Version:  2.0.40 2.2.26 2.4.37 3.8 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

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/latencytop.h>
 24 #include <linux/sched.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 static inline void
742 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
743 {
744         schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
745 }
746 
747 /*
748  * Task is being enqueued - update stats:
749  */
750 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
751 {
752         /*
753          * Are we enqueueing a waiting task? (for current tasks
754          * a dequeue/enqueue event is a NOP)
755          */
756         if (se != cfs_rq->curr)
757                 update_stats_wait_start(cfs_rq, se);
758 }
759 
760 static void
761 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
762 {
763         schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
764                         rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
765         schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
766         schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
767                         rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
768 #ifdef CONFIG_SCHEDSTATS
769         if (entity_is_task(se)) {
770                 trace_sched_stat_wait(task_of(se),
771                         rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
772         }
773 #endif
774         schedstat_set(se->statistics.wait_start, 0);
775 }
776 
777 static inline void
778 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
779 {
780         /*
781          * Mark the end of the wait period if dequeueing a
782          * waiting task:
783          */
784         if (se != cfs_rq->curr)
785                 update_stats_wait_end(cfs_rq, se);
786 }
787 
788 /*
789  * We are picking a new current task - update its stats:
790  */
791 static inline void
792 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
793 {
794         /*
795          * We are starting a new run period:
796          */
797         se->exec_start = rq_clock_task(rq_of(cfs_rq));
798 }
799 
800 /**************************************************
801  * Scheduling class queueing methods:
802  */
803 
804 #ifdef CONFIG_NUMA_BALANCING
805 /*
806  * Approximate time to scan a full NUMA task in ms. The task scan period is
807  * calculated based on the tasks virtual memory size and
808  * numa_balancing_scan_size.
809  */
810 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
811 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
812 
813 /* Portion of address space to scan in MB */
814 unsigned int sysctl_numa_balancing_scan_size = 256;
815 
816 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
817 unsigned int sysctl_numa_balancing_scan_delay = 1000;
818 
819 static unsigned int task_nr_scan_windows(struct task_struct *p)
820 {
821         unsigned long rss = 0;
822         unsigned long nr_scan_pages;
823 
824         /*
825          * Calculations based on RSS as non-present and empty pages are skipped
826          * by the PTE scanner and NUMA hinting faults should be trapped based
827          * on resident pages
828          */
829         nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
830         rss = get_mm_rss(p->mm);
831         if (!rss)
832                 rss = nr_scan_pages;
833 
834         rss = round_up(rss, nr_scan_pages);
835         return rss / nr_scan_pages;
836 }
837 
838 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
839 #define MAX_SCAN_WINDOW 2560
840 
841 static unsigned int task_scan_min(struct task_struct *p)
842 {
843         unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
844         unsigned int scan, floor;
845         unsigned int windows = 1;
846 
847         if (scan_size < MAX_SCAN_WINDOW)
848                 windows = MAX_SCAN_WINDOW / scan_size;
849         floor = 1000 / windows;
850 
851         scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
852         return max_t(unsigned int, floor, scan);
853 }
854 
855 static unsigned int task_scan_max(struct task_struct *p)
856 {
857         unsigned int smin = task_scan_min(p);
858         unsigned int smax;
859 
860         /* Watch for min being lower than max due to floor calculations */
861         smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
862         return max(smin, smax);
863 }
864 
865 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
866 {
867         rq->nr_numa_running += (p->numa_preferred_nid != -1);
868         rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
869 }
870 
871 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
872 {
873         rq->nr_numa_running -= (p->numa_preferred_nid != -1);
874         rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
875 }
876 
877 struct numa_group {
878         atomic_t refcount;
879 
880         spinlock_t lock; /* nr_tasks, tasks */
881         int nr_tasks;
882         pid_t gid;
883 
884         struct rcu_head rcu;
885         nodemask_t active_nodes;
886         unsigned long total_faults;
887         /*
888          * Faults_cpu is used to decide whether memory should move
889          * towards the CPU. As a consequence, these stats are weighted
890          * more by CPU use than by memory faults.
891          */
892         unsigned long *faults_cpu;
893         unsigned long faults[0];
894 };
895 
896 /* Shared or private faults. */
897 #define NR_NUMA_HINT_FAULT_TYPES 2
898 
899 /* Memory and CPU locality */
900 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
901 
902 /* Averaged statistics, and temporary buffers. */
903 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
904 
905 pid_t task_numa_group_id(struct task_struct *p)
906 {
907         return p->numa_group ? p->numa_group->gid : 0;
908 }
909 
910 /*
911  * The averaged statistics, shared & private, memory & cpu,
912  * occupy the first half of the array. The second half of the
913  * array is for current counters, which are averaged into the
914  * first set by task_numa_placement.
915  */
916 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
917 {
918         return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
919 }
920 
921 static inline unsigned long task_faults(struct task_struct *p, int nid)
922 {
923         if (!p->numa_faults)
924                 return 0;
925 
926         return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
927                 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
928 }
929 
930 static inline unsigned long group_faults(struct task_struct *p, int nid)
931 {
932         if (!p->numa_group)
933                 return 0;
934 
935         return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
936                 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
937 }
938 
939 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
940 {
941         return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
942                 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
943 }
944 
945 /* Handle placement on systems where not all nodes are directly connected. */
946 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
947                                         int maxdist, bool task)
948 {
949         unsigned long score = 0;
950         int node;
951 
952         /*
953          * All nodes are directly connected, and the same distance
954          * from each other. No need for fancy placement algorithms.
955          */
956         if (sched_numa_topology_type == NUMA_DIRECT)
957                 return 0;
958 
959         /*
960          * This code is called for each node, introducing N^2 complexity,
961          * which should be ok given the number of nodes rarely exceeds 8.
962          */
963         for_each_online_node(node) {
964                 unsigned long faults;
965                 int dist = node_distance(nid, node);
966 
967                 /*
968                  * The furthest away nodes in the system are not interesting
969                  * for placement; nid was already counted.
970                  */
971                 if (dist == sched_max_numa_distance || node == nid)
972                         continue;
973 
974                 /*
975                  * On systems with a backplane NUMA topology, compare groups
976                  * of nodes, and move tasks towards the group with the most
977                  * memory accesses. When comparing two nodes at distance
978                  * "hoplimit", only nodes closer by than "hoplimit" are part
979                  * of each group. Skip other nodes.
980                  */
981                 if (sched_numa_topology_type == NUMA_BACKPLANE &&
982                                         dist > maxdist)
983                         continue;
984 
985                 /* Add up the faults from nearby nodes. */
986                 if (task)
987                         faults = task_faults(p, node);
988                 else
989                         faults = group_faults(p, node);
990 
991                 /*
992                  * On systems with a glueless mesh NUMA topology, there are
993                  * no fixed "groups of nodes". Instead, nodes that are not
994                  * directly connected bounce traffic through intermediate
995                  * nodes; a numa_group can occupy any set of nodes.
996                  * The further away a node is, the less the faults count.
997                  * This seems to result in good task placement.
998                  */
999                 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1000                         faults *= (sched_max_numa_distance - dist);
1001                         faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1002                 }
1003 
1004                 score += faults;
1005         }
1006 
1007         return score;
1008 }
1009 
1010 /*
1011  * These return the fraction of accesses done by a particular task, or
1012  * task group, on a particular numa node.  The group weight is given a
1013  * larger multiplier, in order to group tasks together that are almost
1014  * evenly spread out between numa nodes.
1015  */
1016 static inline unsigned long task_weight(struct task_struct *p, int nid,
1017                                         int dist)
1018 {
1019         unsigned long faults, total_faults;
1020 
1021         if (!p->numa_faults)
1022                 return 0;
1023 
1024         total_faults = p->total_numa_faults;
1025 
1026         if (!total_faults)
1027                 return 0;
1028 
1029         faults = task_faults(p, nid);
1030         faults += score_nearby_nodes(p, nid, dist, true);
1031 
1032         return 1000 * faults / total_faults;
1033 }
1034 
1035 static inline unsigned long group_weight(struct task_struct *p, int nid,
1036                                          int dist)
1037 {
1038         unsigned long faults, total_faults;
1039 
1040         if (!p->numa_group)
1041                 return 0;
1042 
1043         total_faults = p->numa_group->total_faults;
1044 
1045         if (!total_faults)
1046                 return 0;
1047 
1048         faults = group_faults(p, nid);
1049         faults += score_nearby_nodes(p, nid, dist, false);
1050 
1051         return 1000 * faults / total_faults;
1052 }
1053 
1054 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1055                                 int src_nid, int dst_cpu)
1056 {
1057         struct numa_group *ng = p->numa_group;
1058         int dst_nid = cpu_to_node(dst_cpu);
1059         int last_cpupid, this_cpupid;
1060 
1061         this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1062 
1063         /*
1064          * Multi-stage node selection is used in conjunction with a periodic
1065          * migration fault to build a temporal task<->page relation. By using
1066          * a two-stage filter we remove short/unlikely relations.
1067          *
1068          * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1069          * a task's usage of a particular page (n_p) per total usage of this
1070          * page (n_t) (in a given time-span) to a probability.
1071          *
1072          * Our periodic faults will sample this probability and getting the
1073          * same result twice in a row, given these samples are fully
1074          * independent, is then given by P(n)^2, provided our sample period
1075          * is sufficiently short compared to the usage pattern.
1076          *
1077          * This quadric squishes small probabilities, making it less likely we
1078          * act on an unlikely task<->page relation.
1079          */
1080         last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1081         if (!cpupid_pid_unset(last_cpupid) &&
1082                                 cpupid_to_nid(last_cpupid) != dst_nid)
1083                 return false;
1084 
1085         /* Always allow migrate on private faults */
1086         if (cpupid_match_pid(p, last_cpupid))
1087                 return true;
1088 
1089         /* A shared fault, but p->numa_group has not been set up yet. */
1090         if (!ng)
1091                 return true;
1092 
1093         /*
1094          * Do not migrate if the destination is not a node that
1095          * is actively used by this numa group.
1096          */
1097         if (!node_isset(dst_nid, ng->active_nodes))
1098                 return false;
1099 
1100         /*
1101          * Source is a node that is not actively used by this
1102          * numa group, while the destination is. Migrate.
1103          */
1104         if (!node_isset(src_nid, ng->active_nodes))
1105                 return true;
1106 
1107         /*
1108          * Both source and destination are nodes in active
1109          * use by this numa group. Maximize memory bandwidth
1110          * by migrating from more heavily used groups, to less
1111          * heavily used ones, spreading the load around.
1112          * Use a 1/4 hysteresis to avoid spurious page movement.
1113          */
1114         return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1115 }
1116 
1117 static unsigned long weighted_cpuload(const int cpu);
1118 static unsigned long source_load(int cpu, int type);
1119 static unsigned long target_load(int cpu, int type);
1120 static unsigned long capacity_of(int cpu);
1121 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1122 
1123 /* Cached statistics for all CPUs within a node */
1124 struct numa_stats {
1125         unsigned long nr_running;
1126         unsigned long load;
1127 
1128         /* Total compute capacity of CPUs on a node */
1129         unsigned long compute_capacity;
1130 
1131         /* Approximate capacity in terms of runnable tasks on a node */
1132         unsigned long task_capacity;
1133         int has_free_capacity;
1134 };
1135 
1136 /*
1137  * XXX borrowed from update_sg_lb_stats
1138  */
1139 static void update_numa_stats(struct numa_stats *ns, int nid)
1140 {
1141         int smt, cpu, cpus = 0;
1142         unsigned long capacity;
1143 
1144         memset(ns, 0, sizeof(*ns));
1145         for_each_cpu(cpu, cpumask_of_node(nid)) {
1146                 struct rq *rq = cpu_rq(cpu);
1147 
1148                 ns->nr_running += rq->nr_running;
1149                 ns->load += weighted_cpuload(cpu);
1150                 ns->compute_capacity += capacity_of(cpu);
1151 
1152                 cpus++;
1153         }
1154 
1155         /*
1156          * If we raced with hotplug and there are no CPUs left in our mask
1157          * the @ns structure is NULL'ed and task_numa_compare() will
1158          * not find this node attractive.
1159          *
1160          * We'll either bail at !has_free_capacity, or we'll detect a huge
1161          * imbalance and bail there.
1162          */
1163         if (!cpus)
1164                 return;
1165 
1166         /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1167         smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1168         capacity = cpus / smt; /* cores */
1169 
1170         ns->task_capacity = min_t(unsigned, capacity,
1171                 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1172         ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1173 }
1174 
1175 struct task_numa_env {
1176         struct task_struct *p;
1177 
1178         int src_cpu, src_nid;
1179         int dst_cpu, dst_nid;
1180 
1181         struct numa_stats src_stats, dst_stats;
1182 
1183         int imbalance_pct;
1184         int dist;
1185 
1186         struct task_struct *best_task;
1187         long best_imp;
1188         int best_cpu;
1189 };
1190 
1191 static void task_numa_assign(struct task_numa_env *env,
1192                              struct task_struct *p, long imp)
1193 {
1194         if (env->best_task)
1195                 put_task_struct(env->best_task);
1196         if (p)
1197                 get_task_struct(p);
1198 
1199         env->best_task = p;
1200         env->best_imp = imp;
1201         env->best_cpu = env->dst_cpu;
1202 }
1203 
1204 static bool load_too_imbalanced(long src_load, long dst_load,
1205                                 struct task_numa_env *env)
1206 {
1207         long imb, old_imb;
1208         long orig_src_load, orig_dst_load;
1209         long src_capacity, dst_capacity;
1210 
1211         /*
1212          * The load is corrected for the CPU capacity available on each node.
1213          *
1214          * src_load        dst_load
1215          * ------------ vs ---------
1216          * src_capacity    dst_capacity
1217          */
1218         src_capacity = env->src_stats.compute_capacity;
1219         dst_capacity = env->dst_stats.compute_capacity;
1220 
1221         /* We care about the slope of the imbalance, not the direction. */
1222         if (dst_load < src_load)
1223                 swap(dst_load, src_load);
1224 
1225         /* Is the difference below the threshold? */
1226         imb = dst_load * src_capacity * 100 -
1227               src_load * dst_capacity * env->imbalance_pct;
1228         if (imb <= 0)
1229                 return false;
1230 
1231         /*
1232          * The imbalance is above the allowed threshold.
1233          * Compare it with the old imbalance.
1234          */
1235         orig_src_load = env->src_stats.load;
1236         orig_dst_load = env->dst_stats.load;
1237 
1238         if (orig_dst_load < orig_src_load)
1239                 swap(orig_dst_load, orig_src_load);
1240 
1241         old_imb = orig_dst_load * src_capacity * 100 -
1242                   orig_src_load * dst_capacity * env->imbalance_pct;
1243 
1244         /* Would this change make things worse? */
1245         return (imb > old_imb);
1246 }
1247 
1248 /*
1249  * This checks if the overall compute and NUMA accesses of the system would
1250  * be improved if the source tasks was migrated to the target dst_cpu taking
1251  * into account that it might be best if task running on the dst_cpu should
1252  * be exchanged with the source task
1253  */
1254 static void task_numa_compare(struct task_numa_env *env,
1255                               long taskimp, long groupimp)
1256 {
1257         struct rq *src_rq = cpu_rq(env->src_cpu);
1258         struct rq *dst_rq = cpu_rq(env->dst_cpu);
1259         struct task_struct *cur;
1260         long src_load, dst_load;
1261         long load;
1262         long imp = env->p->numa_group ? groupimp : taskimp;
1263         long moveimp = imp;
1264         int dist = env->dist;
1265 
1266         rcu_read_lock();
1267 
1268         raw_spin_lock_irq(&dst_rq->lock);
1269         cur = dst_rq->curr;
1270         /*
1271          * No need to move the exiting task, and this ensures that ->curr
1272          * wasn't reaped and thus get_task_struct() in task_numa_assign()
1273          * is safe under RCU read lock.
1274          * Note that rcu_read_lock() itself can't protect from the final
1275          * put_task_struct() after the last schedule().
1276          */
1277         if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1278                 cur = NULL;
1279         raw_spin_unlock_irq(&dst_rq->lock);
1280 
1281         /*
1282          * Because we have preemption enabled we can get migrated around and
1283          * end try selecting ourselves (current == env->p) as a swap candidate.
1284          */
1285         if (cur == env->p)
1286                 goto unlock;
1287 
1288         /*
1289          * "imp" is the fault differential for the source task between the
1290          * source and destination node. Calculate the total differential for
1291          * the source task and potential destination task. The more negative
1292          * the value is, the more rmeote accesses that would be expected to
1293          * be incurred if the tasks were swapped.
1294          */
1295         if (cur) {
1296                 /* Skip this swap candidate if cannot move to the source cpu */
1297                 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1298                         goto unlock;
1299 
1300                 /*
1301                  * If dst and source tasks are in the same NUMA group, or not
1302                  * in any group then look only at task weights.
1303                  */
1304                 if (cur->numa_group == env->p->numa_group) {
1305                         imp = taskimp + task_weight(cur, env->src_nid, dist) -
1306                               task_weight(cur, env->dst_nid, dist);
1307                         /*
1308                          * Add some hysteresis to prevent swapping the
1309                          * tasks within a group over tiny differences.
1310                          */
1311                         if (cur->numa_group)
1312                                 imp -= imp/16;
1313                 } else {
1314                         /*
1315                          * Compare the group weights. If a task is all by
1316                          * itself (not part of a group), use the task weight
1317                          * instead.
1318                          */
1319                         if (cur->numa_group)
1320                                 imp += group_weight(cur, env->src_nid, dist) -
1321                                        group_weight(cur, env->dst_nid, dist);
1322                         else
1323                                 imp += task_weight(cur, env->src_nid, dist) -
1324                                        task_weight(cur, env->dst_nid, dist);
1325                 }
1326         }
1327 
1328         if (imp <= env->best_imp && moveimp <= env->best_imp)
1329                 goto unlock;
1330 
1331         if (!cur) {
1332                 /* Is there capacity at our destination? */
1333                 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1334                     !env->dst_stats.has_free_capacity)
1335                         goto unlock;
1336 
1337                 goto balance;
1338         }
1339 
1340         /* Balance doesn't matter much if we're running a task per cpu */
1341         if (imp > env->best_imp && src_rq->nr_running == 1 &&
1342                         dst_rq->nr_running == 1)
1343                 goto assign;
1344 
1345         /*
1346          * In the overloaded case, try and keep the load balanced.
1347          */
1348 balance:
1349         load = task_h_load(env->p);
1350         dst_load = env->dst_stats.load + load;
1351         src_load = env->src_stats.load - load;
1352 
1353         if (moveimp > imp && moveimp > env->best_imp) {
1354                 /*
1355                  * If the improvement from just moving env->p direction is
1356                  * better than swapping tasks around, check if a move is
1357                  * possible. Store a slightly smaller score than moveimp,
1358                  * so an actually idle CPU will win.
1359                  */
1360                 if (!load_too_imbalanced(src_load, dst_load, env)) {
1361                         imp = moveimp - 1;
1362                         cur = NULL;
1363                         goto assign;
1364                 }
1365         }
1366 
1367         if (imp <= env->best_imp)
1368                 goto unlock;
1369 
1370         if (cur) {
1371                 load = task_h_load(cur);
1372                 dst_load -= load;
1373                 src_load += load;
1374         }
1375 
1376         if (load_too_imbalanced(src_load, dst_load, env))
1377                 goto unlock;
1378 
1379         /*
1380          * One idle CPU per node is evaluated for a task numa move.
1381          * Call select_idle_sibling to maybe find a better one.
