                         goto unlock;
1368 
1369                 /*
1370                  * If dst and source tasks are in the same NUMA group, or not
1371                  * in any group then look only at task weights.
1372                  */
1373                 if (cur->numa_group == env->p->numa_group) {
1374                         imp = taskimp + task_weight(cur, env->src_nid, dist) -
1375                               task_weight(cur, env->dst_nid, dist);
1376                         /*
1377                          * Add some hysteresis to prevent swapping the
1378                          * tasks within a group over tiny differences.
1379                          */
1380                         if (cur->numa_group)
1381                                 imp -= imp/16;
1382                 } else {
1383                         /*
1384                          * Compare the group weights. If a task is all by
1385                          * itself (not part of a group), use the task weight
1386                          * instead.
1387                          */
1388                         if (cur->numa_group)
1389                                 imp += group_weight(cur, env->src_nid, dist) -
1390                                        group_weight(cur, env->dst_nid, dist);
1391                         else
1392                                 imp += task_weight(cur, env->src_nid, dist) -
1393                                        task_weight(cur, env->dst_nid, dist);
1394                 }
1395         }
1396 
1397         if (imp <= env->best_imp && moveimp <= env->best_imp)
1398                 goto unlock;
1399 
1400         if (!cur) {
1401                 /* Is there capacity at our destination? */
1402                 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1403                     !env->dst_stats.has_free_capacity)
1404                         goto unlock;
1405 
1406                 goto balance;
1407         }
1408 
1409         /* Balance doesn't matter much if we're running a task per cpu */
1410         if (imp > env->best_imp && src_rq->nr_running == 1 &&
1411                         dst_rq->nr_running == 1)
1412                 goto assign;
1413 
1414         /*
1415          * In the overloaded case, try and keep the load balanced.
1416          */
1417 balance:
1418         load = task_h_load(env->p);
1419         dst_load = env->dst_stats.load + load;
1420         src_load = env->src_stats.load - load;
1421 
1422         if (moveimp > imp && moveimp > env->best_imp) {
1423                 /*
1424                  * If the improvement from just moving env->p direction is
1425                  * better than swapping tasks around, check if a move is
1426                  * possible. Store a slightly smaller score than moveimp,
1427                  * so an actually idle CPU will win.
1428                  */
1429                 if (!load_too_imbalanced(src_load, dst_load, env)) {
1430                         imp = moveimp - 1;
1431                         put_task_struct(cur);
1432                         cur = NULL;
1433                         goto assign;
1434                 }
1435         }
1436 
1437         if (imp <= env->best_imp)
1438                 goto unlock;
1439 
1440         if (cur) {
1441                 load = task_h_load(cur);
1442                 dst_load -= load;
1443                 src_load += load;
1444         }
1445 
1446         if (load_too_imbalanced(src_load, dst_load, env))
1447                 goto unlock;
1448 
1449         /*
1450          * One idle CPU per node is evaluated for a task numa move.
1451          * Call select_idle_sibling to maybe find a better one.
1452          */
1453         if (!cur)
1454                 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1455 
1456 assign:
1457         assigned = true;
1458         task_numa_assign(env, cur, imp);
1459 unlock:
1460         rcu_read_unlock();
1461         /*
1462          * The dst_rq->curr isn't assigned. The protection for task_struct is
1463          * finished.
1464          */
1465         if (cur && !assigned)
1466                 put_task_struct(cur);
1467 }
1468 
1469 static void task_numa_find_cpu(struct task_numa_env *env,
1470                                 long taskimp, long groupimp)
1471 {
1472         int cpu;
1473 
1474         for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1475                 /* Skip this CPU if the source task cannot migrate */
1476                 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1477                         continue;
1478 
1479                 env->dst_cpu = cpu;
1480                 task_numa_compare(env, taskimp, groupimp);
1481         }
1482 }
1483 
1484 /* Only move tasks to a NUMA node less busy than the current node. */
1485 static bool numa_has_capacity(struct task_numa_env *env)
1486 {
1487         struct numa_stats *src = &env->src_stats;
1488         struct numa_stats *dst = &env->dst_stats;
1489 
1490         if (src->has_free_capacity && !dst->has_free_capacity)
1491                 return false;
1492 
1493         /*
1494          * Only consider a task move if the source has a higher load
1495          * than the destination, corrected for CPU capacity on each node.
1496          *
1497          *      src->load                dst->load
1498          * --------------------- vs ---------------------
1499          * src->compute_capacity    dst->compute_capacity
1500          */
1501         if (src->load * dst->compute_capacity * env->imbalance_pct >
1502 
1503             dst->load * src->compute_capacity * 100)
1504                 return true;
1505 
1506         return false;
1507 }
1508 
1509 static int task_numa_migrate(struct task_struct *p)
1510 {
1511         struct task_numa_env env = {
1512                 .p = p,
1513 
1514                 .src_cpu = task_cpu(p),
1515                 .src_nid = task_node(p),
1516 
1517                 .imbalance_pct = 112,
1518 
1519                 .best_task = NULL,
1520                 .best_imp = 0,
1521                 .best_cpu = -1,
1522         };
1523         struct sched_domain *sd;
1524         unsigned long taskweight, groupweight;
1525         int nid, ret, dist;
1526         long taskimp, groupimp;
1527 
1528         /*
1529          * Pick the lowest SD_NUMA domain, as that would have the smallest
1530          * imbalance and would be the first to start moving tasks about.
1531          *
1532          * And we want to avoid any moving of tasks about, as that would create
1533          * random movement of tasks -- counter the numa conditions we're trying
1534          * to satisfy here.
1535          */
1536         rcu_read_lock();
1537         sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1538         if (sd)
1539                 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1540         rcu_read_unlock();
1541 
1542         /*
1543          * Cpusets can break the scheduler domain tree into smaller
1544          * balance domains, some of which do not cross NUMA boundaries.
1545          * Tasks that are "trapped" in such domains cannot be migrated
1546          * elsewhere, so there is no point in (re)trying.
1547          */
1548         if (unlikely(!sd)) {
1549                 p->numa_preferred_nid = task_node(p);
1550                 return -EINVAL;
1551         }
1552 
1553         env.dst_nid = p->numa_preferred_nid;
1554         dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1555         taskweight = task_weight(p, env.src_nid, dist);
1556         groupweight = group_weight(p, env.src_nid, dist);
1557         update_numa_stats(&env.src_stats, env.src_nid);
1558         taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1559         groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1560         update_numa_stats(&env.dst_stats, env.dst_nid);
1561 
1562         /* Try to find a spot on the preferred nid. */
1563         if (numa_has_capacity(&env))
1564                 task_numa_find_cpu(&env, taskimp, groupimp);
1565 
1566         /*
1567          * Look at other nodes in these cases:
1568          * - there is no space available on the preferred_nid
1569          * - the task is part of a numa_group that is interleaved across
1570          *   multiple NUMA nodes; in order to better consolidate the group,
1571          *   we need to check other locations.
1572          */
1573         if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1574                 for_each_online_node(nid) {
1575                         if (nid == env.src_nid || nid == p->numa_preferred_nid)
1576                                 continue;
1577 
1578                         dist = node_distance(env.src_nid, env.dst_nid);
1579                         if (sched_numa_topology_type == NUMA_BACKPLANE &&
1580                                                 dist != env.dist) {
1581                                 taskweight = task_weight(p, env.src_nid, dist);
1582                                 groupweight = group_weight(p, env.src_nid, dist);
1583                         }
1584 
1585                         /* Only consider nodes where both task and groups benefit */
1586                         taskimp = task_weight(p, nid, dist) - taskweight;
1587                         groupimp = group_weight(p, nid, dist) - groupweight;
1588                         if (taskimp < 0 && groupimp < 0)
1589                                 continue;
1590 
1591                         env.dist = dist;
1592                         env.dst_nid = nid;
1593                         update_numa_stats(&env.dst_stats, env.dst_nid);
1594                         if (numa_has_capacity(&env))
1595                                 task_numa_find_cpu(&env, taskimp, groupimp);
1596                 }
1597         }
1598 
1599         /*
1600          * If the task is part of a workload that spans multiple NUMA nodes,
1601          * and is migrating into one of the workload's active nodes, remember
1602          * this node as the task's preferred numa node, so the workload can
1603          * settle down.
1604          * A task that migrated to a second choice node will be better off
1605          * trying for a better one later. Do not set the preferred node here.
1606          */
1607         if (p->numa_group) {
1608                 struct numa_group *ng = p->numa_group;
1609 
1610                 if (env.best_cpu == -1)
1611                         nid = env.src_nid;
1612                 else
1613                         nid = env.dst_nid;
1614 
1615                 if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng))
1616                         sched_setnuma(p, env.dst_nid);
1617         }
1618 
1619         /* No better CPU than the current one was found. */
1620         if (env.best_cpu == -1)
1621                 return -EAGAIN;
1622 
1623         /*
1624          * Reset the scan period if the task is being rescheduled on an
1625          * alternative node to recheck if the tasks is now properly placed.
1626          */
1627         p->numa_scan_period = task_scan_min(p);
1628 
1629         if (env.best_task == NULL) {
1630                 ret = migrate_task_to(p, env.best_cpu);
1631                 if (ret != 0)
1632                         trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1633                 return ret;
1634         }
1635 
1636         ret = migrate_swap(p, env.best_task);
1637         if (ret != 0)
1638                 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1639         put_task_struct(env.best_task);
1640         return ret;
1641 }
1642 
1643 /* Attempt to migrate a task to a CPU on the preferred node. */
1644 static void numa_migrate_preferred(struct task_struct *p)
1645 {
1646         unsigned long interval = HZ;
1647 
1648         /* This task has no NUMA fault statistics yet */
1649         if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1650                 return;
1651 
1652         /* Periodically retry migrating the task to the preferred node */
1653         interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1654         p->numa_migrate_retry = jiffies + interval;
1655 
1656         /* Success if task is already running on preferred CPU */
1657         if (task_node(p) == p->numa_preferred_nid)
1658                 return;
1659 
1660         /* Otherwise, try migrate to a CPU on the preferred node */
1661         task_numa_migrate(p);
1662 }
1663 
1664 /*
1665  * Find out how many nodes on the workload is actively running on. Do this by
1666  * tracking the nodes from which NUMA hinting faults are triggered. This can
1667  * be different from the set of nodes where the workload's memory is currently
1668  * located.
1669  */
1670 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1671 {
1672         unsigned long faults, max_faults = 0;
1673         int nid, active_nodes = 0;
1674 
1675         for_each_online_node(nid) {
1676                 faults = group_faults_cpu(numa_group, nid);
1677                 if (faults > max_faults)
1678                         max_faults = faults;
1679         }
1680 
1681         for_each_online_node(nid) {
1682                 faults = group_faults_cpu(numa_group, nid);
1683                 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1684                         active_nodes++;
1685         }
1686 
1687         numa_group->max_faults_cpu = max_faults;
1688         numa_group->active_nodes = active_nodes;
1689 }
1690 
1691 /*
1692  * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1693  * increments. The more local the fault statistics are, the higher the scan
1694  * period will be for the next scan window. If local/(local+remote) ratio is
1695  * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1696  * the scan period will decrease. Aim for 70% local accesses.
1697  */
1698 #define NUMA_PERIOD_SLOTS 10
1699 #define NUMA_PERIOD_THRESHOLD 7
1700 
1701 /*
1702  * Increase the scan period (slow down scanning) if the majority of
1703  * our memory is already on our local node, or if the majority of
1704  * the page accesses are shared with other processes.
1705  * Otherwise, decrease the scan period.
1706  */
1707 static void update_task_scan_period(struct task_struct *p,
1708                         unsigned long shared, unsigned long private)
1709 {
1710         unsigned int period_slot;
1711         int ratio;
1712         int diff;
1713 
1714         unsigned long remote = p->numa_faults_locality[0];
1715         unsigned long local = p->numa_faults_locality[1];
1716 
1717         /*
1718          * If there were no record hinting faults then either the task is
1719          * completely idle or all activity is areas that are not of interest
1720          * to automatic numa balancing. Related to that, if there were failed
1721          * migration then it implies we are migrating too quickly or the local
1722          * node is overloaded. In either case, scan slower
1723          */
1724         if (local + shared == 0 || p->numa_faults_locality[2]) {
1725                 p->numa_scan_period = min(p->numa_scan_period_max,
1726                         p->numa_scan_period << 1);
1727 
1728                 p->mm->numa_next_scan = jiffies +
1729                         msecs_to_jiffies(p->numa_scan_period);
1730 
1731                 return;
1732         }
1733 
1734         /*
1735          * Prepare to scale scan period relative to the current period.
1736          *       == NUMA_PERIOD_THRESHOLD scan period stays the same
1737          *       <  NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1738          *       >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1739          */
1740         period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1741         ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1742         if (ratio >= NUMA_PERIOD_THRESHOLD) {
1743                 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1744                 if (!slot)
1745                         slot = 1;
1746                 diff = slot * period_slot;
1747         } else {
1748                 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1749 
1750                 /*
1751                  * Scale scan rate increases based on sharing. There is an
1752                  * inverse relationship between the degree of sharing and
1753                  * the adjustment made to the scanning period. Broadly
1754                  * speaking the intent is that there is little point
1755                  * scanning faster if shared accesses dominate as it may
1756                  * simply bounce migrations uselessly
1757                  */
1758                 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1759                 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1760         }
1761 
1762         p->numa_scan_period = clamp(p->numa_scan_period + diff,
1763                         task_scan_min(p), task_scan_max(p));
1764         memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1765 }
1766 
1767 /*
1768  * Get the fraction of time the task has been running since the last
1769  * NUMA placement cycle. The scheduler keeps similar statistics, but
1770  * decays those on a 32ms period, which is orders of magnitude off
1771  * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1772  * stats only if the task is so new there are no NUMA statistics yet.
1773  */
1774 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1775 {
1776         u64 runtime, delta, now;
1777         /* Use the start of this time slice to avoid calculations. */
1778         now = p->se.exec_start;
1779         runtime = p->se.sum_exec_runtime;
1780 
1781         if (p->last_task_numa_placement) {
1782                 delta = runtime - p->last_sum_exec_runtime;
1783                 *period = now - p->last_task_numa_placement;
1784         } else {
1785                 delta = p->se.avg.load_sum / p->se.load.weight;
1786                 *period = LOAD_AVG_MAX;
1787         }
1788 
1789         p->last_sum_exec_runtime = runtime;
1790         p->last_task_numa_placement = now;
1791 
1792         return delta;
1793 }
1794 
1795 /*
1796  * Determine the preferred nid for a task in a numa_group. This needs to
1797  * be done in a way that produces consistent results with group_weight,
1798  * otherwise workloads might not converge.
1799  */
1800 static int preferred_group_nid(struct task_struct *p, int nid)
1801 {
1802         nodemask_t nodes;
1803         int dist;
1804 
1805         /* Direct connections between all NUMA nodes. */
1806         if (sched_numa_topology_type == NUMA_DIRECT)
1807                 return nid;
1808 
1809         /*
1810          * On a system with glueless mesh NUMA topology, group_weight
1811          * scores nodes according to the number of NUMA hinting faults on
1812          * both the node itself, and on nearby nodes.
1813          */
1814         if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1815                 unsigned long score, max_score = 0;
1816                 int node, max_node = nid;
1817 
1818                 dist = sched_max_numa_distance;
1819 
1820                 for_each_online_node(node) {
1821                         score = group_weight(p, node, dist);
1822                         if (score > max_score) {
1823                                 max_score = score;
1824                                 max_node = node;
1825                         }
1826                 }
1827                 return max_node;
1828         }
1829 
1830         /*
1831          * Finding the preferred nid in a system with NUMA backplane
1832          * interconnect topology is more involved. The goal is to locate
1833          * tasks from numa_groups near each other in the system, and
1834          * untangle workloads from different sides of the system. This requires
1835          * searching down the hierarchy of node groups, recursively searching
1836          * inside the highest scoring group of nodes. The nodemask tricks
1837          * keep the complexity of the search down.
1838          */
1839         nodes = node_online_map;
1840         for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1841                 unsigned long max_faults = 0;
1842                 nodemask_t max_group = NODE_MASK_NONE;
1843                 int a, b;
1844 
1845                 /* Are there nodes at this distance from each other? */
1846                 if (!find_numa_distance(dist))
1847                         continue;
1848 
1849                 for_each_node_mask(a, nodes) {
1850                         unsigned long faults = 0;
1851                         nodemask_t this_group;
1852                         nodes_clear(this_group);
1853 
1854                         /* Sum group's NUMA faults; includes a==b case. */
1855                         for_each_node_mask(b, nodes) {
1856                                 if (node_distance(a, b) < dist) {
1857                                         faults += group_faults(p, b);
1858                                         node_set(b, this_group);
1859                                         node_clear(b, nodes);
1860                                 }
1861                         }
1862 
1863                         /* Remember the top group. */
1864                         if (faults > max_faults) {
1865                                 max_faults = faults;
1866                                 max_group = this_group;
1867                                 /*
1868                                  * subtle: at the smallest distance there is
1869                                  * just one node left in each "group", the
1870                                  * winner is the preferred nid.
1871                                  */
1872                                 nid = a;
1873                         }
1874                 }
1875                 /* Next round, evaluate the nodes within max_group. */
1876                 if (!max_faults)
1877                         break;
1878                 nodes = max_group;
1879         }
1880         return nid;
1881 }
1882 
1883 static void task_numa_placement(struct task_struct *p)
1884 {
1885         int seq, nid, max_nid = -1, max_group_nid = -1;
1886         unsigned long max_faults = 0, max_group_faults = 0;
1887         unsigned long fault_types[2] = { 0, 0 };
1888         unsigned long total_faults;
1889         u64 runtime, period;
1890         spinlock_t *group_lock = NULL;
1891 
1892         /*
1893          * The p->mm->numa_scan_seq field gets updated without
1894          * exclusive access. Use READ_ONCE() here to ensure
1895          * that the field is read in a single access:
1896          */
1897         seq = READ_ONCE(p->mm->numa_scan_seq);
1898         if (p->numa_scan_seq == seq)
1899                 return;
1900         p->numa_scan_seq = seq;
1901         p->numa_scan_period_max = task_scan_max(p);
1902 
1903         total_faults = p->numa_faults_locality[0] +
1904                        p->numa_faults_locality[1];
1905         runtime = numa_get_avg_runtime(p, &period);
1906 
1907         /* If the task is part of a group prevent parallel updates to group stats */
1908         if (p->numa_group) {
1909                 group_lock = &p->numa_group->lock;
1910                 spin_lock_irq(group_lock);
1911         }
1912 
1913         /* Find the node with the highest number of faults */
1914         for_each_online_node(nid) {
1915                 /* Keep track of the offsets in numa_faults array */
1916                 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1917                 unsigned long faults = 0, group_faults = 0;
1918                 int priv;
1919 
1920                 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1921                         long diff, f_diff, f_weight;
1922 
1923                         mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1924                         membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1925                         cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1926                         cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1927 
1928                         /* Decay existing window, copy faults since last scan */
1929                         diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1930                         fault_types[priv] += p->numa_faults[membuf_idx];
1931                         p->numa_faults[membuf_idx] = 0;
1932 
1933                         /*
1934                          * Normalize the faults_from, so all tasks in a group
1935                          * count according to CPU use, instead of by the raw
1936                          * number of faults. Tasks with little runtime have
1937                          * little over-all impact on throughput, and thus their
1938                          * faults are less important.
1939                          */
1940                         f_weight = div64_u64(runtime << 16, period + 1);
1941                         f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1942                                    (total_faults + 1);
1943                         f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1944                         p->numa_faults[cpubuf_idx] = 0;
1945 
1946                         p->numa_faults[mem_idx] += diff;
1947                         p->numa_faults[cpu_idx] += f_diff;
1948                         faults += p->numa_faults[mem_idx];
1949                         p->total_numa_faults += diff;
1950                         if (p->numa_group) {
1951                                 /*
1952                                  * safe because we can only change our own group
1953                                  *
1954                                  * mem_idx represents the offset for a given
1955                                  * nid and priv in a specific region because it
1956                                  * is at the beginning of the numa_faults array.
1957                                  */
1958                                 p->numa_group->faults[mem_idx] += diff;
1959                                 p->numa_group->faults_cpu[mem_idx] += f_diff;
1960                                 p->numa_group->total_faults += diff;
1961                                 group_faults += p->numa_group->faults[mem_idx];
1962                         }
1963                 }
1964 
1965                 if (faults > max_faults) {
1966                         max_faults = faults;
1967                         max_nid = nid;
1968                 }
1969 
1970                 if (group_faults > max_group_faults) {
1971                         max_group_faults = group_faults;
1972                         max_group_nid = nid;
1973                 }
1974         }
1975 
1976         update_task_scan_period(p, fault_types[0], fault_types[1]);
1977 
1978         if (p->numa_group) {
1979                 numa_group_count_active_nodes(p->numa_group);
1980                 spin_unlock_irq(group_lock);
1981                 max_nid = preferred_group_nid(p, max_group_nid);
1982         }
1983 
1984         if (max_faults) {
1985                 /* Set the new preferred node */
1986                 if (max_nid != p->numa_preferred_nid)
1987                         sched_setnuma(p, max_nid);
1988 
1989                 if (task_node(p) != p->numa_preferred_nid)
1990                         numa_migrate_preferred(p);
1991         }
1992 }
1993 
1994 static inline int get_numa_group(struct numa_group *grp)
1995 {
1996         return atomic_inc_not_zero(&grp->refcount);
1997 }
1998 
1999 static inline void put_numa_group(struct numa_group *grp)
2000 {
2001         if (atomic_dec_and_test(&grp->refcount))
2002                 kfree_rcu(grp, rcu);
2003 }
2004 
2005 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2006                         int *priv)
2007 {
2008         struct numa_group *grp, *my_grp;
2009         struct task_struct *tsk;
2010         bool join = false;
2011         int cpu = cpupid_to_cpu(cpupid);
2012         int i;
2013 
2014         if (unlikely(!p->numa_group)) {
2015                 unsigned int size = sizeof(struct numa_group) +
2016                                     4*nr_node_ids*sizeof(unsigned long);
2017 
2018                 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2019                 if (!grp)
2020                         return;
2021 
2022                 atomic_set(&grp->refcount, 1);
2023                 grp->active_nodes = 1;
2024                 grp->max_faults_cpu = 0;
2025                 spin_lock_init(&grp->lock);
2026                 grp->gid = p->pid;
2027                 /* Second half of the array tracks nids where faults happen */
2028                 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2029                                                 nr_node_ids;
2030 
2031                 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2032                         grp->faults[i] = p->numa_faults[i];
2033 
2034                 grp->total_faults = p->total_numa_faults;
2035 
2036                 grp->nr_tasks++;
2037                 rcu_assign_pointer(p->numa_group, grp);
2038         }
2039 
2040         rcu_read_lock();
2041         tsk = READ_ONCE(cpu_rq(cpu)->curr);
2042 
2043         if (!cpupid_match_pid(tsk, cpupid))
2044                 goto no_join;
2045 
2046         grp = rcu_dereference(tsk->numa_group);
2047         if (!grp)
2048                 goto no_join;
2049 
2050         my_grp = p->numa_group;
2051         if (grp == my_grp)
2052                 goto no_join;
2053 
2054         /*
2055          * Only join the other group if its bigger; if we're the bigger group,
2056          * the other task will join us.
2057          */
2058         if (my_grp->nr_tasks > grp->nr_tasks)
2059                 goto no_join;
2060 
2061         /*
2062          * Tie-break on the grp address.
2063          */
2064         if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2065                 goto no_join;
2066 
2067         /* Always join threads in the same process. */
2068         if (tsk->mm == current->mm)
2069                 join = true;
2070 
2071         /* Simple filter to avoid false positives due to PID collisions */
2072         if (flags & TNF_SHARED)
2073                 join = true;
2074 
2075         /* Update priv based on whether false sharing was detected */
2076         *priv = !join;
2077 
2078         if (join && !get_numa_group(grp))
2079                 goto no_join;
2080 
2081         rcu_read_unlock();
2082 
2083         if (!join)
2084                 return;
2085 
2086         BUG_ON(irqs_disabled());
2087         double_lock_irq(&my_grp->lock, &grp->lock);
2088 
2089         for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2090                 my_grp->faults[i] -= p->numa_faults[i];
2091                 grp->faults[i] += p->numa_faults[i];
2092         }
2093         my_grp->total_faults -= p->total_numa_faults;
2094         grp->total_faults += p->total_numa_faults;
2095 
2096         my_grp->nr_tasks--;
2097         grp->nr_tasks++;
2098 
2099         spin_unlock(&my_grp->lock);
2100         spin_unlock_irq(&grp->lock);
2101 
2102         rcu_assign_pointer(p->numa_group, grp);
2103 
2104         put_numa_group(my_grp);
2105         return;
2106 
2107 no_join:
2108         rcu_read_unlock();
2109         return;
2110 }
2111 
2112 void task_numa_free(struct task_struct *p)
2113 {
2114         struct numa_group *grp = p->numa_group;
2115         void *numa_faults = p->numa_faults;
2116         unsigned long flags;
2117         int i;
2118 
2119         if (grp) {
2120                 spin_lock_irqsave(&grp->lock, flags);
2121                 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2122                         grp->faults[i] -= p->numa_faults[i];
2123                 grp->total_faults -= p->total_numa_faults;
2124 
2125                 grp->nr_tasks--;
2126                 spin_unlock_irqrestore(&grp->lock, flags);
2127                 RCU_INIT_POINTER(p->numa_group, NULL);
2128                 put_numa_group(grp);
2129         }
2130 
2131         p->numa_faults = NULL;
2132         kfree(numa_faults);
2133 }
2134 
2135 /*
2136  * Got a PROT_NONE fault for a page on @node.
2137  */
2138 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2139 {
2140         struct task_struct *p = current;
2141         bool migrated = flags & TNF_MIGRATED;
2142         int cpu_node = task_node(current);
2143         int local = !!(flags & TNF_FAULT_LOCAL);
2144         struct numa_group *ng;
2145         int priv;
2146 
2147         if (!static_branch_likely(&sched_numa_balancing))
2148                 return;
2149 
2150         /* for example, ksmd faulting in a user's mm */
2151         if (!p->mm)
2152                 return;
2153 
2154         /* Allocate buffer to track faults on a per-node basis */
2155         if (unlikely(!p->numa_faults)) {
2156                 int size = sizeof(*p->numa_faults) *
2157                            NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2158 
2159                 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2160                 if (!p->numa_faults)
2161                         return;
2162 
2163                 p->total_numa_faults = 0;
2164                 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2165         }
2166 
2167         /*
2168          * First accesses are treated as private, otherwise consider accesses
2169          * to be private if the accessing pid has not changed
2170          */
2171         if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2172                 priv = 1;
2173         } else {
2174                 priv = cpupid_match_pid(p, last_cpupid);
2175                 if (!priv && !(flags & TNF_NO_GROUP))
2176                         task_numa_group(p, last_cpupid, flags, &priv);
2177         }
2178 
2179         /*
2180          * If a workload spans multiple NUMA nodes, a shared fault that
2181          * occurs wholly within the set of nodes that the workload is
2182          * actively using should be counted as local. This allows the
2183          * scan rate to slow down when a workload has settled down.
2184          */
2185         ng = p->numa_group;
2186         if (!priv && !local && ng && ng->active_nodes > 1 &&
2187                                 numa_is_active_node(cpu_node, ng) &&
2188                                 numa_is_active_node(mem_node, ng))
2189                 local = 1;
2190 
2191         task_numa_placement(p);
2192 
2193         /*
2194          * Retry task to preferred node migration periodically, in case it
2195          * case it previously failed, or the scheduler moved us.