1382          */
1383         if (!cur)
1384                 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1385 
1386 assign:
1387         task_numa_assign(env, cur, imp);
1388 unlock:
1389         rcu_read_unlock();
1390 }
1391 
1392 static void task_numa_find_cpu(struct task_numa_env *env,
1393                                 long taskimp, long groupimp)
1394 {
1395         int cpu;
1396 
1397         for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1398                 /* Skip this CPU if the source task cannot migrate */
1399                 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1400                         continue;
1401 
1402                 env->dst_cpu = cpu;
1403                 task_numa_compare(env, taskimp, groupimp);
1404         }
1405 }
1406 
1407 /* Only move tasks to a NUMA node less busy than the current node. */
1408 static bool numa_has_capacity(struct task_numa_env *env)
1409 {
1410         struct numa_stats *src = &env->src_stats;
1411         struct numa_stats *dst = &env->dst_stats;
1412 
1413         if (src->has_free_capacity && !dst->has_free_capacity)
1414                 return false;
1415 
1416         /*
1417          * Only consider a task move if the source has a higher load
1418          * than the destination, corrected for CPU capacity on each node.
1419          *
1420          *      src->load                dst->load
1421          * --------------------- vs ---------------------
1422          * src->compute_capacity    dst->compute_capacity
1423          */
1424         if (src->load * dst->compute_capacity * env->imbalance_pct >
1425 
1426             dst->load * src->compute_capacity * 100)
1427                 return true;
1428 
1429         return false;
1430 }
1431 
1432 static int task_numa_migrate(struct task_struct *p)
1433 {
1434         struct task_numa_env env = {
1435                 .p = p,
1436 
1437                 .src_cpu = task_cpu(p),
1438                 .src_nid = task_node(p),
1439 
1440                 .imbalance_pct = 112,
1441 
1442                 .best_task = NULL,
1443                 .best_imp = 0,
1444                 .best_cpu = -1
1445         };
1446         struct sched_domain *sd;
1447         unsigned long taskweight, groupweight;
1448         int nid, ret, dist;
1449         long taskimp, groupimp;
1450 
1451         /*
1452          * Pick the lowest SD_NUMA domain, as that would have the smallest
1453          * imbalance and would be the first to start moving tasks about.
1454          *
1455          * And we want to avoid any moving of tasks about, as that would create
1456          * random movement of tasks -- counter the numa conditions we're trying
1457          * to satisfy here.
1458          */
1459         rcu_read_lock();
1460         sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1461         if (sd)
1462                 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1463         rcu_read_unlock();
1464 
1465         /*
1466          * Cpusets can break the scheduler domain tree into smaller
1467          * balance domains, some of which do not cross NUMA boundaries.
1468          * Tasks that are "trapped" in such domains cannot be migrated
1469          * elsewhere, so there is no point in (re)trying.
1470          */
1471         if (unlikely(!sd)) {
1472                 p->numa_preferred_nid = task_node(p);
1473                 return -EINVAL;
1474         }
1475 
1476         env.dst_nid = p->numa_preferred_nid;
1477         dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1478         taskweight = task_weight(p, env.src_nid, dist);
1479         groupweight = group_weight(p, env.src_nid, dist);
1480         update_numa_stats(&env.src_stats, env.src_nid);
1481         taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1482         groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1483         update_numa_stats(&env.dst_stats, env.dst_nid);
1484 
1485         /* Try to find a spot on the preferred nid. */
1486         if (numa_has_capacity(&env))
1487                 task_numa_find_cpu(&env, taskimp, groupimp);
1488 
1489         /*
1490          * Look at other nodes in these cases:
1491          * - there is no space available on the preferred_nid
1492          * - the task is part of a numa_group that is interleaved across
1493          *   multiple NUMA nodes; in order to better consolidate the group,
1494          *   we need to check other locations.
1495          */
1496         if (env.best_cpu == -1 || (p->numa_group &&
1497                         nodes_weight(p->numa_group->active_nodes) > 1)) {
1498                 for_each_online_node(nid) {
1499                         if (nid == env.src_nid || nid == p->numa_preferred_nid)
1500                                 continue;
1501 
1502                         dist = node_distance(env.src_nid, env.dst_nid);
1503                         if (sched_numa_topology_type == NUMA_BACKPLANE &&
1504                                                 dist != env.dist) {
1505                                 taskweight = task_weight(p, env.src_nid, dist);
1506                                 groupweight = group_weight(p, env.src_nid, dist);
1507                         }
1508 
1509                         /* Only consider nodes where both task and groups benefit */
1510                         taskimp = task_weight(p, nid, dist) - taskweight;
1511                         groupimp = group_weight(p, nid, dist) - groupweight;
1512                         if (taskimp < 0 && groupimp < 0)
1513                                 continue;
1514 
1515                         env.dist = dist;
1516                         env.dst_nid = nid;
1517                         update_numa_stats(&env.dst_stats, env.dst_nid);
1518                         if (numa_has_capacity(&env))
1519                                 task_numa_find_cpu(&env, taskimp, groupimp);
1520                 }
1521         }
1522 
1523         /*
1524          * If the task is part of a workload that spans multiple NUMA nodes,
1525          * and is migrating into one of the workload's active nodes, remember
1526          * this node as the task's preferred numa node, so the workload can
1527          * settle down.
1528          * A task that migrated to a second choice node will be better off
1529          * trying for a better one later. Do not set the preferred node here.
1530          */
1531         if (p->numa_group) {
1532                 if (env.best_cpu == -1)
1533                         nid = env.src_nid;
1534                 else
1535                         nid = env.dst_nid;
1536 
1537                 if (node_isset(nid, p->numa_group->active_nodes))
1538                         sched_setnuma(p, env.dst_nid);
1539         }
1540 
1541         /* No better CPU than the current one was found. */
1542         if (env.best_cpu == -1)
1543                 return -EAGAIN;
1544 
1545         /*
1546          * Reset the scan period if the task is being rescheduled on an
1547          * alternative node to recheck if the tasks is now properly placed.
1548          */
1549         p->numa_scan_period = task_scan_min(p);
1550 
1551         if (env.best_task == NULL) {
1552                 ret = migrate_task_to(p, env.best_cpu);
1553                 if (ret != 0)
1554                         trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1555                 return ret;
1556         }
1557 
1558         ret = migrate_swap(p, env.best_task);
1559         if (ret != 0)
1560                 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1561         put_task_struct(env.best_task);
1562         return ret;
1563 }
1564 
1565 /* Attempt to migrate a task to a CPU on the preferred node. */
1566 static void numa_migrate_preferred(struct task_struct *p)
1567 {
1568         unsigned long interval = HZ;
1569 
1570         /* This task has no NUMA fault statistics yet */
1571         if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1572                 return;
1573 
1574         /* Periodically retry migrating the task to the preferred node */
1575         interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1576         p->numa_migrate_retry = jiffies + interval;
1577 
1578         /* Success if task is already running on preferred CPU */
1579         if (task_node(p) == p->numa_preferred_nid)
1580                 return;
1581 
1582         /* Otherwise, try migrate to a CPU on the preferred node */
1583         task_numa_migrate(p);
1584 }
1585 
1586 /*
1587  * Find the nodes on which the workload is actively running. We do this by
1588  * tracking the nodes from which NUMA hinting faults are triggered. This can
1589  * be different from the set of nodes where the workload's memory is currently
1590  * located.
1591  *
1592  * The bitmask is used to make smarter decisions on when to do NUMA page
1593  * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1594  * are added when they cause over 6/16 of the maximum number of faults, but
1595  * only removed when they drop below 3/16.
1596  */
1597 static void update_numa_active_node_mask(struct numa_group *numa_group)
1598 {
1599         unsigned long faults, max_faults = 0;
1600         int nid;
1601 
1602         for_each_online_node(nid) {
1603                 faults = group_faults_cpu(numa_group, nid);
1604                 if (faults > max_faults)
1605                         max_faults = faults;
1606         }
1607 
1608         for_each_online_node(nid) {
1609                 faults = group_faults_cpu(numa_group, nid);
1610                 if (!node_isset(nid, numa_group->active_nodes)) {
1611                         if (faults > max_faults * 6 / 16)
1612                                 node_set(nid, numa_group->active_nodes);
1613                 } else if (faults < max_faults * 3 / 16)
1614                         node_clear(nid, numa_group->active_nodes);
1615         }
1616 }
1617 
1618 /*
1619  * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1620  * increments. The more local the fault statistics are, the higher the scan
1621  * period will be for the next scan window. If local/(local+remote) ratio is
1622  * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1623  * the scan period will decrease. Aim for 70% local accesses.
1624  */
1625 #define NUMA_PERIOD_SLOTS 10
1626 #define NUMA_PERIOD_THRESHOLD 7
1627 
1628 /*
1629  * Increase the scan period (slow down scanning) if the majority of
1630  * our memory is already on our local node, or if the majority of
1631  * the page accesses are shared with other processes.
1632  * Otherwise, decrease the scan period.
1633  */
1634 static void update_task_scan_period(struct task_struct *p,
1635                         unsigned long shared, unsigned long private)
1636 {
1637         unsigned int period_slot;
1638         int ratio;
1639         int diff;
1640 
1641         unsigned long remote = p->numa_faults_locality[0];
1642         unsigned long local = p->numa_faults_locality[1];
1643 
1644         /*
1645          * If there were no record hinting faults then either the task is
1646          * completely idle or all activity is areas that are not of interest
1647          * to automatic numa balancing. Related to that, if there were failed
1648          * migration then it implies we are migrating too quickly or the local
1649          * node is overloaded. In either case, scan slower
1650          */
1651         if (local + shared == 0 || p->numa_faults_locality[2]) {
1652                 p->numa_scan_period = min(p->numa_scan_period_max,
1653                         p->numa_scan_period << 1);
1654 
1655                 p->mm->numa_next_scan = jiffies +
1656                         msecs_to_jiffies(p->numa_scan_period);
1657 
1658                 return;
1659         }
1660 
1661         /*
1662          * Prepare to scale scan period relative to the current period.
1663          *       == NUMA_PERIOD_THRESHOLD scan period stays the same
1664          *       <  NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1665          *       >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1666          */
1667         period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1668         ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1669         if (ratio >= NUMA_PERIOD_THRESHOLD) {
1670                 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1671                 if (!slot)
1672                         slot = 1;
1673                 diff = slot * period_slot;
1674         } else {
1675                 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1676 
1677                 /*
1678                  * Scale scan rate increases based on sharing. There is an
1679                  * inverse relationship between the degree of sharing and
1680                  * the adjustment made to the scanning period. Broadly
1681                  * speaking the intent is that there is little point
1682                  * scanning faster if shared accesses dominate as it may
1683                  * simply bounce migrations uselessly
1684                  */
1685                 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1686                 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1687         }
1688 
1689         p->numa_scan_period = clamp(p->numa_scan_period + diff,
1690                         task_scan_min(p), task_scan_max(p));
1691         memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1692 }
1693 
1694 /*
1695  * Get the fraction of time the task has been running since the last
1696  * NUMA placement cycle. The scheduler keeps similar statistics, but
1697  * decays those on a 32ms period, which is orders of magnitude off
1698  * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1699  * stats only if the task is so new there are no NUMA statistics yet.
1700  */
1701 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1702 {
1703         u64 runtime, delta, now;
1704         /* Use the start of this time slice to avoid calculations. */
1705         now = p->se.exec_start;
1706         runtime = p->se.sum_exec_runtime;
1707 
1708         if (p->last_task_numa_placement) {
1709                 delta = runtime - p->last_sum_exec_runtime;
1710                 *period = now - p->last_task_numa_placement;
1711         } else {
1712                 delta = p->se.avg.load_sum / p->se.load.weight;
1713                 *period = LOAD_AVG_MAX;
1714         }
1715 
1716         p->last_sum_exec_runtime = runtime;
1717         p->last_task_numa_placement = now;
1718 
1719         return delta;
1720 }
1721 
1722 /*
1723  * Determine the preferred nid for a task in a numa_group. This needs to
1724  * be done in a way that produces consistent results with group_weight,
1725  * otherwise workloads might not converge.
1726  */
1727 static int preferred_group_nid(struct task_struct *p, int nid)
1728 {
1729         nodemask_t nodes;
1730         int dist;
1731 
1732         /* Direct connections between all NUMA nodes. */
1733         if (sched_numa_topology_type == NUMA_DIRECT)
1734                 return nid;
1735 
1736         /*
1737          * On a system with glueless mesh NUMA topology, group_weight
1738          * scores nodes according to the number of NUMA hinting faults on
1739          * both the node itself, and on nearby nodes.
1740          */
1741         if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1742                 unsigned long score, max_score = 0;
1743                 int node, max_node = nid;
1744 
1745                 dist = sched_max_numa_distance;
1746 
1747                 for_each_online_node(node) {
1748                         score = group_weight(p, node, dist);
1749                         if (score > max_score) {
1750                                 max_score = score;
1751                                 max_node = node;
1752                         }
1753                 }
1754                 return max_node;
1755         }
1756 
1757         /*
1758          * Finding the preferred nid in a system with NUMA backplane
1759          * interconnect topology is more involved. The goal is to locate
1760          * tasks from numa_groups near each other in the system, and
1761          * untangle workloads from different sides of the system. This requires
1762          * searching down the hierarchy of node groups, recursively searching
1763          * inside the highest scoring group of nodes. The nodemask tricks
1764          * keep the complexity of the search down.
1765          */
1766         nodes = node_online_map;
1767         for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1768                 unsigned long max_faults = 0;
1769                 nodemask_t max_group = NODE_MASK_NONE;
1770                 int a, b;
1771 
1772                 /* Are there nodes at this distance from each other? */
1773                 if (!find_numa_distance(dist))
1774                         continue;
1775 
1776                 for_each_node_mask(a, nodes) {
1777                         unsigned long faults = 0;
1778                         nodemask_t this_group;
1779                         nodes_clear(this_group);
1780 
1781                         /* Sum group's NUMA faults; includes a==b case. */
1782                         for_each_node_mask(b, nodes) {
1783                                 if (node_distance(a, b) < dist) {
1784                                         faults += group_faults(p, b);
1785                                         node_set(b, this_group);
1786                                         node_clear(b, nodes);
1787                                 }
1788                         }
1789 
1790                         /* Remember the top group. */
1791                         if (faults > max_faults) {
1792                                 max_faults = faults;
1793                                 max_group = this_group;
1794                                 /*
1795                                  * subtle: at the smallest distance there is
1796                                  * just one node left in each "group", the
1797                                  * winner is the preferred nid.
1798                                  */
1799                                 nid = a;
1800                         }
1801                 }
1802                 /* Next round, evaluate the nodes within max_group. */
1803                 if (!max_faults)
1804                         break;
1805                 nodes = max_group;
1806         }
1807         return nid;
1808 }
1809 
1810 static void task_numa_placement(struct task_struct *p)
1811 {
1812         int seq, nid, max_nid = -1, max_group_nid = -1;
1813         unsigned long max_faults = 0, max_group_faults = 0;
1814         unsigned long fault_types[2] = { 0, 0 };
1815         unsigned long total_faults;
1816         u64 runtime, period;
1817         spinlock_t *group_lock = NULL;
1818 
1819         /*
1820          * The p->mm->numa_scan_seq field gets updated without
1821          * exclusive access. Use READ_ONCE() here to ensure
1822          * that the field is read in a single access:
1823          */
1824         seq = READ_ONCE(p->mm->numa_scan_seq);
1825         if (p->numa_scan_seq == seq)
1826                 return;
1827         p->numa_scan_seq = seq;
1828         p->numa_scan_period_max = task_scan_max(p);
1829 
1830         total_faults = p->numa_faults_locality[0] +
1831                        p->numa_faults_locality[1];
1832         runtime = numa_get_avg_runtime(p, &period);
1833 
1834         /* If the task is part of a group prevent parallel updates to group stats */
1835         if (p->numa_group) {
1836                 group_lock = &p->numa_group->lock;
1837                 spin_lock_irq(group_lock);
1838         }
1839 
1840         /* Find the node with the highest number of faults */
1841         for_each_online_node(nid) {
1842                 /* Keep track of the offsets in numa_faults array */
1843                 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1844                 unsigned long faults = 0, group_faults = 0;
1845                 int priv;
1846 
1847                 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1848                         long diff, f_diff, f_weight;
1849 
1850                         mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1851                         membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1852                         cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1853                         cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1854 
1855                         /* Decay existing window, copy faults since last scan */
1856                         diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1857                         fault_types[priv] += p->numa_faults[membuf_idx];
1858                         p->numa_faults[membuf_idx] = 0;
1859 
1860                         /*
1861                          * Normalize the faults_from, so all tasks in a group
1862                          * count according to CPU use, instead of by the raw
1863                          * number of faults. Tasks with little runtime have
1864                          * little over-all impact on throughput, and thus their
1865                          * faults are less important.
1866                          */
1867                         f_weight = div64_u64(runtime << 16, period + 1);
1868                         f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1869                                    (total_faults + 1);
1870                         f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1871                         p->numa_faults[cpubuf_idx] = 0;
1872 
1873                         p->numa_faults[mem_idx] += diff;
1874                         p->numa_faults[cpu_idx] += f_diff;
1875                         faults += p->numa_faults[mem_idx];
1876                         p->total_numa_faults += diff;
1877                         if (p->numa_group) {
1878                                 /*
1879                                  * safe because we can only change our own group
1880                                  *
1881                                  * mem_idx represents the offset for a given
1882                                  * nid and priv in a specific region because it
1883                                  * is at the beginning of the numa_faults array.
1884                                  */
1885                                 p->numa_group->faults[mem_idx] += diff;
1886                                 p->numa_group->faults_cpu[mem_idx] += f_diff;
1887                                 p->numa_group->total_faults += diff;
1888                                 group_faults += p->numa_group->faults[mem_idx];
1889                         }
1890                 }
1891 
1892                 if (faults > max_faults) {
1893                         max_faults = faults;
1894                         max_nid = nid;
1895                 }
1896 
1897                 if (group_faults > max_group_faults) {
1898                         max_group_faults = group_faults;
1899                         max_group_nid = nid;
1900                 }
1901         }
1902 
1903         update_task_scan_period(p, fault_types[0], fault_types[1]);
1904 
1905         if (p->numa_group) {
1906                 update_numa_active_node_mask(p->numa_group);
1907                 spin_unlock_irq(group_lock);
1908                 max_nid = preferred_group_nid(p, max_group_nid);
1909         }
1910 
1911         if (max_faults) {
1912                 /* Set the new preferred node */
1913                 if (max_nid != p->numa_preferred_nid)
1914                         sched_setnuma(p, max_nid);
1915 
1916                 if (task_node(p) != p->numa_preferred_nid)
1917                         numa_migrate_preferred(p);
1918         }
1919 }
1920 
1921 static inline int get_numa_group(struct numa_group *grp)
1922 {
1923         return atomic_inc_not_zero(&grp->refcount);
1924 }
1925 
1926 static inline void put_numa_group(struct numa_group *grp)
1927 {
1928         if (atomic_dec_and_test(&grp->refcount))
1929                 kfree_rcu(grp, rcu);
1930 }
1931 
1932 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1933                         int *priv)
1934 {
1935         struct numa_group *grp, *my_grp;
1936         struct task_struct *tsk;
1937         bool join = false;
1938         int cpu = cpupid_to_cpu(cpupid);
1939         int i;
1940 
1941         if (unlikely(!p->numa_group)) {
1942                 unsigned int size = sizeof(struct numa_group) +
1943                                     4*nr_node_ids*sizeof(unsigned long);
1944 
1945                 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1946                 if (!grp)
1947                         return;
1948 
1949                 atomic_set(&grp->refcount, 1);
1950                 spin_lock_init(&grp->lock);
1951                 grp->gid = p->pid;
1952                 /* Second half of the array tracks nids where faults happen */
1953                 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1954                                                 nr_node_ids;
1955 
1956                 node_set(task_node(current), grp->active_nodes);
1957 
1958                 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1959                         grp->faults[i] = p->numa_faults[i];
1960 
1961                 grp->total_faults = p->total_numa_faults;
1962 
1963                 grp->nr_tasks++;
1964                 rcu_assign_pointer(p->numa_group, grp);
1965         }
1966 
1967         rcu_read_lock();
1968         tsk = READ_ONCE(cpu_rq(cpu)->curr);
1969 
1970         if (!cpupid_match_pid(tsk, cpupid))
1971                 goto no_join;
1972 
1973         grp = rcu_dereference(tsk->numa_group);
1974         if (!grp)
1975                 goto no_join;
1976 
1977         my_grp = p->numa_group;
1978         if (grp == my_grp)
1979                 goto no_join;
1980 
1981         /*
1982          * Only join the other group if its bigger; if we're the bigger group,
1983          * the other task will join us.