2196          */
2197         if (time_after(jiffies, p->numa_migrate_retry))
2198                 numa_migrate_preferred(p);
2199 
2200         if (migrated)
2201                 p->numa_pages_migrated += pages;
2202         if (flags & TNF_MIGRATE_FAIL)
2203                 p->numa_faults_locality[2] += pages;
2204 
2205         p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2206         p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2207         p->numa_faults_locality[local] += pages;
2208 }
2209 
2210 static void reset_ptenuma_scan(struct task_struct *p)
2211 {
2212         /*
2213          * We only did a read acquisition of the mmap sem, so
2214          * p->mm->numa_scan_seq is written to without exclusive access
2215          * and the update is not guaranteed to be atomic. That's not
2216          * much of an issue though, since this is just used for
2217          * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2218          * expensive, to avoid any form of compiler optimizations:
2219          */
2220         WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2221         p->mm->numa_scan_offset = 0;
2222 }
2223 
2224 /*
2225  * The expensive part of numa migration is done from task_work context.
2226  * Triggered from task_tick_numa().
2227  */
2228 void task_numa_work(struct callback_head *work)
2229 {
2230         unsigned long migrate, next_scan, now = jiffies;
2231         struct task_struct *p = current;
2232         struct mm_struct *mm = p->mm;
2233         u64 runtime = p->se.sum_exec_runtime;
2234         struct vm_area_struct *vma;
2235         unsigned long start, end;
2236         unsigned long nr_pte_updates = 0;
2237         long pages, virtpages;
2238 
2239         WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2240 
2241         work->next = work; /* protect against double add */
2242         /*
2243          * Who cares about NUMA placement when they're dying.
2244          *
2245          * NOTE: make sure not to dereference p->mm before this check,
2246          * exit_task_work() happens _after_ exit_mm() so we could be called
2247          * without p->mm even though we still had it when we enqueued this
2248          * work.
2249          */
2250         if (p->flags & PF_EXITING)
2251                 return;
2252 
2253         if (!mm->numa_next_scan) {
2254                 mm->numa_next_scan = now +
2255                         msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2256         }
2257 
2258         /*
2259          * Enforce maximal scan/migration frequency..
2260          */
2261         migrate = mm->numa_next_scan;
2262         if (time_before(now, migrate))
2263                 return;
2264 
2265         if (p->numa_scan_period == 0) {
2266                 p->numa_scan_period_max = task_scan_max(p);
2267                 p->numa_scan_period = task_scan_min(p);
2268         }
2269 
2270         next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2271         if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2272                 return;
2273 
2274         /*
2275          * Delay this task enough that another task of this mm will likely win
2276          * the next time around.
2277          */
2278         p->node_stamp += 2 * TICK_NSEC;
2279 
2280         start = mm->numa_scan_offset;
2281         pages = sysctl_numa_balancing_scan_size;
2282         pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2283         virtpages = pages * 8;     /* Scan up to this much virtual space */
2284         if (!pages)
2285                 return;
2286 
2287 
2288         down_read(&mm->mmap_sem);
2289         vma = find_vma(mm, start);
2290         if (!vma) {
2291                 reset_ptenuma_scan(p);
2292                 start = 0;
2293                 vma = mm->mmap;
2294         }
2295         for (; vma; vma = vma->vm_next) {
2296                 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2297                         is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2298                         continue;
2299                 }
2300 
2301                 /*
2302                  * Shared library pages mapped by multiple processes are not
2303                  * migrated as it is expected they are cache replicated. Avoid
2304                  * hinting faults in read-only file-backed mappings or the vdso
2305                  * as migrating the pages will be of marginal benefit.
2306                  */
2307                 if (!vma->vm_mm ||
2308                     (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2309                         continue;
2310 
2311                 /*
2312                  * Skip inaccessible VMAs to avoid any confusion between
2313                  * PROT_NONE and NUMA hinting ptes
2314                  */
2315                 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2316                         continue;
2317 
2318                 do {
2319                         start = max(start, vma->vm_start);
2320                         end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2321                         end = min(end, vma->vm_end);
2322                         nr_pte_updates = change_prot_numa(vma, start, end);
2323 
2324                         /*
2325                          * Try to scan sysctl_numa_balancing_size worth of
2326                          * hpages that have at least one present PTE that
2327                          * is not already pte-numa. If the VMA contains
2328                          * areas that are unused or already full of prot_numa
2329                          * PTEs, scan up to virtpages, to skip through those
2330                          * areas faster.
2331                          */
2332                         if (nr_pte_updates)
2333                                 pages -= (end - start) >> PAGE_SHIFT;
2334                         virtpages -= (end - start) >> PAGE_SHIFT;
2335 
2336                         start = end;
2337                         if (pages <= 0 || virtpages <= 0)
2338                                 goto out;
2339 
2340                         cond_resched();
2341                 } while (end != vma->vm_end);
2342         }
2343 
2344 out:
2345         /*
2346          * It is possible to reach the end of the VMA list but the last few
2347          * VMAs are not guaranteed to the vma_migratable. If they are not, we
2348          * would find the !migratable VMA on the next scan but not reset the
2349          * scanner to the start so check it now.
2350          */
2351         if (vma)
2352                 mm->numa_scan_offset = start;
2353         else
2354                 reset_ptenuma_scan(p);
2355         up_read(&mm->mmap_sem);
2356 
2357         /*
2358          * Make sure tasks use at least 32x as much time to run other code
2359          * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2360          * Usually update_task_scan_period slows down scanning enough; on an
2361          * overloaded system we need to limit overhead on a per task basis.
2362          */
2363         if (unlikely(p->se.sum_exec_runtime != runtime)) {
2364                 u64 diff = p->se.sum_exec_runtime - runtime;
2365                 p->node_stamp += 32 * diff;
2366         }
2367 }
2368 
2369 /*
2370  * Drive the periodic memory faults..
2371  */
2372 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2373 {
2374         struct callback_head *work = &curr->numa_work;
2375         u64 period, now;
2376 
2377         /*
2378          * We don't care about NUMA placement if we don't have memory.
2379          */
2380         if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2381                 return;
2382 
2383         /*
2384          * Using runtime rather than walltime has the dual advantage that
2385          * we (mostly) drive the selection from busy threads and that the
2386          * task needs to have done some actual work before we bother with
2387          * NUMA placement.
2388          */
2389         now = curr->se.sum_exec_runtime;
2390         period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2391 
2392         if (now > curr->node_stamp + period) {
2393                 if (!curr->node_stamp)
2394                         curr->numa_scan_period = task_scan_min(curr);
2395                 curr->node_stamp += period;
2396 
2397                 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2398                         init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2399                         task_work_add(curr, work, true);
2400                 }
2401         }
2402 }
2403 #else
2404 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2405 {
2406 }
2407 
2408 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2409 {
2410 }
2411 
2412 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2413 {
2414 }
2415 #endif /* CONFIG_NUMA_BALANCING */
2416 
2417 static void
2418 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2419 {
2420         update_load_add(&cfs_rq->load, se->load.weight);
2421         if (!parent_entity(se))
2422                 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2423 #ifdef CONFIG_SMP
2424         if (entity_is_task(se)) {
2425                 struct rq *rq = rq_of(cfs_rq);
2426 
2427                 account_numa_enqueue(rq, task_of(se));
2428                 list_add(&se->group_node, &rq->cfs_tasks);
2429         }
2430 #endif
2431         cfs_rq->nr_running++;
2432 }
2433 
2434 static void
2435 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2436 {
2437         update_load_sub(&cfs_rq->load, se->load.weight);
2438         if (!parent_entity(se))
2439                 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2440         if (entity_is_task(se)) {
2441                 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2442                 list_del_init(&se->group_node);
2443         }
2444         cfs_rq->nr_running--;
2445 }
2446 
2447 #ifdef CONFIG_FAIR_GROUP_SCHED
2448 # ifdef CONFIG_SMP
2449 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2450 {
2451         long tg_weight;
2452 
2453         /*
2454          * Use this CPU's real-time load instead of the last load contribution
2455          * as the updating of the contribution is delayed, and we will use the
2456          * the real-time load to calc the share. See update_tg_load_avg().
2457          */
2458         tg_weight = atomic_long_read(&tg->load_avg);
2459         tg_weight -= cfs_rq->tg_load_avg_contrib;
2460         tg_weight += cfs_rq->load.weight;
2461 
2462         return tg_weight;
2463 }
2464 
2465 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2466 {
2467         long tg_weight, load, shares;
2468 
2469         tg_weight = calc_tg_weight(tg, cfs_rq);
2470         load = cfs_rq->load.weight;
2471 
2472         shares = (tg->shares * load);
2473         if (tg_weight)
2474                 shares /= tg_weight;
2475 
2476         if (shares < MIN_SHARES)
2477                 shares = MIN_SHARES;
2478         if (shares > tg->shares)
2479                 shares = tg->shares;
2480 
2481         return shares;
2482 }
2483 # else /* CONFIG_SMP */
2484 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2485 {
2486         return tg->shares;
2487 }
2488 # endif /* CONFIG_SMP */
2489 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2490                             unsigned long weight)
2491 {
2492         if (se->on_rq) {
2493                 /* commit outstanding execution time */
2494                 if (cfs_rq->curr == se)
2495                         update_curr(cfs_rq);
2496                 account_entity_dequeue(cfs_rq, se);
2497         }
2498 
2499         update_load_set(&se->load, weight);
2500 
2501         if (se->on_rq)
2502                 account_entity_enqueue(cfs_rq, se);
2503 }
2504 
2505 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2506 
2507 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2508 {
2509         struct task_group *tg;
2510         struct sched_entity *se;
2511         long shares;
2512 
2513         tg = cfs_rq->tg;
2514         se = tg->se[cpu_of(rq_of(cfs_rq))];
2515         if (!se || throttled_hierarchy(cfs_rq))
2516                 return;
2517 #ifndef CONFIG_SMP
2518         if (likely(se->load.weight == tg->shares))
2519                 return;
2520 #endif
2521         shares = calc_cfs_shares(cfs_rq, tg);
2522 
2523         reweight_entity(cfs_rq_of(se), se, shares);
2524 }
2525 #else /* CONFIG_FAIR_GROUP_SCHED */
2526 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2527 {
2528 }
2529 #endif /* CONFIG_FAIR_GROUP_SCHED */
2530 
2531 #ifdef CONFIG_SMP
2532 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2533 static const u32 runnable_avg_yN_inv[] = {
2534         0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2535         0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2536         0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2537         0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2538         0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2539         0x85aac367, 0x82cd8698,
2540 };
2541 
2542 /*
2543  * Precomputed \Sum y^k { 1<=k<=n }.  These are floor(true_value) to prevent
2544  * over-estimates when re-combining.
2545  */
2546 static const u32 runnable_avg_yN_sum[] = {
2547             0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2548          9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2549         17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2550 };
2551 
2552 /*
2553  * Approximate:
2554  *   val * y^n,    where y^32 ~= 0.5 (~1 scheduling period)
2555  */
2556 static __always_inline u64 decay_load(u64 val, u64 n)
2557 {
2558         unsigned int local_n;
2559 
2560         if (!n)
2561                 return val;
2562         else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2563                 return 0;
2564 
2565         /* after bounds checking we can collapse to 32-bit */
2566         local_n = n;
2567 
2568         /*
2569          * As y^PERIOD = 1/2, we can combine
2570          *    y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2571          * With a look-up table which covers y^n (n<PERIOD)
2572          *
2573          * To achieve constant time decay_load.
2574          */
2575         if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2576                 val >>= local_n / LOAD_AVG_PERIOD;
2577                 local_n %= LOAD_AVG_PERIOD;
2578         }
2579 
2580         val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2581         return val;
2582 }
2583 
2584 /*
2585  * For updates fully spanning n periods, the contribution to runnable
2586  * average will be: \Sum 1024*y^n
2587  *
2588  * We can compute this reasonably efficiently by combining:
2589  *   y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for  n <PERIOD}
2590  */
2591 static u32 __compute_runnable_contrib(u64 n)
2592 {
2593         u32 contrib = 0;
2594 
2595         if (likely(n <= LOAD_AVG_PERIOD))
2596                 return runnable_avg_yN_sum[n];
2597         else if (unlikely(n >= LOAD_AVG_MAX_N))
2598                 return LOAD_AVG_MAX;
2599 
2600         /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2601         do {
2602                 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2603                 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2604 
2605                 n -= LOAD_AVG_PERIOD;
2606         } while (n > LOAD_AVG_PERIOD);
2607 
2608         contrib = decay_load(contrib, n);
2609         return contrib + runnable_avg_yN_sum[n];
2610 }
2611 
2612 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2613 #error "load tracking assumes 2^10 as unit"
2614 #endif
2615 
2616 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2617 
2618 /*
2619  * We can represent the historical contribution to runnable average as the
2620  * coefficients of a geometric series.  To do this we sub-divide our runnable
2621  * history into segments of approximately 1ms (1024us); label the segment that
2622  * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2623  *
2624  * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2625  *      p0            p1           p2
2626  *     (now)       (~1ms ago)  (~2ms ago)
2627  *
2628  * Let u_i denote the fraction of p_i that the entity was runnable.
2629  *
2630  * We then designate the fractions u_i as our co-efficients, yielding the
2631  * following representation of historical load:
2632  *   u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2633  *
2634  * We choose y based on the with of a reasonably scheduling period, fixing:
2635  *   y^32 = 0.5
2636  *
2637  * This means that the contribution to load ~32ms ago (u_32) will be weighted
2638  * approximately half as much as the contribution to load within the last ms
2639  * (u_0).
2640  *
2641  * When a period "rolls over" and we have new u_0`, multiplying the previous
2642  * sum again by y is sufficient to update:
2643  *   load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2644  *            = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2645  */
2646 static __always_inline int
2647 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2648                   unsigned long weight, int running, struct cfs_rq *cfs_rq)
2649 {
2650         u64 delta, scaled_delta, periods;
2651         u32 contrib;
2652         unsigned int delta_w, scaled_delta_w, decayed = 0;
2653         unsigned long scale_freq, scale_cpu;
2654 
2655         delta = now - sa->last_update_time;
2656         /*
2657          * This should only happen when time goes backwards, which it
2658          * unfortunately does during sched clock init when we swap over to TSC.
2659          */
2660         if ((s64)delta < 0) {
2661                 sa->last_update_time = now;
2662                 return 0;
2663         }
2664 
2665         /*
2666          * Use 1024ns as the unit of measurement since it's a reasonable
2667          * approximation of 1us and fast to compute.
2668          */
2669         delta >>= 10;
2670         if (!delta)
2671                 return 0;
2672         sa->last_update_time = now;
2673 
2674         scale_freq = arch_scale_freq_capacity(NULL, cpu);
2675         scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2676 
2677         /* delta_w is the amount already accumulated against our next period */
2678         delta_w = sa->period_contrib;
2679         if (delta + delta_w >= 1024) {
2680                 decayed = 1;
2681 
2682                 /* how much left for next period will start over, we don't know yet */
2683                 sa->period_contrib = 0;
2684 
2685                 /*
2686                  * Now that we know we're crossing a period boundary, figure
2687                  * out how much from delta we need to complete the current
2688                  * period and accrue it.
2689                  */
2690                 delta_w = 1024 - delta_w;
2691                 scaled_delta_w = cap_scale(delta_w, scale_freq);
2692                 if (weight) {
2693                         sa->load_sum += weight * scaled_delta_w;
2694                         if (cfs_rq) {
2695                                 cfs_rq->runnable_load_sum +=
2696                                                 weight * scaled_delta_w;
2697                         }
2698                 }
2699                 if (running)
2700                         sa->util_sum += scaled_delta_w * scale_cpu;
2701 
2702                 delta -= delta_w;
2703 
2704                 /* Figure out how many additional periods this update spans */
2705                 periods = delta / 1024;
2706                 delta %= 1024;
2707 
2708                 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2709                 if (cfs_rq) {
2710                         cfs_rq->runnable_load_sum =
2711                                 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2712                 }
2713                 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2714 
2715                 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2716                 contrib = __compute_runnable_contrib(periods);
2717                 contrib = cap_scale(contrib, scale_freq);
2718                 if (weight) {
2719                         sa->load_sum += weight * contrib;
2720                         if (cfs_rq)
2721                                 cfs_rq->runnable_load_sum += weight * contrib;
2722                 }
2723                 if (running)
2724                         sa->util_sum += contrib * scale_cpu;
2725         }
2726 
2727         /* Remainder of delta accrued against u_0` */
2728         scaled_delta = cap_scale(delta, scale_freq);
2729         if (weight) {
2730                 sa->load_sum += weight * scaled_delta;
2731                 if (cfs_rq)
2732                         cfs_rq->runnable_load_sum += weight * scaled_delta;
2733         }
2734         if (running)
2735                 sa->util_sum += scaled_delta * scale_cpu;
2736 
2737         sa->period_contrib += delta;
2738 
2739         if (decayed) {
2740                 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2741                 if (cfs_rq) {
2742                         cfs_rq->runnable_load_avg =
2743                                 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2744                 }
2745                 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2746         }
2747 
2748         return decayed;
2749 }
2750 
2751 #ifdef CONFIG_FAIR_GROUP_SCHED
2752 /*
2753  * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2754  * and effective_load (which is not done because it is too costly).
2755  */
2756 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2757 {
2758         long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2759 
2760         /*
2761          * No need to update load_avg for root_task_group as it is not used.
2762          */
2763         if (cfs_rq->tg == &root_task_group)
2764                 return;
2765 
2766         if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2767                 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2768                 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2769         }
2770 }
2771 
2772 /*
2773  * Called within set_task_rq() right before setting a task's cpu. The
2774  * caller only guarantees p->pi_lock is held; no other assumptions,
2775  * including the state of rq->lock, should be made.
2776  */
2777 void set_task_rq_fair(struct sched_entity *se,
2778                       struct cfs_rq *prev, struct cfs_rq *next)
2779 {
2780         if (!sched_feat(ATTACH_AGE_LOAD))
2781                 return;
2782 
2783         /*
2784          * We are supposed to update the task to "current" time, then its up to
2785          * date and ready to go to new CPU/cfs_rq. But we have difficulty in
2786          * getting what current time is, so simply throw away the out-of-date
2787          * time. This will result in the wakee task is less decayed, but giving
2788          * the wakee more load sounds not bad.
2789          */
2790         if (se->avg.last_update_time && prev) {
2791                 u64 p_last_update_time;
2792                 u64 n_last_update_time;
2793 
2794 #ifndef CONFIG_64BIT
2795                 u64 p_last_update_time_copy;
2796                 u64 n_last_update_time_copy;
2797 
2798                 do {
2799                         p_last_update_time_copy = prev->load_last_update_time_copy;
2800                         n_last_update_time_copy = next->load_last_update_time_copy;
2801 
2802                         smp_rmb();
2803 
2804                         p_last_update_time = prev->avg.last_update_time;
2805                         n_last_update_time = next->avg.last_update_time;
2806 
2807                 } while (p_last_update_time != p_last_update_time_copy ||
2808                          n_last_update_time != n_last_update_time_copy);
2809 #else
2810                 p_last_update_time = prev->avg.last_update_time;
2811                 n_last_update_time = next->avg.last_update_time;
2812 #endif
2813                 __update_load_avg(p_last_update_time, cpu_of(rq_of(prev)),
2814                                   &se->avg, 0, 0, NULL);
2815                 se->avg.last_update_time = n_last_update_time;
2816         }
2817 }
2818 #else /* CONFIG_FAIR_GROUP_SCHED */
2819 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2820 #endif /* CONFIG_FAIR_GROUP_SCHED */
2821 
2822 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2823 
2824 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2825 static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
2826 {
2827         struct sched_avg *sa = &cfs_rq->avg;
2828         int decayed, removed = 0;
2829 
2830         if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2831                 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2832                 sa->load_avg = max_t(long, sa->load_avg - r, 0);
2833                 sa->load_sum = max_t(s64, sa->load_sum - r * LOAD_AVG_MAX, 0);
2834                 removed = 1;
2835         }
2836 
2837         if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2838                 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2839                 sa->util_avg = max_t(long, sa->util_avg - r, 0);
2840                 sa->util_sum = max_t(s32, sa->util_sum - r * LOAD_AVG_MAX, 0);
2841         }
2842 
2843         decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2844                 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2845 
2846 #ifndef CONFIG_64BIT
2847         smp_wmb();
2848         cfs_rq->load_last_update_time_copy = sa->last_update_time;
2849 #endif
2850 
2851         return decayed || removed;
2852 }
2853 
2854 /* Update task and its cfs_rq load average */
2855 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2856 {
2857         struct cfs_rq *cfs_rq = cfs_rq_of(se);
2858         u64 now = cfs_rq_clock_task(cfs_rq);
2859         struct rq *rq = rq_of(cfs_rq);
2860         int cpu = cpu_of(rq);
2861 
2862         /*
2863          * Track task load average for carrying it to new CPU after migrated, and
2864          * track group sched_entity load average for task_h_load calc in migration
2865          */
2866         __update_load_avg(now, cpu, &se->avg,
2867                           se->on_rq * scale_load_down(se->load.weight),
2868                           cfs_rq->curr == se, NULL);
2869 
2870         if (update_cfs_rq_load_avg(now, cfs_rq) && update_tg)
2871                 update_tg_load_avg(cfs_rq, 0);
2872 
2873         if (cpu == smp_processor_id() && &rq->cfs == cfs_rq) {
2874                 unsigned long max = rq->cpu_capacity_orig;
2875 
2876                 /*
2877                  * There are a few boundary cases this might miss but it should
2878                  * get called often enough that that should (hopefully) not be
2879                  * a real problem -- added to that it only calls on the local
2880                  * CPU, so if we enqueue remotely we'll miss an update, but
2881                  * the next tick/schedule should update.
2882                  *
2883                  * It will not get called when we go idle, because the idle
2884                  * thread is a different class (!fair), nor will the utilization
2885                  * number include things like RT tasks.
2886                  *
2887                  * As is, the util number is not freq-invariant (we'd have to
2888                  * implement arch_scale_freq_capacity() for that).
2889                  *
2890                  * See cpu_util().
2891                  */
2892                 cpufreq_update_util(rq_clock(rq),
2893                                     min(cfs_rq->avg.util_avg, max), max);
2894         }
2895 }
2896 
2897 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2898 {
2899         if (!sched_feat(ATTACH_AGE_LOAD))
2900                 goto skip_aging;
2901 
2902         /*
2903          * If we got migrated (either between CPUs or between cgroups) we'll
2904          * have aged the average right before clearing @last_update_time.
2905          */
2906         if (se->avg.last_update_time) {
2907                 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2908                                   &se->avg, 0, 0, NULL);
2909 
2910                 /*
2911                  * XXX: we could have just aged the entire load away if we've been
2912                  * absent from the fair class for too long.
2913                  */
2914         }
2915 
2916 skip_aging:
2917         se->avg.last_update_time = cfs_rq->avg.last_update_time;
2918         cfs_rq->avg.load_avg += se->avg.load_avg;
2919         cfs_rq->avg.load_sum += se->avg.load_sum;
2920         cfs_rq->avg.util_avg += se->avg.util_avg;
2921         cfs_rq->avg.util_sum += se->avg.util_sum;
2922 }
2923 
2924 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2925 {
2926         __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2927                           &se->avg, se->on_rq * scale_load_down(se->load.weight),
2928                           cfs_rq->curr == se, NULL);
2929 
2930         cfs_rq->avg.load_avg = max_t(long, cfs_rq->avg.load_avg - se->avg.load_avg, 0);
2931         cfs_rq->avg.load_sum = max_t(s64,  cfs_rq->avg.load_sum - se->avg.load_sum, 0);
2932         cfs_rq->avg.util_avg = max_t(long, cfs_rq->avg.util_avg - se->avg.util_avg, 0);
2933         cfs_rq->avg.util_sum = max_t(s32,  cfs_rq->avg.util_sum - se->avg.util_sum, 0);
2934 }
2935 
2936 /* Add the load generated by se into cfs_rq's load average */
2937 static inline void
2938 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2939 {
2940         struct sched_avg *sa = &se->avg;
2941         u64 now = cfs_rq_clock_task(cfs_rq);
2942         int migrated, decayed;
2943 
2944         migrated = !sa->last_update_time;
2945         if (!migrated) {
2946                 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2947                         se->on_rq * scale_load_down(se->load.weight),
2948                         cfs_rq->curr == se, NULL);
2949         }
2950 
2951         decayed = update_cfs_rq_load_avg(now, cfs_rq);
2952 
2953         cfs_rq->runnable_load_avg += sa->load_avg;
2954         cfs_rq->runnable_load_sum += sa->load_sum;
2955 
2956         if (migrated)
2957                 attach_entity_load_avg(cfs_rq, se);
2958 
2959         if (decayed || migrated)
2960                 update_tg_load_avg(cfs_rq, 0);
2961 }
2962 
2963 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2964 static inline void
2965 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2966 {
2967         update_load_avg(se, 1);
2968 
2969         cfs_rq->runnable_load_avg =
2970                 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2971         cfs_rq->runnable_load_sum =
2972                 max_t(s64,  cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2973 }
2974 
2975 #ifndef CONFIG_64BIT
2976 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2977 {
2978         u64 last_update_time_copy;
2979         u64 last_update_time;
2980 
2981         do {
2982                 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2983                 smp_rmb();
2984                 last_update_time = cfs_rq->avg.last_update_time;
2985         } while (last_update_time != last_update_time_copy);
2986 
2987         return last_update_time;
2988 }
2989 #else
2990 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2991 {
2992         return cfs_rq->avg.last_update_time;
2993 }
2994 #endif
2995 
2996 /*
2997  * Task first catches up with cfs_rq, and then subtract
2998  * itself from the cfs_rq (task must be off the queue now).
2999  */
3000 void remove_entity_load_avg(struct sched_entity *se)
3001 {
3002         struct cfs_rq *cfs_rq = cfs_rq_of(se);
3003         u64 last_update_time;
3004 
3005         /*
3006          * Newly created task or never used group entity should not be removed
3007          * from its (source) cfs_rq
3008          */
3009         if (se->avg.last_update_time == 0)
3010                 return;
3011 
3012         last_update_time = cfs_rq_last_update_time(cfs_rq);
3013 
3014         __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
3015         atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3016         atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3017 }
3018 
3019 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3020 {
3021         return cfs_rq->runnable_load_avg;
3022 }
3023 
3024 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3025 {
3026         return cfs_rq->avg.load_avg;
3027 }
3028 
3029 static int idle_balance(struct rq *this_rq);
3030 
3031 #else /* CONFIG_SMP */
3032 
3033 static inline void update_load_avg(struct sched_entity *se, int not_used)
3034 {
3035         struct cfs_rq *cfs_rq = cfs_rq_of(se);
3036         struct rq *rq = rq_of(cfs_rq);
3037 
3038         cpufreq_trigger_update(rq_clock(rq));
3039 }
3040 
3041 static inline void
3042 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3043 static inline void
3044 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3045 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3046 
3047 static inline void
3048 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3049 static inline void
3050 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3051 
3052 static inline int idle_balance(struct rq *rq)
3053 {
3054         return 0;
3055 }
3056 
3057 #endif /* CONFIG_SMP */
3058 
3059 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
3060 {
3061 #ifdef CONFIG_SCHEDSTATS
3062         struct task_struct *tsk = NULL;
3063 
3064         if (entity_is_task(se))
3065                 tsk = task_of(se);
3066 
3067         if (se->statistics.sleep_start) {
3068                 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
3069 
3070                 if ((s64)delta < 0)
3071                         delta = 0;
3072 
3073                 if (unlikely(delta > se->statistics.sleep_max))
3074                         se->statistics.sleep_max = delta;
3075 
3076                 se->statistics.sleep_start = 0;
3077                 se->statistics.sum_sleep_runtime += delta;
3078 
3079                 if (tsk) {
3080                         account_scheduler_latency(tsk, delta >> 10, 1);
3081                         trace_sched_stat_sleep(tsk, delta);
3082                 }
3083         }
3084         if (se->statistics.block_start) {
3085                 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
3086 
3087                 if ((s64)delta < 0)
3088                         delta = 0;
3089 
3090                 if (unlikely(delta > se->statistics.block_max))
3091                         se->statistics.block_max = delta;
3092 
3093                 se->statistics.block_start = 0;
3094                 se->statistics.sum_sleep_runtime += delta;
3095 
3096                 if (tsk) {
3097                         if (tsk->in_iowait) {
3098                                 se->statistics.iowait_sum += delta;
3099                                 se->statistics.iowait_count++;
3100                                 trace_sched_stat_iowait(tsk, delta);
3101                         }
3102 
3103                         trace_sched_stat_blocked(tsk, delta);
3104 
3105                         /*
3106                          * Blocking time is in units of nanosecs, so shift by
3107                          * 20 to get a milliseconds-range estimation of the
3108                          * amount of time that the task spent sleeping:
3109                          */
3110                         if (unlikely(prof_on == SLEEP_PROFILING)) {
3111                                 profile_hits(SLEEP_PROFILING,
3112                                                 (void *)get_wchan(tsk),
3113                                                 delta >> 20);
3114                         }
3115                         account_scheduler_latency(tsk, delta >> 10, 0);
3116                 }
3117         }
3118 #endif
3119 }
3120 
3121 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3122 {
3123 #ifdef CONFIG_SCHED_DEBUG
3124         s64 d = se->vruntime - cfs_rq->min_vruntime;
3125 
3126         if (d < 0)
3127                 d = -d;
3128 
3129         if (d > 3*sysctl_sched_latency)
3130                 schedstat_inc(cfs_rq, nr_spread_over);
3131 #endif
3132 }
3133 
3134 static void
3135 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3136 {
3137         u64 vruntime = cfs_rq->min_vruntime;
3138 
3139         /*
3140          * The 'current' period is already promised to the current tasks,
3141          * however the extra weight of the new task will slow them down a
3142          * little, place the new task so that it fits in the slot that
3143          * stays open at the end.
3144          */
3145         if (initial && sched_feat(START_DEBIT))
3146                 vruntime += sched_vslice(cfs_rq, se);
3147 
3148         /* sleeps up to a single latency don't count. */
3149         if (!initial) {
3150                 unsigned long thresh = sysctl_sched_latency;
3151 
3152                 /*
3153                  * Halve their sleep time's effect, to allow
3154                  * for a gentler effect of sleepers:
3155                  */
3156                 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3157                         thresh >>= 1;
3158 
3159                 vruntime -= thresh;
3160         }
3161 
3162         /* ensure we never gain time by being placed backwards. */
3163         se->vruntime = max_vruntime(se->vruntime, vruntime);
3164 }
3165 
3166 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3167 
3168 static inline void check_schedstat_required(void)
3169 {
3170 #ifdef CONFIG_SCHEDSTATS
3171         if (schedstat_enabled())
3172                 return;
3173 
3174         /* Force schedstat enabled if a dependent tracepoint is active */
3175         if (trace_sched_stat_wait_enabled()    ||
3176                         trace_sched_stat_sleep_enabled()   ||
3177                         trace_sched_stat_iowait_enabled()  ||
3178                         trace_sched_stat_blocked_enabled() ||
3179                         trace_sched_stat_runtime_enabled())  {
3180                 pr_warn_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3181                              "stat_blocked and stat_runtime require the "
3182                              "kernel parameter schedstats=enabled or "
3183                              "kernel.sched_schedstats=1\n");
3184         }
3185 #endif
3186 }
3187 
3188 static void
3189 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3190 {
3191         /*
3192          * Update the normalized vruntime before updating min_vruntime
3193          * through calling update_curr().
3194          */
3195         if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3196                 se->vruntime += cfs_rq->min_vruntime;
3197 
3198         /*
3199          * Update run-time statistics of the 'current'.
3200          */
3201         update_curr(cfs_rq);
3202         enqueue_entity_load_avg(cfs_rq, se);
3203         account_entity_enqueue(cfs_rq, se);
3204         update_cfs_shares(cfs_rq);
3205 
3206         if (flags & ENQUEUE_WAKEUP) {
3207                 place_entity(cfs_rq, se, 0);
3208                 if (schedstat_enabled())
3209                         enqueue_sleeper(cfs_rq, se);
3210         }
3211 
3212         check_schedstat_required();
3213         if (schedstat_enabled()) {
3214                 update_stats_enqueue(cfs_rq, se);
3215                 check_spread(cfs_rq, se);
3216         }
3217         if (se != cfs_rq->curr)
3218                 __enqueue_entity(cfs_rq, se);
3219         se->on_rq = 1;
3220 
3221         if (cfs_rq->nr_running == 1) {
3222                 list_add_leaf_cfs_rq(cfs_rq);
3223                 check_enqueue_throttle(cfs_rq);
3224         }
3225 }
3226 
3227 static void __clear_buddies_last(struct sched_entity *se)
3228 {
3229         for_each_sched_entity(se) {
3230                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3231                 if (cfs_rq->last != se)
3232                         break;
3233 
3234                 cfs_rq->last = NULL;
3235         }
3236 }
3237 
3238 static void __clear_buddies_next(struct sched_entity *se)
3239 {
3240         for_each_sched_entity(se) {
3241                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3242                 if (cfs_rq->next != se)
3243                         break;
3244 
3245                 cfs_rq->next = NULL;
3246         }
3247 }
3248 
3249 static void __clear_buddies_skip(struct sched_entity *se)
3250 {
3251         for_each_sched_entity(se) {
3252                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3253                 if (cfs_rq->skip != se)
3254                         break;
3255 
3256                 cfs_rq->skip = NULL;
3257         }
3258 }
3259 
3260 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3261 {
3262         if (cfs_rq->last == se)
3263                 __clear_buddies_last(se);
3264 
3265         if (cfs_rq->next == se)
3266                 __clear_buddies_next(se);
3267 
3268         if (cfs_rq->skip == se)
3269                 __clear_buddies_skip(se);
3270 }
3271 
3272 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3273 
3274 static void
3275 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3276 {
3277         /*
3278          * Update run-time statistics of the 'current'.
3279          */
3280         update_curr(cfs_rq);
3281         dequeue_entity_load_avg(cfs_rq, se);
3282 
3283         if (schedstat_enabled())
3284                 update_stats_dequeue(cfs_rq, se, flags);
3285 
3286         clear_buddies(cfs_rq, se);
3287 
3288         if (se != cfs_rq->curr)
3289                 __dequeue_entity(cfs_rq, se);
3290         se->on_rq = 0;
3291         account_entity_dequeue(cfs_rq, se);
3292 
3293         /*
3294          * Normalize the entity after updating the min_vruntime because the
3295          * update can refer to the ->curr item and we need to reflect this
3296          * movement in our normalized position.
3297          */
3298         if (!(flags & DEQUEUE_SLEEP))
3299                 se->vruntime -= cfs_rq->min_vruntime;
3300 
3301         /* return excess runtime on last dequeue */
3302         return_cfs_rq_runtime(cfs_rq);
3303 
3304         update_min_vruntime(cfs_rq);
3305         update_cfs_shares(cfs_rq);
3306 }
3307 
3308 /*
3309  * Preempt the current task with a newly woken task if needed:
3310  */
3311 static void
3312 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3313 {
3314         unsigned long ideal_runtime, delta_exec;
3315         struct sched_entity *se;
3316         s64 delta;
3317 
3318         ideal_runtime = sched_slice(cfs_rq, curr);
3319         delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3320         if (delta_exec > ideal_runtime) {
3321                 resched_curr(rq_of(cfs_rq));
3322                 /*
3323                  * The current task ran long enough, ensure it doesn't get
3324                  * re-elected due to buddy favours.
3325                  */
3326                 clear_buddies(cfs_rq, curr);
3327                 return;
3328         }
3329 
3330         /*
3331          * Ensure that a task that missed wakeup preemption by a
3332          * narrow margin doesn't have to wait for a full slice.
3333          * This also mitigates buddy induced latencies under load.
3334          */
3335         if (delta_exec < sysctl_sched_min_granularity)
3336                 return;
3337 
3338         se = __pick_first_entity(cfs_rq);
3339         delta = curr->vruntime - se->vruntime;
3340 
3341         if (delta < 0)
3342                 return;
3343 
3344         if (delta > ideal_runtime)
3345                 resched_curr(rq_of(cfs_rq));
3346 }
3347 
3348 static void
3349 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3350 {
3351         /* 'current' is not kept within the tree. */
3352         if (se->on_rq) {
3353                 /*
3354                  * Any task has to be enqueued before it get to execute on
3355                  * a CPU. So account for the time it spent waiting on the
3356                  * runqueue.
3357                  */
3358                 if (schedstat_enabled())
3359                         update_stats_wait_end(cfs_rq, se);
3360                 __dequeue_entity(cfs_rq, se);
3361                 update_load_avg(se, 1);
3362         }
3363 
3364         update_stats_curr_start(cfs_rq, se);
3365         cfs_rq->curr = se;
3366 #ifdef CONFIG_SCHEDSTATS
3367         /*
3368          * Track our maximum slice length, if the CPU's load is at
3369          * least twice that of our own weight (i.e. dont track it
3370          * when there are only lesser-weight tasks around):
3371          */
3372         if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3373                 se->statistics.slice_max = max(se->statistics.slice_max,
3374                         se->sum_exec_runtime - se->prev_sum_exec_runtime);
3375         }
3376 #endif
3377         se->prev_sum_exec_runtime = se->sum_exec_runtime;
3378 }
3379 
3380 static int
3381 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3382 
3383 /*
3384  * Pick the next process, keeping these things in mind, in this order:
3385  * 1) keep things fair between processes/task groups
3386  * 2) pick the "next" process, since someone really wants that to run
3387  * 3) pick the "last" process, for cache locality
3388  * 4) do not run the "skip" process, if something else is available
3389  */
3390 static struct sched_entity *
3391 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3392 {
3393         struct sched_entity *left = __pick_first_entity(cfs_rq);
3394         struct sched_entity *se;
3395 
3396         /*
3397          * If curr is set we have to see if its left of the leftmost entity
3398          * still in the tree, provided there was anything in the tree at all.
3399          */
3400         if (!left || (curr && entity_before(curr, left)))
3401                 left = curr;
3402 
3403         se = left; /* ideally we run the leftmost entity */
3404 
3405         /*
3406          * Avoid running the skip buddy, if running something else can
3407          * be done without getting too unfair.
3408          */
3409         if (cfs_rq->skip == se) {
3410                 struct sched_entity *second;
3411 
3412                 if (se == curr) {
3413                         second = __pick_first_entity(cfs_rq);
3414                 } else {
3415                         second = __pick_next_entity(se);
3416                         if (!second || (curr && entity_before(curr, second)))
3417                                 second = curr;
3418                 }
3419 
3420                 if (second && wakeup_preempt_entity(second, left) < 1)
3421                         se = second;
3422         }
3423 
3424         /*
3425          * Prefer last buddy, try to return the CPU to a preempted task.
3426          */
3427         if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3428                 se = cfs_rq->last;
3429 
3430         /*
3431          * Someone really wants this to run. If it's not unfair, run it.
3432          */
3433         if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3434                 se = cfs_rq->next;
3435 
3436         clear_buddies(cfs_rq, se);
3437 
3438         return se;
3439 }
3440 
3441 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3442 
3443 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3444 {
3445         /*
3446          * If still on the runqueue then deactivate_task()
3447          * was not called and update_curr() has to be done:
3448          */
3449         if (prev->on_rq)
3450                 update_curr(cfs_rq);
3451 
3452         /* throttle cfs_rqs exceeding runtime */
3453         check_cfs_rq_runtime(cfs_rq);
3454 
3455         if (schedstat_enabled()) {
3456                 check_spread(cfs_rq, prev);
3457                 if (prev->on_rq)
3458                         update_stats_wait_start(cfs_rq, prev);
3459         }
3460 
3461         if (prev->on_rq) {
3462                 /* Put 'current' back into the tree. */
3463                 __enqueue_entity(cfs_rq, prev);
3464                 /* in !on_rq case, update occurred at dequeue */
3465                 update_load_avg(prev, 0);
3466         }
3467         cfs_rq->curr = NULL;
3468 }
3469 
3470 static void
3471 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3472 {
3473         /*
3474          * Update run-time statistics of the 'current'.
3475          */
3476         update_curr(cfs_rq);
3477 
3478         /*
3479          * Ensure that runnable average is periodically updated.
3480          */
3481         update_load_avg(curr, 1);
3482         update_cfs_shares(cfs_rq);
3483 
3484 #ifdef CONFIG_SCHED_HRTICK
3485         /*
3486          * queued ticks are scheduled to match the slice, so don't bother
3487          * validating it and just reschedule.
3488          */
3489         if (queued) {
3490                 resched_curr(rq_of(cfs_rq));
3491                 return;
3492         }
3493         /*
3494          * don't let the period tick interfere with the hrtick preemption
3495          */
3496         if (!sched_feat(DOUBLE_TICK) &&
3497                         hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3498                 return;
3499 #endif
3500 
3501         if (cfs_rq->nr_running > 1)
3502                 check_preempt_tick(cfs_rq, curr);
3503 }
3504 
3505 
3506 /**************************************************
3507  * CFS bandwidth control machinery
3508  */
3509 
3510 #ifdef CONFIG_CFS_BANDWIDTH
3511 
3512 #ifdef HAVE_JUMP_LABEL
3513 static struct static_key __cfs_bandwidth_used;
3514 
3515 static inline bool cfs_bandwidth_used(void)
3516 {
3517         return static_key_false(&__cfs_bandwidth_used);
3518 }
3519 
3520 void cfs_bandwidth_usage_inc(void)
3521 {
3522         static_key_slow_inc(&__cfs_bandwidth_used);
3523 }
3524 
3525 void cfs_bandwidth_usage_dec(void)
3526 {
3527         static_key_slow_dec(&__cfs_bandwidth_used);
3528 }
3529 #else /* HAVE_JUMP_LABEL */
3530 static bool cfs_bandwidth_used(void)
3531 {
3532         return true;
3533 }
3534 
3535 void cfs_bandwidth_usage_inc(void) {}
3536 void cfs_bandwidth_usage_dec(void) {}
3537 #endif /* HAVE_JUMP_LABEL */
3538 
3539 /*
3540  * default period for cfs group bandwidth.
3541  * default: 0.1s, units: nanoseconds
3542  */
3543 static inline u64 default_cfs_period(void)
3544 {
3545         return 100000000ULL;
3546 }
3547 
3548 static inline u64 sched_cfs_bandwidth_slice(void)
3549 {
3550         return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3551 }
3552 
3553 /*
3554  * Replenish runtime according to assigned quota and update expiration time.
3555  * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3556  * additional synchronization around rq->lock.
3557  *
3558  * requires cfs_b->lock
3559  */
3560 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3561 {
3562         u64 now;
3563 
3564         if (cfs_b->quota == RUNTIME_INF)
3565                 return;
3566 
3567         now = sched_clock_cpu(smp_processor_id());
3568         cfs_b->runtime = cfs_b->quota;
3569         cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3570 }
3571 
3572 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3573 {
3574         return &tg->cfs_bandwidth;
3575 }
3576 
3577 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3578 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3579 {
3580         if (unlikely(cfs_rq->throttle_count))
3581                 return cfs_rq->throttled_clock_task;
3582 
3583         return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3584 }
3585 
3586 /* returns 0 on failure to allocate runtime */
3587 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3588 {
3589         struct task_group *tg = cfs_rq->tg;
3590         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3591         u64 amount = 0, min_amount, expires;
3592 
3593         /* note: this is a positive sum as runtime_remaining <= 0 */
3594         min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3595 
3596         raw_spin_lock(&cfs_b->lock);
3597         if (cfs_b->quota == RUNTIME_INF)
3598                 amount = min_amount;
3599         else {
3600                 start_cfs_bandwidth(cfs_b);
3601 
3602                 if (cfs_b->runtime > 0) {
3603                         amount = min(cfs_b->runtime, min_amount);
3604                         cfs_b->runtime -= amount;
3605                         cfs_b->idle = 0;
3606                 }
3607         }
3608         expires = cfs_b->runtime_expires;
3609         raw_spin_unlock(&cfs_b->lock);
3610 
3611         cfs_rq->runtime_remaining += amount;
3612         /*
3613          * we may have advanced our local expiration to account for allowed
3614          * spread between our sched_clock and the one on which runtime was
3615          * issued.
3616          */
3617         if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3618                 cfs_rq->runtime_expires = expires;
3619 
3620         return cfs_rq->runtime_remaining > 0;
3621 }
3622 
3623 /*
3624  * Note: This depends on the synchronization provided by sched_clock and the
3625  * fact that rq->clock snapshots this value.
3626  */
3627 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3628 {
3629         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3630 
3631         /* if the deadline is ahead of our clock, nothing to do */
3632         if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3633                 return;
3634 
3635         if (cfs_rq->runtime_remaining < 0)
3636                 return;
3637 
3638         /*
3639          * If the local deadline has passed we have to consider the
3640          * possibility that our sched_clock is 'fast' and the global deadline
3641          * has not truly expired.
3642          *
3643          * Fortunately we can check determine whether this the case by checking
3644          * whether the global deadline has advanced. It is valid to compare
3645          * cfs_b->runtime_expires without any locks since we only care about
3646          * exact equality, so a partial write will still work.
3647          */
3648 
3649         if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3650                 /* extend local deadline, drift is bounded above by 2 ticks */
3651                 cfs_rq->runtime_expires += TICK_NSEC;
3652         } else {
3653                 /* global deadline is ahead, expiration has passed */
3654                 cfs_rq->runtime_remaining = 0;
3655         }
3656 }
3657 
3658 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3659 {
3660         /* dock delta_exec before expiring quota (as it could span periods) */
3661         cfs_rq->runtime_remaining -= delta_exec;
3662         expire_cfs_rq_runtime(cfs_rq);
3663 
3664         if (likely(cfs_rq->runtime_remaining > 0))
3665                 return;
3666 
3667         /*
3668          * if we're unable to extend our runtime we resched so that the active
3669          * hierarchy can be throttled
3670          */
3671         if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3672                 resched_curr(rq_of(cfs_rq));
3673 }
3674 
3675 static __always_inline
3676 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3677 {
3678         if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3679                 return;
3680 
3681         __account_cfs_rq_runtime(cfs_rq, delta_exec);
3682 }
3683 
3684 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3685 {
3686         return cfs_bandwidth_used() && cfs_rq->throttled;
3687 }
3688 
3689 /* check whether cfs_rq, or any parent, is throttled */
3690 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3691 {
3692         return cfs_bandwidth_used() && cfs_rq->throttle_count;
3693 }
3694 
3695 /*
3696  * Ensure that neither of the group entities corresponding to src_cpu or
3697  * dest_cpu are members of a throttled hierarchy when performing group
3698  * load-balance operations.
3699  */
3700 static inline int throttled_lb_pair(struct task_group *tg,
3701                                     int src_cpu, int dest_cpu)
3702 {
3703         struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3704 
3705         src_cfs_rq = tg->cfs_rq[src_cpu];
3706         dest_cfs_rq = tg->cfs_rq[dest_cpu];
3707 
3708         return throttled_hierarchy(src_cfs_rq) ||
3709                throttled_hierarchy(dest_cfs_rq);
3710 }
3711 
3712 /* updated child weight may affect parent so we have to do this bottom up */
3713 static int tg_unthrottle_up(struct task_group *tg, void *data)
3714 {
3715         struct rq *rq = data;
3716         struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3717 
3718         cfs_rq->throttle_count--;
3719 #ifdef CONFIG_SMP
3720         if (!cfs_rq->throttle_count) {
3721                 /* adjust cfs_rq_clock_task() */
3722                 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3723                                              cfs_rq->throttled_clock_task;
3724         }
3725 #endif
3726 
3727         return 0;
3728 }
3729 
3730 static int tg_throttle_down(struct task_group *tg, void *data)
3731 {
3732         struct rq *rq = data;
3733         struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3734 
3735         /* group is entering throttled state, stop time */
3736         if (!cfs_rq->throttle_count)
3737                 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3738         cfs_rq->throttle_count++;
3739 
3740         return 0;
3741 }
3742 
3743 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3744 {
3745         struct rq *rq = rq_of(cfs_rq);
3746         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3747         struct sched_entity *se;
3748         long task_delta, dequeue = 1;
3749         bool empty;
3750 
3751         se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3752 
3753         /* freeze hierarchy runnable averages while throttled */
3754         rcu_read_lock();
3755         walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3756         rcu_read_unlock();
3757 
3758         task_delta = cfs_rq->h_nr_running;
3759         for_each_sched_entity(se) {
3760                 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3761                 /* throttled entity or throttle-on-deactivate */
3762                 if (!se->on_rq)
3763                         break;
3764 
3765                 if (dequeue)
3766                         dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3767                 qcfs_rq->h_nr_running -= task_delta;
3768 
3769                 if (qcfs_rq->load.weight)
3770                         dequeue = 0;
3771         }
3772 
3773         if (!se)
3774                 sub_nr_running(rq, task_delta);
3775 
3776         cfs_rq->throttled = 1;
3777         cfs_rq->throttled_clock = rq_clock(rq);
3778         raw_spin_lock(&cfs_b->lock);
3779         empty = list_empty(&cfs_b->throttled_cfs_rq);
3780 
3781         /*
3782          * Add to the _head_ of the list, so that an already-started
3783          * distribute_cfs_runtime will not see us
3784          */
3785         list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3786 
3787         /*
3788          * If we're the first throttled task, make sure the bandwidth
3789          * timer is running.
3790          */
3791         if (empty)
3792                 start_cfs_bandwidth(cfs_b);
3793 
3794         raw_spin_unlock(&cfs_b->lock);
3795 }
3796 
3797 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3798 {
3799         struct rq *rq = rq_of(cfs_rq);
3800         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3801         struct sched_entity *se;
3802         int enqueue = 1;
3803         long task_delta;
3804 
3805         se = cfs_rq->tg->se[cpu_of(rq)];
3806 
3807         cfs_rq->throttled = 0;
3808 
3809         update_rq_clock(rq);
3810 
3811         raw_spin_lock(&cfs_b->lock);
3812         cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3813         list_del_rcu(&cfs_rq->throttled_list);
3814         raw_spin_unlock(&cfs_b->lock);
3815 
3816         /* update hierarchical throttle state */
3817         walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3818 
3819         if (!cfs_rq->load.weight)
3820                 return;
3821 
3822         task_delta = cfs_rq->h_nr_running;
3823         for_each_sched_entity(se) {
3824                 if (se->on_rq)
3825                         enqueue = 0;
3826 
3827                 cfs_rq = cfs_rq_of(se);
3828                 if (enqueue)
3829                         enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3830                 cfs_rq->h_nr_running += task_delta;
3831 
3832                 if (cfs_rq_throttled(cfs_rq))
3833                         break;
3834         }
3835 
3836         if (!se)
3837                 add_nr_running(rq, task_delta);
3838 
3839         /* determine whether we need to wake up potentially idle cpu */
3840         if (rq->curr == rq->idle && rq->cfs.nr_running)
3841                 resched_curr(rq);
3842 }
3843 
3844 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3845                 u64 remaining, u64 expires)
3846 {
3847         struct cfs_rq *cfs_rq;
3848         u64 runtime;
3849         u64 starting_runtime = remaining;
3850 
3851         rcu_read_lock();
3852         list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3853                                 throttled_list) {
3854                 struct rq *rq = rq_of(cfs_rq);
3855 
3856                 raw_spin_lock(&rq->lock);
3857                 if (!cfs_rq_throttled(cfs_rq))
3858                         goto next;
3859 
3860                 runtime = -cfs_rq->runtime_remaining + 1;
3861                 if (runtime > remaining)
3862                         runtime = remaining;
3863                 remaining -= runtime;
3864 
3865                 cfs_rq->runtime_remaining += runtime;
3866                 cfs_rq->runtime_expires = expires;
3867 
3868                 /* we check whether we're throttled above */
3869                 if (cfs_rq->runtime_remaining > 0)
3870                         unthrottle_cfs_rq(cfs_rq);
3871 
3872 next:
3873                 raw_spin_unlock(&rq->lock);
3874 
3875                 if (!remaining)
3876                         break;
3877         }
3878         rcu_read_unlock();
3879 
3880         return starting_runtime - remaining;
3881 }
3882 
3883 /*
3884  * Responsible for refilling a task_group's bandwidth and unthrottling its
3885  * cfs_rqs as appropriate. If there has been no activity within the last
3886  * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3887  * used to track this state.
3888  */
3889 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3890 {
3891         u64 runtime, runtime_expires;
3892         int throttled;
3893 
3894         /* no need to continue the timer with no bandwidth constraint */
3895         if (cfs_b->quota == RUNTIME_INF)
3896                 goto out_deactivate;
3897 
3898         throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3899         cfs_b->nr_periods += overrun;
3900 
3901         /*
3902          * idle depends on !throttled (for the case of a large deficit), and if
3903          * we're going inactive then everything else can be deferred
3904          */
3905         if (cfs_b->idle && !throttled)
3906                 goto out_deactivate;
3907 
3908         __refill_cfs_bandwidth_runtime(cfs_b);
3909 
3910         if (!throttled) {
3911                 /* mark as potentially idle for the upcoming period */
3912                 cfs_b->idle = 1;
3913                 return 0;
3914         }
3915 
3916         /* account preceding periods in which throttling occurred */
3917         cfs_b->nr_throttled += overrun;
3918 
3919         runtime_expires = cfs_b->runtime_expires;
3920 
3921         /*
3922          * This check is repeated as we are holding onto the new bandwidth while
3923          * we unthrottle. This can potentially race with an unthrottled group
3924          * trying to acquire new bandwidth from the global pool. This can result
3925          * in us over-using our runtime if it is all used during this loop, but
3926          * only by limited amounts in that extreme case.
3927          */
3928         while (throttled && cfs_b->runtime > 0) {
3929                 runtime = cfs_b->runtime;
3930                 raw_spin_unlock(&cfs_b->lock);
3931                 /* we can't nest cfs_b->lock while distributing bandwidth */
3932                 runtime = distribute_cfs_runtime(cfs_b, runtime,
3933                                                  runtime_expires);
3934                 raw_spin_lock(&cfs_b->lock);
3935 
3936                 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3937 
3938                 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3939         }
3940 
3941         /*
3942          * While we are ensured activity in the period following an
3943          * unthrottle, this also covers the case in which the new bandwidth is
3944          * insufficient to cover the existing bandwidth deficit.  (Forcing the
3945          * timer to remain active while there are any throttled entities.)
3946          */
3947         cfs_b->idle = 0;
3948 
3949         return 0;
3950 
3951 out_deactivate:
3952         return 1;
3953 }
3954 
3955 /* a cfs_rq won't donate quota below this amount */
3956 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3957 /* minimum remaining period time to redistribute slack quota */
3958 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3959 /* how long we wait to gather additional slack before distributing */
3960 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3961 
3962 /*
3963  * Are we near the end of the current quota period?
3964  *
3965  * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3966  * hrtimer base being cleared by hrtimer_start. In the case of
3967  * migrate_hrtimers, base is never cleared, so we are fine.