1984          */
1985         if (my_grp->nr_tasks > grp->nr_tasks)
1986                 goto no_join;
1987 
1988         /*
1989          * Tie-break on the grp address.
1990          */
1991         if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1992                 goto no_join;
1993 
1994         /* Always join threads in the same process. */
1995         if (tsk->mm == current->mm)
1996                 join = true;
1997 
1998         /* Simple filter to avoid false positives due to PID collisions */
1999         if (flags & TNF_SHARED)
2000                 join = true;
2001 
2002         /* Update priv based on whether false sharing was detected */
2003         *priv = !join;
2004 
2005         if (join && !get_numa_group(grp))
2006                 goto no_join;
2007 
2008         rcu_read_unlock();
2009 
2010         if (!join)
2011                 return;
2012 
2013         BUG_ON(irqs_disabled());
2014         double_lock_irq(&my_grp->lock, &grp->lock);
2015 
2016         for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2017                 my_grp->faults[i] -= p->numa_faults[i];
2018                 grp->faults[i] += p->numa_faults[i];
2019         }
2020         my_grp->total_faults -= p->total_numa_faults;
2021         grp->total_faults += p->total_numa_faults;
2022 
2023         my_grp->nr_tasks--;
2024         grp->nr_tasks++;
2025 
2026         spin_unlock(&my_grp->lock);
2027         spin_unlock_irq(&grp->lock);
2028 
2029         rcu_assign_pointer(p->numa_group, grp);
2030 
2031         put_numa_group(my_grp);
2032         return;
2033 
2034 no_join:
2035         rcu_read_unlock();
2036         return;
2037 }
2038 
2039 void task_numa_free(struct task_struct *p)
2040 {
2041         struct numa_group *grp = p->numa_group;
2042         void *numa_faults = p->numa_faults;
2043         unsigned long flags;
2044         int i;
2045 
2046         if (grp) {
2047                 spin_lock_irqsave(&grp->lock, flags);
2048                 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2049                         grp->faults[i] -= p->numa_faults[i];
2050                 grp->total_faults -= p->total_numa_faults;
2051 
2052                 grp->nr_tasks--;
2053                 spin_unlock_irqrestore(&grp->lock, flags);
2054                 RCU_INIT_POINTER(p->numa_group, NULL);
2055                 put_numa_group(grp);
2056         }
2057 
2058         p->numa_faults = NULL;
2059         kfree(numa_faults);
2060 }
2061 
2062 /*
2063  * Got a PROT_NONE fault for a page on @node.
2064  */
2065 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2066 {
2067         struct task_struct *p = current;
2068         bool migrated = flags & TNF_MIGRATED;
2069         int cpu_node = task_node(current);
2070         int local = !!(flags & TNF_FAULT_LOCAL);
2071         int priv;
2072 
2073         if (!static_branch_likely(&sched_numa_balancing))
2074                 return;
2075 
2076         /* for example, ksmd faulting in a user's mm */
2077         if (!p->mm)
2078                 return;
2079 
2080         /* Allocate buffer to track faults on a per-node basis */
2081         if (unlikely(!p->numa_faults)) {
2082                 int size = sizeof(*p->numa_faults) *
2083                            NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2084 
2085                 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2086                 if (!p->numa_faults)
2087                         return;
2088 
2089                 p->total_numa_faults = 0;
2090                 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2091         }
2092 
2093         /*
2094          * First accesses are treated as private, otherwise consider accesses
2095          * to be private if the accessing pid has not changed
2096          */
2097         if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2098                 priv = 1;
2099         } else {
2100                 priv = cpupid_match_pid(p, last_cpupid);
2101                 if (!priv && !(flags & TNF_NO_GROUP))
2102                         task_numa_group(p, last_cpupid, flags, &priv);
2103         }
2104 
2105         /*
2106          * If a workload spans multiple NUMA nodes, a shared fault that
2107          * occurs wholly within the set of nodes that the workload is
2108          * actively using should be counted as local. This allows the
2109          * scan rate to slow down when a workload has settled down.
2110          */
2111         if (!priv && !local && p->numa_group &&
2112                         node_isset(cpu_node, p->numa_group->active_nodes) &&
2113                         node_isset(mem_node, p->numa_group->active_nodes))
2114                 local = 1;
2115 
2116         task_numa_placement(p);
2117 
2118         /*
2119          * Retry task to preferred node migration periodically, in case it
2120          * case it previously failed, or the scheduler moved us.
2121          */
2122         if (time_after(jiffies, p->numa_migrate_retry))
2123                 numa_migrate_preferred(p);
2124 
2125         if (migrated)
2126                 p->numa_pages_migrated += pages;
2127         if (flags & TNF_MIGRATE_FAIL)
2128                 p->numa_faults_locality[2] += pages;
2129 
2130         p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2131         p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2132         p->numa_faults_locality[local] += pages;
2133 }
2134 
2135 static void reset_ptenuma_scan(struct task_struct *p)
2136 {
2137         /*
2138          * We only did a read acquisition of the mmap sem, so
2139          * p->mm->numa_scan_seq is written to without exclusive access
2140          * and the update is not guaranteed to be atomic. That's not
2141          * much of an issue though, since this is just used for
2142          * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2143          * expensive, to avoid any form of compiler optimizations:
2144          */
2145         WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2146         p->mm->numa_scan_offset = 0;
2147 }
2148 
2149 /*
2150  * The expensive part of numa migration is done from task_work context.
2151  * Triggered from task_tick_numa().
2152  */
2153 void task_numa_work(struct callback_head *work)
2154 {
2155         unsigned long migrate, next_scan, now = jiffies;
2156         struct task_struct *p = current;
2157         struct mm_struct *mm = p->mm;
2158         struct vm_area_struct *vma;
2159         unsigned long start, end;
2160         unsigned long nr_pte_updates = 0;
2161         long pages, virtpages;
2162 
2163         WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2164 
2165         work->next = work; /* protect against double add */
2166         /*
2167          * Who cares about NUMA placement when they're dying.
2168          *
2169          * NOTE: make sure not to dereference p->mm before this check,
2170          * exit_task_work() happens _after_ exit_mm() so we could be called
2171          * without p->mm even though we still had it when we enqueued this
2172          * work.
2173          */
2174         if (p->flags & PF_EXITING)
2175                 return;
2176 
2177         if (!mm->numa_next_scan) {
2178                 mm->numa_next_scan = now +
2179                         msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2180         }
2181 
2182         /*
2183          * Enforce maximal scan/migration frequency..
2184          */
2185         migrate = mm->numa_next_scan;
2186         if (time_before(now, migrate))
2187                 return;
2188 
2189         if (p->numa_scan_period == 0) {
2190                 p->numa_scan_period_max = task_scan_max(p);
2191                 p->numa_scan_period = task_scan_min(p);
2192         }
2193 
2194         next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2195         if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2196                 return;
2197 
2198         /*
2199          * Delay this task enough that another task of this mm will likely win
2200          * the next time around.
2201          */
2202         p->node_stamp += 2 * TICK_NSEC;
2203 
2204         start = mm->numa_scan_offset;
2205         pages = sysctl_numa_balancing_scan_size;
2206         pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2207         virtpages = pages * 8;     /* Scan up to this much virtual space */
2208         if (!pages)
2209                 return;
2210 
2211 
2212         down_read(&mm->mmap_sem);
2213         vma = find_vma(mm, start);
2214         if (!vma) {
2215                 reset_ptenuma_scan(p);
2216                 start = 0;
2217                 vma = mm->mmap;
2218         }
2219         for (; vma; vma = vma->vm_next) {
2220                 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2221                         is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2222                         continue;
2223                 }
2224 
2225                 /*
2226                  * Shared library pages mapped by multiple processes are not
2227                  * migrated as it is expected they are cache replicated. Avoid
2228                  * hinting faults in read-only file-backed mappings or the vdso
2229                  * as migrating the pages will be of marginal benefit.
2230                  */
2231                 if (!vma->vm_mm ||
2232                     (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2233                         continue;
2234 
2235                 /*
2236                  * Skip inaccessible VMAs to avoid any confusion between
2237                  * PROT_NONE and NUMA hinting ptes
2238                  */
2239                 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2240                         continue;
2241 
2242                 do {
2243                         start = max(start, vma->vm_start);
2244                         end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2245                         end = min(end, vma->vm_end);
2246                         nr_pte_updates = change_prot_numa(vma, start, end);
2247 
2248                         /*
2249                          * Try to scan sysctl_numa_balancing_size worth of
2250                          * hpages that have at least one present PTE that
2251                          * is not already pte-numa. If the VMA contains
2252                          * areas that are unused or already full of prot_numa
2253                          * PTEs, scan up to virtpages, to skip through those
2254                          * areas faster.
2255                          */
2256                         if (nr_pte_updates)
2257                                 pages -= (end - start) >> PAGE_SHIFT;
2258                         virtpages -= (end - start) >> PAGE_SHIFT;
2259 
2260                         start = end;
2261                         if (pages <= 0 || virtpages <= 0)
2262                                 goto out;
2263 
2264                         cond_resched();
2265                 } while (end != vma->vm_end);
2266         }
2267 
2268 out:
2269         /*
2270          * It is possible to reach the end of the VMA list but the last few
2271          * VMAs are not guaranteed to the vma_migratable. If they are not, we
2272          * would find the !migratable VMA on the next scan but not reset the
2273          * scanner to the start so check it now.
2274          */
2275         if (vma)
2276                 mm->numa_scan_offset = start;
2277         else
2278                 reset_ptenuma_scan(p);
2279         up_read(&mm->mmap_sem);
2280 }
2281 
2282 /*
2283  * Drive the periodic memory faults..
2284  */
2285 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2286 {
2287         struct callback_head *work = &curr->numa_work;
2288         u64 period, now;
2289 
2290         /*
2291          * We don't care about NUMA placement if we don't have memory.
2292          */
2293         if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2294                 return;
2295 
2296         /*
2297          * Using runtime rather than walltime has the dual advantage that
2298          * we (mostly) drive the selection from busy threads and that the
2299          * task needs to have done some actual work before we bother with
2300          * NUMA placement.
2301          */
2302         now = curr->se.sum_exec_runtime;
2303         period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2304 
2305         if (now > curr->node_stamp + period) {
2306                 if (!curr->node_stamp)
2307                         curr->numa_scan_period = task_scan_min(curr);
2308                 curr->node_stamp += period;
2309 
2310                 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2311                         init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2312                         task_work_add(curr, work, true);
2313                 }
2314         }
2315 }
2316 #else
2317 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2318 {
2319 }
2320 
2321 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2322 {
2323 }
2324 
2325 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2326 {
2327 }
2328 #endif /* CONFIG_NUMA_BALANCING */
2329 
2330 static void
2331 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2332 {
2333         update_load_add(&cfs_rq->load, se->load.weight);
2334         if (!parent_entity(se))
2335                 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2336 #ifdef CONFIG_SMP
2337         if (entity_is_task(se)) {
2338                 struct rq *rq = rq_of(cfs_rq);
2339 
2340                 account_numa_enqueue(rq, task_of(se));
2341                 list_add(&se->group_node, &rq->cfs_tasks);
2342         }
2343 #endif
2344         cfs_rq->nr_running++;
2345 }
2346 
2347 static void
2348 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2349 {
2350         update_load_sub(&cfs_rq->load, se->load.weight);
2351         if (!parent_entity(se))
2352                 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2353         if (entity_is_task(se)) {
2354                 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2355                 list_del_init(&se->group_node);
2356         }
2357         cfs_rq->nr_running--;
2358 }
2359 
2360 #ifdef CONFIG_FAIR_GROUP_SCHED
2361 # ifdef CONFIG_SMP
2362 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2363 {
2364         long tg_weight;
2365 
2366         /*
2367          * Use this CPU's real-time load instead of the last load contribution
2368          * as the updating of the contribution is delayed, and we will use the
2369          * the real-time load to calc the share. See update_tg_load_avg().
2370          */
2371         tg_weight = atomic_long_read(&tg->load_avg);
2372         tg_weight -= cfs_rq->tg_load_avg_contrib;
2373         tg_weight += cfs_rq->load.weight;
2374 
2375         return tg_weight;
2376 }
2377 
2378 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2379 {
2380         long tg_weight, load, shares;
2381 
2382         tg_weight = calc_tg_weight(tg, cfs_rq);
2383         load = cfs_rq->load.weight;
2384 
2385         shares = (tg->shares * load);
2386         if (tg_weight)
2387                 shares /= tg_weight;
2388 
2389         if (shares < MIN_SHARES)
2390                 shares = MIN_SHARES;
2391         if (shares > tg->shares)
2392                 shares = tg->shares;
2393 
2394         return shares;
2395 }
2396 # else /* CONFIG_SMP */
2397 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2398 {
2399         return tg->shares;
2400 }
2401 # endif /* CONFIG_SMP */
2402 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2403                             unsigned long weight)
2404 {
2405         if (se->on_rq) {
2406                 /* commit outstanding execution time */
2407                 if (cfs_rq->curr == se)
2408                         update_curr(cfs_rq);
2409                 account_entity_dequeue(cfs_rq, se);
2410         }
2411 
2412         update_load_set(&se->load, weight);
2413 
2414         if (se->on_rq)
2415                 account_entity_enqueue(cfs_rq, se);
2416 }
2417 
2418 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2419 
2420 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2421 {
2422         struct task_group *tg;
2423         struct sched_entity *se;
2424         long shares;
2425 
2426         tg = cfs_rq->tg;
2427         se = tg->se[cpu_of(rq_of(cfs_rq))];
2428         if (!se || throttled_hierarchy(cfs_rq))
2429                 return;
2430 #ifndef CONFIG_SMP
2431         if (likely(se->load.weight == tg->shares))
2432                 return;
2433 #endif
2434         shares = calc_cfs_shares(cfs_rq, tg);
2435 
2436         reweight_entity(cfs_rq_of(se), se, shares);
2437 }
2438 #else /* CONFIG_FAIR_GROUP_SCHED */
2439 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2440 {
2441 }
2442 #endif /* CONFIG_FAIR_GROUP_SCHED */
2443 
2444 #ifdef CONFIG_SMP
2445 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2446 static const u32 runnable_avg_yN_inv[] = {
2447         0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2448         0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2449         0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2450         0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2451         0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2452         0x85aac367, 0x82cd8698,
2453 };
2454 
2455 /*
2456  * Precomputed \Sum y^k { 1<=k<=n }.  These are floor(true_value) to prevent
2457  * over-estimates when re-combining.
2458  */
2459 static const u32 runnable_avg_yN_sum[] = {
2460             0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2461          9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2462         17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2463 };
2464 
2465 /*
2466  * Approximate:
2467  *   val * y^n,    where y^32 ~= 0.5 (~1 scheduling period)
2468  */
2469 static __always_inline u64 decay_load(u64 val, u64 n)
2470 {
2471         unsigned int local_n;
2472 
2473         if (!n)
2474                 return val;
2475         else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2476                 return 0;
2477 
2478         /* after bounds checking we can collapse to 32-bit */
2479         local_n = n;
2480 
2481         /*
2482          * As y^PERIOD = 1/2, we can combine
2483          *    y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2484          * With a look-up table which covers y^n (n<PERIOD)
2485          *
2486          * To achieve constant time decay_load.
2487          */
2488         if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2489                 val >>= local_n / LOAD_AVG_PERIOD;
2490                 local_n %= LOAD_AVG_PERIOD;
2491         }
2492 
2493         val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2494         return val;
2495 }
2496 
2497 /*
2498  * For updates fully spanning n periods, the contribution to runnable
2499  * average will be: \Sum 1024*y^n
2500  *
2501  * We can compute this reasonably efficiently by combining:
2502  *   y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for  n <PERIOD}
2503  */
2504 static u32 __compute_runnable_contrib(u64 n)
2505 {
2506         u32 contrib = 0;
2507 
2508         if (likely(n <= LOAD_AVG_PERIOD))
2509                 return runnable_avg_yN_sum[n];
2510         else if (unlikely(n >= LOAD_AVG_MAX_N))
2511                 return LOAD_AVG_MAX;
2512 
2513         /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2514         do {
2515                 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2516                 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2517 
2518                 n -= LOAD_AVG_PERIOD;
2519         } while (n > LOAD_AVG_PERIOD);
2520 
2521         contrib = decay_load(contrib, n);
2522         return contrib + runnable_avg_yN_sum[n];
2523 }
2524 
2525 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2526 #error "load tracking assumes 2^10 as unit"
2527 #endif
2528 
2529 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2530 
2531 /*
2532  * We can represent the historical contribution to runnable average as the
2533  * coefficients of a geometric series.  To do this we sub-divide our runnable
2534  * history into segments of approximately 1ms (1024us); label the segment that
2535  * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2536  *
2537  * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2538  *      p0            p1           p2
2539  *     (now)       (~1ms ago)  (~2ms ago)
2540  *
2541  * Let u_i denote the fraction of p_i that the entity was runnable.
2542  *
2543  * We then designate the fractions u_i as our co-efficients, yielding the
2544  * following representation of historical load:
2545  *   u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2546  *
2547  * We choose y based on the with of a reasonably scheduling period, fixing:
2548  *   y^32 = 0.5
2549  *
2550  * This means that the contribution to load ~32ms ago (u_32) will be weighted
2551  * approximately half as much as the contribution to load within the last ms
2552  * (u_0).
2553  *
2554  * When a period "rolls over" and we have new u_0`, multiplying the previous
2555  * sum again by y is sufficient to update:
2556  *   load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2557  *            = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2558  */
2559 static __always_inline int
2560 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2561                   unsigned long weight, int running, struct cfs_rq *cfs_rq)
2562 {
2563         u64 delta, scaled_delta, periods;
2564         u32 contrib;
2565         unsigned int delta_w, scaled_delta_w, decayed = 0;
2566         unsigned long scale_freq, scale_cpu;
2567 
2568         delta = now - sa->last_update_time;
2569         /*
2570          * This should only happen when time goes backwards, which it
2571          * unfortunately does during sched clock init when we swap over to TSC.
2572          */
2573         if ((s64)delta < 0) {
2574                 sa->last_update_time = now;
2575                 return 0;
2576         }
2577 
2578         /*
2579          * Use 1024ns as the unit of measurement since it's a reasonable
2580          * approximation of 1us and fast to compute.
2581          */
2582         delta >>= 10;
2583         if (!delta)
2584                 return 0;
2585         sa->last_update_time = now;
2586 
2587         scale_freq = arch_scale_freq_capacity(NULL, cpu);
2588         scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2589 
2590         /* delta_w is the amount already accumulated against our next period */
2591         delta_w = sa->period_contrib;
2592         if (delta + delta_w >= 1024) {
2593                 decayed = 1;
2594 
2595                 /* how much left for next period will start over, we don't know yet */
2596                 sa->period_contrib = 0;
2597 
2598                 /*
2599                  * Now that we know we're crossing a period boundary, figure
2600                  * out how much from delta we need to complete the current
2601                  * period and accrue it.