3968  */
3969 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3970 {
3971         struct hrtimer *refresh_timer = &cfs_b->period_timer;
3972         u64 remaining;
3973 
3974         /* if the call-back is running a quota refresh is already occurring */
3975         if (hrtimer_callback_running(refresh_timer))
3976                 return 1;
3977 
3978         /* is a quota refresh about to occur? */
3979         remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3980         if (remaining < min_expire)
3981                 return 1;
3982 
3983         return 0;
3984 }
3985 
3986 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3987 {
3988         u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3989 
3990         /* if there's a quota refresh soon don't bother with slack */
3991         if (runtime_refresh_within(cfs_b, min_left))
3992                 return;
3993 
3994         hrtimer_start(&cfs_b->slack_timer,
3995                         ns_to_ktime(cfs_bandwidth_slack_period),
3996                         HRTIMER_MODE_REL);
3997 }
3998 
3999 /* we know any runtime found here is valid as update_curr() precedes return */
4000 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4001 {
4002         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4003         s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4004 
4005         if (slack_runtime <= 0)
4006                 return;
4007 
4008         raw_spin_lock(&cfs_b->lock);
4009         if (cfs_b->quota != RUNTIME_INF &&
4010             cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4011                 cfs_b->runtime += slack_runtime;
4012 
4013                 /* we are under rq->lock, defer unthrottling using a timer */
4014                 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4015                     !list_empty(&cfs_b->throttled_cfs_rq))
4016                         start_cfs_slack_bandwidth(cfs_b);
4017         }
4018         raw_spin_unlock(&cfs_b->lock);
4019 
4020         /* even if it's not valid for return we don't want to try again */
4021         cfs_rq->runtime_remaining -= slack_runtime;
4022 }
4023 
4024 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4025 {
4026         if (!cfs_bandwidth_used())
4027                 return;
4028 
4029         if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4030                 return;
4031 
4032         __return_cfs_rq_runtime(cfs_rq);
4033 }
4034 
4035 /*
4036  * This is done with a timer (instead of inline with bandwidth return) since
4037  * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4038  */
4039 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4040 {
4041         u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4042         u64 expires;
4043 
4044         /* confirm we're still not at a refresh boundary */
4045         raw_spin_lock(&cfs_b->lock);
4046         if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4047                 raw_spin_unlock(&cfs_b->lock);
4048                 return;
4049         }
4050 
4051         if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4052                 runtime = cfs_b->runtime;
4053 
4054         expires = cfs_b->runtime_expires;
4055         raw_spin_unlock(&cfs_b->lock);
4056 
4057         if (!runtime)
4058                 return;
4059 
4060         runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4061 
4062         raw_spin_lock(&cfs_b->lock);
4063         if (expires == cfs_b->runtime_expires)
4064                 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4065         raw_spin_unlock(&cfs_b->lock);
4066 }
4067 
4068 /*
4069  * When a group wakes up we want to make sure that its quota is not already
4070  * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4071  * runtime as update_curr() throttling can not not trigger until it's on-rq.
4072  */
4073 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4074 {
4075         if (!cfs_bandwidth_used())
4076                 return;
4077 
4078         /* an active group must be handled by the update_curr()->put() path */
4079         if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4080                 return;
4081 
4082         /* ensure the group is not already throttled */
4083         if (cfs_rq_throttled(cfs_rq))
4084                 return;
4085 
4086         /* update runtime allocation */
4087         account_cfs_rq_runtime(cfs_rq, 0);
4088         if (cfs_rq->runtime_remaining <= 0)
4089                 throttle_cfs_rq(cfs_rq);
4090 }
4091 
4092 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4093 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4094 {
4095         if (!cfs_bandwidth_used())
4096                 return false;
4097 
4098         if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4099                 return false;
4100 
4101         /*
4102          * it's possible for a throttled entity to be forced into a running
4103          * state (e.g. set_curr_task), in this case we're finished.
4104          */
4105         if (cfs_rq_throttled(cfs_rq))
4106                 return true;
4107 
4108         throttle_cfs_rq(cfs_rq);
4109         return true;
4110 }
4111 
4112 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4113 {
4114         struct cfs_bandwidth *cfs_b =
4115                 container_of(timer, struct cfs_bandwidth, slack_timer);
4116 
4117         do_sched_cfs_slack_timer(cfs_b);
4118 
4119         return HRTIMER_NORESTART;
4120 }
4121 
4122 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4123 {
4124         struct cfs_bandwidth *cfs_b =
4125                 container_of(timer, struct cfs_bandwidth, period_timer);
4126         int overrun;
4127         int idle = 0;
4128 
4129         raw_spin_lock(&cfs_b->lock);
4130         for (;;) {
4131                 overrun = hrtimer_forward_now(timer, cfs_b->period);
4132                 if (!overrun)
4133                         break;
4134 
4135                 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4136         }
4137         if (idle)
4138                 cfs_b->period_active = 0;
4139         raw_spin_unlock(&cfs_b->lock);
4140 
4141         return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4142 }
4143 
4144 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4145 {
4146         raw_spin_lock_init(&cfs_b->lock);
4147         cfs_b->runtime = 0;
4148         cfs_b->quota = RUNTIME_INF;
4149         cfs_b->period = ns_to_ktime(default_cfs_period());
4150 
4151         INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4152         hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4153         cfs_b->period_timer.function = sched_cfs_period_timer;
4154         hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4155         cfs_b->slack_timer.function = sched_cfs_slack_timer;
4156 }
4157 
4158 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4159 {
4160         cfs_rq->runtime_enabled = 0;
4161         INIT_LIST_HEAD(&cfs_rq->throttled_list);
4162 }
4163 
4164 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4165 {
4166         lockdep_assert_held(&cfs_b->lock);
4167 
4168         if (!cfs_b->period_active) {
4169                 cfs_b->period_active = 1;
4170                 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4171                 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4172         }
4173 }
4174 
4175 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4176 {
4177         /* init_cfs_bandwidth() was not called */
4178         if (!cfs_b->throttled_cfs_rq.next)
4179                 return;
4180 
4181         hrtimer_cancel(&cfs_b->period_timer);
4182         hrtimer_cancel(&cfs_b->slack_timer);
4183 }
4184 
4185 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4186 {
4187         struct cfs_rq *cfs_rq;
4188 
4189         for_each_leaf_cfs_rq(rq, cfs_rq) {
4190                 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4191 
4192                 raw_spin_lock(&cfs_b->lock);
4193                 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4194                 raw_spin_unlock(&cfs_b->lock);
4195         }
4196 }
4197 
4198 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4199 {
4200         struct cfs_rq *cfs_rq;
4201 
4202         for_each_leaf_cfs_rq(rq, cfs_rq) {
4203                 if (!cfs_rq->runtime_enabled)
4204                         continue;
4205 
4206                 /*
4207                  * clock_task is not advancing so we just need to make sure
4208                  * there's some valid quota amount
4209                  */
4210                 cfs_rq->runtime_remaining = 1;
4211                 /*
4212                  * Offline rq is schedulable till cpu is completely disabled
4213                  * in take_cpu_down(), so we prevent new cfs throttling here.
4214                  */
4215                 cfs_rq->runtime_enabled = 0;
4216 
4217                 if (cfs_rq_throttled(cfs_rq))
4218                         unthrottle_cfs_rq(cfs_rq);
4219         }
4220 }
4221 
4222 #else /* CONFIG_CFS_BANDWIDTH */
4223 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4224 {
4225         return rq_clock_task(rq_of(cfs_rq));
4226 }
4227 
4228 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4229 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4230 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4231 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4232 
4233 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4234 {
4235         return 0;
4236 }
4237 
4238 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4239 {
4240         return 0;
4241 }
4242 
4243 static inline int throttled_lb_pair(struct task_group *tg,
4244                                     int src_cpu, int dest_cpu)
4245 {
4246         return 0;
4247 }
4248 
4249 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4250 
4251 #ifdef CONFIG_FAIR_GROUP_SCHED
4252 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4253 #endif
4254 
4255 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4256 {
4257         return NULL;
4258 }
4259 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4260 static inline void update_runtime_enabled(struct rq *rq) {}
4261 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4262 
4263 #endif /* CONFIG_CFS_BANDWIDTH */
4264 
4265 /**************************************************
4266  * CFS operations on tasks:
4267  */
4268 
4269 #ifdef CONFIG_SCHED_HRTICK
4270 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4271 {
4272         struct sched_entity *se = &p->se;
4273         struct cfs_rq *cfs_rq = cfs_rq_of(se);
4274 
4275         WARN_ON(task_rq(p) != rq);
4276 
4277         if (cfs_rq->nr_running > 1) {
4278                 u64 slice = sched_slice(cfs_rq, se);
4279                 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4280                 s64 delta = slice - ran;
4281 
4282                 if (delta < 0) {
4283                         if (rq->curr == p)
4284                                 resched_curr(rq);
4285                         return;
4286                 }
4287                 hrtick_start(rq, delta);
4288         }
4289 }
4290 
4291 /*
4292  * called from enqueue/dequeue and updates the hrtick when the
4293  * current task is from our class and nr_running is low enough
4294  * to matter.
4295  */
4296 static void hrtick_update(struct rq *rq)
4297 {
4298         struct task_struct *curr = rq->curr;
4299 
4300         if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4301                 return;
4302 
4303         if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4304                 hrtick_start_fair(rq, curr);
4305 }
4306 #else /* !CONFIG_SCHED_HRTICK */
4307 static inline void
4308 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4309 {
4310 }
4311 
4312 static inline void hrtick_update(struct rq *rq)
4313 {
4314 }
4315 #endif
4316 
4317 /*
4318  * The enqueue_task method is called before nr_running is
4319  * increased. Here we update the fair scheduling stats and
4320  * then put the task into the rbtree:
4321  */
4322 static void
4323 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4324 {
4325         struct cfs_rq *cfs_rq;
4326         struct sched_entity *se = &p->se;
4327 
4328         for_each_sched_entity(se) {
4329                 if (se->on_rq)
4330                         break;
4331                 cfs_rq = cfs_rq_of(se);
4332                 enqueue_entity(cfs_rq, se, flags);
4333 
4334                 /*
4335                  * end evaluation on encountering a throttled cfs_rq
4336                  *
4337                  * note: in the case of encountering a throttled cfs_rq we will
4338                  * post the final h_nr_running increment below.
4339                 */
4340                 if (cfs_rq_throttled(cfs_rq))
4341                         break;
4342                 cfs_rq->h_nr_running++;
4343 
4344                 flags = ENQUEUE_WAKEUP;
4345         }
4346 
4347         for_each_sched_entity(se) {
4348                 cfs_rq = cfs_rq_of(se);
4349                 cfs_rq->h_nr_running++;
4350 
4351                 if (cfs_rq_throttled(cfs_rq))
4352                         break;
4353 
4354                 update_load_avg(se, 1);
4355                 update_cfs_shares(cfs_rq);
4356         }
4357 
4358         if (!se)
4359                 add_nr_running(rq, 1);
4360 
4361         hrtick_update(rq);
4362 }
4363 
4364 static void set_next_buddy(struct sched_entity *se);
4365 
4366 /*
4367  * The dequeue_task method is called before nr_running is
4368  * decreased. We remove the task from the rbtree and
4369  * update the fair scheduling stats:
4370  */
4371 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4372 {
4373         struct cfs_rq *cfs_rq;
4374         struct sched_entity *se = &p->se;
4375         int task_sleep = flags & DEQUEUE_SLEEP;
4376 
4377         for_each_sched_entity(se) {
4378                 cfs_rq = cfs_rq_of(se);
4379                 dequeue_entity(cfs_rq, se, flags);
4380 
4381                 /*
4382                  * end evaluation on encountering a throttled cfs_rq
4383                  *
4384                  * note: in the case of encountering a throttled cfs_rq we will
4385                  * post the final h_nr_running decrement below.
4386                 */
4387                 if (cfs_rq_throttled(cfs_rq))
4388                         break;
4389                 cfs_rq->h_nr_running--;
4390 
4391                 /* Don't dequeue parent if it has other entities besides us */
4392                 if (cfs_rq->load.weight) {
4393                         /*
4394                          * Bias pick_next to pick a task from this cfs_rq, as
4395                          * p is sleeping when it is within its sched_slice.
4396                          */
4397                         if (task_sleep && parent_entity(se))
4398                                 set_next_buddy(parent_entity(se));
4399 
4400                         /* avoid re-evaluating load for this entity */
4401                         se = parent_entity(se);
4402                         break;
4403                 }
4404                 flags |= DEQUEUE_SLEEP;
4405         }
4406 
4407         for_each_sched_entity(se) {
4408                 cfs_rq = cfs_rq_of(se);
4409                 cfs_rq->h_nr_running--;
4410 
4411                 if (cfs_rq_throttled(cfs_rq))
4412                         break;
4413 
4414                 update_load_avg(se, 1);
4415                 update_cfs_shares(cfs_rq);
4416         }
4417 
4418         if (!se)
4419                 sub_nr_running(rq, 1);
4420 
4421         hrtick_update(rq);
4422 }
4423 
4424 #ifdef CONFIG_SMP
4425 
4426 /*
4427  * per rq 'load' arrray crap; XXX kill this.
4428  */
4429 
4430 /*
4431  * The exact cpuload calculated at every tick would be:
4432  *
4433  *   load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
4434  *
4435  * If a cpu misses updates for n ticks (as it was idle) and update gets
4436  * called on the n+1-th tick when cpu may be busy, then we have:
4437  *
4438  *   load_n   = (1 - 1/2^i)^n * load_0
4439  *   load_n+1 = (1 - 1/2^i)   * load_n + (1/2^i) * cur_load
4440  *
4441  * decay_load_missed() below does efficient calculation of
4442  *
4443  *   load' = (1 - 1/2^i)^n * load
4444  *
4445  * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
4446  * This allows us to precompute the above in said factors, thereby allowing the
4447  * reduction of an arbitrary n in O(log_2 n) steps. (See also
4448  * fixed_power_int())
4449  *
4450  * The calculation is approximated on a 128 point scale.
4451  */
4452 #define DEGRADE_SHIFT           7
4453 
4454 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4455 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4456         {   0,   0,  0,  0,  0,  0, 0, 0 },
4457         {  64,  32,  8,  0,  0,  0, 0, 0 },
4458         {  96,  72, 40, 12,  1,  0, 0, 0 },
4459         { 112,  98, 75, 43, 15,  1, 0, 0 },
4460         { 120, 112, 98, 76, 45, 16, 2, 0 }
4461 };
4462 
4463 /*
4464  * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4465  * would be when CPU is idle and so we just decay the old load without
4466  * adding any new load.
4467  */
4468 static unsigned long
4469 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4470 {
4471         int j = 0;
4472 
4473         if (!missed_updates)
4474                 return load;
4475 
4476         if (missed_updates >= degrade_zero_ticks[idx])
4477                 return 0;
4478 
4479         if (idx == 1)
4480                 return load >> missed_updates;
4481 
4482         while (missed_updates) {
4483                 if (missed_updates % 2)
4484                         load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4485 
4486                 missed_updates >>= 1;
4487                 j++;
4488         }
4489         return load;
4490 }
4491 
4492 /**
4493  * __update_cpu_load - update the rq->cpu_load[] statistics
4494  * @this_rq: The rq to update statistics for
4495  * @this_load: The current load
4496  * @pending_updates: The number of missed updates
4497  * @active: !0 for NOHZ_FULL
4498  *
4499  * Update rq->cpu_load[] statistics. This function is usually called every
4500  * scheduler tick (TICK_NSEC).
4501  *
4502  * This function computes a decaying average:
4503  *
4504  *   load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
4505  *
4506  * Because of NOHZ it might not get called on every tick which gives need for
4507  * the @pending_updates argument.
4508  *
4509  *   load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
4510  *             = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
4511  *             = A * (A * load[i]_n-2 + B) + B
4512  *             = A * (A * (A * load[i]_n-3 + B) + B) + B
4513  *             = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
4514  *             = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
4515  *             = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
4516  *             = (1 - 1/2^i)^n * (load[i]_0 - load) + load
4517  *
4518  * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
4519  * any change in load would have resulted in the tick being turned back on.
4520  *
4521  * For regular NOHZ, this reduces to:
4522  *
4523  *   load[i]_n = (1 - 1/2^i)^n * load[i]_0
4524  *
4525  * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
4526  * term. See the @active paramter.
4527  */
4528 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4529                               unsigned long pending_updates, int active)
4530 {
4531         unsigned long tickless_load = active ? this_rq->cpu_load[0] : 0;
4532         int i, scale;
4533 
4534         this_rq->nr_load_updates++;
4535 
4536         /* Update our load: */
4537         this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4538         for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4539                 unsigned long old_load, new_load;
4540 
4541                 /* scale is effectively 1 << i now, and >> i divides by scale */
4542 
4543                 old_load = this_rq->cpu_load[i];
4544                 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4545                 if (tickless_load) {
4546                         old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
4547                         /*
4548                          * old_load can never be a negative value because a
4549                          * decayed tickless_load cannot be greater than the
4550                          * original tickless_load.
4551                          */
4552                         old_load += tickless_load;
4553                 }
4554                 new_load = this_load;
4555                 /*
4556                  * Round up the averaging division if load is increasing. This
4557                  * prevents us from getting stuck on 9 if the load is 10, for
4558                  * example.
4559                  */
4560                 if (new_load > old_load)
4561                         new_load += scale - 1;
4562 
4563                 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4564         }
4565 
4566         sched_avg_update(this_rq);
4567 }
4568 
4569 /* Used instead of source_load when we know the type == 0 */
4570 static unsigned long weighted_cpuload(const int cpu)
4571 {
4572         return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4573 }
4574 
4575 #ifdef CONFIG_NO_HZ_COMMON
4576 static void __update_cpu_load_nohz(struct rq *this_rq,
4577                                    unsigned long curr_jiffies,
4578                                    unsigned long load,
4579                                    int active)
4580 {
4581         unsigned long pending_updates;
4582 
4583         pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4584         if (pending_updates) {
4585                 this_rq->last_load_update_tick = curr_jiffies;
4586                 /*
4587                  * In the regular NOHZ case, we were idle, this means load 0.
4588                  * In the NOHZ_FULL case, we were non-idle, we should consider
4589                  * its weighted load.
4590                  */
4591                 __update_cpu_load(this_rq, load, pending_updates, active);
4592         }
4593 }
4594 
4595 /*
4596  * There is no sane way to deal with nohz on smp when using jiffies because the
4597  * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4598  * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4599  *
4600  * Therefore we cannot use the delta approach from the regular tick since that
4601  * would seriously skew the load calculation. However we'll make do for those
4602  * updates happening while idle (nohz_idle_balance) or coming out of idle
4603  * (tick_nohz_idle_exit).
4604  *
4605  * This means we might still be one tick off for nohz periods.
4606  */
4607 
4608 /*
4609  * Called from nohz_idle_balance() to update the load ratings before doing the
4610  * idle balance.
4611  */
4612 static void update_cpu_load_idle(struct rq *this_rq)
4613 {
4614         /*
4615          * bail if there's load or we're actually up-to-date.
4616          */
4617         if (weighted_cpuload(cpu_of(this_rq)))
4618                 return;
4619 
4620         __update_cpu_load_nohz(this_rq, READ_ONCE(jiffies), 0, 0);
4621 }
4622 
4623 /*
4624  * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4625  */
4626 void update_cpu_load_nohz(int active)
4627 {
4628         struct rq *this_rq = this_rq();
4629         unsigned long curr_jiffies = READ_ONCE(jiffies);
4630         unsigned long load = active ? weighted_cpuload(cpu_of(this_rq)) : 0;
4631 
4632         if (curr_jiffies == this_rq->last_load_update_tick)
4633                 return;
4634 
4635         raw_spin_lock(&this_rq->lock);
4636         __update_cpu_load_nohz(this_rq, curr_jiffies, load, active);
4637         raw_spin_unlock(&this_rq->lock);
4638 }
4639 #endif /* CONFIG_NO_HZ */
4640 
4641 /*
4642  * Called from scheduler_tick()
4643  */
4644 void update_cpu_load_active(struct rq *this_rq)
4645 {
4646         unsigned long load = weighted_cpuload(cpu_of(this_rq));
4647         /*
4648          * See the mess around update_cpu_load_idle() / update_cpu_load_nohz().
4649          */
4650         this_rq->last_load_update_tick = jiffies;
4651         __update_cpu_load(this_rq, load, 1, 1);
4652 }
4653 
4654 /*
4655  * Return a low guess at the load of a migration-source cpu weighted
4656  * according to the scheduling class and "nice" value.
4657  *
4658  * We want to under-estimate the load of migration sources, to
4659  * balance conservatively.
4660  */
4661 static unsigned long source_load(int cpu, int type)
4662 {
4663         struct rq *rq = cpu_rq(cpu);
4664         unsigned long total = weighted_cpuload(cpu);
4665 
4666         if (type == 0 || !sched_feat(LB_BIAS))
4667                 return total;
4668 
4669         return min(rq->cpu_load[type-1], total);
4670 }
4671 
4672 /*
4673  * Return a high guess at the load of a migration-target cpu weighted
4674  * according to the scheduling class and "nice" value.
4675  */
4676 static unsigned long target_load(int cpu, int type)
4677 {
4678         struct rq *rq = cpu_rq(cpu);
4679         unsigned long total = weighted_cpuload(cpu);
4680 
4681         if (type == 0 || !sched_feat(LB_BIAS))
4682                 return total;
4683 
4684         return max(rq->cpu_load[type-1], total);
4685 }
4686 
4687 static unsigned long capacity_of(int cpu)
4688 {
4689         return cpu_rq(cpu)->cpu_capacity;
4690 }
4691 
4692 static unsigned long capacity_orig_of(int cpu)
4693 {
4694         return cpu_rq(cpu)->cpu_capacity_orig;
4695 }
4696 
4697 static unsigned long cpu_avg_load_per_task(int cpu)
4698 {
4699         struct rq *rq = cpu_rq(cpu);
4700         unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4701         unsigned long load_avg = weighted_cpuload(cpu);
4702 
4703         if (nr_running)
4704                 return load_avg / nr_running;
4705 
4706         return 0;
4707 }
4708 
4709 static void record_wakee(struct task_struct *p)
4710 {
4711         /*
4712          * Rough decay (wiping) for cost saving, don't worry
4713          * about the boundary, really active task won't care
4714          * about the loss.
4715          */
4716         if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4717                 current->wakee_flips >>= 1;
4718                 current->wakee_flip_decay_ts = jiffies;
4719         }
4720 
4721         if (current->last_wakee != p) {
4722                 current->last_wakee = p;
4723                 current->wakee_flips++;
4724         }
4725 }
4726 
4727 static void task_waking_fair(struct task_struct *p)
4728 {
4729         struct sched_entity *se = &p->se;
4730         struct cfs_rq *cfs_rq = cfs_rq_of(se);
4731         u64 min_vruntime;
4732 
4733 #ifndef CONFIG_64BIT
4734         u64 min_vruntime_copy;
4735 
4736         do {
4737                 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4738                 smp_rmb();
4739                 min_vruntime = cfs_rq->min_vruntime;
4740         } while (min_vruntime != min_vruntime_copy);
4741 #else
4742         min_vruntime = cfs_rq->min_vruntime;
4743 #endif
4744 
4745         se->vruntime -= min_vruntime;
4746         record_wakee(p);
4747 }
4748 
4749 #ifdef CONFIG_FAIR_GROUP_SCHED
4750 /*
4751  * effective_load() calculates the load change as seen from the root_task_group
4752  *
4753  * Adding load to a group doesn't make a group heavier, but can cause movement
4754  * of group shares between cpus. Assuming the shares were perfectly aligned one
4755  * can calculate the shift in shares.
4756  *
4757  * Calculate the effective load difference if @wl is added (subtracted) to @tg
4758  * on this @cpu and results in a total addition (subtraction) of @wg to the
4759  * total group weight.
4760  *
4761  * Given a runqueue weight distribution (rw_i) we can compute a shares
4762  * distribution (s_i) using:
4763  *
4764  *   s_i = rw_i / \Sum rw_j                                             (1)
4765  *
4766  * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4767  * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4768  * shares distribution (s_i):
4769  *
4770  *   rw_i = {   2,   4,   1,   0 }
4771  *   s_i  = { 2/7, 4/7, 1/7,   0 }
4772  *
4773  * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4774  * task used to run on and the CPU the waker is running on), we need to
4775  * compute the effect of waking a task on either CPU and, in case of a sync
4776  * wakeup, compute the effect of the current task going to sleep.
4777  *
4778  * So for a change of @wl to the local @cpu with an overall group weight change
4779  * of @wl we can compute the new shares distribution (s'_i) using:
4780  *
4781  *   s'_i = (rw_i + @wl) / (@wg + \Sum rw_j)                            (2)
4782  *
4783  * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4784  * differences in waking a task to CPU 0. The additional task changes the
4785  * weight and shares distributions like:
4786  *
4787  *   rw'_i = {   3,   4,   1,   0 }
4788  *   s'_i  = { 3/8, 4/8, 1/8,   0 }
4789  *
4790  * We can then compute the difference in effective weight by using:
4791  *
4792  *   dw_i = S * (s'_i - s_i)                                            (3)
4793  *
4794  * Where 'S' is the group weight as seen by its parent.
4795  *
4796  * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4797  * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4798  * 4/7) times the weight of the group.
4799  */
4800 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4801 {
4802         struct sched_entity *se = tg->se[cpu];
4803 
4804         if (!tg->parent)        /* the trivial, non-cgroup case */
4805                 return wl;
4806 
4807         for_each_sched_entity(se) {
4808                 long w, W;
4809 
4810                 tg = se->my_q->tg;
4811 
4812                 /*
4813                  * W = @wg + \Sum rw_j
4814                  */
4815                 W = wg + calc_tg_weight(tg, se->my_q);
4816 
4817                 /*
4818                  * w = rw_i + @wl
4819                  */
4820                 w = cfs_rq_load_avg(se->my_q) + wl;
4821 
4822                 /*
4823                  * wl = S * s'_i; see (2)
4824                  */
4825                 if (W > 0 && w < W)
4826                         wl = (w * (long)tg->shares) / W;
4827                 else
4828                         wl = tg->shares;
4829 
4830                 /*
4831                  * Per the above, wl is the new se->load.weight value; since
4832                  * those are clipped to [MIN_SHARES, ...) do so now. See
4833                  * calc_cfs_shares().
4834                  */
4835                 if (wl < MIN_SHARES)
4836                         wl = MIN_SHARES;
4837 
4838                 /*
4839                  * wl = dw_i = S * (s'_i - s_i); see (3)
4840                  */
4841                 wl -= se->avg.load_avg;
4842 
4843                 /*
4844                  * Recursively apply this logic to all parent groups to compute
4845                  * the final effective load change on the root group. Since
4846                  * only the @tg group gets extra weight, all parent groups can
4847                  * only redistribute existing shares. @wl is the shift in shares
4848                  * resulting from this level per the above.
4849                  */
4850                 wg = 0;
4851         }
4852 
4853         return wl;
4854 }
4855 #else
4856 
4857 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4858 {
4859         return wl;
4860 }
4861 
4862 #endif
4863 
4864 /*
4865  * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
4866  * A waker of many should wake a different task than the one last awakened
4867  * at a frequency roughly N times higher than one of its wakees.  In order
4868  * to determine whether we should let the load spread vs consolodating to
4869  * shared cache, we look for a minimum 'flip' frequency of llc_size in one
4870  * partner, and a factor of lls_size higher frequency in the other.  With
4871  * both conditions met, we can be relatively sure that the relationship is
4872  * non-monogamous, with partner count exceeding socket size.  Waker/wakee
4873  * being client/server, worker/dispatcher, interrupt source or whatever is
4874  * irrelevant, spread criteria is apparent partner count exceeds socket size.
4875  */
4876 static int wake_wide(struct task_struct *p)
4877 {
4878         unsigned int master = current->wakee_flips;
4879         unsigned int slave = p->wakee_flips;
4880         int factor = this_cpu_read(sd_llc_size);
4881 
4882         if (master < slave)
4883                 swap(master, slave);
4884         if (slave < factor || master < slave * factor)
4885                 return 0;
4886         return 1;
4887 }
4888 
4889 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4890 {
4891         s64 this_load, load;
4892         s64 this_eff_load, prev_eff_load;
4893         int idx, this_cpu, prev_cpu;
4894         struct task_group *tg;
4895         unsigned long weight;
4896         int balanced;
4897 
4898         idx       = sd->wake_idx;
4899         this_cpu  = smp_processor_id();
4900         prev_cpu  = task_cpu(p);
4901         load      = source_load(prev_cpu, idx);
4902         this_load = target_load(this_cpu, idx);
4903 
4904         /*
4905          * If sync wakeup then subtract the (maximum possible)
4906          * effect of the currently running task from the load
4907          * of the current CPU:
4908          */
4909         if (sync) {
4910                 tg = task_group(current);
4911                 weight = current->se.avg.load_avg;
4912 
4913                 this_load += effective_load(tg, this_cpu, -weight, -weight);
4914                 load += effective_load(tg, prev_cpu, 0, -weight);
4915         }
4916 
4917         tg = task_group(p);
4918         weight = p->se.avg.load_avg;
4919 
4920         /*
4921          * In low-load situations, where prev_cpu is idle and this_cpu is idle
4922          * due to the sync cause above having dropped this_load to 0, we'll
4923          * always have an imbalance, but there's really nothing you can do
4924          * about that, so that's good too.
4925          *
4926          * Otherwise check if either cpus are near enough in load to allow this
4927          * task to be woken on this_cpu.
4928          */
4929         this_eff_load = 100;
4930         this_eff_load *= capacity_of(prev_cpu);
4931 
4932         prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4933         prev_eff_load *= capacity_of(this_cpu);
4934 
4935         if (this_load > 0) {
4936                 this_eff_load *= this_load +
4937                         effective_load(tg, this_cpu, weight, weight);
4938 
4939                 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4940         }
4941 
4942         balanced = this_eff_load <= prev_eff_load;
4943 
4944         schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4945 
4946         if (!balanced)
4947                 return 0;
4948 
4949         schedstat_inc(sd, ttwu_move_affine);
4950         schedstat_inc(p, se.statistics.nr_wakeups_affine);
4951 
4952         return 1;
4953 }
4954 
4955 /*
4956  * find_idlest_group finds and returns the least busy CPU group within the
4957  * domain.
4958  */
4959 static struct sched_group *
4960 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4961                   int this_cpu, int sd_flag)
4962 {
4963         struct sched_group *idlest = NULL, *group = sd->groups;
4964         unsigned long min_load = ULONG_MAX, this_load = 0;
4965         int load_idx = sd->forkexec_idx;
4966         int imbalance = 100 + (sd->imbalance_pct-100)/2;
4967 
4968         if (sd_flag & SD_BALANCE_WAKE)
4969                 load_idx = sd->wake_idx;
4970 
4971         do {
4972                 unsigned long load, avg_load;
4973                 int local_group;
4974                 int i;
4975 
4976                 /* Skip over this group if it has no CPUs allowed */
4977                 if (!cpumask_intersects(sched_group_cpus(group),
4978                                         tsk_cpus_allowed(p)))
4979                         continue;
4980 
4981                 local_group = cpumask_test_cpu(this_cpu,
4982                                                sched_group_cpus(group));
4983 
4984                 /* Tally up the load of all CPUs in the group */
4985                 avg_load = 0;
4986 
4987                 for_each_cpu(i, sched_group_cpus(group)) {
4988                         /* Bias balancing toward cpus of our domain */
4989                         if (local_group)
4990                                 load = source_load(i, load_idx);
4991                         else
4992                                 load = target_load(i, load_idx);
4993 
4994                         avg_load += load;
4995                 }
4996 
4997                 /* Adjust by relative CPU capacity of the group */
4998                 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
4999 
5000                 if (local_group) {
5001                         this_load = avg_load;
5002                 } else if (avg_load < min_load) {
5003                         min_load = avg_load;
5004                         idlest = group;
5005                 }
5006         } while (group = group->next, group != sd->groups);
5007 
5008         if (!idlest || 100*this_load < imbalance*min_load)
5009                 return NULL;
5010         return idlest;
5011 }
5012 
5013 /*
5014  * find_idlest_cpu - find the idlest cpu among the cpus in group.
5015  */
5016 static int
5017 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5018 {
5019         unsigned long load, min_load = ULONG_MAX;
5020         unsigned int min_exit_latency = UINT_MAX;
5021         u64 latest_idle_timestamp = 0;
5022         int least_loaded_cpu = this_cpu;
5023         int shallowest_idle_cpu = -1;
5024         int i;
5025 
5026         /* Traverse only the allowed CPUs */
5027         for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5028                 if (idle_cpu(i)) {
5029                         struct rq *rq = cpu_rq(i);
5030                         struct cpuidle_state *idle = idle_get_state(rq);
5031                         if (idle && idle->exit_latency < min_exit_latency) {
5032                                 /*
5033                                  * We give priority to a CPU whose idle state
5034                                  * has the smallest exit latency irrespective
5035                                  * of any idle timestamp.
5036                                  */
5037                                 min_exit_latency = idle->exit_latency;
5038                                 latest_idle_timestamp = rq->idle_stamp;
5039                                 shallowest_idle_cpu = i;
5040                         } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5041                                    rq->idle_stamp > latest_idle_timestamp) {
5042                                 /*
5043                                  * If equal or no active idle state, then
5044                                  * the most recently idled CPU might have
5045                                  * a warmer cache.
5046                                  */
5047                                 latest_idle_timestamp = rq->idle_stamp;
5048                                 shallowest_idle_cpu = i;
5049                         }
5050                 } else if (shallowest_idle_cpu == -1) {
5051                         load = weighted_cpuload(i);
5052                         if (load < min_load || (load == min_load && i == this_cpu)) {
5053                                 min_load = load;
5054                                 least_loaded_cpu = i;
5055                         }
5056                 }
5057         }
5058 
5059         return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5060 }
5061 
5062 /*
5063  * Try and locate an idle CPU in the sched_domain.
5064  */
5065 static int select_idle_sibling(struct task_struct *p, int target)
5066 {
5067         struct sched_domain *sd;
5068         struct sched_group *sg;
5069         int i = task_cpu(p);
5070 
5071         if (idle_cpu(target))
5072                 return target;
5073 
5074         /*
5075          * If the prevous cpu is cache affine and idle, don't be stupid.
5076          */
5077         if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
5078                 return i;
5079 
5080         /*
5081          * Otherwise, iterate the domains and find an eligible idle cpu.
5082          *
5083          * A completely idle sched group at higher domains is more
5084          * desirable than an idle group at a lower level, because lower
5085          * domains have smaller groups and usually share hardware
5086          * resources which causes tasks to contend on them, e.g. x86
5087          * hyperthread siblings in the lowest domain (SMT) can contend
5088          * on the shared cpu pipeline.
5089          *
5090          * However, while we prefer idle groups at higher domains
5091          * finding an idle cpu at the lowest domain is still better than
5092          * returning 'target', which we've already established, isn't
5093          * idle.
5094          */
5095         sd = rcu_dereference(per_cpu(sd_llc, target));
5096         for_each_lower_domain(sd) {
5097                 sg = sd->groups;
5098                 do {
5099                         if (!cpumask_intersects(sched_group_cpus(sg),
5100                                                 tsk_cpus_allowed(p)))
5101                                 goto next;
5102 
5103                         /* Ensure the entire group is idle */
5104                         for_each_cpu(i, sched_group_cpus(sg)) {
5105                                 if (i == target || !idle_cpu(i))
5106                                         goto next;
5107                         }
5108 
5109                         /*
5110                          * It doesn't matter which cpu we pick, the
5111                          * whole group is idle.
5112                          */
5113                         target = cpumask_first_and(sched_group_cpus(sg),
5114                                         tsk_cpus_allowed(p));
5115                         goto done;
5116 next:
5117                         sg = sg->next;
5118                 } while (sg != sd->groups);
5119         }
5120 done:
5121         return target;
5122 }
5123 
5124 /*
5125  * cpu_util returns the amount of capacity of a CPU that is used by CFS
5126  * tasks. The unit of the return value must be the one of capacity so we can
5127  * compare the utilization with the capacity of the CPU that is available for
5128  * CFS task (ie cpu_capacity).
5129  *
5130  * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5131  * recent utilization of currently non-runnable tasks on a CPU. It represents
5132  * the amount of utilization of a CPU in the range [0..capacity_orig] where
5133  * capacity_orig is the cpu_capacity available at the highest frequency
5134  * (arch_scale_freq_capacity()).
5135  * The utilization of a CPU converges towards a sum equal to or less than the
5136  * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5137  * the running time on this CPU scaled by capacity_curr.
5138  *
5139  * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5140  * higher than capacity_orig because of unfortunate rounding in
5141  * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5142  * the average stabilizes with the new running time. We need to check that the
5143  * utilization stays within the range of [0..capacity_orig] and cap it if
5144  * necessary. Without utilization capping, a group could be seen as overloaded
5145  * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5146  * available capacity. We allow utilization to overshoot capacity_curr (but not
5147  * capacity_orig) as it useful for predicting the capacity required after task
5148  * migrations (scheduler-driven DVFS).
5149  */
5150 static int cpu_util(int cpu)
5151 {
5152         unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
5153         unsigned long capacity = capacity_orig_of(cpu);
5154 
5155         return (util >= capacity) ? capacity : util;
5156 }
5157 
5158 /*
5159  * select_task_rq_fair: Select target runqueue for the waking task in domains
5160  * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5161  * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5162  *
5163  * Balances load by selecting the idlest cpu in the idlest group, or under
5164  * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5165  *
5166  * Returns the target cpu number.
5167  *
5168  * preempt must be disabled.
5169  */
5170 static int
5171 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5172 {
5173         struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5174         int cpu = smp_processor_id();
5175         int new_cpu = prev_cpu;
5176         int want_affine = 0;
5177         int sync = wake_flags & WF_SYNC;
5178 
5179         if (sd_flag & SD_BALANCE_WAKE)
5180                 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
5181 
5182         rcu_read_lock();
5183         for_each_domain(cpu, tmp) {
5184                 if (!(tmp->flags & SD_LOAD_BALANCE))
5185                         break;
5186 
5187                 /*
5188                  * If both cpu and prev_cpu are part of this domain,
5189                  * cpu is a valid SD_WAKE_AFFINE target.
5190                  */
5191                 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5192                     cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5193                         affine_sd = tmp;
5194                         break;
5195                 }
5196 
5197                 if (tmp->flags & sd_flag)
5198                         sd = tmp;
5199                 else if (!want_affine)
5200                         break;
5201         }
5202 
5203         if (affine_sd) {
5204                 sd = NULL; /* Prefer wake_affine over balance flags */
5205                 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5206                         new_cpu = cpu;
5207         }
5208 
5209         if (!sd) {
5210                 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5211                         new_cpu = select_idle_sibling(p, new_cpu);
5212 
5213         } else while (sd) {
5214                 struct sched_group *group;
5215                 int weight;
5216 
5217                 if (!(sd->flags & sd_flag)) {
5218                         sd = sd->child;
5219                         continue;
5220                 }
5221 
5222                 group = find_idlest_group(sd, p, cpu, sd_flag);
5223                 if (!group) {
5224                         sd = sd->child;
5225                         continue;
5226                 }
5227 
5228                 new_cpu = find_idlest_cpu(group, p, cpu);
5229                 if (new_cpu == -1 || new_cpu == cpu) {
5230                         /* Now try balancing at a lower domain level of cpu */
5231                         sd = sd->child;
5232                         continue;
5233                 }
5234 
5235                 /* Now try balancing at a lower domain level of new_cpu */
5236                 cpu = new_cpu;
5237                 weight = sd->span_weight;
5238                 sd = NULL;
5239                 for_each_domain(cpu, tmp) {
5240                         if (weight <= tmp->span_weight)
5241                                 break;
5242                         if (tmp->flags & sd_flag)
5243                                 sd = tmp;
5244                 }
5245                 /* while loop will break here if sd == NULL */
5246         }
5247         rcu_read_unlock();
5248 
5249         return new_cpu;
5250 }
5251 
5252 /*
5253  * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5254  * cfs_rq_of(p) references at time of call are still valid and identify the
5255  * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
5256  */
5257 static void migrate_task_rq_fair(struct task_struct *p)
5258 {
5259         /*
5260          * We are supposed to update the task to "current" time, then its up to date
5261          * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5262          * what current time is, so simply throw away the out-of-date time. This
5263          * will result in the wakee task is less decayed, but giving the wakee more
5264          * load sounds not bad.
5265          */
5266         remove_entity_load_avg(&p->se);
5267 
5268         /* Tell new CPU we are migrated */
5269         p->se.avg.last_update_time = 0;
5270 
5271         /* We have migrated, no longer consider this task hot */
5272         p->se.exec_start = 0;
5273 }
5274 
5275 static void task_dead_fair(struct task_struct *p)
5276 {
5277         remove_entity_load_avg(&p->se);
5278 }
5279 #endif /* CONFIG_SMP */
5280 
5281 static unsigned long
5282 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5283 {
5284         unsigned long gran = sysctl_sched_wakeup_granularity;
5285 
5286         /*
5287          * Since its curr running now, convert the gran from real-time
5288          * to virtual-time in his units.
5289          *
5290          * By using 'se' instead of 'curr' we penalize light tasks, so
5291          * they get preempted easier. That is, if 'se' < 'curr' then
5292          * the resulting gran will be larger, therefore penalizing the
5293          * lighter, if otoh 'se' > 'curr' then the resulting gran will
5294          * be smaller, again penalizing the lighter task.
5295          *
5296          * This is especially important for buddies when the leftmost
5297          * task is higher priority than the buddy.
5298          */
5299         return calc_delta_fair(gran, se);
5300 }
5301 
5302 /*
5303  * Should 'se' preempt 'curr'.
5304  *
5305  *             |s1
5306  *        |s2
5307  *   |s3
5308  *         g
5309  *      |<--->|c
5310  *
5311  *  w(c, s1) = -1
5312  *  w(c, s2) =  0
5313  *  w(c, s3) =  1
5314  *
5315  */
5316 static int
5317 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5318 {
5319         s64 gran, vdiff = curr->vruntime - se->vruntime;
5320 
5321         if (vdiff <= 0)
5322                 return -1;
5323 
5324         gran = wakeup_gran(curr, se);
5325         if (vdiff > gran)
5326                 return 1;
5327 
5328         return 0;
5329 }
5330 
5331 static void set_last_buddy(struct sched_entity *se)
5332 {
5333         if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5334                 return;
5335 
5336         for_each_sched_entity(se)
5337                 cfs_rq_of(se)->last = se;
5338 }
5339 
5340 static void set_next_buddy(struct sched_entity *se)
5341 {
5342         if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5343                 return;
5344 
5345         for_each_sched_entity(se)
5346                 cfs_rq_of(se)->next = se;
5347 }
5348 
5349 static void set_skip_buddy(struct sched_entity *se)
5350 {
5351         for_each_sched_entity(se)
5352                 cfs_rq_of(se)->skip = se;
5353 }
5354 
5355 /*
5356  * Preempt the current task with a newly woken task if needed:
5357  */
5358 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5359 {
5360         struct task_struct *curr = rq->curr;
5361         struct sched_entity *se = &curr->se, *pse = &p->se;
5362         struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5363         int scale = cfs_rq->nr_running >= sched_nr_latency;
5364         int next_buddy_marked = 0;
5365 
5366         if (unlikely(se == pse))
5367                 return;
5368 
5369         /*
5370          * This is possible from callers such as attach_tasks(), in which we
5371          * unconditionally check_prempt_curr() after an enqueue (which may have
5372          * lead to a throttle).  This both saves work and prevents false
5373          * next-buddy nomination below.
5374          */
5375         if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5376                 return;
5377 
5378         if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5379                 set_next_buddy(pse);
5380                 next_buddy_marked = 1;
5381         }
5382 
5383         /*
5384          * We can come here with TIF_NEED_RESCHED already set from new task
5385          * wake up path.
5386          *
5387          * Note: this also catches the edge-case of curr being in a throttled
5388          * group (e.g. via set_curr_task), since update_curr() (in the
5389          * enqueue of curr) will have resulted in resched being set.  This
5390          * prevents us from potentially nominating it as a false LAST_BUDDY
5391          * below.
5392          */
5393         if (test_tsk_need_resched(curr))
5394                 return;
5395 
5396         /* Idle tasks are by definition preempted by non-idle tasks. */
5397         if (unlikely(curr->policy == SCHED_IDLE) &&
5398             likely(p->policy != SCHED_IDLE))
5399                 goto preempt;
5400 
5401         /*
5402          * Batch and idle tasks do not preempt non-idle tasks (their preemption
5403          * is driven by the tick):
5404          */
5405         if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5406                 return;
5407 
5408         find_matching_se(&se, &pse);
5409         update_curr(cfs_rq_of(se));
5410         BUG_ON(!pse);
5411         if (wakeup_preempt_entity(se, pse) == 1) {
5412                 /*
5413                  * Bias pick_next to pick the sched entity that is
5414                  * triggering this preemption.
5415                  */
5416                 if (!next_buddy_marked)
5417                         set_next_buddy(pse);
5418                 goto preempt;
5419         }
5420 
5421         return;
5422 
5423 preempt:
5424         resched_curr(rq);
5425         /*
5426          * Only set the backward buddy when the current task is still
5427          * on the rq. This can happen when a wakeup gets interleaved
5428          * with schedule on the ->pre_schedule() or idle_balance()
5429          * point, either of which can * drop the rq lock.
5430          *
5431          * Also, during early boot the idle thread is in the fair class,
5432          * for obvious reasons its a bad idea to schedule back to it.
5433          */
5434         if (unlikely(!se->on_rq || curr == rq->idle))
5435                 return;
5436 
5437         if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5438                 set_last_buddy(se);
5439 }
5440 
5441 static struct task_struct *
5442 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
5443 {
5444         struct cfs_rq *cfs_rq = &rq->cfs;
5445         struct sched_entity *se;
5446         struct task_struct *p;
5447         int new_tasks;
5448 
5449 again:
5450 #ifdef CONFIG_FAIR_GROUP_SCHED
5451         if (!cfs_rq->nr_running)
5452                 goto idle;
5453 
5454         if (prev->sched_class != &fair_sched_class)
5455                 goto simple;
5456 
5457         /*
5458          * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5459          * likely that a next task is from the same cgroup as the current.
5460          *
5461          * Therefore attempt to avoid putting and setting the entire cgroup
5462          * hierarchy, only change the part that actually changes.
5463          */
5464 
5465         do {
5466                 struct sched_entity *curr = cfs_rq->curr;
5467 
5468                 /*
5469                  * Since we got here without doing put_prev_entity() we also
5470                  * have to consider cfs_rq->curr. If it is still a runnable
5471                  * entity, update_curr() will update its vruntime, otherwise
5472                  * forget we've ever seen it.
5473                  */
5474                 if (curr) {
5475                         if (curr->on_rq)
5476                                 update_curr(cfs_rq);
5477                         else
5478                                 curr = NULL;
5479 
5480                         /*
5481                          * This call to check_cfs_rq_runtime() will do the
5482                          * throttle and dequeue its entity in the parent(s).
5483                          * Therefore the 'simple' nr_running test will indeed
5484                          * be correct.
5485                          */
5486                         if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5487                                 goto simple;
5488                 }
5489 
5490                 se = pick_next_entity(cfs_rq, curr);
5491                 cfs_rq = group_cfs_rq(se);
5492         } while (cfs_rq);
5493 
5494         p = task_of(se);
5495 
5496         /*
5497          * Since we haven't yet done put_prev_entity and if the selected task
5498          * is a different task than we started out with, try and touch the
5499          * least amount of cfs_rqs.
5500          */
5501         if (prev != p) {
5502                 struct sched_entity *pse = &prev->se;
5503 
5504                 while (!(cfs_rq = is_same_group(se, pse))) {
5505                         int se_depth = se->depth;
5506                         int pse_depth = pse->depth;
5507 
5508                         if (se_depth <= pse_depth) {
5509                                 put_prev_entity(cfs_rq_of(pse), pse);
5510                                 pse = parent_entity(pse);
5511                         }
5512                         if (se_depth >= pse_depth) {
5513                                 set_next_entity(cfs_rq_of(se), se);
5514                                 se = parent_entity(se);
5515                         }
5516                 }
5517 
5518                 put_prev_entity(cfs_rq, pse);
5519                 set_next_entity(cfs_rq, se);
5520         }
5521 
5522         if (hrtick_enabled(rq))
5523                 hrtick_start_fair(rq, p);
5524 
5525         return p;
5526 simple:
5527         cfs_rq = &rq->cfs;
5528 #endif
5529 
5530         if (!cfs_rq->nr_running)
5531                 goto idle;
5532 
5533         put_prev_task(rq, prev);
5534 
5535         do {
5536                 se = pick_next_entity(cfs_rq, NULL);
5537                 set_next_entity(cfs_rq, se);
5538                 cfs_rq = group_cfs_rq(se);
5539         } while (cfs_rq);
5540 
5541         p = task_of(se);
5542 
5543         if (hrtick_enabled(rq))
5544                 hrtick_start_fair(rq, p);
5545 
5546         return p;
5547 
5548 idle:
5549         /*
5550          * This is OK, because current is on_cpu, which avoids it being picked
5551          * for load-balance and preemption/IRQs are still disabled avoiding
5552          * further scheduler activity on it and we're being very careful to
5553          * re-start the picking loop.
5554          */
5555         lockdep_unpin_lock(&rq->lock);
5556         new_tasks = idle_balance(rq);
5557         lockdep_pin_lock(&rq->lock);
5558         /*
5559          * Because idle_balance() releases (and re-acquires) rq->lock, it is
5560          * possible for any higher priority task to appear. In that case we
5561          * must re-start the pick_next_entity() loop.
5562          */
5563         if (new_tasks < 0)
5564                 return RETRY_TASK;
5565 
5566         if (new_tasks > 0)
5567                 goto again;
5568 
5569         return NULL;
5570 }
5571 
5572 /*
5573  * Account for a descheduled task:
5574  */
5575 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5576 {
5577         struct sched_entity *se = &prev->se;
5578         struct cfs_rq *cfs_rq;
5579 
5580         for_each_sched_entity(se) {
5581                 cfs_rq = cfs_rq_of(se);
5582                 put_prev_entity(cfs_rq, se);
5583         }
5584 }
5585 
5586 /*
5587  * sched_yield() is very simple
5588  *
5589  * The magic of dealing with the ->skip buddy is in pick_next_entity.
5590  */
5591 static void yield_task_fair(struct rq *rq)
5592 {
5593         struct task_struct *curr = rq->curr;
5594         struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5595         struct sched_entity *se = &curr->se;
5596 
5597         /*
5598          * Are we the only task in the tree?
5599          */
5600         if (unlikely(rq->nr_running == 1))
5601                 return;
5602 
5603         clear_buddies(cfs_rq, se);
5604 
5605         if (curr->policy != SCHED_BATCH) {
5606                 update_rq_clock(rq);
5607                 /*
5608                  * Update run-time statistics of the 'current'.
5609                  */
5610                 update_curr(cfs_rq);
5611                 /*
5612                  * Tell update_rq_clock() that we've just updated,
5613                  * so we don't do microscopic update in schedule()
5614                  * and double the fastpath cost.
5615                  */
5616                 rq_clock_skip_update(rq, true);
5617         }
5618 
5619         set_skip_buddy(se);
5620 }
5621 
5622 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5623 {
5624         struct sched_entity *se = &p->se;
5625 
5626         /* throttled hierarchies are not runnable */
5627         if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5628                 return false;
5629 
5630         /* Tell the scheduler that we'd really like pse to run next. */
5631         set_next_buddy(se);
5632 
5633         yield_task_fair(rq);
5634 
5635         return true;
5636 }
5637 
5638 #ifdef CONFIG_SMP
5639 /**************************************************
5640  * Fair scheduling class load-balancing methods.
5641  *
5642  * BASICS
5643  *
5644  * The purpose of load-balancing is to achieve the same basic fairness the
5645  * per-cpu scheduler provides, namely provide a proportional amount of compute
5646  * time to each task. This is expressed in the following equation:
5647  *
5648  *   W_i,n/P_i == W_j,n/P_j for all i,j                               (1)
5649  *
5650  * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5651  * W_i,0 is defined as:
5652  *
5653  *   W_i,0 = \Sum_j w_i,j                                             (2)
5654  *
5655  * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5656  * is derived from the nice value as per prio_to_weight[].
5657  *
5658  * The weight average is an exponential decay average of the instantaneous
5659  * weight:
5660  *
5661  *   W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0               (3)
5662  *
5663  * C_i is the compute capacity of cpu i, typically it is the
5664  * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5665  * can also include other factors [XXX].
5666  *
5667  * To achieve this balance we define a measure of imbalance which follows
5668  * directly from (1):
5669  *
5670  *   imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j }    (4)
5671  *
5672  * We them move tasks around to minimize the imbalance. In the continuous
5673  * function space it is obvious this converges, in the discrete case we get
5674  * a few fun cases generally called infeasible weight scenarios.
5675  *
5676  * [XXX expand on:
5677  *     - infeasible weights;
5678  *     - local vs global optima in the discrete case. ]
5679  *
5680  *
5681  * SCHED DOMAINS
5682  *
5683  * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5684  * for all i,j solution, we create a tree of cpus that follows the hardware
5685  * topology where each level pairs two lower groups (or better). This results
5686  * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5687  * tree to only the first of the previous level and we decrease the frequency
5688  * of load-balance at each level inv. proportional to the number of cpus in
5689  * the groups.
5690  *
5691  * This yields:
5692  *
5693  *     log_2 n     1     n
5694  *   \Sum       { --- * --- * 2^i } = O(n)                            (5)
5695  *     i = 0      2^i   2^i
5696  *                               `- size of each group
5697  *         |         |     `- number of cpus doing load-balance
5698  *         |         `- freq
5699  *         `- sum over all levels
5700  *
5701  * Coupled with a limit on how many tasks we can migrate every balance pass,
5702  * this makes (5) the runtime complexity of the balancer.
5703  *
5704  * An important property here is that each CPU is still (indirectly) connected
5705  * to every other cpu in at most O(log n) steps:
5706  *
5707  * The adjacency matrix of the resulting graph is given by:
5708  *
5709  *             log_2 n     
5710  *   A_i,j = \Union     (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1)  (6)
5711  *             k = 0
5712  *
5713  * And you'll find that:
5714  *
5715  *   A^(log_2 n)_i,j != 0  for all i,j                                (7)
5716  *
5717  * Showing there's indeed a path between every cpu in at most O(log n) steps.
5718  * The task movement gives a factor of O(m), giving a convergence complexity
5719  * of:
5720  *
5721  *   O(nm log n),  n := nr_cpus, m := nr_tasks                        (8)
5722  *
5723  *
5724  * WORK CONSERVING
5725  *
5726  * In order to avoid CPUs going idle while there's still work to do, new idle
5727  * balancing is more aggressive and has the newly idle cpu iterate up the domain
5728  * tree itself instead of relying on other CPUs to bring it work.
5729  *
5730  * This adds some complexity to both (5) and (8) but it reduces the total idle
5731  * time.
5732  *
5733  * [XXX more?]
5734  *
5735  *
5736  * CGROUPS
5737  *
5738  * Cgroups make a horror show out of (2), instead of a simple sum we get:
5739  *
5740  *                                s_k,i
5741  *   W_i,0 = \Sum_j \Prod_k w_k * -----                               (9)
5742  *                                 S_k
5743  *
5744  * Where
5745  *
5746  *   s_k,i = \Sum_j w_i,j,k  and  S_k = \Sum_i s_k,i                 (10)
5747  *
5748  * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5749  *
5750  * The big problem is S_k, its a global sum needed to compute a local (W_i)
5751  * property.
5752  *
5753  * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5754  *      rewrite all of this once again.]