2602                  */
2603                 delta_w = 1024 - delta_w;
2604                 scaled_delta_w = cap_scale(delta_w, scale_freq);
2605                 if (weight) {
2606                         sa->load_sum += weight * scaled_delta_w;
2607                         if (cfs_rq) {
2608                                 cfs_rq->runnable_load_sum +=
2609                                                 weight * scaled_delta_w;
2610                         }
2611                 }
2612                 if (running)
2613                         sa->util_sum += scaled_delta_w * scale_cpu;
2614 
2615                 delta -= delta_w;
2616 
2617                 /* Figure out how many additional periods this update spans */
2618                 periods = delta / 1024;
2619                 delta %= 1024;
2620 
2621                 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2622                 if (cfs_rq) {
2623                         cfs_rq->runnable_load_sum =
2624                                 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2625                 }
2626                 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2627 
2628                 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2629                 contrib = __compute_runnable_contrib(periods);
2630                 contrib = cap_scale(contrib, scale_freq);
2631                 if (weight) {
2632                         sa->load_sum += weight * contrib;
2633                         if (cfs_rq)
2634                                 cfs_rq->runnable_load_sum += weight * contrib;
2635                 }
2636                 if (running)
2637                         sa->util_sum += contrib * scale_cpu;
2638         }
2639 
2640         /* Remainder of delta accrued against u_0` */
2641         scaled_delta = cap_scale(delta, scale_freq);
2642         if (weight) {
2643                 sa->load_sum += weight * scaled_delta;
2644                 if (cfs_rq)
2645                         cfs_rq->runnable_load_sum += weight * scaled_delta;
2646         }
2647         if (running)
2648                 sa->util_sum += scaled_delta * scale_cpu;
2649 
2650         sa->period_contrib += delta;
2651 
2652         if (decayed) {
2653                 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2654                 if (cfs_rq) {
2655                         cfs_rq->runnable_load_avg =
2656                                 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2657                 }
2658                 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2659         }
2660 
2661         return decayed;
2662 }
2663 
2664 #ifdef CONFIG_FAIR_GROUP_SCHED
2665 /*
2666  * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2667  * and effective_load (which is not done because it is too costly).
2668  */
2669 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2670 {
2671         long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2672 
2673         if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2674                 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2675                 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2676         }
2677 }
2678 
2679 #else /* CONFIG_FAIR_GROUP_SCHED */
2680 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2681 #endif /* CONFIG_FAIR_GROUP_SCHED */
2682 
2683 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2684 
2685 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2686 static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
2687 {
2688         struct sched_avg *sa = &cfs_rq->avg;
2689         int decayed, removed = 0;
2690 
2691         if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2692                 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2693                 sa->load_avg = max_t(long, sa->load_avg - r, 0);
2694                 sa->load_sum = max_t(s64, sa->load_sum - r * LOAD_AVG_MAX, 0);
2695                 removed = 1;
2696         }
2697 
2698         if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2699                 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2700                 sa->util_avg = max_t(long, sa->util_avg - r, 0);
2701                 sa->util_sum = max_t(s32, sa->util_sum - r * LOAD_AVG_MAX, 0);
2702         }
2703 
2704         decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2705                 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2706 
2707 #ifndef CONFIG_64BIT
2708         smp_wmb();
2709         cfs_rq->load_last_update_time_copy = sa->last_update_time;
2710 #endif
2711 
2712         return decayed || removed;
2713 }
2714 
2715 /* Update task and its cfs_rq load average */
2716 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2717 {
2718         struct cfs_rq *cfs_rq = cfs_rq_of(se);
2719         u64 now = cfs_rq_clock_task(cfs_rq);
2720         int cpu = cpu_of(rq_of(cfs_rq));
2721 
2722         /*
2723          * Track task load average for carrying it to new CPU after migrated, and
2724          * track group sched_entity load average for task_h_load calc in migration
2725          */
2726         __update_load_avg(now, cpu, &se->avg,
2727                           se->on_rq * scale_load_down(se->load.weight),
2728                           cfs_rq->curr == se, NULL);
2729 
2730         if (update_cfs_rq_load_avg(now, cfs_rq) && update_tg)
2731                 update_tg_load_avg(cfs_rq, 0);
2732 }
2733 
2734 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2735 {
2736         if (!sched_feat(ATTACH_AGE_LOAD))
2737                 goto skip_aging;
2738 
2739         /*
2740          * If we got migrated (either between CPUs or between cgroups) we'll
2741          * have aged the average right before clearing @last_update_time.
2742          */
2743         if (se->avg.last_update_time) {
2744                 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2745                                   &se->avg, 0, 0, NULL);
2746 
2747                 /*
2748                  * XXX: we could have just aged the entire load away if we've been
2749                  * absent from the fair class for too long.
2750                  */
2751         }
2752 
2753 skip_aging:
2754         se->avg.last_update_time = cfs_rq->avg.last_update_time;
2755         cfs_rq->avg.load_avg += se->avg.load_avg;
2756         cfs_rq->avg.load_sum += se->avg.load_sum;
2757         cfs_rq->avg.util_avg += se->avg.util_avg;
2758         cfs_rq->avg.util_sum += se->avg.util_sum;
2759 }
2760 
2761 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2762 {
2763         __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2764                           &se->avg, se->on_rq * scale_load_down(se->load.weight),
2765                           cfs_rq->curr == se, NULL);
2766 
2767         cfs_rq->avg.load_avg = max_t(long, cfs_rq->avg.load_avg - se->avg.load_avg, 0);
2768         cfs_rq->avg.load_sum = max_t(s64,  cfs_rq->avg.load_sum - se->avg.load_sum, 0);
2769         cfs_rq->avg.util_avg = max_t(long, cfs_rq->avg.util_avg - se->avg.util_avg, 0);
2770         cfs_rq->avg.util_sum = max_t(s32,  cfs_rq->avg.util_sum - se->avg.util_sum, 0);
2771 }
2772 
2773 /* Add the load generated by se into cfs_rq's load average */
2774 static inline void
2775 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2776 {
2777         struct sched_avg *sa = &se->avg;
2778         u64 now = cfs_rq_clock_task(cfs_rq);
2779         int migrated, decayed;
2780 
2781         migrated = !sa->last_update_time;
2782         if (!migrated) {
2783                 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2784                         se->on_rq * scale_load_down(se->load.weight),
2785                         cfs_rq->curr == se, NULL);
2786         }
2787 
2788         decayed = update_cfs_rq_load_avg(now, cfs_rq);
2789 
2790         cfs_rq->runnable_load_avg += sa->load_avg;
2791         cfs_rq->runnable_load_sum += sa->load_sum;
2792 
2793         if (migrated)
2794                 attach_entity_load_avg(cfs_rq, se);
2795 
2796         if (decayed || migrated)
2797                 update_tg_load_avg(cfs_rq, 0);
2798 }
2799 
2800 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2801 static inline void
2802 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2803 {
2804         update_load_avg(se, 1);
2805 
2806         cfs_rq->runnable_load_avg =
2807                 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2808         cfs_rq->runnable_load_sum =
2809                 max_t(s64,  cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2810 }
2811 
2812 /*
2813  * Task first catches up with cfs_rq, and then subtract
2814  * itself from the cfs_rq (task must be off the queue now).
2815  */
2816 void remove_entity_load_avg(struct sched_entity *se)
2817 {
2818         struct cfs_rq *cfs_rq = cfs_rq_of(se);
2819         u64 last_update_time;
2820 
2821 #ifndef CONFIG_64BIT
2822         u64 last_update_time_copy;
2823 
2824         do {
2825                 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2826                 smp_rmb();
2827                 last_update_time = cfs_rq->avg.last_update_time;
2828         } while (last_update_time != last_update_time_copy);
2829 #else
2830         last_update_time = cfs_rq->avg.last_update_time;
2831 #endif
2832 
2833         __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
2834         atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2835         atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2836 }
2837 
2838 /*
2839  * Update the rq's load with the elapsed running time before entering
2840  * idle. if the last scheduled task is not a CFS task, idle_enter will
2841  * be the only way to update the runnable statistic.
2842  */
2843 void idle_enter_fair(struct rq *this_rq)
2844 {
2845 }
2846 
2847 /*
2848  * Update the rq's load with the elapsed idle time before a task is
2849  * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2850  * be the only way to update the runnable statistic.
2851  */
2852 void idle_exit_fair(struct rq *this_rq)
2853 {
2854 }
2855 
2856 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
2857 {
2858         return cfs_rq->runnable_load_avg;
2859 }
2860 
2861 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
2862 {
2863         return cfs_rq->avg.load_avg;
2864 }
2865 
2866 static int idle_balance(struct rq *this_rq);
2867 
2868 #else /* CONFIG_SMP */
2869 
2870 static inline void update_load_avg(struct sched_entity *se, int update_tg) {}
2871 static inline void
2872 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2873 static inline void
2874 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2875 static inline void remove_entity_load_avg(struct sched_entity *se) {}
2876 
2877 static inline void
2878 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2879 static inline void
2880 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2881 
2882 static inline int idle_balance(struct rq *rq)
2883 {
2884         return 0;
2885 }
2886 
2887 #endif /* CONFIG_SMP */
2888 
2889 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2890 {
2891 #ifdef CONFIG_SCHEDSTATS
2892         struct task_struct *tsk = NULL;
2893 
2894         if (entity_is_task(se))
2895                 tsk = task_of(se);
2896 
2897         if (se->statistics.sleep_start) {
2898                 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2899 
2900                 if ((s64)delta < 0)
2901                         delta = 0;
2902 
2903                 if (unlikely(delta > se->statistics.sleep_max))
2904                         se->statistics.sleep_max = delta;
2905 
2906                 se->statistics.sleep_start = 0;
2907                 se->statistics.sum_sleep_runtime += delta;
2908 
2909                 if (tsk) {
2910                         account_scheduler_latency(tsk, delta >> 10, 1);
2911                         trace_sched_stat_sleep(tsk, delta);
2912                 }
2913         }
2914         if (se->statistics.block_start) {
2915                 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2916 
2917                 if ((s64)delta < 0)
2918                         delta = 0;
2919 
2920                 if (unlikely(delta > se->statistics.block_max))
2921                         se->statistics.block_max = delta;
2922 
2923                 se->statistics.block_start = 0;
2924                 se->statistics.sum_sleep_runtime += delta;
2925 
2926                 if (tsk) {
2927                         if (tsk->in_iowait) {
2928                                 se->statistics.iowait_sum += delta;
2929                                 se->statistics.iowait_count++;
2930                                 trace_sched_stat_iowait(tsk, delta);
2931                         }
2932 
2933                         trace_sched_stat_blocked(tsk, delta);
2934 
2935                         /*
2936                          * Blocking time is in units of nanosecs, so shift by
2937                          * 20 to get a milliseconds-range estimation of the
2938                          * amount of time that the task spent sleeping:
2939                          */
2940                         if (unlikely(prof_on == SLEEP_PROFILING)) {
2941                                 profile_hits(SLEEP_PROFILING,
2942                                                 (void *)get_wchan(tsk),
2943                                                 delta >> 20);
2944                         }
2945                         account_scheduler_latency(tsk, delta >> 10, 0);
2946                 }
2947         }
2948 #endif
2949 }
2950 
2951 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2952 {
2953 #ifdef CONFIG_SCHED_DEBUG
2954         s64 d = se->vruntime - cfs_rq->min_vruntime;
2955 
2956         if (d < 0)
2957                 d = -d;
2958 
2959         if (d > 3*sysctl_sched_latency)
2960                 schedstat_inc(cfs_rq, nr_spread_over);
2961 #endif
2962 }
2963 
2964 static void
2965 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2966 {
2967         u64 vruntime = cfs_rq->min_vruntime;
2968 
2969         /*
2970          * The 'current' period is already promised to the current tasks,
2971          * however the extra weight of the new task will slow them down a
2972          * little, place the new task so that it fits in the slot that
2973          * stays open at the end.
2974          */
2975         if (initial && sched_feat(START_DEBIT))
2976                 vruntime += sched_vslice(cfs_rq, se);
2977 
2978         /* sleeps up to a single latency don't count. */
2979         if (!initial) {
2980                 unsigned long thresh = sysctl_sched_latency;
2981 
2982                 /*
2983                  * Halve their sleep time's effect, to allow
2984                  * for a gentler effect of sleepers:
2985                  */
2986                 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2987                         thresh >>= 1;
2988 
2989                 vruntime -= thresh;
2990         }
2991 
2992         /* ensure we never gain time by being placed backwards. */
2993         se->vruntime = max_vruntime(se->vruntime, vruntime);
2994 }
2995 
2996 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2997 
2998 static void
2999 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3000 {
3001         /*
3002          * Update the normalized vruntime before updating min_vruntime
3003          * through calling update_curr().
3004          */
3005         if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3006                 se->vruntime += cfs_rq->min_vruntime;
3007 
3008         /*
3009          * Update run-time statistics of the 'current'.
3010          */
3011         update_curr(cfs_rq);
3012         enqueue_entity_load_avg(cfs_rq, se);
3013         account_entity_enqueue(cfs_rq, se);
3014         update_cfs_shares(cfs_rq);
3015 
3016         if (flags & ENQUEUE_WAKEUP) {
3017                 place_entity(cfs_rq, se, 0);
3018                 enqueue_sleeper(cfs_rq, se);
3019         }
3020 
3021         update_stats_enqueue(cfs_rq, se);
3022         check_spread(cfs_rq, se);
3023         if (se != cfs_rq->curr)
3024                 __enqueue_entity(cfs_rq, se);
3025         se->on_rq = 1;
3026 
3027         if (cfs_rq->nr_running == 1) {
3028                 list_add_leaf_cfs_rq(cfs_rq);
3029                 check_enqueue_throttle(cfs_rq);
3030         }
3031 }
3032 
3033 static void __clear_buddies_last(struct sched_entity *se)
3034 {
3035         for_each_sched_entity(se) {
3036                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3037                 if (cfs_rq->last != se)
3038                         break;
3039 
3040                 cfs_rq->last = NULL;
3041         }
3042 }
3043 
3044 static void __clear_buddies_next(struct sched_entity *se)
3045 {
3046         for_each_sched_entity(se) {
3047                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3048                 if (cfs_rq->next != se)
3049                         break;
3050 
3051                 cfs_rq->next = NULL;
3052         }
3053 }
3054 
3055 static void __clear_buddies_skip(struct sched_entity *se)
3056 {
3057         for_each_sched_entity(se) {
3058                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3059                 if (cfs_rq->skip != se)
3060                         break;
3061 
3062                 cfs_rq->skip = NULL;
3063         }
3064 }
3065 
3066 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3067 {
3068         if (cfs_rq->last == se)
3069                 __clear_buddies_last(se);
3070 
3071         if (cfs_rq->next == se)
3072                 __clear_buddies_next(se);
3073 
3074         if (cfs_rq->skip == se)
3075                 __clear_buddies_skip(se);
3076 }
3077 
3078 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3079 
3080 static void
3081 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3082 {
3083         /*
3084          * Update run-time statistics of the 'current'.
3085          */
3086         update_curr(cfs_rq);
3087         dequeue_entity_load_avg(cfs_rq, se);
3088 
3089         update_stats_dequeue(cfs_rq, se);
3090         if (flags & DEQUEUE_SLEEP) {
3091 #ifdef CONFIG_SCHEDSTATS
3092                 if (entity_is_task(se)) {
3093                         struct task_struct *tsk = task_of(se);
3094 
3095                         if (tsk->state & TASK_INTERRUPTIBLE)
3096                                 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3097                         if (tsk->state & TASK_UNINTERRUPTIBLE)
3098                                 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3099                 }
3100 #endif
3101         }
3102 
3103         clear_buddies(cfs_rq, se);
3104 
3105         if (se != cfs_rq->curr)
3106                 __dequeue_entity(cfs_rq, se);
3107         se->on_rq = 0;
3108         account_entity_dequeue(cfs_rq, se);
3109 
3110         /*
3111          * Normalize the entity after updating the min_vruntime because the
3112          * update can refer to the ->curr item and we need to reflect this
3113          * movement in our normalized position.
3114          */
3115         if (!(flags & DEQUEUE_SLEEP))
3116                 se->vruntime -= cfs_rq->min_vruntime;
3117 
3118         /* return excess runtime on last dequeue */
3119         return_cfs_rq_runtime(cfs_rq);
3120 
3121         update_min_vruntime(cfs_rq);
3122         update_cfs_shares(cfs_rq);
3123 }
3124 
3125 /*
3126  * Preempt the current task with a newly woken task if needed:
3127  */
3128 static void
3129 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3130 {
3131         unsigned long ideal_runtime, delta_exec;
3132         struct sched_entity *se;
3133         s64 delta;
3134 
3135         ideal_runtime = sched_slice(cfs_rq, curr);
3136         delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3137         if (delta_exec > ideal_runtime) {
3138                 resched_curr(rq_of(cfs_rq));
3139                 /*
3140                  * The current task ran long enough, ensure it doesn't get
3141                  * re-elected due to buddy favours.
3142                  */
3143                 clear_buddies(cfs_rq, curr);
3144                 return;
3145         }
3146 
3147         /*
3148          * Ensure that a task that missed wakeup preemption by a
3149          * narrow margin doesn't have to wait for a full slice.
3150          * This also mitigates buddy induced latencies under load.
3151          */
3152         if (delta_exec < sysctl_sched_min_granularity)
3153                 return;
3154 
3155         se = __pick_first_entity(cfs_rq);
3156         delta = curr->vruntime - se->vruntime;
3157 
3158         if (delta < 0)
3159                 return;
3160 
3161         if (delta > ideal_runtime)
3162                 resched_curr(rq_of(cfs_rq));
3163 }
3164 
3165 static void
3166 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3167 {
3168         /* 'current' is not kept within the tree. */
3169         if (se->on_rq) {
3170                 /*
3171                  * Any task has to be enqueued before it get to execute on
3172                  * a CPU. So account for the time it spent waiting on the
3173                  * runqueue.
3174                  */
3175                 update_stats_wait_end(cfs_rq, se);
3176                 __dequeue_entity(cfs_rq, se);
3177                 update_load_avg(se, 1);
3178         }
3179 
3180         update_stats_curr_start(cfs_rq, se);
3181         cfs_rq->curr = se;
3182 #ifdef CONFIG_SCHEDSTATS
3183         /*
3184          * Track our maximum slice length, if the CPU's load is at
3185          * least twice that of our own weight (i.e. dont track it
3186          * when there are only lesser-weight tasks around):
3187          */
3188         if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3189                 se->statistics.slice_max = max(se->statistics.slice_max,
3190                         se->sum_exec_runtime - se->prev_sum_exec_runtime);
3191         }
3192 #endif
3193         se->prev_sum_exec_runtime = se->sum_exec_runtime;
3194 }
3195 
3196 static int
3197 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3198 
3199 /*
3200  * Pick the next process, keeping these things in mind, in this order:
3201  * 1) keep things fair between processes/task groups
3202  * 2) pick the "next" process, since someone really wants that to run
3203  * 3) pick the "last" process, for cache locality
3204  * 4) do not run the "skip" process, if something else is available
3205  */
3206 static struct sched_entity *
3207 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3208 {
3209         struct sched_entity *left = __pick_first_entity(cfs_rq);
3210         struct sched_entity *se;
3211 
3212         /*
3213          * If curr is set we have to see if its left of the leftmost entity
3214          * still in the tree, provided there was anything in the tree at all.
3215          */
3216         if (!left || (curr && entity_before(curr, left)))
3217                 left = curr;
3218 
3219         se = left; /* ideally we run the leftmost entity */
3220 
3221         /*
3222          * Avoid running the skip buddy, if running something else can
3223          * be done without getting too unfair.
3224          */
3225         if (cfs_rq->skip == se) {
3226                 struct sched_entity *second;
3227 
3228                 if (se == curr) {
3229                         second = __pick_first_entity(cfs_rq);
3230                 } else {
3231                         second = __pick_next_entity(se);
3232                         if (!second || (curr && entity_before(curr, second)))
3233                                 second = curr;
3234                 }
3235 
3236                 if (second && wakeup_preempt_entity(second, left) < 1)
3237                         se = second;
3238         }
3239 
3240         /*
3241          * Prefer last buddy, try to return the CPU to a preempted task.
3242          */
3243         if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3244                 se = cfs_rq->last;
3245 
3246         /*
3247          * Someone really wants this to run. If it's not unfair, run it.