5755  */ 
5756 
5757 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5758 
5759 enum fbq_type { regular, remote, all };
5760 
5761 #define LBF_ALL_PINNED  0x01
5762 #define LBF_NEED_BREAK  0x02
5763 #define LBF_DST_PINNED  0x04
5764 #define LBF_SOME_PINNED 0x08
5765 
5766 struct lb_env {
5767         struct sched_domain     *sd;
5768 
5769         struct rq               *src_rq;
5770         int                     src_cpu;
5771 
5772         int                     dst_cpu;
5773         struct rq               *dst_rq;
5774 
5775         struct cpumask          *dst_grpmask;
5776         int                     new_dst_cpu;
5777         enum cpu_idle_type      idle;
5778         long                    imbalance;
5779         /* The set of CPUs under consideration for load-balancing */
5780         struct cpumask          *cpus;
5781 
5782         unsigned int            flags;
5783 
5784         unsigned int            loop;
5785         unsigned int            loop_break;
5786         unsigned int            loop_max;
5787 
5788         enum fbq_type           fbq_type;
5789         struct list_head        tasks;
5790 };
5791 
5792 /*
5793  * Is this task likely cache-hot:
5794  */
5795 static int task_hot(struct task_struct *p, struct lb_env *env)
5796 {
5797         s64 delta;
5798 
5799         lockdep_assert_held(&env->src_rq->lock);
5800 
5801         if (p->sched_class != &fair_sched_class)
5802                 return 0;
5803 
5804         if (unlikely(p->policy == SCHED_IDLE))
5805                 return 0;
5806 
5807         /*
5808          * Buddy candidates are cache hot:
5809          */
5810         if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
5811                         (&p->se == cfs_rq_of(&p->se)->next ||
5812                          &p->se == cfs_rq_of(&p->se)->last))
5813                 return 1;
5814 
5815         if (sysctl_sched_migration_cost == -1)
5816                 return 1;
5817         if (sysctl_sched_migration_cost == 0)
5818                 return 0;
5819 
5820         delta = rq_clock_task(env->src_rq) - p->se.exec_start;
5821 
5822         return delta < (s64)sysctl_sched_migration_cost;
5823 }
5824 
5825 #ifdef CONFIG_NUMA_BALANCING
5826 /*
5827  * Returns 1, if task migration degrades locality
5828  * Returns 0, if task migration improves locality i.e migration preferred.
5829  * Returns -1, if task migration is not affected by locality.
5830  */
5831 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5832 {
5833         struct numa_group *numa_group = rcu_dereference(p->numa_group);
5834         unsigned long src_faults, dst_faults;
5835         int src_nid, dst_nid;
5836 
5837         if (!static_branch_likely(&sched_numa_balancing))
5838                 return -1;
5839 
5840         if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
5841                 return -1;
5842 
5843         src_nid = cpu_to_node(env->src_cpu);
5844         dst_nid = cpu_to_node(env->dst_cpu);
5845 
5846         if (src_nid == dst_nid)
5847                 return -1;
5848 
5849         /* Migrating away from the preferred node is always bad. */
5850         if (src_nid == p->numa_preferred_nid) {
5851                 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
5852                         return 1;
5853                 else
5854                         return -1;
5855         }
5856 
5857         /* Encourage migration to the preferred node. */
5858         if (dst_nid == p->numa_preferred_nid)
5859                 return 0;
5860 
5861         if (numa_group) {
5862                 src_faults = group_faults(p, src_nid);
5863                 dst_faults = group_faults(p, dst_nid);
5864         } else {
5865                 src_faults = task_faults(p, src_nid);
5866                 dst_faults = task_faults(p, dst_nid);
5867         }
5868 
5869         return dst_faults < src_faults;
5870 }
5871 
5872 #else
5873 static inline int migrate_degrades_locality(struct task_struct *p,
5874                                              struct lb_env *env)
5875 {
5876         return -1;
5877 }
5878 #endif
5879 
5880 /*
5881  * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5882  */
5883 static
5884 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5885 {
5886         int tsk_cache_hot;
5887 
5888         lockdep_assert_held(&env->src_rq->lock);
5889 
5890         /*
5891          * We do not migrate tasks that are:
5892          * 1) throttled_lb_pair, or
5893          * 2) cannot be migrated to this CPU due to cpus_allowed, or
5894          * 3) running (obviously), or
5895          * 4) are cache-hot on their current CPU.
5896          */
5897         if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5898                 return 0;
5899 
5900         if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5901                 int cpu;
5902 
5903                 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5904 
5905                 env->flags |= LBF_SOME_PINNED;
5906 
5907                 /*
5908                  * Remember if this task can be migrated to any other cpu in
5909                  * our sched_group. We may want to revisit it if we couldn't
5910                  * meet load balance goals by pulling other tasks on src_cpu.
5911                  *
5912                  * Also avoid computing new_dst_cpu if we have already computed
5913                  * one in current iteration.
5914                  */
5915                 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5916                         return 0;
5917 
5918                 /* Prevent to re-select dst_cpu via env's cpus */
5919                 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5920                         if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5921                                 env->flags |= LBF_DST_PINNED;
5922                                 env->new_dst_cpu = cpu;
5923                                 break;
5924                         }
5925                 }
5926 
5927                 return 0;
5928         }
5929 
5930         /* Record that we found atleast one task that could run on dst_cpu */
5931         env->flags &= ~LBF_ALL_PINNED;
5932 
5933         if (task_running(env->src_rq, p)) {
5934                 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5935                 return 0;
5936         }
5937 
5938         /*
5939          * Aggressive migration if:
5940          * 1) destination numa is preferred
5941          * 2) task is cache cold, or
5942          * 3) too many balance attempts have failed.
5943          */
5944         tsk_cache_hot = migrate_degrades_locality(p, env);
5945         if (tsk_cache_hot == -1)
5946                 tsk_cache_hot = task_hot(p, env);
5947 
5948         if (tsk_cache_hot <= 0 ||
5949             env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5950                 if (tsk_cache_hot == 1) {
5951                         schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5952                         schedstat_inc(p, se.statistics.nr_forced_migrations);
5953                 }
5954                 return 1;
5955         }
5956 
5957         schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5958         return 0;
5959 }
5960 
5961 /*
5962  * detach_task() -- detach the task for the migration specified in env
5963  */
5964 static void detach_task(struct task_struct *p, struct lb_env *env)
5965 {
5966         lockdep_assert_held(&env->src_rq->lock);
5967 
5968         p->on_rq = TASK_ON_RQ_MIGRATING;
5969         deactivate_task(env->src_rq, p, 0);
5970         set_task_cpu(p, env->dst_cpu);
5971 }
5972 
5973 /*
5974  * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5975  * part of active balancing operations within "domain".
5976  *
5977  * Returns a task if successful and NULL otherwise.
5978  */
5979 static struct task_struct *detach_one_task(struct lb_env *env)
5980 {
5981         struct task_struct *p, *n;
5982 
5983         lockdep_assert_held(&env->src_rq->lock);
5984 
5985         list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5986                 if (!can_migrate_task(p, env))
5987                         continue;
5988 
5989                 detach_task(p, env);
5990 
5991                 /*
5992                  * Right now, this is only the second place where
5993                  * lb_gained[env->idle] is updated (other is detach_tasks)
5994                  * so we can safely collect stats here rather than
5995                  * inside detach_tasks().
5996                  */
5997                 schedstat_inc(env->sd, lb_gained[env->idle]);
5998                 return p;
5999         }
6000         return NULL;
6001 }
6002 
6003 static const unsigned int sched_nr_migrate_break = 32;
6004 
6005 /*
6006  * detach_tasks() -- tries to detach up to imbalance weighted load from
6007  * busiest_rq, as part of a balancing operation within domain "sd".
6008  *
6009  * Returns number of detached tasks if successful and 0 otherwise.
6010  */
6011 static int detach_tasks(struct lb_env *env)
6012 {
6013         struct list_head *tasks = &env->src_rq->cfs_tasks;
6014         struct task_struct *p;
6015         unsigned long load;
6016         int detached = 0;
6017 
6018         lockdep_assert_held(&env->src_rq->lock);
6019 
6020         if (env->imbalance <= 0)
6021                 return 0;
6022 
6023         while (!list_empty(tasks)) {
6024                 /*
6025                  * We don't want to steal all, otherwise we may be treated likewise,
6026                  * which could at worst lead to a livelock crash.
6027                  */
6028                 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6029                         break;
6030 
6031                 p = list_first_entry(tasks, struct task_struct, se.group_node);
6032 
6033                 env->loop++;
6034                 /* We've more or less seen every task there is, call it quits */
6035                 if (env->loop > env->loop_max)
6036                         break;
6037 
6038                 /* take a breather every nr_migrate tasks */
6039                 if (env->loop > env->loop_break) {
6040                         env->loop_break += sched_nr_migrate_break;
6041                         env->flags |= LBF_NEED_BREAK;
6042                         break;
6043                 }
6044 
6045                 if (!can_migrate_task(p, env))
6046                         goto next;
6047 
6048                 load = task_h_load(p);
6049 
6050                 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6051                         goto next;
6052 
6053                 if ((load / 2) > env->imbalance)
6054                         goto next;
6055 
6056                 detach_task(p, env);
6057                 list_add(&p->se.group_node, &env->tasks);
6058 
6059                 detached++;
6060                 env->imbalance -= load;
6061 
6062 #ifdef CONFIG_PREEMPT
6063                 /*
6064                  * NEWIDLE balancing is a source of latency, so preemptible
6065                  * kernels will stop after the first task is detached to minimize
6066                  * the critical section.
6067                  */
6068                 if (env->idle == CPU_NEWLY_IDLE)
6069                         break;
6070 #endif
6071 
6072                 /*
6073                  * We only want to steal up to the prescribed amount of
6074                  * weighted load.
6075                  */
6076                 if (env->imbalance <= 0)
6077                         break;
6078 
6079                 continue;
6080 next:
6081                 list_move_tail(&p->se.group_node, tasks);
6082         }
6083 
6084         /*
6085          * Right now, this is one of only two places we collect this stat
6086          * so we can safely collect detach_one_task() stats here rather
6087          * than inside detach_one_task().
6088          */
6089         schedstat_add(env->sd, lb_gained[env->idle], detached);
6090 
6091         return detached;
6092 }
6093 
6094 /*
6095  * attach_task() -- attach the task detached by detach_task() to its new rq.
6096  */
6097 static void attach_task(struct rq *rq, struct task_struct *p)
6098 {
6099         lockdep_assert_held(&rq->lock);
6100 
6101         BUG_ON(task_rq(p) != rq);
6102         activate_task(rq, p, 0);
6103         p->on_rq = TASK_ON_RQ_QUEUED;
6104         check_preempt_curr(rq, p, 0);
6105 }
6106 
6107 /*
6108  * attach_one_task() -- attaches the task returned from detach_one_task() to
6109  * its new rq.
6110  */
6111 static void attach_one_task(struct rq *rq, struct task_struct *p)
6112 {
6113         raw_spin_lock(&rq->lock);
6114         attach_task(rq, p);
6115         raw_spin_unlock(&rq->lock);
6116 }
6117 
6118 /*
6119  * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6120  * new rq.
6121  */
6122 static void attach_tasks(struct lb_env *env)
6123 {
6124         struct list_head *tasks = &env->tasks;
6125         struct task_struct *p;
6126 
6127         raw_spin_lock(&env->dst_rq->lock);
6128 
6129         while (!list_empty(tasks)) {
6130                 p = list_first_entry(tasks, struct task_struct, se.group_node);
6131                 list_del_init(&p->se.group_node);
6132 
6133                 attach_task(env->dst_rq, p);
6134         }
6135 
6136         raw_spin_unlock(&env->dst_rq->lock);
6137 }
6138 
6139 #ifdef CONFIG_FAIR_GROUP_SCHED
6140 static void update_blocked_averages(int cpu)
6141 {
6142         struct rq *rq = cpu_rq(cpu);
6143         struct cfs_rq *cfs_rq;
6144         unsigned long flags;
6145 
6146         raw_spin_lock_irqsave(&rq->lock, flags);
6147         update_rq_clock(rq);
6148 
6149         /*
6150          * Iterates the task_group tree in a bottom up fashion, see
6151          * list_add_leaf_cfs_rq() for details.
6152          */
6153         for_each_leaf_cfs_rq(rq, cfs_rq) {
6154                 /* throttled entities do not contribute to load */
6155                 if (throttled_hierarchy(cfs_rq))
6156                         continue;
6157 
6158                 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
6159                         update_tg_load_avg(cfs_rq, 0);
6160         }
6161         raw_spin_unlock_irqrestore(&rq->lock, flags);
6162 }
6163 
6164 /*
6165  * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6166  * This needs to be done in a top-down fashion because the load of a child
6167  * group is a fraction of its parents load.
6168  */
6169 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6170 {
6171         struct rq *rq = rq_of(cfs_rq);
6172         struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6173         unsigned long now = jiffies;
6174         unsigned long load;
6175 
6176         if (cfs_rq->last_h_load_update == now)
6177                 return;
6178 
6179         cfs_rq->h_load_next = NULL;
6180         for_each_sched_entity(se) {
6181                 cfs_rq = cfs_rq_of(se);
6182                 cfs_rq->h_load_next = se;
6183                 if (cfs_rq->last_h_load_update == now)
6184                         break;
6185         }
6186 
6187         if (!se) {
6188                 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6189                 cfs_rq->last_h_load_update = now;
6190         }
6191 
6192         while ((se = cfs_rq->h_load_next) != NULL) {
6193                 load = cfs_rq->h_load;
6194                 load = div64_ul(load * se->avg.load_avg,
6195                         cfs_rq_load_avg(cfs_rq) + 1);
6196                 cfs_rq = group_cfs_rq(se);
6197                 cfs_rq->h_load = load;
6198                 cfs_rq->last_h_load_update = now;
6199         }
6200 }
6201 
6202 static unsigned long task_h_load(struct task_struct *p)
6203 {
6204         struct cfs_rq *cfs_rq = task_cfs_rq(p);
6205 
6206         update_cfs_rq_h_load(cfs_rq);
6207         return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6208                         cfs_rq_load_avg(cfs_rq) + 1);
6209 }
6210 #else
6211 static inline void update_blocked_averages(int cpu)
6212 {
6213         struct rq *rq = cpu_rq(cpu);
6214         struct cfs_rq *cfs_rq = &rq->cfs;
6215         unsigned long flags;
6216 
6217         raw_spin_lock_irqsave(&rq->lock, flags);
6218         update_rq_clock(rq);
6219         update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
6220         raw_spin_unlock_irqrestore(&rq->lock, flags);
6221 }
6222 
6223 static unsigned long task_h_load(struct task_struct *p)
6224 {
6225         return p->se.avg.load_avg;
6226 }
6227 #endif
6228 
6229 /********** Helpers for find_busiest_group ************************/
6230 
6231 enum group_type {
6232         group_other = 0,
6233         group_imbalanced,
6234         group_overloaded,
6235 };
6236 
6237 /*
6238  * sg_lb_stats - stats of a sched_group required for load_balancing
6239  */
6240 struct sg_lb_stats {
6241         unsigned long avg_load; /*Avg load across the CPUs of the group */
6242         unsigned long group_load; /* Total load over the CPUs of the group */
6243         unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6244         unsigned long load_per_task;
6245         unsigned long group_capacity;
6246         unsigned long group_util; /* Total utilization of the group */
6247         unsigned int sum_nr_running; /* Nr tasks running in the group */
6248         unsigned int idle_cpus;
6249         unsigned int group_weight;
6250         enum group_type group_type;
6251         int group_no_capacity;
6252 #ifdef CONFIG_NUMA_BALANCING
6253         unsigned int nr_numa_running;
6254         unsigned int nr_preferred_running;
6255 #endif
6256 };
6257 
6258 /*
6259  * sd_lb_stats - Structure to store the statistics of a sched_domain
6260  *               during load balancing.
6261  */
6262 struct sd_lb_stats {
6263         struct sched_group *busiest;    /* Busiest group in this sd */
6264         struct sched_group *local;      /* Local group in this sd */
6265         unsigned long total_load;       /* Total load of all groups in sd */
6266         unsigned long total_capacity;   /* Total capacity of all groups in sd */
6267         unsigned long avg_load; /* Average load across all groups in sd */
6268 
6269         struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6270         struct sg_lb_stats local_stat;  /* Statistics of the local group */
6271 };
6272 
6273 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6274 {
6275         /*
6276          * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6277          * local_stat because update_sg_lb_stats() does a full clear/assignment.
6278          * We must however clear busiest_stat::avg_load because
6279          * update_sd_pick_busiest() reads this before assignment.
6280          */
6281         *sds = (struct sd_lb_stats){
6282                 .busiest = NULL,
6283                 .local = NULL,
6284                 .total_load = 0UL,
6285                 .total_capacity = 0UL,
6286                 .busiest_stat = {
6287                         .avg_load = 0UL,
6288                         .sum_nr_running = 0,
6289                         .group_type = group_other,
6290                 },
6291         };
6292 }
6293 
6294 /**
6295  * get_sd_load_idx - Obtain the load index for a given sched domain.
6296  * @sd: The sched_domain whose load_idx is to be obtained.
6297  * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6298  *
6299  * Return: The load index.
6300  */
6301 static inline int get_sd_load_idx(struct sched_domain *sd,
6302                                         enum cpu_idle_type idle)
6303 {
6304         int load_idx;
6305 
6306         switch (idle) {
6307         case CPU_NOT_IDLE:
6308                 load_idx = sd->busy_idx;
6309                 break;
6310 
6311         case CPU_NEWLY_IDLE:
6312                 load_idx = sd->newidle_idx;
6313                 break;
6314         default:
6315                 load_idx = sd->idle_idx;
6316                 break;
6317         }
6318 
6319         return load_idx;
6320 }
6321 
6322 static unsigned long scale_rt_capacity(int cpu)
6323 {
6324         struct rq *rq = cpu_rq(cpu);
6325         u64 total, used, age_stamp, avg;
6326         s64 delta;
6327 
6328         /*
6329          * Since we're reading these variables without serialization make sure
6330          * we read them once before doing sanity checks on them.
6331          */
6332         age_stamp = READ_ONCE(rq->age_stamp);
6333         avg = READ_ONCE(rq->rt_avg);
6334         delta = __rq_clock_broken(rq) - age_stamp;
6335 
6336         if (unlikely(delta < 0))
6337                 delta = 0;
6338 
6339         total = sched_avg_period() + delta;
6340 
6341         used = div_u64(avg, total);
6342 
6343         if (likely(used < SCHED_CAPACITY_SCALE))
6344                 return SCHED_CAPACITY_SCALE - used;
6345 
6346         return 1;
6347 }
6348 
6349 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6350 {
6351         unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6352         struct sched_group *sdg = sd->groups;
6353 
6354         cpu_rq(cpu)->cpu_capacity_orig = capacity;
6355 
6356         capacity *= scale_rt_capacity(cpu);
6357         capacity >>= SCHED_CAPACITY_SHIFT;
6358 
6359         if (!capacity)
6360                 capacity = 1;
6361 
6362         cpu_rq(cpu)->cpu_capacity = capacity;
6363         sdg->sgc->capacity = capacity;
6364 }
6365 
6366 void update_group_capacity(struct sched_domain *sd, int cpu)
6367 {
6368         struct sched_domain *child = sd->child;
6369         struct sched_group *group, *sdg = sd->groups;
6370         unsigned long capacity;
6371         unsigned long interval;
6372 
6373         interval = msecs_to_jiffies(sd->balance_interval);
6374         interval = clamp(interval, 1UL, max_load_balance_interval);
6375         sdg->sgc->next_update = jiffies + interval;
6376 
6377         if (!child) {
6378                 update_cpu_capacity(sd, cpu);
6379                 return;
6380         }
6381 
6382         capacity = 0;
6383 
6384         if (child->flags & SD_OVERLAP) {
6385                 /*
6386                  * SD_OVERLAP domains cannot assume that child groups
6387                  * span the current group.
6388                  */
6389 
6390                 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6391                         struct sched_group_capacity *sgc;
6392                         struct rq *rq = cpu_rq(cpu);
6393 
6394                         /*
6395                          * build_sched_domains() -> init_sched_groups_capacity()
6396                          * gets here before we've attached the domains to the
6397                          * runqueues.
6398                          *
6399                          * Use capacity_of(), which is set irrespective of domains
6400                          * in update_cpu_capacity().
6401                          *
6402                          * This avoids capacity from being 0 and
6403                          * causing divide-by-zero issues on boot.
6404                          */
6405                         if (unlikely(!rq->sd)) {
6406                                 capacity += capacity_of(cpu);
6407                                 continue;
6408                         }
6409 
6410                         sgc = rq->sd->groups->sgc;
6411                         capacity += sgc->capacity;
6412                 }
6413         } else  {
6414                 /*
6415                  * !SD_OVERLAP domains can assume that child groups
6416                  * span the current group.
6417                  */ 
6418 
6419                 group = child->groups;
6420                 do {
6421                         capacity += group->sgc->capacity;
6422                         group = group->next;
6423                 } while (group != child->groups);
6424         }
6425 
6426         sdg->sgc->capacity = capacity;
6427 }
6428 
6429 /*
6430  * Check whether the capacity of the rq has been noticeably reduced by side
6431  * activity. The imbalance_pct is used for the threshold.
6432  * Return true is the capacity is reduced
6433  */
6434 static inline int
6435 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6436 {
6437         return ((rq->cpu_capacity * sd->imbalance_pct) <
6438                                 (rq->cpu_capacity_orig * 100));
6439 }
6440 
6441 /*
6442  * Group imbalance indicates (and tries to solve) the problem where balancing
6443  * groups is inadequate due to tsk_cpus_allowed() constraints.
6444  *
6445  * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6446  * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6447  * Something like:
6448  *
6449  *      { 0 1 2 3 } { 4 5 6 7 }
6450  *              *     * * *
6451  *
6452  * If we were to balance group-wise we'd place two tasks in the first group and
6453  * two tasks in the second group. Clearly this is undesired as it will overload
6454  * cpu 3 and leave one of the cpus in the second group unused.
6455  *
6456  * The current solution to this issue is detecting the skew in the first group
6457  * by noticing the lower domain failed to reach balance and had difficulty
6458  * moving tasks due to affinity constraints.
6459  *
6460  * When this is so detected; this group becomes a candidate for busiest; see
6461  * update_sd_pick_busiest(). And calculate_imbalance() and
6462  * find_busiest_group() avoid some of the usual balance conditions to allow it
6463  * to create an effective group imbalance.
6464  *
6465  * This is a somewhat tricky proposition since the next run might not find the
6466  * group imbalance and decide the groups need to be balanced again. A most
6467  * subtle and fragile situation.
6468  */
6469 
6470 static inline int sg_imbalanced(struct sched_group *group)
6471 {
6472         return group->sgc->imbalance;
6473 }
6474 
6475 /*
6476  * group_has_capacity returns true if the group has spare capacity that could
6477  * be used by some tasks.
6478  * We consider that a group has spare capacity if the  * number of task is
6479  * smaller than the number of CPUs or if the utilization is lower than the
6480  * available capacity for CFS tasks.
6481  * For the latter, we use a threshold to stabilize the state, to take into
6482  * account the variance of the tasks' load and to return true if the available
6483  * capacity in meaningful for the load balancer.
6484  * As an example, an available capacity of 1% can appear but it doesn't make
6485  * any benefit for the load balance.
6486  */
6487 static inline bool
6488 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6489 {
6490         if (sgs->sum_nr_running < sgs->group_weight)
6491                 return true;
6492 
6493         if ((sgs->group_capacity * 100) >
6494                         (sgs->group_util * env->sd->imbalance_pct))
6495                 return true;
6496 
6497         return false;
6498 }
6499 
6500 /*
6501  *  group_is_overloaded returns true if the group has more tasks than it can
6502  *  handle.
6503  *  group_is_overloaded is not equals to !group_has_capacity because a group
6504  *  with the exact right number of tasks, has no more spare capacity but is not
6505  *  overloaded so both group_has_capacity and group_is_overloaded return
6506  *  false.
6507  */
6508 static inline bool
6509 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6510 {
6511         if (sgs->sum_nr_running <= sgs->group_weight)
6512                 return false;
6513 
6514         if ((sgs->group_capacity * 100) <
6515                         (sgs->group_util * env->sd->imbalance_pct))
6516                 return true;
6517 
6518         return false;
6519 }
6520 
6521 static inline enum
6522 group_type group_classify(struct sched_group *group,
6523                           struct sg_lb_stats *sgs)
6524 {
6525         if (sgs->group_no_capacity)
6526                 return group_overloaded;
6527 
6528         if (sg_imbalanced(group))
6529                 return group_imbalanced;
6530 
6531         return group_other;
6532 }
6533 
6534 /**
6535  * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6536  * @env: The load balancing environment.
6537  * @group: sched_group whose statistics are to be updated.
6538  * @load_idx: Load index of sched_domain of this_cpu for load calc.
6539  * @local_group: Does group contain this_cpu.
6540  * @sgs: variable to hold the statistics for this group.
6541  * @overload: Indicate more than one runnable task for any CPU.
6542  */
6543 static inline void update_sg_lb_stats(struct lb_env *env,
6544                         struct sched_group *group, int load_idx,
6545                         int local_group, struct sg_lb_stats *sgs,
6546                         bool *overload)
6547 {
6548         unsigned long load;
6549         int i, nr_running;
6550 
6551         memset(sgs, 0, sizeof(*sgs));
6552 
6553         for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6554                 struct rq *rq = cpu_rq(i);
6555 
6556                 /* Bias balancing toward cpus of our domain */
6557                 if (local_group)
6558                         load = target_load(i, load_idx);
6559                 else
6560                         load = source_load(i, load_idx);
6561 
6562                 sgs->group_load += load;
6563                 sgs->group_util += cpu_util(i);
6564                 sgs->sum_nr_running += rq->cfs.h_nr_running;
6565 
6566                 nr_running = rq->nr_running;
6567                 if (nr_running > 1)
6568                         *overload = true;
6569 
6570 #ifdef CONFIG_NUMA_BALANCING
6571                 sgs->nr_numa_running += rq->nr_numa_running;
6572                 sgs->nr_preferred_running += rq->nr_preferred_running;
6573 #endif
6574                 sgs->sum_weighted_load += weighted_cpuload(i);
6575                 /*
6576                  * No need to call idle_cpu() if nr_running is not 0
6577                  */
6578                 if (!nr_running && idle_cpu(i))
6579                         sgs->idle_cpus++;
6580         }
6581 
6582         /* Adjust by relative CPU capacity of the group */
6583         sgs->group_capacity = group->sgc->capacity;
6584         sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6585 
6586         if (sgs->sum_nr_running)
6587                 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6588 
6589         sgs->group_weight = group->group_weight;
6590 
6591         sgs->group_no_capacity = group_is_overloaded(env, sgs);
6592         sgs->group_type = group_classify(group, sgs);
6593 }
6594 
6595 /**
6596  * update_sd_pick_busiest - return 1 on busiest group
6597  * @env: The load balancing environment.
6598  * @sds: sched_domain statistics
6599  * @sg: sched_group candidate to be checked for being the busiest
6600  * @sgs: sched_group statistics
6601  *
6602  * Determine if @sg is a busier group than the previously selected
6603  * busiest group.
6604  *
6605  * Return: %true if @sg is a busier group than the previously selected
6606  * busiest group. %false otherwise.
6607  */
6608 static bool update_sd_pick_busiest(struct lb_env *env,
6609                                    struct sd_lb_stats *sds,
6610                                    struct sched_group *sg,
6611                                    struct sg_lb_stats *sgs)
6612 {
6613         struct sg_lb_stats *busiest = &sds->busiest_stat;
6614 
6615         if (sgs->group_type > busiest->group_type)
6616                 return true;
6617 
6618         if (sgs->group_type < busiest->group_type)
6619                 return false;
6620 
6621         if (sgs->avg_load <= busiest->avg_load)
6622                 return false;
6623 
6624         /* This is the busiest node in its class. */
6625         if (!(env->sd->flags & SD_ASYM_PACKING))
6626                 return true;
6627 
6628         /*
6629          * ASYM_PACKING needs to move all the work to the lowest
6630          * numbered CPUs in the group, therefore mark all groups
6631          * higher than ourself as busy.