3248          */
3249         if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3250                 se = cfs_rq->next;
3251 
3252         clear_buddies(cfs_rq, se);
3253 
3254         return se;
3255 }
3256 
3257 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3258 
3259 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3260 {
3261         /*
3262          * If still on the runqueue then deactivate_task()
3263          * was not called and update_curr() has to be done:
3264          */
3265         if (prev->on_rq)
3266                 update_curr(cfs_rq);
3267 
3268         /* throttle cfs_rqs exceeding runtime */
3269         check_cfs_rq_runtime(cfs_rq);
3270 
3271         check_spread(cfs_rq, prev);
3272         if (prev->on_rq) {
3273                 update_stats_wait_start(cfs_rq, prev);
3274                 /* Put 'current' back into the tree. */
3275                 __enqueue_entity(cfs_rq, prev);
3276                 /* in !on_rq case, update occurred at dequeue */
3277                 update_load_avg(prev, 0);
3278         }
3279         cfs_rq->curr = NULL;
3280 }
3281 
3282 static void
3283 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3284 {
3285         /*
3286          * Update run-time statistics of the 'current'.
3287          */
3288         update_curr(cfs_rq);
3289 
3290         /*
3291          * Ensure that runnable average is periodically updated.
3292          */
3293         update_load_avg(curr, 1);
3294         update_cfs_shares(cfs_rq);
3295 
3296 #ifdef CONFIG_SCHED_HRTICK
3297         /*
3298          * queued ticks are scheduled to match the slice, so don't bother
3299          * validating it and just reschedule.
3300          */
3301         if (queued) {
3302                 resched_curr(rq_of(cfs_rq));
3303                 return;
3304         }
3305         /*
3306          * don't let the period tick interfere with the hrtick preemption
3307          */
3308         if (!sched_feat(DOUBLE_TICK) &&
3309                         hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3310                 return;
3311 #endif
3312 
3313         if (cfs_rq->nr_running > 1)
3314                 check_preempt_tick(cfs_rq, curr);
3315 }
3316 
3317 
3318 /**************************************************
3319  * CFS bandwidth control machinery
3320  */
3321 
3322 #ifdef CONFIG_CFS_BANDWIDTH
3323 
3324 #ifdef HAVE_JUMP_LABEL
3325 static struct static_key __cfs_bandwidth_used;
3326 
3327 static inline bool cfs_bandwidth_used(void)
3328 {
3329         return static_key_false(&__cfs_bandwidth_used);
3330 }
3331 
3332 void cfs_bandwidth_usage_inc(void)
3333 {
3334         static_key_slow_inc(&__cfs_bandwidth_used);
3335 }
3336 
3337 void cfs_bandwidth_usage_dec(void)
3338 {
3339         static_key_slow_dec(&__cfs_bandwidth_used);
3340 }
3341 #else /* HAVE_JUMP_LABEL */
3342 static bool cfs_bandwidth_used(void)
3343 {
3344         return true;
3345 }
3346 
3347 void cfs_bandwidth_usage_inc(void) {}
3348 void cfs_bandwidth_usage_dec(void) {}
3349 #endif /* HAVE_JUMP_LABEL */
3350 
3351 /*
3352  * default period for cfs group bandwidth.
3353  * default: 0.1s, units: nanoseconds
3354  */
3355 static inline u64 default_cfs_period(void)
3356 {
3357         return 100000000ULL;
3358 }
3359 
3360 static inline u64 sched_cfs_bandwidth_slice(void)
3361 {
3362         return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3363 }
3364 
3365 /*
3366  * Replenish runtime according to assigned quota and update expiration time.
3367  * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3368  * additional synchronization around rq->lock.
3369  *
3370  * requires cfs_b->lock
3371  */
3372 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3373 {
3374         u64 now;
3375 
3376         if (cfs_b->quota == RUNTIME_INF)
3377                 return;
3378 
3379         now = sched_clock_cpu(smp_processor_id());
3380         cfs_b->runtime = cfs_b->quota;
3381         cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3382 }
3383 
3384 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3385 {
3386         return &tg->cfs_bandwidth;
3387 }
3388 
3389 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3390 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3391 {
3392         if (unlikely(cfs_rq->throttle_count))
3393                 return cfs_rq->throttled_clock_task;
3394 
3395         return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3396 }
3397 
3398 /* returns 0 on failure to allocate runtime */
3399 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3400 {
3401         struct task_group *tg = cfs_rq->tg;
3402         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3403         u64 amount = 0, min_amount, expires;
3404 
3405         /* note: this is a positive sum as runtime_remaining <= 0 */
3406         min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3407 
3408         raw_spin_lock(&cfs_b->lock);
3409         if (cfs_b->quota == RUNTIME_INF)
3410                 amount = min_amount;
3411         else {
3412                 start_cfs_bandwidth(cfs_b);
3413 
3414                 if (cfs_b->runtime > 0) {
3415                         amount = min(cfs_b->runtime, min_amount);
3416                         cfs_b->runtime -= amount;
3417                         cfs_b->idle = 0;
3418                 }
3419         }
3420         expires = cfs_b->runtime_expires;
3421         raw_spin_unlock(&cfs_b->lock);
3422 
3423         cfs_rq->runtime_remaining += amount;
3424         /*
3425          * we may have advanced our local expiration to account for allowed
3426          * spread between our sched_clock and the one on which runtime was
3427          * issued.
3428          */
3429         if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3430                 cfs_rq->runtime_expires = expires;
3431 
3432         return cfs_rq->runtime_remaining > 0;
3433 }
3434 
3435 /*
3436  * Note: This depends on the synchronization provided by sched_clock and the
3437  * fact that rq->clock snapshots this value.
3438  */
3439 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3440 {
3441         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3442 
3443         /* if the deadline is ahead of our clock, nothing to do */
3444         if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3445                 return;
3446 
3447         if (cfs_rq->runtime_remaining < 0)
3448                 return;
3449 
3450         /*
3451          * If the local deadline has passed we have to consider the
3452          * possibility that our sched_clock is 'fast' and the global deadline
3453          * has not truly expired.
3454          *
3455          * Fortunately we can check determine whether this the case by checking
3456          * whether the global deadline has advanced. It is valid to compare
3457          * cfs_b->runtime_expires without any locks since we only care about
3458          * exact equality, so a partial write will still work.
3459          */
3460 
3461         if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3462                 /* extend local deadline, drift is bounded above by 2 ticks */
3463                 cfs_rq->runtime_expires += TICK_NSEC;
3464         } else {
3465                 /* global deadline is ahead, expiration has passed */
3466                 cfs_rq->runtime_remaining = 0;
3467         }
3468 }
3469 
3470 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3471 {
3472         /* dock delta_exec before expiring quota (as it could span periods) */
3473         cfs_rq->runtime_remaining -= delta_exec;
3474         expire_cfs_rq_runtime(cfs_rq);
3475 
3476         if (likely(cfs_rq->runtime_remaining > 0))
3477                 return;
3478 
3479         /*
3480          * if we're unable to extend our runtime we resched so that the active
3481          * hierarchy can be throttled
3482          */
3483         if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3484                 resched_curr(rq_of(cfs_rq));
3485 }
3486 
3487 static __always_inline
3488 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3489 {
3490         if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3491                 return;
3492 
3493         __account_cfs_rq_runtime(cfs_rq, delta_exec);
3494 }
3495 
3496 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3497 {
3498         return cfs_bandwidth_used() && cfs_rq->throttled;
3499 }
3500 
3501 /* check whether cfs_rq, or any parent, is throttled */
3502 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3503 {
3504         return cfs_bandwidth_used() && cfs_rq->throttle_count;
3505 }
3506 
3507 /*
3508  * Ensure that neither of the group entities corresponding to src_cpu or
3509  * dest_cpu are members of a throttled hierarchy when performing group
3510  * load-balance operations.
3511  */
3512 static inline int throttled_lb_pair(struct task_group *tg,
3513                                     int src_cpu, int dest_cpu)
3514 {
3515         struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3516 
3517         src_cfs_rq = tg->cfs_rq[src_cpu];
3518         dest_cfs_rq = tg->cfs_rq[dest_cpu];
3519 
3520         return throttled_hierarchy(src_cfs_rq) ||
3521                throttled_hierarchy(dest_cfs_rq);
3522 }
3523 
3524 /* updated child weight may affect parent so we have to do this bottom up */
3525 static int tg_unthrottle_up(struct task_group *tg, void *data)
3526 {
3527         struct rq *rq = data;
3528         struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3529 
3530         cfs_rq->throttle_count--;
3531 #ifdef CONFIG_SMP
3532         if (!cfs_rq->throttle_count) {
3533                 /* adjust cfs_rq_clock_task() */
3534                 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3535                                              cfs_rq->throttled_clock_task;
3536         }
3537 #endif
3538 
3539         return 0;
3540 }
3541 
3542 static int tg_throttle_down(struct task_group *tg, void *data)
3543 {
3544         struct rq *rq = data;
3545         struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3546 
3547         /* group is entering throttled state, stop time */
3548         if (!cfs_rq->throttle_count)
3549                 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3550         cfs_rq->throttle_count++;
3551 
3552         return 0;
3553 }
3554 
3555 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3556 {
3557         struct rq *rq = rq_of(cfs_rq);
3558         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3559         struct sched_entity *se;
3560         long task_delta, dequeue = 1;
3561         bool empty;
3562 
3563         se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3564 
3565         /* freeze hierarchy runnable averages while throttled */
3566         rcu_read_lock();
3567         walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3568         rcu_read_unlock();
3569 
3570         task_delta = cfs_rq->h_nr_running;
3571         for_each_sched_entity(se) {
3572                 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3573                 /* throttled entity or throttle-on-deactivate */
3574                 if (!se->on_rq)
3575                         break;
3576 
3577                 if (dequeue)
3578                         dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3579                 qcfs_rq->h_nr_running -= task_delta;
3580 
3581                 if (qcfs_rq->load.weight)
3582                         dequeue = 0;
3583         }
3584 
3585         if (!se)
3586                 sub_nr_running(rq, task_delta);
3587 
3588         cfs_rq->throttled = 1;
3589         cfs_rq->throttled_clock = rq_clock(rq);
3590         raw_spin_lock(&cfs_b->lock);
3591         empty = list_empty(&cfs_b->throttled_cfs_rq);
3592 
3593         /*
3594          * Add to the _head_ of the list, so that an already-started
3595          * distribute_cfs_runtime will not see us
3596          */
3597         list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3598 
3599         /*
3600          * If we're the first throttled task, make sure the bandwidth
3601          * timer is running.
3602          */
3603         if (empty)
3604                 start_cfs_bandwidth(cfs_b);
3605 
3606         raw_spin_unlock(&cfs_b->lock);
3607 }
3608 
3609 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3610 {
3611         struct rq *rq = rq_of(cfs_rq);
3612         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3613         struct sched_entity *se;
3614         int enqueue = 1;
3615         long task_delta;
3616 
3617         se = cfs_rq->tg->se[cpu_of(rq)];
3618 
3619         cfs_rq->throttled = 0;
3620 
3621         update_rq_clock(rq);
3622 
3623         raw_spin_lock(&cfs_b->lock);
3624         cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3625         list_del_rcu(&cfs_rq->throttled_list);
3626         raw_spin_unlock(&cfs_b->lock);
3627 
3628         /* update hierarchical throttle state */
3629         walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3630 
3631         if (!cfs_rq->load.weight)
3632                 return;
3633 
3634         task_delta = cfs_rq->h_nr_running;
3635         for_each_sched_entity(se) {
3636                 if (se->on_rq)
3637                         enqueue = 0;
3638 
3639                 cfs_rq = cfs_rq_of(se);
3640                 if (enqueue)
3641                         enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3642                 cfs_rq->h_nr_running += task_delta;
3643 
3644                 if (cfs_rq_throttled(cfs_rq))
3645                         break;
3646         }
3647 
3648         if (!se)
3649                 add_nr_running(rq, task_delta);
3650 
3651         /* determine whether we need to wake up potentially idle cpu */
3652         if (rq->curr == rq->idle && rq->cfs.nr_running)
3653                 resched_curr(rq);
3654 }
3655 
3656 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3657                 u64 remaining, u64 expires)
3658 {
3659         struct cfs_rq *cfs_rq;
3660         u64 runtime;
3661         u64 starting_runtime = remaining;
3662 
3663         rcu_read_lock();
3664         list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3665                                 throttled_list) {
3666                 struct rq *rq = rq_of(cfs_rq);
3667 
3668                 raw_spin_lock(&rq->lock);
3669                 if (!cfs_rq_throttled(cfs_rq))
3670                         goto next;
3671 
3672                 runtime = -cfs_rq->runtime_remaining + 1;
3673                 if (runtime > remaining)
3674                         runtime = remaining;
3675                 remaining -= runtime;
3676 
3677                 cfs_rq->runtime_remaining += runtime;
3678                 cfs_rq->runtime_expires = expires;
3679 
3680                 /* we check whether we're throttled above */
3681                 if (cfs_rq->runtime_remaining > 0)
3682                         unthrottle_cfs_rq(cfs_rq);
3683 
3684 next:
3685                 raw_spin_unlock(&rq->lock);
3686 
3687                 if (!remaining)
3688                         break;
3689         }
3690         rcu_read_unlock();
3691 
3692         return starting_runtime - remaining;
3693 }
3694 
3695 /*
3696  * Responsible for refilling a task_group's bandwidth and unthrottling its
3697  * cfs_rqs as appropriate. If there has been no activity within the last
3698  * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3699  * used to track this state.
3700  */
3701 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3702 {
3703         u64 runtime, runtime_expires;
3704         int throttled;
3705 
3706         /* no need to continue the timer with no bandwidth constraint */
3707         if (cfs_b->quota == RUNTIME_INF)
3708                 goto out_deactivate;
3709 
3710         throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3711         cfs_b->nr_periods += overrun;
3712 
3713         /*
3714          * idle depends on !throttled (for the case of a large deficit), and if
3715          * we're going inactive then everything else can be deferred
3716          */
3717         if (cfs_b->idle && !throttled)
3718                 goto out_deactivate;
3719 
3720         __refill_cfs_bandwidth_runtime(cfs_b);
3721 
3722         if (!throttled) {
3723                 /* mark as potentially idle for the upcoming period */
3724                 cfs_b->idle = 1;
3725                 return 0;
3726         }
3727 
3728         /* account preceding periods in which throttling occurred */
3729         cfs_b->nr_throttled += overrun;
3730 
3731         runtime_expires = cfs_b->runtime_expires;
3732 
3733         /*
3734          * This check is repeated as we are holding onto the new bandwidth while
3735          * we unthrottle. This can potentially race with an unthrottled group
3736          * trying to acquire new bandwidth from the global pool. This can result
3737          * in us over-using our runtime if it is all used during this loop, but
3738          * only by limited amounts in that extreme case.
3739          */
3740         while (throttled && cfs_b->runtime > 0) {
3741                 runtime = cfs_b->runtime;
3742                 raw_spin_unlock(&cfs_b->lock);
3743                 /* we can't nest cfs_b->lock while distributing bandwidth */
3744                 runtime = distribute_cfs_runtime(cfs_b, runtime,
3745                                                  runtime_expires);
3746                 raw_spin_lock(&cfs_b->lock);
3747 
3748                 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3749 
3750                 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3751         }
3752 
3753         /*
3754          * While we are ensured activity in the period following an
3755          * unthrottle, this also covers the case in which the new bandwidth is
3756          * insufficient to cover the existing bandwidth deficit.  (Forcing the
3757          * timer to remain active while there are any throttled entities.)
3758          */
3759         cfs_b->idle = 0;
3760 
3761         return 0;
3762 
3763 out_deactivate:
3764         return 1;
3765 }
3766 
3767 /* a cfs_rq won't donate quota below this amount */
3768 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3769 /* minimum remaining period time to redistribute slack quota */
3770 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3771 /* how long we wait to gather additional slack before distributing */
3772 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3773 
3774 /*
3775  * Are we near the end of the current quota period?
3776  *
3777  * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3778  * hrtimer base being cleared by hrtimer_start. In the case of
3779  * migrate_hrtimers, base is never cleared, so we are fine.
3780  */
3781 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3782 {
3783         struct hrtimer *refresh_timer = &cfs_b->period_timer;
3784         u64 remaining;
3785 
3786         /* if the call-back is running a quota refresh is already occurring */
3787         if (hrtimer_callback_running(refresh_timer))
3788                 return 1;
3789 
3790         /* is a quota refresh about to occur? */
3791         remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3792         if (remaining < min_expire)
3793                 return 1;
3794 
3795         return 0;
3796 }
3797 
3798 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3799 {
3800         u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3801 
3802         /* if there's a quota refresh soon don't bother with slack */
3803         if (runtime_refresh_within(cfs_b, min_left))
3804                 return;
3805 
3806         hrtimer_start(&cfs_b->slack_timer,
3807                         ns_to_ktime(cfs_bandwidth_slack_period),
3808                         HRTIMER_MODE_REL);
3809 }
3810 
3811 /* we know any runtime found here is valid as update_curr() precedes return */
3812 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3813 {
3814         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3815         s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3816 
3817         if (slack_runtime <= 0)
3818                 return;
3819 
3820         raw_spin_lock(&cfs_b->lock);
3821         if (cfs_b->quota != RUNTIME_INF &&
3822             cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3823                 cfs_b->runtime += slack_runtime;
3824 
3825                 /* we are under rq->lock, defer unthrottling using a timer */
3826                 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3827                     !list_empty(&cfs_b->throttled_cfs_rq))
3828                         start_cfs_slack_bandwidth(cfs_b);
3829         }
3830         raw_spin_unlock(&cfs_b->lock);
3831 
3832         /* even if it's not valid for return we don't want to try again */
3833         cfs_rq->runtime_remaining -= slack_runtime;
3834 }
3835 
3836 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3837 {
3838         if (!cfs_bandwidth_used())
3839                 return;
3840 
3841         if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3842                 return;
3843 
3844         __return_cfs_rq_runtime(cfs_rq);
3845 }
3846 
3847 /*
3848  * This is done with a timer (instead of inline with bandwidth return) since
3849  * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3850  */
3851 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3852 {
3853         u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3854         u64 expires;
3855 
3856         /* confirm we're still not at a refresh boundary */
3857         raw_spin_lock(&cfs_b->lock);
3858         if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3859                 raw_spin_unlock(&cfs_b->lock);
3860                 return;
3861         }
3862 
3863         if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3864                 runtime = cfs_b->runtime;
3865 
3866         expires = cfs_b->runtime_expires;
3867         raw_spin_unlock(&cfs_b->lock);
3868 
3869         if (!runtime)
3870                 return;
3871 
3872         runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3873 
3874         raw_spin_lock(&cfs_b->lock);
3875         if (expires == cfs_b->runtime_expires)
3876                 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3877         raw_spin_unlock(&cfs_b->lock);
3878 }
3879 
3880 /*
3881  * When a group wakes up we want to make sure that its quota is not already
3882  * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3883  * runtime as update_curr() throttling can not not trigger until it's on-rq.
3884  */
3885 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3886 {
3887         if (!cfs_bandwidth_used())
3888                 return;
3889 
3890         /* an active group must be handled by the update_curr()->put() path */
3891         if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3892                 return;
3893 
3894         /* ensure the group is not already throttled */
3895         if (cfs_rq_throttled(cfs_rq))
3896                 return;
3897 
3898         /* update runtime allocation */
3899         account_cfs_rq_runtime(cfs_rq, 0);
3900         if (cfs_rq->runtime_remaining <= 0)
3901                 throttle_cfs_rq(cfs_rq);
3902 }
3903 
3904 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3905 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3906 {
3907         if (!cfs_bandwidth_used())
3908                 return false;
3909 
3910         if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3911                 return false;
3912 
3913         /*
3914          * it's possible for a throttled entity to be forced into a running
3915          * state (e.g. set_curr_task), in this case we're finished.
3916          */
3917         if (cfs_rq_throttled(cfs_rq))
3918                 return true;
3919 
3920         throttle_cfs_rq(cfs_rq);
3921         return true;
3922 }
3923 
3924 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3925 {
3926         struct cfs_bandwidth *cfs_b =
3927                 container_of(timer, struct cfs_bandwidth, slack_timer);
3928 
3929         do_sched_cfs_slack_timer(cfs_b);
3930 
3931         return HRTIMER_NORESTART;
3932 }
3933 
3934 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3935 {
3936         struct cfs_bandwidth *cfs_b =
3937                 container_of(timer, struct cfs_bandwidth, period_timer);
3938         int overrun;
3939         int idle = 0;
3940 
3941         raw_spin_lock(&cfs_b->lock);
3942         for (;;) {
3943                 overrun = hrtimer_forward_now(timer, cfs_b->period);
3944                 if (!overrun)
3945                         break;
3946 
3947                 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3948         }
3949         if (idle)
3950                 cfs_b->period_active = 0;
3951         raw_spin_unlock(&cfs_b->lock);
3952 
3953         return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3954 }
3955 
3956 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3957 {
3958         raw_spin_lock_init(&cfs_b->lock);
3959         cfs_b->runtime = 0;
3960         cfs_b->quota = RUNTIME_INF;
3961         cfs_b->period = ns_to_ktime(default_cfs_period());
3962 
3963         INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3964         hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
3965         cfs_b->period_timer.function = sched_cfs_period_timer;
3966         hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3967         cfs_b->slack_timer.function = sched_cfs_slack_timer;
3968 }
3969 
3970 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3971 {
3972         cfs_rq->runtime_enabled = 0;
3973         INIT_LIST_HEAD(&cfs_rq->throttled_list);
3974 }
3975 
3976 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3977 {
3978         lockdep_assert_held(&cfs_b->lock);
3979 
3980         if (!cfs_b->period_active) {
3981                 cfs_b->period_active = 1;
3982                 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
3983                 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
3984         }
3985 }
3986 
3987 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3988 {
3989         /* init_cfs_bandwidth() was not called */
3990         if (!cfs_b->throttled_cfs_rq.next)
3991                 return;
3992 
3993         hrtimer_cancel(&cfs_b->period_timer);
3994         hrtimer_cancel(&cfs_b->slack_timer);
3995 }
3996 
3997 static void __maybe_unused update_runtime_enabled(struct rq *rq)
3998 {
3999         struct cfs_rq *cfs_rq;
4000 
4001         for_each_leaf_cfs_rq(rq, cfs_rq) {
4002                 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4003 
4004                 raw_spin_lock(&cfs_b->lock);
4005                 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4006                 raw_spin_unlock(&cfs_b->lock);
4007         }
4008 }
4009 
4010 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4011 {
4012         struct cfs_rq *cfs_rq;
4013 
4014         for_each_leaf_cfs_rq(rq, cfs_rq) {
4015                 if (!cfs_rq->runtime_enabled)
4016                         continue;
4017 
4018                 /*
4019                  * clock_task is not advancing so we just need to make sure
4020                  * there's some valid quota amount
4021                  */
4022                 cfs_rq->runtime_remaining = 1;
4023                 /*
4024                  * Offline rq is schedulable till cpu is completely disabled
4025                  * in take_cpu_down(), so we prevent new cfs throttling here.
4026                  */
4027                 cfs_rq->runtime_enabled = 0;
4028 
4029                 if (cfs_rq_throttled(cfs_rq))
4030                         unthrottle_cfs_rq(cfs_rq);
4031         }
4032 }
4033 
4034 #else /* CONFIG_CFS_BANDWIDTH */
4035 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4036 {
4037         return rq_clock_task(rq_of(cfs_rq));
4038 }
4039 
4040 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4041 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4042 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4043 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4044 
4045 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4046 {
4047         return 0;
4048 }
4049 
4050 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4051 {
4052         return 0;
4053 }
4054 
4055 static inline int throttled_lb_pair(struct task_group *tg,
4056                                     int src_cpu, int dest_cpu)
4057 {
4058         return 0;
4059 }
4060 
4061 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4062 
4063 #ifdef CONFIG_FAIR_GROUP_SCHED
4064 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4065 #endif
4066 
4067 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4068 {
4069         return NULL;
4070 }
4071 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4072 static inline void update_runtime_enabled(struct rq *rq) {}
4073 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4074 
4075 #endif /* CONFIG_CFS_BANDWIDTH */
4076 
4077 /**************************************************
4078  * CFS operations on tasks:
4079  */
4080 
4081 #ifdef CONFIG_SCHED_HRTICK
4082 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4083 {
4084         struct sched_entity *se = &p->se;
4085         struct cfs_rq *cfs_rq = cfs_rq_of(se);
4086 
4087         WARN_ON(task_rq(p) != rq);
4088 
4089         if (cfs_rq->nr_running > 1) {
4090                 u64 slice = sched_slice(cfs_rq, se);
4091                 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4092                 s64 delta = slice - ran;
4093 
4094                 if (delta < 0) {
4095                         if (rq->curr == p)
4096                                 resched_curr(rq);
4097                         return;
4098                 }
4099                 hrtick_start(rq, delta);
4100         }
4101 }
4102 
4103 /*
4104  * called from enqueue/dequeue and updates the hrtick when the
4105  * current task is from our class and nr_running is low enough
4106  * to matter.
4107  */
4108 static void hrtick_update(struct rq *rq)
4109 {
4110         struct task_struct *curr = rq->curr;
4111 
4112         if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4113                 return;
4114 
4115         if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4116                 hrtick_start_fair(rq, curr);
4117 }
4118 #else /* !CONFIG_SCHED_HRTICK */
4119 static inline void
4120 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4121 {
4122 }
4123 
4124 static inline void hrtick_update(struct rq *rq)
4125 {
4126 }
4127 #endif
4128 
4129 /*
4130  * The enqueue_task method is called before nr_running is
4131  * increased. Here we update the fair scheduling stats and
4132  * then put the task into the rbtree:
4133  */
4134 static void
4135 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4136 {
4137         struct cfs_rq *cfs_rq;
4138         struct sched_entity *se = &p->se;
4139 
4140         for_each_sched_entity(se) {
4141                 if (se->on_rq)
4142                         break;
4143                 cfs_rq = cfs_rq_of(se);
4144                 enqueue_entity(cfs_rq, se, flags);
4145 
4146                 /*
4147                  * end evaluation on encountering a throttled cfs_rq
4148                  *
4149                  * note: in the case of encountering a throttled cfs_rq we will
4150                  * post the final h_nr_running increment below.
4151                 */
4152                 if (cfs_rq_throttled(cfs_rq))
4153                         break;
4154                 cfs_rq->h_nr_running++;
4155 
4156                 flags = ENQUEUE_WAKEUP;
4157         }
4158 
4159         for_each_sched_entity(se) {
4160                 cfs_rq = cfs_rq_of(se);
4161                 cfs_rq->h_nr_running++;
4162 
4163                 if (cfs_rq_throttled(cfs_rq))
4164                         break;
4165 
4166                 update_load_avg(se, 1);
4167                 update_cfs_shares(cfs_rq);
4168         }
4169 
4170         if (!se)
4171                 add_nr_running(rq, 1);
4172 
4173         hrtick_update(rq);
4174 }
4175 
4176 static void set_next_buddy(struct sched_entity *se);
4177 
4178 /*
4179  * The dequeue_task method is called before nr_running is
4180  * decreased. We remove the task from the rbtree and
4181  * update the fair scheduling stats:
4182  */
4183 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4184 {
4185         struct cfs_rq *cfs_rq;
4186         struct sched_entity *se = &p->se;
4187         int task_sleep = flags & DEQUEUE_SLEEP;
4188 
4189         for_each_sched_entity(se) {
4190                 cfs_rq = cfs_rq_of(se);
4191                 dequeue_entity(cfs_rq, se, flags);
4192 
4193                 /*
4194                  * end evaluation on encountering a throttled cfs_rq
4195                  *
4196                  * note: in the case of encountering a throttled cfs_rq we will
4197                  * post the final h_nr_running decrement below.
4198                 */
4199                 if (cfs_rq_throttled(cfs_rq))
4200                         break;
4201                 cfs_rq->h_nr_running--;
4202 
4203                 /* Don't dequeue parent if it has other entities besides us */
4204                 if (cfs_rq->load.weight) {
4205                         /*
4206                          * Bias pick_next to pick a task from this cfs_rq, as
4207                          * p is sleeping when it is within its sched_slice.
4208                          */
4209                         if (task_sleep && parent_entity(se))
4210                                 set_next_buddy(parent_entity(se));
4211 
4212                         /* avoid re-evaluating load for this entity */
4213                         se = parent_entity(se);
4214                         break;
4215                 }
4216                 flags |= DEQUEUE_SLEEP;
4217         }
4218 
4219         for_each_sched_entity(se) {
4220                 cfs_rq = cfs_rq_of(se);
4221                 cfs_rq->h_nr_running--;
4222 
4223                 if (cfs_rq_throttled(cfs_rq))
4224                         break;
4225 
4226                 update_load_avg(se, 1);
4227                 update_cfs_shares(cfs_rq);
4228         }
4229 
4230         if (!se)
4231                 sub_nr_running(rq, 1);
4232 
4233         hrtick_update(rq);
4234 }
4235 
4236 #ifdef CONFIG_SMP
4237 
4238 /*
4239  * per rq 'load' arrray crap; XXX kill this.
4240  */
4241 
4242 /*
4243  * The exact cpuload at various idx values, calculated at every tick would be
4244  * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4245  *
4246  * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4247  * on nth tick when cpu may be busy, then we have:
4248  * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4249  * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4250  *
4251  * decay_load_missed() below does efficient calculation of
4252  * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4253  * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4254  *
4255  * The calculation is approximated on a 128 point scale.
4256  * degrade_zero_ticks is the number of ticks after which load at any
4257  * particular idx is approximated to be zero.
4258  * degrade_factor is a precomputed table, a row for each load idx.
4259  * Each column corresponds to degradation factor for a power of two ticks,
4260  * based on 128 point scale.
4261  * Example:
4262  * row 2, col 3 (=12) says that the degradation at load idx 2 after
4263  * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4264  *
4265  * With this power of 2 load factors, we can degrade the load n times
4266  * by looking at 1 bits in n and doing as many mult/shift instead of
4267  * n mult/shifts needed by the exact degradation.
4268  */
4269 #define DEGRADE_SHIFT           7
4270 static const unsigned char
4271                 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4272 static const unsigned char
4273                 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4274                                         {0, 0, 0, 0, 0, 0, 0, 0},
4275                                         {64, 32, 8, 0, 0, 0, 0, 0},
4276                                         {96, 72, 40, 12, 1, 0, 0},
4277                                         {112, 98, 75, 43, 15, 1, 0},
4278                                         {120, 112, 98, 76, 45, 16, 2} };
4279 
4280 /*
4281  * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4282  * would be when CPU is idle and so we just decay the old load without
4283  * adding any new load.
4284  */
4285 static unsigned long
4286 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4287 {
4288         int j = 0;
4289 
4290         if (!missed_updates)
4291                 return load;
4292 
4293         if (missed_updates >= degrade_zero_ticks[idx])
4294                 return 0;
4295 
4296         if (idx == 1)
4297                 return load >> missed_updates;
4298 
4299         while (missed_updates) {
4300                 if (missed_updates % 2)
4301                         load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4302 
4303                 missed_updates >>= 1;
4304                 j++;
4305         }
4306         return load;
4307 }
4308 
4309 /*
4310  * Update rq->cpu_load[] statistics. This function is usually called every
4311  * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4312  * every tick. We fix it up based on jiffies.
4313  */
4314 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4315                               unsigned long pending_updates)
4316 {
4317         int i, scale;
4318 
4319         this_rq->nr_load_updates++;
4320 
4321         /* Update our load: */
4322         this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4323         for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4324                 unsigned long old_load, new_load;
4325 
4326                 /* scale is effectively 1 << i now, and >> i divides by scale */
4327 
4328                 old_load = this_rq->cpu_load[i];
4329                 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4330                 new_load = this_load;
4331                 /*
4332                  * Round up the averaging division if load is increasing. This
4333                  * prevents us from getting stuck on 9 if the load is 10, for
4334                  * example.
4335                  */
4336                 if (new_load > old_load)
4337                         new_load += scale - 1;
4338 
4339                 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4340         }
4341 
4342         sched_avg_update(this_rq);
4343 }
4344 
4345 /* Used instead of source_load when we know the type == 0 */
4346 static unsigned long weighted_cpuload(const int cpu)
4347 {
4348         return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4349 }
4350 
4351 #ifdef CONFIG_NO_HZ_COMMON
4352 /*
4353  * There is no sane way to deal with nohz on smp when using jiffies because the
4354  * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4355  * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4356  *
4357  * Therefore we cannot use the delta approach from the regular tick since that
4358  * would seriously skew the load calculation. However we'll make do for those
4359  * updates happening while idle (nohz_idle_balance) or coming out of idle
4360  * (tick_nohz_idle_exit).
4361  *
4362  * This means we might still be one tick off for nohz periods.
4363  */
4364 
4365 /*
4366  * Called from nohz_idle_balance() to update the load ratings before doing the
4367  * idle balance.
4368  */
4369 static void update_idle_cpu_load(struct rq *this_rq)
4370 {
4371         unsigned long curr_jiffies = READ_ONCE(jiffies);
4372         unsigned long load = weighted_cpuload(cpu_of(this_rq));
4373         unsigned long pending_updates;
4374 
4375         /*
4376          * bail if there's load or we're actually up-to-date.
4377          */
4378         if (load || curr_jiffies == this_rq->last_load_update_tick)
4379                 return;
4380 
4381         pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4382         this_rq->last_load_update_tick = curr_jiffies;
4383 
4384         __update_cpu_load(this_rq, load, pending_updates);
4385 }
4386 
4387 /*
4388  * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4389  */
4390 void update_cpu_load_nohz(void)
4391 {
4392         struct rq *this_rq = this_rq();
4393         unsigned long curr_jiffies = READ_ONCE(jiffies);
4394         unsigned long pending_updates;
4395 
4396         if (curr_jiffies == this_rq->last_load_update_tick)
4397                 return;
4398 
4399         raw_spin_lock(&this_rq->lock);
4400         pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4401         if (pending_updates) {
4402                 this_rq->last_load_update_tick = curr_jiffies;
4403                 /*
4404                  * We were idle, this means load 0, the current load might be
4405                  * !0 due to remote wakeups and the sort.
4406                  */
4407                 __update_cpu_load(this_rq, 0, pending_updates);
4408         }
4409         raw_spin_unlock(&this_rq->lock);
4410 }
4411 #endif /* CONFIG_NO_HZ */
4412 
4413 /*
4414  * Called from scheduler_tick()
4415  */
4416 void update_cpu_load_active(struct rq *this_rq)
4417 {
4418         unsigned long load = weighted_cpuload(cpu_of(this_rq));
4419         /*
4420          * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4421          */
4422         this_rq->last_load_update_tick = jiffies;
4423         __update_cpu_load(this_rq, load, 1);
4424 }
4425 
4426 /*
4427  * Return a low guess at the load of a migration-source cpu weighted
4428  * according to the scheduling class and "nice" value.
4429  *
4430  * We want to under-estimate the load of migration sources, to
4431  * balance conservatively.
4432  */
4433 static unsigned long source_load(int cpu, int type)
4434 {
4435         struct rq *rq = cpu_rq(cpu);
4436         unsigned long total = weighted_cpuload(cpu);
4437 
4438         if (type == 0 || !sched_feat(LB_BIAS))
4439                 return total;
4440 
4441         return min(rq->cpu_load[type-1], total);
4442 }
4443 
4444 /*
4445  * Return a high guess at the load of a migration-target cpu weighted
4446  * according to the scheduling class and "nice" value.
4447  */
4448 static unsigned long target_load(int cpu, int type)
4449 {
4450         struct rq *rq = cpu_rq(cpu);
4451         unsigned long total = weighted_cpuload(cpu);
4452 
4453         if (type == 0 || !sched_feat(LB_BIAS))
4454                 return total;
4455 
4456         return max(rq->cpu_load[type-1], total);
4457 }
4458 
4459 static unsigned long capacity_of(int cpu)
4460 {
4461         return cpu_rq(cpu)->cpu_capacity;
4462 }
4463 
4464 static unsigned long capacity_orig_of(int cpu)
4465 {
4466         return cpu_rq(cpu)->cpu_capacity_orig;
4467 }
4468 
4469 static unsigned long cpu_avg_load_per_task(int cpu)
4470 {
4471         struct rq *rq = cpu_rq(cpu);
4472         unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4473         unsigned long load_avg = weighted_cpuload(cpu);
4474 
4475         if (nr_running)
4476                 return load_avg / nr_running;
4477 
4478         return 0;
4479 }
4480 
4481 static void record_wakee(struct task_struct *p)
4482 {
4483         /*
4484          * Rough decay (wiping) for cost saving, don't worry
4485          * about the boundary, really active task won't care
4486          * about the loss.
4487          */
4488         if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4489                 current->wakee_flips >>= 1;
4490                 current->wakee_flip_decay_ts = jiffies;
4491         }
4492 
4493         if (current->last_wakee != p) {
4494                 current->last_wakee = p;
4495                 current->wakee_flips++;
4496         }
4497 }
4498 
4499 static void task_waking_fair(struct task_struct *p)
4500 {
4501         struct sched_entity *se = &p->se;
4502         struct cfs_rq *cfs_rq = cfs_rq_of(se);
4503         u64 min_vruntime;
4504 
4505 #ifndef CONFIG_64BIT
4506         u64 min_vruntime_copy;
4507 
4508         do {
4509                 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4510                 smp_rmb();
4511                 min_vruntime = cfs_rq->min_vruntime;
4512         } while (min_vruntime != min_vruntime_copy);
4513 #else
4514         min_vruntime = cfs_rq->min_vruntime;
4515 #endif
4516 
4517         se->vruntime -= min_vruntime;
4518         record_wakee(p);
4519 }
4520 
4521 #ifdef CONFIG_FAIR_GROUP_SCHED
4522 /*
4523  * effective_load() calculates the load change as seen from the root_task_group
4524  *
4525  * Adding load to a group doesn't make a group heavier, but can cause movement
4526  * of group shares between cpus. Assuming the shares were perfectly aligned one
4527  * can calculate the shift in shares.
4528  *
4529  * Calculate the effective load difference if @wl is added (subtracted) to @tg
4530  * on this @cpu and results in a total addition (subtraction) of @wg to the
4531  * total group weight.
4532  *
4533  * Given a runqueue weight distribution (rw_i) we can compute a shares
4534  * distribution (s_i) using:
4535  *
4536  *   s_i = rw_i / \Sum rw_j                                             (1)
4537  *
4538  * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4539  * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4540  * shares distribution (s_i):
4541  *
4542  *   rw_i = {   2,   4,   1,   0 }
4543  *   s_i  = { 2/7, 4/7, 1/7,   0 }
4544  *
4545  * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4546  * task used to run on and the CPU the waker is running on), we need to
4547  * compute the effect of waking a task on either CPU and, in case of a sync
4548  * wakeup, compute the effect of the current task going to sleep.
4549  *
4550  * So for a change of @wl to the local @cpu with an overall group weight change
4551  * of @wl we can compute the new shares distribution (s'_i) using:
4552  *
4553  *   s'_i = (rw_i + @wl) / (@wg + \Sum rw_j)                            (2)
4554  *
4555  * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4556  * differences in waking a task to CPU 0. The additional task changes the
4557  * weight and shares distributions like:
4558  *
4559  *   rw'_i = {   3,   4,   1,   0 }
4560  *   s'_i  = { 3/8, 4/8, 1/8,   0 }
4561  *
4562  * We can then compute the difference in effective weight by using:
4563  *
4564  *   dw_i = S * (s'_i - s_i)                                            (3)
4565  *
4566  * Where 'S' is the group weight as seen by its parent.