6632          */
6633         if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6634                 if (!sds->busiest)
6635                         return true;
6636 
6637                 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
6638                         return true;
6639         }
6640 
6641         return false;
6642 }
6643 
6644 #ifdef CONFIG_NUMA_BALANCING
6645 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6646 {
6647         if (sgs->sum_nr_running > sgs->nr_numa_running)
6648                 return regular;
6649         if (sgs->sum_nr_running > sgs->nr_preferred_running)
6650                 return remote;
6651         return all;
6652 }
6653 
6654 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6655 {
6656         if (rq->nr_running > rq->nr_numa_running)
6657                 return regular;
6658         if (rq->nr_running > rq->nr_preferred_running)
6659                 return remote;
6660         return all;
6661 }
6662 #else
6663 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6664 {
6665         return all;
6666 }
6667 
6668 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6669 {
6670         return regular;
6671 }
6672 #endif /* CONFIG_NUMA_BALANCING */
6673 
6674 /**
6675  * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6676  * @env: The load balancing environment.
6677  * @sds: variable to hold the statistics for this sched_domain.
6678  */
6679 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6680 {
6681         struct sched_domain *child = env->sd->child;
6682         struct sched_group *sg = env->sd->groups;
6683         struct sg_lb_stats tmp_sgs;
6684         int load_idx, prefer_sibling = 0;
6685         bool overload = false;
6686 
6687         if (child && child->flags & SD_PREFER_SIBLING)
6688                 prefer_sibling = 1;
6689 
6690         load_idx = get_sd_load_idx(env->sd, env->idle);
6691 
6692         do {
6693                 struct sg_lb_stats *sgs = &tmp_sgs;
6694                 int local_group;
6695 
6696                 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6697                 if (local_group) {
6698                         sds->local = sg;
6699                         sgs = &sds->local_stat;
6700 
6701                         if (env->idle != CPU_NEWLY_IDLE ||
6702                             time_after_eq(jiffies, sg->sgc->next_update))
6703                                 update_group_capacity(env->sd, env->dst_cpu);
6704                 }
6705 
6706                 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6707                                                 &overload);
6708 
6709                 if (local_group)
6710                         goto next_group;
6711 
6712                 /*
6713                  * In case the child domain prefers tasks go to siblings
6714                  * first, lower the sg capacity so that we'll try
6715                  * and move all the excess tasks away. We lower the capacity
6716                  * of a group only if the local group has the capacity to fit
6717                  * these excess tasks. The extra check prevents the case where
6718                  * you always pull from the heaviest group when it is already
6719                  * under-utilized (possible with a large weight task outweighs
6720                  * the tasks on the system).
6721                  */
6722                 if (prefer_sibling && sds->local &&
6723                     group_has_capacity(env, &sds->local_stat) &&
6724                     (sgs->sum_nr_running > 1)) {
6725                         sgs->group_no_capacity = 1;
6726                         sgs->group_type = group_classify(sg, sgs);
6727                 }
6728 
6729                 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6730                         sds->busiest = sg;
6731                         sds->busiest_stat = *sgs;
6732                 }
6733 
6734 next_group:
6735                 /* Now, start updating sd_lb_stats */
6736                 sds->total_load += sgs->group_load;
6737                 sds->total_capacity += sgs->group_capacity;
6738 
6739                 sg = sg->next;
6740         } while (sg != env->sd->groups);
6741 
6742         if (env->sd->flags & SD_NUMA)
6743                 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6744 
6745         if (!env->sd->parent) {
6746                 /* update overload indicator if we are at root domain */
6747                 if (env->dst_rq->rd->overload != overload)
6748                         env->dst_rq->rd->overload = overload;
6749         }
6750 
6751 }
6752 
6753 /**
6754  * check_asym_packing - Check to see if the group is packed into the
6755  *                      sched doman.
6756  *
6757  * This is primarily intended to used at the sibling level.  Some
6758  * cores like POWER7 prefer to use lower numbered SMT threads.  In the
6759  * case of POWER7, it can move to lower SMT modes only when higher
6760  * threads are idle.  When in lower SMT modes, the threads will
6761  * perform better since they share less core resources.  Hence when we
6762  * have idle threads, we want them to be the higher ones.
6763  *
6764  * This packing function is run on idle threads.  It checks to see if
6765  * the busiest CPU in this domain (core in the P7 case) has a higher
6766  * CPU number than the packing function is being run on.  Here we are
6767  * assuming lower CPU number will be equivalent to lower a SMT thread
6768  * number.
6769  *
6770  * Return: 1 when packing is required and a task should be moved to
6771  * this CPU.  The amount of the imbalance is returned in *imbalance.
6772  *
6773  * @env: The load balancing environment.
6774  * @sds: Statistics of the sched_domain which is to be packed
6775  */
6776 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6777 {
6778         int busiest_cpu;
6779 
6780         if (!(env->sd->flags & SD_ASYM_PACKING))
6781                 return 0;
6782 
6783         if (!sds->busiest)
6784                 return 0;
6785 
6786         busiest_cpu = group_first_cpu(sds->busiest);
6787         if (env->dst_cpu > busiest_cpu)
6788                 return 0;
6789 
6790         env->imbalance = DIV_ROUND_CLOSEST(
6791                 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6792                 SCHED_CAPACITY_SCALE);
6793 
6794         return 1;
6795 }
6796 
6797 /**
6798  * fix_small_imbalance - Calculate the minor imbalance that exists
6799  *                      amongst the groups of a sched_domain, during
6800  *                      load balancing.
6801  * @env: The load balancing environment.
6802  * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6803  */
6804 static inline
6805 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6806 {
6807         unsigned long tmp, capa_now = 0, capa_move = 0;
6808         unsigned int imbn = 2;
6809         unsigned long scaled_busy_load_per_task;
6810         struct sg_lb_stats *local, *busiest;
6811 
6812         local = &sds->local_stat;
6813         busiest = &sds->busiest_stat;
6814 
6815         if (!local->sum_nr_running)
6816                 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6817         else if (busiest->load_per_task > local->load_per_task)
6818                 imbn = 1;
6819 
6820         scaled_busy_load_per_task =
6821                 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6822                 busiest->group_capacity;
6823 
6824         if (busiest->avg_load + scaled_busy_load_per_task >=
6825             local->avg_load + (scaled_busy_load_per_task * imbn)) {
6826                 env->imbalance = busiest->load_per_task;
6827                 return;
6828         }
6829 
6830         /*
6831          * OK, we don't have enough imbalance to justify moving tasks,
6832          * however we may be able to increase total CPU capacity used by
6833          * moving them.
6834          */
6835 
6836         capa_now += busiest->group_capacity *
6837                         min(busiest->load_per_task, busiest->avg_load);
6838         capa_now += local->group_capacity *
6839                         min(local->load_per_task, local->avg_load);
6840         capa_now /= SCHED_CAPACITY_SCALE;
6841 
6842         /* Amount of load we'd subtract */
6843         if (busiest->avg_load > scaled_busy_load_per_task) {
6844                 capa_move += busiest->group_capacity *
6845                             min(busiest->load_per_task,
6846                                 busiest->avg_load - scaled_busy_load_per_task);
6847         }
6848 
6849         /* Amount of load we'd add */
6850         if (busiest->avg_load * busiest->group_capacity <
6851             busiest->load_per_task * SCHED_CAPACITY_SCALE) {
6852                 tmp = (busiest->avg_load * busiest->group_capacity) /
6853                       local->group_capacity;
6854         } else {
6855                 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6856                       local->group_capacity;
6857         }
6858         capa_move += local->group_capacity *
6859                     min(local->load_per_task, local->avg_load + tmp);
6860         capa_move /= SCHED_CAPACITY_SCALE;
6861 
6862         /* Move if we gain throughput */
6863         if (capa_move > capa_now)
6864                 env->imbalance = busiest->load_per_task;
6865 }
6866 
6867 /**
6868  * calculate_imbalance - Calculate the amount of imbalance present within the
6869  *                       groups of a given sched_domain during load balance.
6870  * @env: load balance environment
6871  * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6872  */
6873 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6874 {
6875         unsigned long max_pull, load_above_capacity = ~0UL;
6876         struct sg_lb_stats *local, *busiest;
6877 
6878         local = &sds->local_stat;
6879         busiest = &sds->busiest_stat;
6880 
6881         if (busiest->group_type == group_imbalanced) {
6882                 /*
6883                  * In the group_imb case we cannot rely on group-wide averages
6884                  * to ensure cpu-load equilibrium, look at wider averages. XXX
6885                  */
6886                 busiest->load_per_task =
6887                         min(busiest->load_per_task, sds->avg_load);
6888         }
6889 
6890         /*
6891          * In the presence of smp nice balancing, certain scenarios can have
6892          * max load less than avg load(as we skip the groups at or below
6893          * its cpu_capacity, while calculating max_load..)
6894          */
6895         if (busiest->avg_load <= sds->avg_load ||
6896             local->avg_load >= sds->avg_load) {
6897                 env->imbalance = 0;
6898                 return fix_small_imbalance(env, sds);
6899         }
6900 
6901         /*
6902          * If there aren't any idle cpus, avoid creating some.
6903          */
6904         if (busiest->group_type == group_overloaded &&
6905             local->group_type   == group_overloaded) {
6906                 load_above_capacity = busiest->sum_nr_running *
6907                                         SCHED_LOAD_SCALE;
6908                 if (load_above_capacity > busiest->group_capacity)
6909                         load_above_capacity -= busiest->group_capacity;
6910                 else
6911                         load_above_capacity = ~0UL;
6912         }
6913 
6914         /*
6915          * We're trying to get all the cpus to the average_load, so we don't
6916          * want to push ourselves above the average load, nor do we wish to
6917          * reduce the max loaded cpu below the average load. At the same time,
6918          * we also don't want to reduce the group load below the group capacity
6919          * (so that we can implement power-savings policies etc). Thus we look
6920          * for the minimum possible imbalance.
6921          */
6922         max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6923 
6924         /* How much load to actually move to equalise the imbalance */
6925         env->imbalance = min(
6926                 max_pull * busiest->group_capacity,
6927                 (sds->avg_load - local->avg_load) * local->group_capacity
6928         ) / SCHED_CAPACITY_SCALE;
6929 
6930         /*
6931          * if *imbalance is less than the average load per runnable task
6932          * there is no guarantee that any tasks will be moved so we'll have
6933          * a think about bumping its value to force at least one task to be
6934          * moved
6935          */
6936         if (env->imbalance < busiest->load_per_task)
6937                 return fix_small_imbalance(env, sds);
6938 }
6939 
6940 /******* find_busiest_group() helpers end here *********************/
6941 
6942 /**
6943  * find_busiest_group - Returns the busiest group within the sched_domain
6944  * if there is an imbalance. If there isn't an imbalance, and
6945  * the user has opted for power-savings, it returns a group whose
6946  * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6947  * such a group exists.
6948  *
6949  * Also calculates the amount of weighted load which should be moved
6950  * to restore balance.
6951  *
6952  * @env: The load balancing environment.
6953  *
6954  * Return:      - The busiest group if imbalance exists.
6955  *              - If no imbalance and user has opted for power-savings balance,
6956  *                 return the least loaded group whose CPUs can be
6957  *                 put to idle by rebalancing its tasks onto our group.
6958  */
6959 static struct sched_group *find_busiest_group(struct lb_env *env)
6960 {
6961         struct sg_lb_stats *local, *busiest;
6962         struct sd_lb_stats sds;
6963 
6964         init_sd_lb_stats(&sds);
6965 
6966         /*
6967          * Compute the various statistics relavent for load balancing at
6968          * this level.
6969          */
6970         update_sd_lb_stats(env, &sds);
6971         local = &sds.local_stat;
6972         busiest = &sds.busiest_stat;
6973 
6974         /* ASYM feature bypasses nice load balance check */
6975         if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6976             check_asym_packing(env, &sds))
6977                 return sds.busiest;
6978 
6979         /* There is no busy sibling group to pull tasks from */
6980         if (!sds.busiest || busiest->sum_nr_running == 0)
6981                 goto out_balanced;
6982 
6983         sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
6984                                                 / sds.total_capacity;
6985 
6986         /*
6987          * If the busiest group is imbalanced the below checks don't
6988          * work because they assume all things are equal, which typically
6989          * isn't true due to cpus_allowed constraints and the like.
6990          */
6991         if (busiest->group_type == group_imbalanced)
6992                 goto force_balance;
6993 
6994         /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6995         if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
6996             busiest->group_no_capacity)
6997                 goto force_balance;
6998 
6999         /*
7000          * If the local group is busier than the selected busiest group
7001          * don't try and pull any tasks.
7002          */
7003         if (local->avg_load >= busiest->avg_load)
7004                 goto out_balanced;
7005 
7006         /*
7007          * Don't pull any tasks if this group is already above the domain
7008          * average load.
7009          */
7010         if (local->avg_load >= sds.avg_load)
7011                 goto out_balanced;
7012 
7013         if (env->idle == CPU_IDLE) {
7014                 /*
7015                  * This cpu is idle. If the busiest group is not overloaded
7016                  * and there is no imbalance between this and busiest group
7017                  * wrt idle cpus, it is balanced. The imbalance becomes
7018                  * significant if the diff is greater than 1 otherwise we
7019                  * might end up to just move the imbalance on another group
7020                  */
7021                 if ((busiest->group_type != group_overloaded) &&
7022                                 (local->idle_cpus <= (busiest->idle_cpus + 1)))
7023                         goto out_balanced;
7024         } else {
7025                 /*
7026                  * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7027                  * imbalance_pct to be conservative.
7028                  */
7029                 if (100 * busiest->avg_load <=
7030                                 env->sd->imbalance_pct * local->avg_load)
7031                         goto out_balanced;
7032         }
7033 
7034 force_balance:
7035         /* Looks like there is an imbalance. Compute it */
7036         calculate_imbalance(env, &sds);
7037         return sds.busiest;
7038 
7039 out_balanced:
7040         env->imbalance = 0;
7041         return NULL;
7042 }
7043 
7044 /*
7045  * find_busiest_queue - find the busiest runqueue among the cpus in group.
7046  */
7047 static struct rq *find_busiest_queue(struct lb_env *env,
7048                                      struct sched_group *group)
7049 {
7050         struct rq *busiest = NULL, *rq;
7051         unsigned long busiest_load = 0, busiest_capacity = 1;
7052         int i;
7053 
7054         for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7055                 unsigned long capacity, wl;
7056                 enum fbq_type rt;
7057 
7058                 rq = cpu_rq(i);
7059                 rt = fbq_classify_rq(rq);
7060 
7061                 /*
7062                  * We classify groups/runqueues into three groups:
7063                  *  - regular: there are !numa tasks
7064                  *  - remote:  there are numa tasks that run on the 'wrong' node
7065                  *  - all:     there is no distinction
7066                  *
7067                  * In order to avoid migrating ideally placed numa tasks,
7068                  * ignore those when there's better options.
7069                  *
7070                  * If we ignore the actual busiest queue to migrate another
7071                  * task, the next balance pass can still reduce the busiest
7072                  * queue by moving tasks around inside the node.
7073                  *
7074                  * If we cannot move enough load due to this classification
7075                  * the next pass will adjust the group classification and
7076                  * allow migration of more tasks.
7077                  *
7078                  * Both cases only affect the total convergence complexity.
7079                  */
7080                 if (rt > env->fbq_type)
7081                         continue;
7082 
7083                 capacity = capacity_of(i);
7084 
7085                 wl = weighted_cpuload(i);
7086 
7087                 /*
7088                  * When comparing with imbalance, use weighted_cpuload()
7089                  * which is not scaled with the cpu capacity.
7090                  */
7091 
7092                 if (rq->nr_running == 1 && wl > env->imbalance &&
7093                     !check_cpu_capacity(rq, env->sd))
7094                         continue;
7095 
7096                 /*
7097                  * For the load comparisons with the other cpu's, consider
7098                  * the weighted_cpuload() scaled with the cpu capacity, so
7099                  * that the load can be moved away from the cpu that is
7100                  * potentially running at a lower capacity.
7101                  *
7102                  * Thus we're looking for max(wl_i / capacity_i), crosswise
7103                  * multiplication to rid ourselves of the division works out
7104                  * to: wl_i * capacity_j > wl_j * capacity_i;  where j is
7105                  * our previous maximum.
7106                  */
7107                 if (wl * busiest_capacity > busiest_load * capacity) {
7108                         busiest_load = wl;
7109                         busiest_capacity = capacity;
7110                         busiest = rq;
7111                 }
7112         }
7113 
7114         return busiest;
7115 }
7116 
7117 /*
7118  * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7119  * so long as it is large enough.
7120  */
7121 #define MAX_PINNED_INTERVAL     512
7122 
7123 /* Working cpumask for load_balance and load_balance_newidle. */
7124 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
7125 
7126 static int need_active_balance(struct lb_env *env)
7127 {
7128         struct sched_domain *sd = env->sd;
7129 
7130         if (env->idle == CPU_NEWLY_IDLE) {
7131 
7132                 /*
7133                  * ASYM_PACKING needs to force migrate tasks from busy but
7134                  * higher numbered CPUs in order to pack all tasks in the
7135                  * lowest numbered CPUs.
7136                  */
7137                 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7138                         return 1;
7139         }
7140 
7141         /*
7142          * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7143          * It's worth migrating the task if the src_cpu's capacity is reduced
7144          * because of other sched_class or IRQs if more capacity stays
7145          * available on dst_cpu.
7146          */
7147         if ((env->idle != CPU_NOT_IDLE) &&
7148             (env->src_rq->cfs.h_nr_running == 1)) {
7149                 if ((check_cpu_capacity(env->src_rq, sd)) &&
7150                     (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7151                         return 1;
7152         }
7153 
7154         return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7155 }
7156 
7157 static int active_load_balance_cpu_stop(void *data);
7158 
7159 static int should_we_balance(struct lb_env *env)
7160 {
7161         struct sched_group *sg = env->sd->groups;
7162         struct cpumask *sg_cpus, *sg_mask;
7163         int cpu, balance_cpu = -1;
7164 
7165         /*
7166          * In the newly idle case, we will allow all the cpu's
7167          * to do the newly idle load balance.
7168          */
7169         if (env->idle == CPU_NEWLY_IDLE)
7170                 return 1;
7171 
7172         sg_cpus = sched_group_cpus(sg);
7173         sg_mask = sched_group_mask(sg);
7174         /* Try to find first idle cpu */
7175         for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7176                 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7177                         continue;
7178 
7179                 balance_cpu = cpu;
7180                 break;
7181         }
7182 
7183         if (balance_cpu == -1)
7184                 balance_cpu = group_balance_cpu(sg);
7185 
7186         /*
7187          * First idle cpu or the first cpu(busiest) in this sched group
7188          * is eligible for doing load balancing at this and above domains.
7189          */
7190         return balance_cpu == env->dst_cpu;
7191 }
7192 
7193 /*
7194  * Check this_cpu to ensure it is balanced within domain. Attempt to move
7195  * tasks if there is an imbalance.
7196  */
7197 static int load_balance(int this_cpu, struct rq *this_rq,
7198                         struct sched_domain *sd, enum cpu_idle_type idle,
7199                         int *continue_balancing)
7200 {
7201         int ld_moved, cur_ld_moved, active_balance = 0;
7202         struct sched_domain *sd_parent = sd->parent;
7203         struct sched_group *group;
7204         struct rq *busiest;
7205         unsigned long flags;
7206         struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
7207 
7208         struct lb_env env = {
7209                 .sd             = sd,
7210                 .dst_cpu        = this_cpu,
7211                 .dst_rq         = this_rq,
7212                 .dst_grpmask    = sched_group_cpus(sd->groups),
7213                 .idle           = idle,
7214                 .loop_break     = sched_nr_migrate_break,
7215                 .cpus           = cpus,
7216                 .fbq_type       = all,
7217                 .tasks          = LIST_HEAD_INIT(env.tasks),
7218         };
7219 
7220         /*
7221          * For NEWLY_IDLE load_balancing, we don't need to consider
7222          * other cpus in our group
7223          */
7224         if (idle == CPU_NEWLY_IDLE)
7225                 env.dst_grpmask = NULL;
7226 
7227         cpumask_copy(cpus, cpu_active_mask);
7228 
7229         schedstat_inc(sd, lb_count[idle]);
7230 
7231 redo:
7232         if (!should_we_balance(&env)) {
7233                 *continue_balancing = 0;
7234                 goto out_balanced;
7235         }
7236 
7237         group = find_busiest_group(&env);
7238         if (!group) {
7239                 schedstat_inc(sd, lb_nobusyg[idle]);
7240                 goto out_balanced;
7241         }
7242 
7243         busiest = find_busiest_queue(&env, group);
7244         if (!busiest) {
7245                 schedstat_inc(sd, lb_nobusyq[idle]);
7246                 goto out_balanced;
7247         }
7248 
7249         BUG_ON(busiest == env.dst_rq);
7250 
7251         schedstat_add(sd, lb_imbalance[idle], env.imbalance);
7252 
7253         env.src_cpu = busiest->cpu;
7254         env.src_rq = busiest;
7255 
7256         ld_moved = 0;
7257         if (busiest->nr_running > 1) {
7258                 /*
7259                  * Attempt to move tasks. If find_busiest_group has found
7260                  * an imbalance but busiest->nr_running <= 1, the group is
7261                  * still unbalanced. ld_moved simply stays zero, so it is
7262                  * correctly treated as an imbalance.
7263                  */
7264                 env.flags |= LBF_ALL_PINNED;
7265                 env.loop_max  = min(sysctl_sched_nr_migrate, busiest->nr_running);
7266 
7267 more_balance:
7268                 raw_spin_lock_irqsave(&busiest->lock, flags);
7269 
7270                 /*
7271                  * cur_ld_moved - load moved in current iteration
7272                  * ld_moved     - cumulative load moved across iterations
7273                  */
7274                 cur_ld_moved = detach_tasks(&env);
7275 
7276                 /*
7277                  * We've detached some tasks from busiest_rq. Every
7278                  * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7279                  * unlock busiest->lock, and we are able to be sure
7280                  * that nobody can manipulate the tasks in parallel.
7281                  * See task_rq_lock() family for the details.
7282                  */
7283 
7284                 raw_spin_unlock(&busiest->lock);
7285 
7286                 if (cur_ld_moved) {
7287                         attach_tasks(&env);
7288                         ld_moved += cur_ld_moved;
7289                 }
7290 
7291                 local_irq_restore(flags);
7292 
7293                 if (env.flags & LBF_NEED_BREAK) {
7294                         env.flags &= ~LBF_NEED_BREAK;
7295                         goto more_balance;
7296                 }
7297 
7298                 /*
7299                  * Revisit (affine) tasks on src_cpu that couldn't be moved to
7300                  * us and move them to an alternate dst_cpu in our sched_group
7301                  * where they can run. The upper limit on how many times we
7302                  * iterate on same src_cpu is dependent on number of cpus in our
7303                  * sched_group.
7304                  *
7305                  * This changes load balance semantics a bit on who can move
7306                  * load to a given_cpu. In addition to the given_cpu itself
7307                  * (or a ilb_cpu acting on its behalf where given_cpu is
7308                  * nohz-idle), we now have balance_cpu in a position to move
7309                  * load to given_cpu. In rare situations, this may cause
7310                  * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7311                  * _independently_ and at _same_ time to move some load to
7312                  * given_cpu) causing exceess load to be moved to given_cpu.
7313                  * This however should not happen so much in practice and
7314                  * moreover subsequent load balance cycles should correct the
7315                  * excess load moved.
7316                  */
7317                 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7318 
7319                         /* Prevent to re-select dst_cpu via env's cpus */
7320                         cpumask_clear_cpu(env.dst_cpu, env.cpus);
7321 
7322                         env.dst_rq       = cpu_rq(env.new_dst_cpu);
7323                         env.dst_cpu      = env.new_dst_cpu;
7324                         env.flags       &= ~LBF_DST_PINNED;
7325                         env.loop         = 0;
7326                         env.loop_break   = sched_nr_migrate_break;
7327 
7328                         /*
7329                          * Go back to "more_balance" rather than "redo" since we
7330                          * need to continue with same src_cpu.
7331                          */
7332                         goto more_balance;
7333                 }
7334 
7335                 /*
7336                  * We failed to reach balance because of affinity.
7337                  */
7338                 if (sd_parent) {
7339                         int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7340 
7341                         if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7342                                 *group_imbalance = 1;
7343                 }
7344 
7345                 /* All tasks on this runqueue were pinned by CPU affinity */
7346                 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7347                         cpumask_clear_cpu(cpu_of(busiest), cpus);
7348                         if (!cpumask_empty(cpus)) {
7349                                 env.loop = 0;
7350                                 env.loop_break = sched_nr_migrate_break;
7351                                 goto redo;
7352                         }
7353                         goto out_all_pinned;
7354                 }
7355         }
7356 
7357         if (!ld_moved) {
7358                 schedstat_inc(sd, lb_failed[idle]);
7359                 /*
7360                  * Increment the failure counter only on periodic balance.
7361                  * We do not want newidle balance, which can be very
7362                  * frequent, pollute the failure counter causing
7363                  * excessive cache_hot migrations and active balances.
7364                  */
7365                 if (idle != CPU_NEWLY_IDLE)
7366                         sd->nr_balance_failed++;
7367 
7368                 if (need_active_balance(&env)) {
7369                         raw_spin_lock_irqsave(&busiest->lock, flags);
7370 
7371                         /* don't kick the active_load_balance_cpu_stop,
7372                          * if the curr task on busiest cpu can't be
7373                          * moved to this_cpu
7374                          */
7375                         if (!cpumask_test_cpu(this_cpu,
7376                                         tsk_cpus_allowed(busiest->curr))) {
7377                                 raw_spin_unlock_irqrestore(&busiest->lock,
7378                                                             flags);
7379                                 env.flags |= LBF_ALL_PINNED;
7380                                 goto out_one_pinned;
7381                         }
7382 
7383                         /*
7384                          * ->active_balance synchronizes accesses to
7385                          * ->active_balance_work.  Once set, it's cleared
7386                          * only after active load balance is finished.
7387                          */
7388                         if (!busiest->active_balance) {
7389                                 busiest->active_balance = 1;
7390                                 busiest->push_cpu = this_cpu;
7391                                 active_balance = 1;
7392                         }
7393                         raw_spin_unlock_irqrestore(&busiest->lock, flags);
7394 
7395                         if (active_balance) {
7396                                 stop_one_cpu_nowait(cpu_of(busiest),
7397                                         active_load_balance_cpu_stop, busiest,
7398                                         &busiest->active_balance_work);
7399                         }
7400 
7401                         /*
7402                          * We've kicked active balancing, reset the failure
7403                          * counter.
7404                          */
7405                         sd->nr_balance_failed = sd->cache_nice_tries+1;
7406                 }
7407         } else
7408                 sd->nr_balance_failed = 0;
7409 
7410         if (likely(!active_balance)) {
7411                 /* We were unbalanced, so reset the balancing interval */
7412                 sd->balance_interval = sd->min_interval;
7413         } else {
7414                 /*
7415                  * If we've begun active balancing, start to back off. This
7416                  * case may not be covered by the all_pinned logic if there
7417                  * is only 1 task on the busy runqueue (because we don't call
7418                  * detach_tasks).
7419                  */
7420                 if (sd->balance_interval < sd->max_interval)
7421                         sd->balance_interval *= 2;
7422         }
7423 
7424         goto out;
7425 
7426 out_balanced:
7427         /*
7428          * We reach balance although we may have faced some affinity
7429          * constraints. Clear the imbalance flag if it was set.
7430          */
7431         if (sd_parent) {
7432                 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7433 
7434                 if (*group_imbalance)
7435                         *group_imbalance = 0;
7436         }
7437 
7438 out_all_pinned:
7439         /*
7440          * We reach balance because all tasks are pinned at this level so
7441          * we can't migrate them. Let the imbalance flag set so parent level
7442          * can try to migrate them.
7443          */
7444         schedstat_inc(sd, lb_balanced[idle]);
7445 
7446         sd->nr_balance_failed = 0;
7447 
7448 out_one_pinned:
7449         /* tune up the balancing interval */
7450         if (((env.flags & LBF_ALL_PINNED) &&
7451                         sd->balance_interval < MAX_PINNED_INTERVAL) ||
7452                         (sd->balance_interval < sd->max_interval))
7453                 sd->balance_interval *= 2;
7454 
7455         ld_moved = 0;
7456 out:
7457         return ld_moved;
7458 }
7459 
7460 static inline unsigned long
7461 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7462 {
7463         unsigned long interval = sd->balance_interval;
7464 
7465         if (cpu_busy)
7466                 interval *= sd->busy_factor;
7467 
7468         /* scale ms to jiffies */
7469         interval = msecs_to_jiffies(interval);
7470         interval = clamp(interval, 1UL, max_load_balance_interval);
7471 
7472         return interval;
7473 }
7474 
7475 static inline void
7476 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
7477 {
7478         unsigned long interval, next;
7479 
7480         interval = get_sd_balance_interval(sd, cpu_busy);
7481         next = sd->last_balance + interval;
7482 
7483         if (time_after(*next_balance, next))
7484                 *next_balance = next;
7485 }
7486 
7487 /*
7488  * idle_balance is called by schedule() if this_cpu is about to become
7489  * idle. Attempts to pull tasks from other CPUs.
7490  */
7491 static int idle_balance(struct rq *this_rq)
7492 {
7493         unsigned long next_balance = jiffies + HZ;
7494         int this_cpu = this_rq->cpu;
7495         struct sched_domain *sd;
7496         int pulled_task = 0;
7497         u64 curr_cost = 0;
7498 
7499         /*
7500          * We must set idle_stamp _before_ calling idle_balance(), such that we
7501          * measure the duration of idle_balance() as idle time.
7502          */
7503         this_rq->idle_stamp = rq_clock(this_rq);
7504 
7505         if (this_rq->avg_idle < sysctl_sched_migration_cost ||
7506             !this_rq->rd->overload) {
7507                 rcu_read_lock();
7508                 sd = rcu_dereference_check_sched_domain(this_rq->sd);
7509                 if (sd)
7510                         update_next_balance(sd, 0, &next_balance);
7511                 rcu_read_unlock();
7512 
7513                 goto out;
7514         }
7515 
7516         raw_spin_unlock(&this_rq->lock);
7517 
7518         update_blocked_averages(this_cpu);
7519         rcu_read_lock();
7520         for_each_domain(this_cpu, sd) {
7521                 int continue_balancing = 1;
7522                 u64 t0, domain_cost;
7523 
7524                 if (!(sd->flags & SD_LOAD_BALANCE))
7525                         continue;
7526 
7527                 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7528                         update_next_balance(sd, 0, &next_balance);
7529                         break;
7530                 }
7531 
7532                 if (sd->flags & SD_BALANCE_NEWIDLE) {
7533                         t0 = sched_clock_cpu(this_cpu);
7534 
7535                         pulled_task = load_balance(this_cpu, this_rq,
7536                                                    sd, CPU_NEWLY_IDLE,
7537                                                    &continue_balancing);
7538 
7539                         domain_cost = sched_clock_cpu(this_cpu) - t0;
7540                         if (domain_cost > sd->max_newidle_lb_cost)
7541                                 sd->max_newidle_lb_cost = domain_cost;
7542 
7543                         curr_cost += domain_cost;
7544                 }
7545 
7546                 update_next_balance(sd, 0, &next_balance);
7547 
7548                 /*
7549                  * Stop searching for tasks to pull if there are
7550                  * now runnable tasks on this rq.
7551                  */
7552                 if (pulled_task || this_rq->nr_running > 0)
7553                         break;
7554         }
7555         rcu_read_unlock();
7556 
7557         raw_spin_lock(&this_rq->lock);
7558 
7559         if (curr_cost > this_rq->max_idle_balance_cost)
7560                 this_rq->max_idle_balance_cost = curr_cost;
7561 
7562         /*
7563          * While browsing the domains, we released the rq lock, a task could
7564          * have been enqueued in the meantime. Since we're not going idle,
7565          * pretend we pulled a task.
7566          */
7567         if (this_rq->cfs.h_nr_running && !pulled_task)
7568                 pulled_task = 1;
7569 
7570 out:
7571         /* Move the next balance forward */
7572         if (time_after(this_rq->next_balance, next_balance))
7573                 this_rq->next_balance = next_balance;
7574 
7575         /* Is there a task of a high priority class? */
7576         if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7577                 pulled_task = -1;
7578 
7579         if (pulled_task)
7580                 this_rq->idle_stamp = 0;
7581 
7582         return pulled_task;
7583 }
7584 
7585 /*
7586  * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7587  * running tasks off the busiest CPU onto idle CPUs. It requires at
7588  * least 1 task to be running on each physical CPU where possible, and
7589  * avoids physical / logical imbalances.
7590  */
7591 static int active_load_balance_cpu_stop(void *data)
7592 {
7593         struct rq *busiest_rq = data;
7594         int busiest_cpu = cpu_of(busiest_rq);
7595         int target_cpu = busiest_rq->push_cpu;
7596         struct rq *target_rq = cpu_rq(target_cpu);
7597         struct sched_domain *sd;
7598         struct task_struct *p = NULL;
7599 
7600         raw_spin_lock_irq(&busiest_rq->lock);
7601 
7602         /* make sure the requested cpu hasn't gone down in the meantime */
7603         if (unlikely(busiest_cpu != smp_processor_id() ||
7604                      !busiest_rq->active_balance))
7605                 goto out_unlock;
7606 
7607         /* Is there any task to move? */
7608         if (busiest_rq->nr_running <= 1)
7609                 goto out_unlock;
7610 
7611         /*
7612          * This condition is "impossible", if it occurs
7613          * we need to fix it. Originally reported by
7614          * Bjorn Helgaas on a 128-cpu setup.
7615          */
7616         BUG_ON(busiest_rq == target_rq);
7617 
7618         /* Search for an sd spanning us and the target CPU. */
7619         rcu_read_lock();
7620         for_each_domain(target_cpu, sd) {
7621                 if ((sd->flags & SD_LOAD_BALANCE) &&
7622                     cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7623                                 break;
7624         }
7625 
7626         if (likely(sd)) {
7627                 struct lb_env env = {
7628                         .sd             = sd,
7629                         .dst_cpu        = target_cpu,
7630                         .dst_rq         = target_rq,
7631                         .src_cpu        = busiest_rq->cpu,
7632                         .src_rq         = busiest_rq,
7633                         .idle           = CPU_IDLE,
7634                 };
7635 
7636                 schedstat_inc(sd, alb_count);
7637 
7638                 p = detach_one_task(&env);
7639                 if (p)
7640                         schedstat_inc(sd, alb_pushed);
7641                 else
7642                         schedstat_inc(sd, alb_failed);
7643         }
7644         rcu_read_unlock();
7645 out_unlock:
7646         busiest_rq->active_balance = 0;
7647         raw_spin_unlock(&busiest_rq->lock);
7648 
7649         if (p)
7650                 attach_one_task(target_rq, p);
7651 
7652         local_irq_enable();
7653 
7654         return 0;
7655 }
7656 
7657 static inline int on_null_domain(struct rq *rq)
7658 {
7659         return unlikely(!rcu_dereference_sched(rq->sd));
7660 }
7661 
7662 #ifdef CONFIG_NO_HZ_COMMON
7663 /*
7664  * idle load balancing details
7665  * - When one of the busy CPUs notice that there may be an idle rebalancing
7666  *   needed, they will kick the idle load balancer, which then does idle
7667  *   load balancing for all the idle CPUs.
7668  */
7669 static struct {
7670         cpumask_var_t idle_cpus_mask;
7671         atomic_t nr_cpus;
7672         unsigned long next_balance;     /* in jiffy units */
7673 } nohz ____cacheline_aligned;
7674 
7675 static inline int find_new_ilb(void)
7676 {
7677         int ilb = cpumask_first(nohz.idle_cpus_mask);
7678 
7679         if (ilb < nr_cpu_ids && idle_cpu(ilb))
7680                 return ilb;
7681 
7682         return nr_cpu_ids;
7683 }
7684 
7685 /*
7686  * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7687  * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7688  * CPU (if there is one).
7689  */
7690 static void nohz_balancer_kick(void)
7691 {
7692         int ilb_cpu;
7693 
7694         nohz.next_balance++;
7695 
7696         ilb_cpu = find_new_ilb();
7697 
7698         if (ilb_cpu >= nr_cpu_ids)
7699                 return;
7700 
7701         if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7702                 return;
7703         /*
7704          * Use smp_send_reschedule() instead of resched_cpu().
7705          * This way we generate a sched IPI on the target cpu which
7706          * is idle. And the softirq performing nohz idle load balance
7707          * will be run before returning from the IPI.
7708          */
7709         smp_send_reschedule(ilb_cpu);
7710         return;
7711 }
7712 
7713 static inline void nohz_balance_exit_idle(int cpu)
7714 {
7715         if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7716                 /*
7717                  * Completely isolated CPUs don't ever set, so we must test.
7718                  */
7719                 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7720                         cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7721                         atomic_dec(&nohz.nr_cpus);
7722                 }
7723                 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7724         }
7725 }
7726 
7727 static inline void set_cpu_sd_state_busy(void)
7728 {
7729         struct sched_domain *sd;
7730         int cpu = smp_processor_id();
7731 
7732         rcu_read_lock();
7733         sd = rcu_dereference(per_cpu(sd_busy, cpu));
7734 
7735         if (!sd || !sd->nohz_idle)
7736                 goto unlock;
7737         sd->nohz_idle = 0;
7738 
7739         atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7740 unlock:
7741         rcu_read_unlock();
7742 }
7743 
7744 void set_cpu_sd_state_idle(void)
7745 {
7746         struct sched_domain *sd;
7747         int cpu = smp_processor_id();
7748 
7749         rcu_read_lock();
7750         sd = rcu_dereference(per_cpu(sd_busy, cpu));
7751 
7752         if (!sd || sd->nohz_idle)
7753                 goto unlock;
7754         sd->nohz_idle = 1;
7755 
7756         atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7757 unlock:
7758         rcu_read_unlock();
7759 }
7760 
7761 /*
7762  * This routine will record that the cpu is going idle with tick stopped.
7763  * This info will be used in performing idle load balancing in the future.
7764  */
7765 void nohz_balance_enter_idle(int cpu)
7766 {
7767         /*
7768          * If this cpu is going down, then nothing needs to be done.
7769          */
7770         if (!cpu_active(cpu))
7771                 return;
7772 
7773         if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7774                 return;
7775 
7776         /*
7777          * If we're a completely isolated CPU, we don't play.
7778          */
7779         if (on_null_domain(cpu_rq(cpu)))
7780                 return;
7781 
7782         cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7783         atomic_inc(&nohz.nr_cpus);
7784         set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7785 }
7786 
7787 static int sched_ilb_notifier(struct notifier_block *nfb,
7788                                         unsigned long action, void *hcpu)
7789 {
7790         switch (action & ~CPU_TASKS_FROZEN) {
7791         case CPU_DYING:
7792                 nohz_balance_exit_idle(smp_processor_id());
7793                 return NOTIFY_OK;
7794         default:
7795                 return NOTIFY_DONE;
7796         }
7797 }
7798 #endif
7799 
7800 static DEFINE_SPINLOCK(balancing);
7801 
7802 /*
7803  * Scale the max load_balance interval with the number of CPUs in the system.
7804  * This trades load-balance latency on larger machines for less cross talk.
7805  */
7806 void update_max_interval(void)
7807 {
7808         max_load_balance_interval = HZ*num_online_cpus()/10;
7809 }
7810 
7811 /*
7812  * It checks each scheduling domain to see if it is due to be balanced,
7813  * and initiates a balancing operation if so.
7814  *
7815  * Balancing parameters are set up in init_sched_domains.
7816  */
7817 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7818 {
7819         int continue_balancing = 1;
7820         int cpu = rq->cpu;
7821         unsigned long interval;
7822         struct sched_domain *sd;
7823         /* Earliest time when we have to do rebalance again */
7824         unsigned long next_balance = jiffies + 60*HZ;
7825         int update_next_balance = 0;
7826         int need_serialize, need_decay = 0;
7827         u64 max_cost = 0;
7828 
7829         update_blocked_averages(cpu);
7830 
7831         rcu_read_lock();
7832         for_each_domain(cpu, sd) {
7833                 /*
7834                  * Decay the newidle max times here because this is a regular
7835                  * visit to all the domains. Decay ~1% per second.
7836                  */
7837                 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7838                         sd->max_newidle_lb_cost =
7839                                 (sd->max_newidle_lb_cost * 253) / 256;
7840                         sd->next_decay_max_lb_cost = jiffies + HZ;
7841                         need_decay = 1;
7842                 }
7843                 max_cost += sd->max_newidle_lb_cost;
7844 
7845                 if (!(sd->flags & SD_LOAD_BALANCE))
7846                         continue;
7847 
7848                 /*
7849                  * Stop the load balance at this level. There is another
7850                  * CPU in our sched group which is doing load balancing more
7851                  * actively.
7852                  */
7853                 if (!continue_balancing) {
7854                         if (need_decay)
7855                                 continue;
7856                         break;
7857                 }
7858 
7859                 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7860 
7861                 need_serialize = sd->flags & SD_SERIALIZE;
7862                 if (need_serialize) {
7863                         if (!spin_trylock(&balancing))
7864                                 goto out;
7865                 }
7866 
7867                 if (time_after_eq(jiffies, sd->last_balance + interval)) {
7868                         if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7869                                 /*
7870                                  * The LBF_DST_PINNED logic could have changed
7871                                  * env->dst_cpu, so we can't know our idle
7872                                  * state even if we migrated tasks. Update it.
7873                                  */
7874                                 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7875                         }
7876                         sd->last_balance = jiffies;
7877                         interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7878                 }
7879                 if (need_serialize)
7880                         spin_unlock(&balancing);
7881 out:
7882                 if (time_after(next_balance, sd->last_balance + interval)) {
7883                         next_balance = sd->last_balance + interval;
7884                         update_next_balance = 1;
7885                 }
7886         }
7887         if (need_decay) {
7888                 /*
7889                  * Ensure the rq-wide value also decays but keep it at a
7890                  * reasonable floor to avoid funnies with rq->avg_idle.
7891                  */
7892                 rq->max_idle_balance_cost =
7893                         max((u64)sysctl_sched_migration_cost, max_cost);
7894         }
7895         rcu_read_unlock();
7896 
7897         /*
7898          * next_balance will be updated only when there is a need.
7899          * When the cpu is attached to null domain for ex, it will not be
7900          * updated.
7901          */
7902         if (likely(update_next_balance)) {
7903                 rq->next_balance = next_balance;
7904 
7905 #ifdef CONFIG_NO_HZ_COMMON
7906                 /*
7907                  * If this CPU has been elected to perform the nohz idle
7908                  * balance. Other idle CPUs have already rebalanced with
7909                  * nohz_idle_balance() and nohz.next_balance has been
7910                  * updated accordingly. This CPU is now running the idle load
7911                  * balance for itself and we need to update the
7912                  * nohz.next_balance accordingly.
7913                  */
7914                 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
7915                         nohz.next_balance = rq->next_balance;
7916 #endif
7917         }
7918 }
7919 
7920 #ifdef CONFIG_NO_HZ_COMMON
7921 /*
7922  * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7923  * rebalancing for all the cpus for whom scheduler ticks are stopped.
7924  */
7925 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7926 {
7927         int this_cpu = this_rq->cpu;
7928         struct rq *rq;
7929         int balance_cpu;
7930         /* Earliest time when we have to do rebalance again */
7931         unsigned long next_balance = jiffies + 60*HZ;
7932         int update_next_balance = 0;
7933 
7934         if (idle != CPU_IDLE ||
7935             !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7936                 goto end;
7937 
7938         for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7939                 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7940                         continue;
7941 
7942                 /*
7943                  * If this cpu gets work to do, stop the load balancing
7944                  * work being done for other cpus. Next load
7945                  * balancing owner will pick it up.
7946                  */
7947                 if (need_resched())
7948                         break;
7949 
7950                 rq = cpu_rq(balance_cpu);
7951 
7952                 /*
7953                  * If time for next balance is due,
7954                  * do the balance.
7955                  */
7956                 if (time_after_eq(jiffies, rq->next_balance)) {
7957                         raw_spin_lock_irq(&rq->lock);
7958                         update_rq_clock(rq);
7959                         update_cpu_load_idle(rq);
7960                         raw_spin_unlock_irq(&rq->lock);
7961                         rebalance_domains(rq, CPU_IDLE);
7962                 }
7963 
7964                 if (time_after(next_balance, rq->next_balance)) {
7965                         next_balance = rq->next_balance;
7966                         update_next_balance = 1;
7967                 }
7968         }
7969 
7970         /*
7971          * next_balance will be updated only when there is a need.
7972          * When the CPU is attached to null domain for ex, it will not be
7973          * updated.
7974          */
7975         if (likely(update_next_balance))
7976                 nohz.next_balance = next_balance;
7977 end:
7978         clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7979 }
7980 
7981 /*
7982  * Current heuristic for kicking the idle load balancer in the presence
7983  * of an idle cpu in the system.
7984  *   - This rq has more than one task.
7985  *   - This rq has at least one CFS task and the capacity of the CPU is
7986  *     significantly reduced because of RT tasks or IRQs.
7987  *   - At parent of LLC scheduler domain level, this cpu's scheduler group has
7988  *     multiple busy cpu.
7989  *   - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7990  *     domain span are idle.
7991  */
7992 static inline bool nohz_kick_needed(struct rq *rq)
7993 {
7994         unsigned long now = jiffies;
7995         struct sched_domain *sd;
7996         struct sched_group_capacity *sgc;
7997         int nr_busy, cpu = rq->cpu;
7998         bool kick = false;
7999 
8000         if (unlikely(rq->idle_balance))
8001                 return false;
8002 
8003        /*
8004         * We may be recently in ticked or tickless idle mode. At the first
8005         * busy tick after returning from idle, we will update the busy stats.
8006         */
8007         set_cpu_sd_state_busy();
8008         nohz_balance_exit_idle(cpu);
8009 
8010         /*
8011          * None are in tickless mode and hence no need for NOHZ idle load
8012          * balancing.
8013          */
8014         if (likely(!atomic_read(&nohz.nr_cpus)))
8015                 return false;
8016 
8017         if (time_before(now, nohz.next_balance))
8018                 return false;
8019 
8020         if (rq->nr_running >= 2)
8021                 return true;
8022 
8023         rcu_read_lock();
8024         sd = rcu_dereference(per_cpu(sd_busy, cpu));
8025         if (sd) {
8026                 sgc = sd->groups->sgc;
8027                 nr_busy = atomic_read(&sgc->nr_busy_cpus);
8028 
8029                 if (nr_busy > 1) {
8030                         kick = true;
8031                         goto unlock;
8032                 }
8033 
8034         }
8035 
8036         sd = rcu_dereference(rq->sd);
8037         if (sd) {
8038                 if ((rq->cfs.h_nr_running >= 1) &&
8039                                 check_cpu_capacity(rq, sd)) {
8040                         kick = true;
8041                         goto unlock;
8042                 }
8043         }
8044 
8045         sd = rcu_dereference(per_cpu(sd_asym, cpu));
8046         if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
8047                                   sched_domain_span(sd)) < cpu)) {
8048                 kick = true;
8049                 goto unlock;
8050         }
8051 
8052 unlock:
8053         rcu_read_unlock();
8054         return kick;
8055 }
8056 #else
8057 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8058 #endif
8059 
8060 /*
8061  * run_rebalance_domains is triggered when needed from the scheduler tick.
8062  * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8063  */
8064 static void run_rebalance_domains(struct softirq_action *h)
8065 {
8066         struct rq *this_rq = this_rq();
8067         enum cpu_idle_type idle = this_rq->idle_balance ?
8068                                                 CPU_IDLE : CPU_NOT_IDLE;
8069 
8070         /*
8071          * If this cpu has a pending nohz_balance_kick, then do the
8072          * balancing on behalf of the other idle cpus whose ticks are
8073          * stopped. Do nohz_idle_balance *before* rebalance_domains to
8074          * give the idle cpus a chance to load balance. Else we may
8075          * load balance only within the local sched_domain hierarchy
8076          * and abort nohz_idle_balance altogether if we pull some load.
8077          */
8078         nohz_idle_balance(this_rq, idle);
8079         rebalance_domains(this_rq, idle);
8080 }
8081 
8082 /*
8083  * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8084  */
8085 void trigger_load_balance(struct rq *rq)
8086 {
8087         /* Don't need to rebalance while attached to NULL domain */
8088         if (unlikely(on_null_domain(rq)))
8089                 return;
8090 
8091         if (time_after_eq(jiffies, rq->next_balance))
8092                 raise_softirq(SCHED_SOFTIRQ);
8093 #ifdef CONFIG_NO_HZ_COMMON
8094         if (nohz_kick_needed(rq))
8095                 nohz_balancer_kick();
8096 #endif
8097 }
8098 
8099 static void rq_online_fair(struct rq *rq)
8100 {
8101         update_sysctl();
8102 
8103         update_runtime_enabled(rq);
8104 }
8105 
8106 static void rq_offline_fair(struct rq *rq)
8107 {
8108         update_sysctl();
8109 
8110         /* Ensure any throttled groups are reachable by pick_next_task */
8111         unthrottle_offline_cfs_rqs(rq);
8112 }
8113 
8114 #endif /* CONFIG_SMP */
8115 
8116 /*
8117  * scheduler tick hitting a task of our scheduling class:
8118  */
8119 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
8120 {
8121         struct cfs_rq *cfs_rq;
8122         struct sched_entity *se = &curr->se;
8123 
8124         for_each_sched_entity(se) {
8125                 cfs_rq = cfs_rq_of(se);
8126                 entity_tick(cfs_rq, se, queued);
8127         }
8128 
8129         if (static_branch_unlikely(&sched_numa_balancing))
8130                 task_tick_numa(rq, curr);
8131 }
8132 
8133 /*
8134  * called on fork with the child task as argument from the parent's context
8135  *  - child not yet on the tasklist
8136  *  - preemption disabled
8137  */
8138 static void task_fork_fair(struct task_struct *p)
8139 {
8140         struct cfs_rq *cfs_rq;
8141         struct sched_entity *se = &p->se, *curr;
8142         int this_cpu = smp_processor_id();
8143         struct rq *rq = this_rq();
8144         unsigned long flags;
8145 
8146         raw_spin_lock_irqsave(&rq->lock, flags);
8147 
8148         update_rq_clock(rq);
8149 
8150         cfs_rq = task_cfs_rq(current);
8151         curr = cfs_rq->curr;
8152 
8153         /*
8154          * Not only the cpu but also the task_group of the parent might have
8155          * been changed after parent->se.parent,cfs_rq were copied to
8156          * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
8157          * of child point to valid ones.
8158          */
8159         rcu_read_lock();
8160         __set_task_cpu(p, this_cpu);
8161         rcu_read_unlock();
8162 
8163         update_curr(cfs_rq);
8164 
8165         if (curr)
8166                 se->vruntime = curr->vruntime;
8167         place_entity(cfs_rq, se, 1);
8168 
8169         if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8170                 /*
8171                  * Upon rescheduling, sched_class::put_prev_task() will place
8172                  * 'current' within the tree based on its new key value.
8173                  */
8174                 swap(curr->vruntime, se->vruntime);
8175                 resched_curr(rq);
8176         }
8177 
8178         se->vruntime -= cfs_rq->min_vruntime;
8179 
8180         raw_spin_unlock_irqrestore(&rq->lock, flags);
8181 }
8182 
8183 /*
8184  * Priority of the task has changed. Check to see if we preempt
8185  * the current task.
8186  */
8187 static void
8188 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
8189 {
8190         if (!task_on_rq_queued(p))
8191                 return;
8192 
8193         /*
8194          * Reschedule if we are currently running on this runqueue and
8195          * our priority decreased, or if we are not currently running on
8196          * this runqueue and our priority is higher than the current's
8197          */
8198         if (rq->curr == p) {
8199                 if (p->prio > oldprio)
8200                         resched_curr(rq);
8201         } else
8202                 check_preempt_curr(rq, p, 0);
8203 }
8204 
8205 static inline bool vruntime_normalized(struct task_struct *p)
8206 {
8207         struct sched_entity *se = &p->se;
8208 
8209         /*
8210          * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
8211          * the dequeue_entity(.flags=0) will already have normalized the
8212          * vruntime.
8213          */
8214         if (p->on_rq)
8215                 return true;
8216 
8217         /*
8218          * When !on_rq, vruntime of the task has usually NOT been normalized.
8219          * But there are some cases where it has already been normalized:
8220          *
8221          * - A forked child which is waiting for being woken up by
8222          *   wake_up_new_task().
8223          * - A task which has been woken up by try_to_wake_up() and
8224          *   waiting for actually being woken up by sched_ttwu_pending().
8225          */
8226         if (!se->sum_exec_runtime || p->state == TASK_WAKING)
8227                 return true;
8228 
8229         return false;
8230 }
8231 
8232 static void detach_task_cfs_rq(struct task_struct *p)
8233 {
8234         struct sched_entity *se = &p->se;
8235         struct cfs_rq *cfs_rq = cfs_rq_of(se);
8236 
8237         if (!vruntime_normalized(p)) {
8238                 /*
8239                  * Fix up our vruntime so that the current sleep doesn't
8240                  * cause 'unlimited' sleep bonus.
8241                  */
8242                 place_entity(cfs_rq, se, 0);
8243                 se->vruntime -= cfs_rq->min_vruntime;
8244         }
8245 
8246         /* Catch up with the cfs_rq and remove our load when we leave */
8247         detach_entity_load_avg(cfs_rq, se);
8248 }
8249 
8250 static void attach_task_cfs_rq(struct task_struct *p)
8251 {
8252         struct sched_entity *se = &p->se;
8253         struct cfs_rq *cfs_rq = cfs_rq_of(se);
8254 
8255 #ifdef CONFIG_FAIR_GROUP_SCHED
8256         /*
8257          * Since the real-depth could have been changed (only FAIR
8258          * class maintain depth value), reset depth properly.
8259          */
8260         se->depth = se->parent ? se->parent->depth + 1 : 0;
8261 #endif
8262 
8263         /* Synchronize task with its cfs_rq */
8264         attach_entity_load_avg(cfs_rq, se);
8265 
8266         if (!vruntime_normalized(p))
8267                 se->vruntime += cfs_rq->min_vruntime;
8268 }
8269 
8270 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8271 {
8272         detach_task_cfs_rq(p);
8273 }
8274 
8275 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8276 {
8277         attach_task_cfs_rq(p);
8278 
8279         if (task_on_rq_queued(p)) {
8280                 /*
8281                  * We were most likely switched from sched_rt, so
8282                  * kick off the schedule if running, otherwise just see
8283                  * if we can still preempt the current task.
8284                  */
8285                 if (rq->curr == p)
8286                         resched_curr(rq);
8287                 else
8288                         check_preempt_curr(rq, p, 0);
8289         }
8290 }
8291 
8292 /* Account for a task changing its policy or group.
8293  *
8294  * This routine is mostly called to set cfs_rq->curr field when a task
8295  * migrates between groups/classes.
8296  */
8297 static void set_curr_task_fair(struct rq *rq)
8298 {
8299         struct sched_entity *se = &rq->curr->se;
8300 
8301         for_each_sched_entity(se) {
8302                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8303 
8304                 set_next_entity(cfs_rq, se);
8305                 /* ensure bandwidth has been allocated on our new cfs_rq */
8306                 account_cfs_rq_runtime(cfs_rq, 0);
8307         }
8308 }