4567  *
4568  * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4569  * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4570  * 4/7) times the weight of the group.
4571  */
4572 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4573 {
4574         struct sched_entity *se = tg->se[cpu];
4575 
4576         if (!tg->parent)        /* the trivial, non-cgroup case */
4577                 return wl;
4578 
4579         for_each_sched_entity(se) {
4580                 long w, W;
4581 
4582                 tg = se->my_q->tg;
4583 
4584                 /*
4585                  * W = @wg + \Sum rw_j
4586                  */
4587                 W = wg + calc_tg_weight(tg, se->my_q);
4588 
4589                 /*
4590                  * w = rw_i + @wl
4591                  */
4592                 w = cfs_rq_load_avg(se->my_q) + wl;
4593 
4594                 /*
4595                  * wl = S * s'_i; see (2)
4596                  */
4597                 if (W > 0 && w < W)
4598                         wl = (w * (long)tg->shares) / W;
4599                 else
4600                         wl = tg->shares;
4601 
4602                 /*
4603                  * Per the above, wl is the new se->load.weight value; since
4604                  * those are clipped to [MIN_SHARES, ...) do so now. See
4605                  * calc_cfs_shares().
4606                  */
4607                 if (wl < MIN_SHARES)
4608                         wl = MIN_SHARES;
4609 
4610                 /*
4611                  * wl = dw_i = S * (s'_i - s_i); see (3)
4612                  */
4613                 wl -= se->avg.load_avg;
4614 
4615                 /*
4616                  * Recursively apply this logic to all parent groups to compute
4617                  * the final effective load change on the root group. Since
4618                  * only the @tg group gets extra weight, all parent groups can
4619                  * only redistribute existing shares. @wl is the shift in shares
4620                  * resulting from this level per the above.
4621                  */
4622                 wg = 0;
4623         }
4624 
4625         return wl;
4626 }
4627 #else
4628 
4629 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4630 {
4631         return wl;
4632 }
4633 
4634 #endif
4635 
4636 /*
4637  * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
4638  * A waker of many should wake a different task than the one last awakened
4639  * at a frequency roughly N times higher than one of its wakees.  In order
4640  * to determine whether we should let the load spread vs consolodating to
4641  * shared cache, we look for a minimum 'flip' frequency of llc_size in one
4642  * partner, and a factor of lls_size higher frequency in the other.  With
4643  * both conditions met, we can be relatively sure that the relationship is
4644  * non-monogamous, with partner count exceeding socket size.  Waker/wakee
4645  * being client/server, worker/dispatcher, interrupt source or whatever is
4646  * irrelevant, spread criteria is apparent partner count exceeds socket size.
4647  */
4648 static int wake_wide(struct task_struct *p)
4649 {
4650         unsigned int master = current->wakee_flips;
4651         unsigned int slave = p->wakee_flips;
4652         int factor = this_cpu_read(sd_llc_size);
4653 
4654         if (master < slave)
4655                 swap(master, slave);
4656         if (slave < factor || master < slave * factor)
4657                 return 0;
4658         return 1;
4659 }
4660 
4661 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4662 {
4663         s64 this_load, load;
4664         s64 this_eff_load, prev_eff_load;
4665         int idx, this_cpu, prev_cpu;
4666         struct task_group *tg;
4667         unsigned long weight;
4668         int balanced;
4669 
4670         idx       = sd->wake_idx;
4671         this_cpu  = smp_processor_id();
4672         prev_cpu  = task_cpu(p);
4673         load      = source_load(prev_cpu, idx);
4674         this_load = target_load(this_cpu, idx);
4675 
4676         /*
4677          * If sync wakeup then subtract the (maximum possible)
4678          * effect of the currently running task from the load
4679          * of the current CPU:
4680          */
4681         if (sync) {
4682                 tg = task_group(current);
4683                 weight = current->se.avg.load_avg;
4684 
4685                 this_load += effective_load(tg, this_cpu, -weight, -weight);
4686                 load += effective_load(tg, prev_cpu, 0, -weight);
4687         }
4688 
4689         tg = task_group(p);
4690         weight = p->se.avg.load_avg;
4691 
4692         /*
4693          * In low-load situations, where prev_cpu is idle and this_cpu is idle
4694          * due to the sync cause above having dropped this_load to 0, we'll
4695          * always have an imbalance, but there's really nothing you can do
4696          * about that, so that's good too.
4697          *
4698          * Otherwise check if either cpus are near enough in load to allow this
4699          * task to be woken on this_cpu.
4700          */
4701         this_eff_load = 100;
4702         this_eff_load *= capacity_of(prev_cpu);
4703 
4704         prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4705         prev_eff_load *= capacity_of(this_cpu);
4706 
4707         if (this_load > 0) {
4708                 this_eff_load *= this_load +
4709                         effective_load(tg, this_cpu, weight, weight);
4710 
4711                 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4712         }
4713 
4714         balanced = this_eff_load <= prev_eff_load;
4715 
4716         schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4717 
4718         if (!balanced)
4719                 return 0;
4720 
4721         schedstat_inc(sd, ttwu_move_affine);
4722         schedstat_inc(p, se.statistics.nr_wakeups_affine);
4723 
4724         return 1;
4725 }
4726 
4727 /*
4728  * find_idlest_group finds and returns the least busy CPU group within the
4729  * domain.
4730  */
4731 static struct sched_group *
4732 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4733                   int this_cpu, int sd_flag)
4734 {
4735         struct sched_group *idlest = NULL, *group = sd->groups;
4736         unsigned long min_load = ULONG_MAX, this_load = 0;
4737         int load_idx = sd->forkexec_idx;
4738         int imbalance = 100 + (sd->imbalance_pct-100)/2;
4739 
4740         if (sd_flag & SD_BALANCE_WAKE)
4741                 load_idx = sd->wake_idx;
4742 
4743         do {
4744                 unsigned long load, avg_load;
4745                 int local_group;
4746                 int i;
4747 
4748                 /* Skip over this group if it has no CPUs allowed */
4749                 if (!cpumask_intersects(sched_group_cpus(group),
4750                                         tsk_cpus_allowed(p)))
4751                         continue;
4752 
4753                 local_group = cpumask_test_cpu(this_cpu,
4754                                                sched_group_cpus(group));
4755 
4756                 /* Tally up the load of all CPUs in the group */
4757                 avg_load = 0;
4758 
4759                 for_each_cpu(i, sched_group_cpus(group)) {
4760                         /* Bias balancing toward cpus of our domain */
4761                         if (local_group)
4762                                 load = source_load(i, load_idx);
4763                         else
4764                                 load = target_load(i, load_idx);
4765 
4766                         avg_load += load;
4767                 }
4768 
4769                 /* Adjust by relative CPU capacity of the group */
4770                 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
4771 
4772                 if (local_group) {
4773                         this_load = avg_load;
4774                 } else if (avg_load < min_load) {
4775                         min_load = avg_load;
4776                         idlest = group;
4777                 }
4778         } while (group = group->next, group != sd->groups);
4779 
4780         if (!idlest || 100*this_load < imbalance*min_load)
4781                 return NULL;
4782         return idlest;
4783 }
4784 
4785 /*
4786  * find_idlest_cpu - find the idlest cpu among the cpus in group.
4787  */
4788 static int
4789 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4790 {
4791         unsigned long load, min_load = ULONG_MAX;
4792         unsigned int min_exit_latency = UINT_MAX;
4793         u64 latest_idle_timestamp = 0;
4794         int least_loaded_cpu = this_cpu;
4795         int shallowest_idle_cpu = -1;
4796         int i;
4797 
4798         /* Traverse only the allowed CPUs */
4799         for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4800                 if (idle_cpu(i)) {
4801                         struct rq *rq = cpu_rq(i);
4802                         struct cpuidle_state *idle = idle_get_state(rq);
4803                         if (idle && idle->exit_latency < min_exit_latency) {
4804                                 /*
4805                                  * We give priority to a CPU whose idle state
4806                                  * has the smallest exit latency irrespective
4807                                  * of any idle timestamp.
4808                                  */
4809                                 min_exit_latency = idle->exit_latency;
4810                                 latest_idle_timestamp = rq->idle_stamp;
4811                                 shallowest_idle_cpu = i;
4812                         } else if ((!idle || idle->exit_latency == min_exit_latency) &&
4813                                    rq->idle_stamp > latest_idle_timestamp) {
4814                                 /*
4815                                  * If equal or no active idle state, then
4816                                  * the most recently idled CPU might have
4817                                  * a warmer cache.
4818                                  */
4819                                 latest_idle_timestamp = rq->idle_stamp;
4820                                 shallowest_idle_cpu = i;
4821                         }
4822                 } else if (shallowest_idle_cpu == -1) {
4823                         load = weighted_cpuload(i);
4824                         if (load < min_load || (load == min_load && i == this_cpu)) {
4825                                 min_load = load;
4826                                 least_loaded_cpu = i;
4827                         }
4828                 }
4829         }
4830 
4831         return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
4832 }
4833 
4834 /*
4835  * Try and locate an idle CPU in the sched_domain.
4836  */
4837 static int select_idle_sibling(struct task_struct *p, int target)
4838 {
4839         struct sched_domain *sd;
4840         struct sched_group *sg;
4841         int i = task_cpu(p);
4842 
4843         if (idle_cpu(target))
4844                 return target;
4845 
4846         /*
4847          * If the prevous cpu is cache affine and idle, don't be stupid.
4848          */
4849         if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4850                 return i;
4851 
4852         /*
4853          * Otherwise, iterate the domains and find an elegible idle cpu.
4854          */
4855         sd = rcu_dereference(per_cpu(sd_llc, target));
4856         for_each_lower_domain(sd) {
4857                 sg = sd->groups;
4858                 do {
4859                         if (!cpumask_intersects(sched_group_cpus(sg),
4860                                                 tsk_cpus_allowed(p)))
4861                                 goto next;
4862 
4863                         for_each_cpu(i, sched_group_cpus(sg)) {
4864                                 if (i == target || !idle_cpu(i))
4865                                         goto next;
4866                         }
4867 
4868                         target = cpumask_first_and(sched_group_cpus(sg),
4869                                         tsk_cpus_allowed(p));
4870                         goto done;
4871 next:
4872                         sg = sg->next;
4873                 } while (sg != sd->groups);
4874         }
4875 done:
4876         return target;
4877 }
4878 
4879 /*
4880  * cpu_util returns the amount of capacity of a CPU that is used by CFS
4881  * tasks. The unit of the return value must be the one of capacity so we can
4882  * compare the utilization with the capacity of the CPU that is available for
4883  * CFS task (ie cpu_capacity).
4884  *
4885  * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
4886  * recent utilization of currently non-runnable tasks on a CPU. It represents
4887  * the amount of utilization of a CPU in the range [0..capacity_orig] where
4888  * capacity_orig is the cpu_capacity available at the highest frequency
4889  * (arch_scale_freq_capacity()).
4890  * The utilization of a CPU converges towards a sum equal to or less than the
4891  * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
4892  * the running time on this CPU scaled by capacity_curr.
4893  *
4894  * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
4895  * higher than capacity_orig because of unfortunate rounding in
4896  * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
4897  * the average stabilizes with the new running time. We need to check that the
4898  * utilization stays within the range of [0..capacity_orig] and cap it if
4899  * necessary. Without utilization capping, a group could be seen as overloaded
4900  * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
4901  * available capacity. We allow utilization to overshoot capacity_curr (but not
4902  * capacity_orig) as it useful for predicting the capacity required after task
4903  * migrations (scheduler-driven DVFS).
4904  */
4905 static int cpu_util(int cpu)
4906 {
4907         unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
4908         unsigned long capacity = capacity_orig_of(cpu);
4909 
4910         return (util >= capacity) ? capacity : util;
4911 }
4912 
4913 /*
4914  * select_task_rq_fair: Select target runqueue for the waking task in domains
4915  * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4916  * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4917  *
4918  * Balances load by selecting the idlest cpu in the idlest group, or under
4919  * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4920  *
4921  * Returns the target cpu number.
4922  *
4923  * preempt must be disabled.
4924  */
4925 static int
4926 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4927 {
4928         struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4929         int cpu = smp_processor_id();
4930         int new_cpu = prev_cpu;
4931         int want_affine = 0;
4932         int sync = wake_flags & WF_SYNC;
4933 
4934         if (sd_flag & SD_BALANCE_WAKE)
4935                 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
4936 
4937         rcu_read_lock();
4938         for_each_domain(cpu, tmp) {
4939                 if (!(tmp->flags & SD_LOAD_BALANCE))
4940                         break;
4941 
4942                 /*
4943                  * If both cpu and prev_cpu are part of this domain,
4944                  * cpu is a valid SD_WAKE_AFFINE target.
4945                  */
4946                 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4947                     cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4948                         affine_sd = tmp;
4949                         break;
4950                 }
4951 
4952                 if (tmp->flags & sd_flag)
4953                         sd = tmp;
4954                 else if (!want_affine)
4955                         break;
4956         }
4957 
4958         if (affine_sd) {
4959                 sd = NULL; /* Prefer wake_affine over balance flags */
4960                 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4961                         new_cpu = cpu;
4962         }
4963 
4964         if (!sd) {
4965                 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
4966                         new_cpu = select_idle_sibling(p, new_cpu);
4967 
4968         } else while (sd) {
4969                 struct sched_group *group;
4970                 int weight;
4971 
4972                 if (!(sd->flags & sd_flag)) {
4973                         sd = sd->child;
4974                         continue;
4975                 }
4976 
4977                 group = find_idlest_group(sd, p, cpu, sd_flag);
4978                 if (!group) {
4979                         sd = sd->child;
4980                         continue;
4981                 }
4982 
4983                 new_cpu = find_idlest_cpu(group, p, cpu);
4984                 if (new_cpu == -1 || new_cpu == cpu) {
4985                         /* Now try balancing at a lower domain level of cpu */
4986                         sd = sd->child;
4987                         continue;
4988                 }
4989 
4990                 /* Now try balancing at a lower domain level of new_cpu */
4991                 cpu = new_cpu;
4992                 weight = sd->span_weight;
4993                 sd = NULL;
4994                 for_each_domain(cpu, tmp) {
4995                         if (weight <= tmp->span_weight)
4996                                 break;
4997                         if (tmp->flags & sd_flag)
4998                                 sd = tmp;
4999                 }
5000                 /* while loop will break here if sd == NULL */
5001         }
5002         rcu_read_unlock();
5003 
5004         return new_cpu;
5005 }
5006 
5007 /*
5008  * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5009  * cfs_rq_of(p) references at time of call are still valid and identify the
5010  * previous cpu.  However, the caller only guarantees p->pi_lock is held; no
5011  * other assumptions, including the state of rq->lock, should be made.
5012  */
5013 static void migrate_task_rq_fair(struct task_struct *p)
5014 {
5015         /*
5016          * We are supposed to update the task to "current" time, then its up to date
5017          * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5018          * what current time is, so simply throw away the out-of-date time. This
5019          * will result in the wakee task is less decayed, but giving the wakee more
5020          * load sounds not bad.
5021          */
5022         remove_entity_load_avg(&p->se);
5023 
5024         /* Tell new CPU we are migrated */
5025         p->se.avg.last_update_time = 0;
5026 
5027         /* We have migrated, no longer consider this task hot */
5028         p->se.exec_start = 0;
5029 }
5030 
5031 static void task_dead_fair(struct task_struct *p)
5032 {
5033         remove_entity_load_avg(&p->se);
5034 }
5035 #endif /* CONFIG_SMP */
5036 
5037 static unsigned long
5038 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5039 {
5040         unsigned long gran = sysctl_sched_wakeup_granularity;
5041 
5042         /*
5043          * Since its curr running now, convert the gran from real-time
5044          * to virtual-time in his units.
5045          *
5046          * By using 'se' instead of 'curr' we penalize light tasks, so
5047          * they get preempted easier. That is, if 'se' < 'curr' then
5048          * the resulting gran will be larger, therefore penalizing the
5049          * lighter, if otoh 'se' > 'curr' then the resulting gran will
5050          * be smaller, again penalizing the lighter task.
5051          *
5052          * This is especially important for buddies when the leftmost
5053          * task is higher priority than the buddy.
5054          */
5055         return calc_delta_fair(gran, se);
5056 }
5057 
5058 /*
5059  * Should 'se' preempt 'curr'.
5060  *
5061  *             |s1
5062  *        |s2
5063  *   |s3
5064  *         g
5065  *      |<--->|c
5066  *
5067  *  w(c, s1) = -1
5068  *  w(c, s2) =  0
5069  *  w(c, s3) =  1
5070  *
5071  */
5072 static int
5073 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5074 {
5075         s64 gran, vdiff = curr->vruntime - se->vruntime;
5076 
5077         if (vdiff <= 0)
5078                 return -1;
5079 
5080         gran = wakeup_gran(curr, se);
5081         if (vdiff > gran)
5082                 return 1;
5083 
5084         return 0;
5085 }
5086 
5087 static void set_last_buddy(struct sched_entity *se)
5088 {
5089         if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5090                 return;
5091 
5092         for_each_sched_entity(se)
5093                 cfs_rq_of(se)->last = se;
5094 }
5095 
5096 static void set_next_buddy(struct sched_entity *se)
5097 {
5098         if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5099                 return;
5100 
5101         for_each_sched_entity(se)
5102                 cfs_rq_of(se)->next = se;
5103 }
5104 
5105 static void set_skip_buddy(struct sched_entity *se)
5106 {
5107         for_each_sched_entity(se)
5108                 cfs_rq_of(se)->skip = se;
5109 }
5110 
5111 /*
5112  * Preempt the current task with a newly woken task if needed:
5113  */
5114 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5115 {
5116         struct task_struct *curr = rq->curr;
5117         struct sched_entity *se = &curr->se, *pse = &p->se;
5118         struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5119         int scale = cfs_rq->nr_running >= sched_nr_latency;
5120         int next_buddy_marked = 0;
5121 
5122         if (unlikely(se == pse))
5123                 return;
5124 
5125         /*
5126          * This is possible from callers such as attach_tasks(), in which we
5127          * unconditionally check_prempt_curr() after an enqueue (which may have
5128          * lead to a throttle).  This both saves work and prevents false
5129          * next-buddy nomination below.
5130          */
5131         if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5132                 return;
5133 
5134         if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5135                 set_next_buddy(pse);
5136                 next_buddy_marked = 1;
5137         }
5138 
5139         /*
5140          * We can come here with TIF_NEED_RESCHED already set from new task
5141          * wake up path.
5142          *
5143          * Note: this also catches the edge-case of curr being in a throttled
5144          * group (e.g. via set_curr_task), since update_curr() (in the
5145          * enqueue of curr) will have resulted in resched being set.  This
5146          * prevents us from potentially nominating it as a false LAST_BUDDY
5147          * below.
5148          */
5149         if (test_tsk_need_resched(curr))
5150                 return;
5151 
5152         /* Idle tasks are by definition preempted by non-idle tasks. */
5153         if (unlikely(curr->policy == SCHED_IDLE) &&
5154             likely(p->policy != SCHED_IDLE))
5155                 goto preempt;
5156 
5157         /*
5158          * Batch and idle tasks do not preempt non-idle tasks (their preemption
5159          * is driven by the tick):
5160          */
5161         if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5162                 return;
5163 
5164         find_matching_se(&se, &pse);
5165         update_curr(cfs_rq_of(se));
5166         BUG_ON(!pse);
5167         if (wakeup_preempt_entity(se, pse) == 1) {
5168                 /*
5169                  * Bias pick_next to pick the sched entity that is
5170                  * triggering this preemption.
5171                  */
5172                 if (!next_buddy_marked)
5173                         set_next_buddy(pse);
5174                 goto preempt;
5175         }
5176 
5177         return;
5178 
5179 preempt:
5180         resched_curr(rq);
5181         /*
5182          * Only set the backward buddy when the current task is still
5183          * on the rq. This can happen when a wakeup gets interleaved
5184          * with schedule on the ->pre_schedule() or idle_balance()
5185          * point, either of which can * drop the rq lock.
5186          *
5187          * Also, during early boot the idle thread is in the fair class,
5188          * for obvious reasons its a bad idea to schedule back to it.
5189          */
5190         if (unlikely(!se->on_rq || curr == rq->idle))
5191                 return;
5192 
5193         if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5194                 set_last_buddy(se);
5195 }
5196 
5197 static struct task_struct *
5198 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
5199 {
5200         struct cfs_rq *cfs_rq = &rq->cfs;
5201         struct sched_entity *se;
5202         struct task_struct *p;
5203         int new_tasks;
5204 
5205 again:
5206 #ifdef CONFIG_FAIR_GROUP_SCHED
5207         if (!cfs_rq->nr_running)
5208                 goto idle;
5209 
5210         if (prev->sched_class != &fair_sched_class)
5211                 goto simple;
5212 
5213         /*
5214          * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5215          * likely that a next task is from the same cgroup as the current.
5216          *
5217          * Therefore attempt to avoid putting and setting the entire cgroup
5218          * hierarchy, only change the part that actually changes.
5219          */
5220 
5221         do {
5222                 struct sched_entity *curr = cfs_rq->curr;
5223 
5224                 /*
5225                  * Since we got here without doing put_prev_entity() we also
5226                  * have to consider cfs_rq->curr. If it is still a runnable
5227                  * entity, update_curr() will update its vruntime, otherwise
5228                  * forget we've ever seen it.
5229                  */
5230                 if (curr) {
5231                         if (curr->on_rq)
5232                                 update_curr(cfs_rq);
5233                         else
5234                                 curr = NULL;
5235 
5236                         /*
5237                          * This call to check_cfs_rq_runtime() will do the
5238                          * throttle and dequeue its entity in the parent(s).
5239                          * Therefore the 'simple' nr_running test will indeed
5240                          * be correct.
5241                          */
5242                         if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5243                                 goto simple;
5244                 }
5245 
5246                 se = pick_next_entity(cfs_rq, curr);
5247                 cfs_rq = group_cfs_rq(se);
5248         } while (cfs_rq);
5249 
5250         p = task_of(se);
5251 
5252         /*
5253          * Since we haven't yet done put_prev_entity and if the selected task
5254          * is a different task than we started out with, try and touch the
5255          * least amount of cfs_rqs.
5256          */
5257         if (prev != p) {
5258                 struct sched_entity *pse = &prev->se;
5259 
5260                 while (!(cfs_rq = is_same_group(se, pse))) {
5261                         int se_depth = se->depth;
5262                         int pse_depth = pse->depth;
5263 
5264                         if (se_depth <= pse_depth) {
5265                                 put_prev_entity(cfs_rq_of(pse), pse);
5266                                 pse = parent_entity(pse);
5267                         }
5268                         if (se_depth >= pse_depth) {
5269                                 set_next_entity(cfs_rq_of(se), se);
5270                                 se = parent_entity(se);
5271                         }
5272                 }
5273 
5274                 put_prev_entity(cfs_rq, pse);
5275                 set_next_entity(cfs_rq, se);
5276         }
5277 
5278         if (hrtick_enabled(rq))
5279                 hrtick_start_fair(rq, p);
5280 
5281         return p;
5282 simple:
5283         cfs_rq = &rq->cfs;
5284 #endif
5285 
5286         if (!cfs_rq->nr_running)
5287                 goto idle;
5288 
5289         put_prev_task(rq, prev);
5290 
5291         do {
5292                 se = pick_next_entity(cfs_rq, NULL);
5293                 set_next_entity(cfs_rq, se);
5294                 cfs_rq = group_cfs_rq(se);
5295         } while (cfs_rq);
5296 
5297         p = task_of(se);
5298 
5299         if (hrtick_enabled(rq))
5300                 hrtick_start_fair(rq, p);
5301 
5302         return p;
5303 
5304 idle:
5305         /*
5306          * This is OK, because current is on_cpu, which avoids it being picked
5307          * for load-balance and preemption/IRQs are still disabled avoiding
5308          * further scheduler activity on it and we're being very careful to
5309          * re-start the picking loop.
5310          */
5311         lockdep_unpin_lock(&rq->lock);
5312         new_tasks = idle_balance(rq);
5313         lockdep_pin_lock(&rq->lock);
5314         /*
5315          * Because idle_balance() releases (and re-acquires) rq->lock, it is
5316          * possible for any higher priority task to appear. In that case we
5317          * must re-start the pick_next_entity() loop.
5318          */
5319         if (new_tasks < 0)
5320                 return RETRY_TASK;
5321 
5322         if (new_tasks > 0)
5323                 goto again;
5324 
5325         return NULL;
5326 }
5327 
5328 /*
5329  * Account for a descheduled task:
5330  */
5331 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5332 {
5333         struct sched_entity *se = &prev->se;
5334         struct cfs_rq *cfs_rq;
5335 
5336         for_each_sched_entity(se) {
5337                 cfs_rq = cfs_rq_of(se);
5338                 put_prev_entity(cfs_rq, se);
5339         }
5340 }
5341 
5342 /*
5343  * sched_yield() is very simple
5344  *
5345  * The magic of dealing with the ->skip buddy is in pick_next_entity.
5346  */
5347 static void yield_task_fair(struct rq *rq)
5348 {
5349         struct task_struct *curr = rq->curr;
5350         struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5351         struct sched_entity *se = &curr->se;
5352 
5353         /*
5354          * Are we the only task in the tree?
5355          */
5356         if (unlikely(rq->nr_running == 1))
5357                 return;
5358 
5359         clear_buddies(cfs_rq, se);
5360 
5361         if (curr->policy != SCHED_BATCH) {
5362                 update_rq_clock(rq);
5363                 /*
5364                  * Update run-time statistics of the 'current'.
5365                  */
5366                 update_curr(cfs_rq);
5367                 /*
5368                  * Tell update_rq_clock() that we've just updated,
5369                  * so we don't do microscopic update in schedule()
5370                  * and double the fastpath cost.
5371                  */
5372                 rq_clock_skip_update(rq, true);
5373         }
5374 
5375         set_skip_buddy(se);
5376 }
5377 
5378 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5379 {
5380         struct sched_entity *se = &p->se;
5381 
5382         /* throttled hierarchies are not runnable */
5383         if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5384                 return false;
5385 
5386         /* Tell the scheduler that we'd really like pse to run next. */
5387         set_next_buddy(se);
5388 
5389         yield_task_fair(rq);
5390 
5391         return true;
5392 }
5393 
5394 #ifdef CONFIG_SMP
5395 /**************************************************
5396  * Fair scheduling class load-balancing methods.
5397  *
5398  * BASICS
5399  *
5400  * The purpose of load-balancing is to achieve the same basic fairness the
5401  * per-cpu scheduler provides, namely provide a proportional amount of compute
5402  * time to each task. This is expressed in the following equation:
5403  *
5404  *   W_i,n/P_i == W_j,n/P_j for all i,j                               (1)
5405  *
5406  * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5407  * W_i,0 is defined as:
5408  *
5409  *   W_i,0 = \Sum_j w_i,j                                             (2)
5410  *
5411  * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5412  * is derived from the nice value as per prio_to_weight[].
5413  *
5414  * The weight average is an exponential decay average of the instantaneous
5415  * weight:
5416  *
5417  *   W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0               (3)
5418  *
5419  * C_i is the compute capacity of cpu i, typically it is the
5420  * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5421  * can also include other factors [XXX].
5422  *
5423  * To achieve this balance we define a measure of imbalance which follows
5424  * directly from (1):
5425  *
5426  *   imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j }    (4)
5427  *
5428  * We them move tasks around to minimize the imbalance. In the continuous
5429  * function space it is obvious this converges, in the discrete case we get
5430  * a few fun cases generally called infeasible weight scenarios.
5431  *
5432  * [XXX expand on:
5433  *     - infeasible weights;
5434  *     - local vs global optima in the discrete case. ]
5435  *
5436  *
5437  * SCHED DOMAINS
5438  *
5439  * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5440  * for all i,j solution, we create a tree of cpus that follows the hardware
5441  * topology where each level pairs two lower groups (or better). This results
5442  * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5443  * tree to only the first of the previous level and we decrease the frequency
5444  * of load-balance at each level inv. proportional to the number of cpus in
5445  * the groups.
5446  *
5447  * This yields:
5448  *
5449  *     log_2 n     1     n
5450  *   \Sum       { --- * --- * 2^i } = O(n)                            (5)
5451  *     i = 0      2^i   2^i
5452  *                               `- size of each group
5453  *         |         |     `- number of cpus doing load-balance
5454  *         |         `- freq
5455  *         `- sum over all levels
5456  *
5457  * Coupled with a limit on how many tasks we can migrate every balance pass,
5458  * this makes (5) the runtime complexity of the balancer.
5459  *
5460  * An important property here is that each CPU is still (indirectly) connected
5461  * to every other cpu in at most O(log n) steps:
5462  *
5463  * The adjacency matrix of the resulting graph is given by:
5464  *
5465  *             log_2 n     
5466  *   A_i,j = \Union     (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1)  (6)
5467  *             k = 0
5468  *
5469  * And you'll find that:
5470  *
5471  *   A^(log_2 n)_i,j != 0  for all i,j                                (7)
5472  *
5473  * Showing there's indeed a path between every cpu in at most O(log n) steps.
5474  * The task movement gives a factor of O(m), giving a convergence complexity
5475  * of:
5476  *
5477  *   O(nm log n),  n := nr_cpus, m := nr_tasks                        (8)
5478  *
5479  *
5480  * WORK CONSERVING
5481  *
5482  * In order to avoid CPUs going idle while there's still work to do, new idle
5483  * balancing is more aggressive and has the newly idle cpu iterate up the domain
5484  * tree itself instead of relying on other CPUs to bring it work.
5485  *
5486  * This adds some complexity to both (5) and (8) but it reduces the total idle
5487  * time.
5488  *
5489  * [XXX more?]
5490  *
5491  *
5492  * CGROUPS
5493  *
5494  * Cgroups make a horror show out of (2), instead of a simple sum we get:
5495  *
5496  *                                s_k,i
5497  *   W_i,0 = \Sum_j \Prod_k w_k * -----                               (9)
5498  *                                 S_k
5499  *
5500  * Where
5501  *
5502  *   s_k,i = \Sum_j w_i,j,k  and  S_k = \Sum_i s_k,i                 (10)
5503  *
5504  * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5505  *
5506  * The big problem is S_k, its a global sum needed to compute a local (W_i)
5507  * property.
5508  *
5509  * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5510  *      rewrite all of this once again.]
5511  */ 
5512 
5513 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5514 
5515 enum fbq_type { regular, remote, all };
5516 
5517 #define LBF_ALL_PINNED  0x01
5518 #define LBF_NEED_BREAK  0x02
5519 #define LBF_DST_PINNED  0x04
5520 #define LBF_SOME_PINNED 0x08
5521 
5522 struct lb_env {
5523         struct sched_domain     *sd;
5524 
5525         struct rq               *src_rq;
5526         int                     src_cpu;
5527 
5528         int                     dst_cpu;
5529         struct rq               *dst_rq;
5530 
5531         struct cpumask          *dst_grpmask;
5532         int                     new_dst_cpu;
5533         enum cpu_idle_type      idle;
5534         long                    imbalance;
5535         /* The set of CPUs under consideration for load-balancing */
5536         struct cpumask          *cpus;
5537 
5538         unsigned int            flags;
5539 
5540         unsigned int            loop;
5541         unsigned int            loop_break;
5542         unsigned int            loop_max;
5543 
5544         enum fbq_type           fbq_type;
5545         struct list_head        tasks;
5546 };
5547 
5548 /*
5549  * Is this task likely cache-hot:
5550  */
5551 static int task_hot(struct task_struct *p, struct lb_env *env)
5552 {
5553         s64 delta;
5554 
5555         lockdep_assert_held(&env->src_rq->lock);
5556 
5557         if (p->sched_class != &fair_sched_class)
5558                 return 0;
5559 
5560         if (unlikely(p->policy == SCHED_IDLE))
5561                 return 0;
5562 
5563         /*
5564          * Buddy candidates are cache hot:
5565          */
5566         if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
5567                         (&p->se == cfs_rq_of(&p->se)->next ||
5568                          &p->se == cfs_rq_of(&p->se)->last))
5569                 return 1;
5570 
5571         if (sysctl_sched_migration_cost == -1)
5572                 return 1;
5573         if (sysctl_sched_migration_cost == 0)
5574                 return 0;
5575 
5576         delta = rq_clock_task(env->src_rq) - p->se.exec_start;
5577 
5578         return delta < (s64)sysctl_sched_migration_cost;
5579 }
5580 
5581 #ifdef CONFIG_NUMA_BALANCING
5582 /*
5583  * Returns 1, if task migration degrades locality
5584  * Returns 0, if task migration improves locality i.e migration preferred.
5585  * Returns -1, if task migration is not affected by locality.
5586  */
5587 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5588 {
5589         struct numa_group *numa_group = rcu_dereference(p->numa_group);
5590         unsigned long src_faults, dst_faults;
5591         int src_nid, dst_nid;
5592 
5593         if (!static_branch_likely(&sched_numa_balancing))
5594                 return -1;
5595 
5596         if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
5597                 return -1;
5598 
5599         src_nid = cpu_to_node(env->src_cpu);
5600         dst_nid = cpu_to_node(env->dst_cpu);
5601 
5602         if (src_nid == dst_nid)
5603                 return -1;
5604 
5605         /* Migrating away from the preferred node is always bad. */
5606         if (src_nid == p->numa_preferred_nid) {
5607                 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
5608                         return 1;
5609                 else
5610                         return -1;
5611         }
5612 
5613         /* Encourage migration to the preferred node. */
5614         if (dst_nid == p->numa_preferred_nid)
5615                 return 0;
5616 
5617         if (numa_group) {
5618                 src_faults = group_faults(p, src_nid);
5619                 dst_faults = group_faults(p, dst_nid);
5620         } else {
5621                 src_faults = task_faults(p, src_nid);
5622                 dst_faults = task_faults(p, dst_nid);
5623         }
5624 
5625         return dst_faults < src_faults;
5626 }
5627 
5628 #else
5629 static inline int migrate_degrades_locality(struct task_struct *p,
5630                                              struct lb_env *env)
5631 {
5632         return -1;
5633 }
5634 #endif
5635 
5636 /*
5637  * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5638  */
5639 static
5640 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5641 {
5642         int tsk_cache_hot;
5643 
5644         lockdep_assert_held(&env->src_rq->lock);
5645 
5646         /*
5647          * We do not migrate tasks that are:
5648          * 1) throttled_lb_pair, or
5649          * 2) cannot be migrated to this CPU due to cpus_allowed, or
5650          * 3) running (obviously), or
5651          * 4) are cache-hot on their current CPU.
5652          */
5653         if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5654                 return 0;
5655 
5656         if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5657                 int cpu;
5658 
5659                 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5660 
5661                 env->flags |= LBF_SOME_PINNED;
5662 
5663                 /*
5664                  * Remember if this task can be migrated to any other cpu in
5665                  * our sched_group. We may want to revisit it if we couldn't
5666                  * meet load balance goals by pulling other tasks on src_cpu.
5667                  *
5668                  * Also avoid computing new_dst_cpu if we have already computed
5669                  * one in current iteration.
5670                  */
5671                 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5672                         return 0;
5673 
5674                 /* Prevent to re-select dst_cpu via env's cpus */
5675                 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5676                         if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5677                                 env->flags |= LBF_DST_PINNED;
5678                                 env->new_dst_cpu = cpu;
5679                                 break;
5680                         }
5681                 }
5682 
5683                 return 0;
5684         }
5685 
5686         /* Record that we found atleast one task that could run on dst_cpu */
5687         env->flags &= ~LBF_ALL_PINNED;
5688 
5689         if (task_running(env->src_rq, p)) {
5690                 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5691                 return 0;
5692         }
5693 
5694         /*
5695          * Aggressive migration if:
5696          * 1) destination numa is preferred
5697          * 2) task is cache cold, or
5698          * 3) too many balance attempts have failed.
5699          */
5700         tsk_cache_hot = migrate_degrades_locality(p, env);
5701         if (tsk_cache_hot == -1)
5702                 tsk_cache_hot = task_hot(p, env);
5703 
5704         if (tsk_cache_hot <= 0 ||
5705             env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5706                 if (tsk_cache_hot == 1) {
5707                         schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5708                         schedstat_inc(p, se.statistics.nr_forced_migrations);
5709                 }
5710                 return 1;
5711         }
5712 
5713         schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5714         return 0;
5715 }
5716 
5717 /*
5718  * detach_task() -- detach the task for the migration specified in env
5719  */
5720 static void detach_task(struct task_struct *p, struct lb_env *env)
5721 {
5722         lockdep_assert_held(&env->src_rq->lock);
5723 
5724         deactivate_task(env->src_rq, p, 0);
5725         p->on_rq = TASK_ON_RQ_MIGRATING;
5726         set_task_cpu(p, env->dst_cpu);
5727 }
5728 
5729 /*
5730  * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5731  * part of active balancing operations within "domain".
5732  *
5733  * Returns a task if successful and NULL otherwise.
5734  */
5735 static struct task_struct *detach_one_task(struct lb_env *env)
5736 {
5737         struct task_struct *p, *n;
5738 
5739         lockdep_assert_held(&env->src_rq->lock);
5740 
5741         list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5742                 if (!can_migrate_task(p, env))
5743                         continue;
5744 
5745                 detach_task(p, env);
5746 
5747                 /*
5748                  * Right now, this is only the second place where
5749                  * lb_gained[env->idle] is updated (other is detach_tasks)
5750                  * so we can safely collect stats here rather than
5751                  * inside detach_tasks().
5752                  */
5753                 schedstat_inc(env->sd, lb_gained[env->idle]);
5754                 return p;
5755         }
5756         return NULL;
5757 }
5758 
5759 static const unsigned int sched_nr_migrate_break = 32;
5760 
5761 /*
5762  * detach_tasks() -- tries to detach up to imbalance weighted load from
5763  * busiest_rq, as part of a balancing operation within domain "sd".
5764  *
5765  * Returns number of detached tasks if successful and 0 otherwise.
5766  */
5767 static int detach_tasks(struct lb_env *env)
5768 {
5769         struct list_head *tasks = &env->src_rq->cfs_tasks;
5770         struct task_struct *p;
5771         unsigned long load;
5772         int detached = 0;
5773 
5774         lockdep_assert_held(&env->src_rq->lock);
5775 
5776         if (env->imbalance <= 0)
5777                 return 0;
5778 
5779         while (!list_empty(tasks)) {
5780                 /*
5781                  * We don't want to steal all, otherwise we may be treated likewise,
5782                  * which could at worst lead to a livelock crash.
5783                  */
5784                 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
5785                         break;
5786 
5787                 p = list_first_entry(tasks, struct task_struct, se.group_node);
5788 
5789                 env->loop++;
5790                 /* We've more or less seen every task there is, call it quits */
5791                 if (env->loop > env->loop_max)
5792                         break;
5793 
5794                 /* take a breather every nr_migrate tasks */
5795                 if (env->loop > env->loop_break) {
5796                         env->loop_break += sched_nr_migrate_break;
5797                         env->flags |= LBF_NEED_BREAK;
5798                         break;
5799                 }
5800 
5801                 if (!can_migrate_task(p, env))
5802                         goto next;
5803 
5804                 load = task_h_load(p);
5805 
5806                 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5807                         goto next;
5808 
5809                 if ((load / 2) > env->imbalance)
5810                         goto next;
5811 
5812                 detach_task(p, env);
5813                 list_add(&p->se.group_node, &env->tasks);
5814 
5815                 detached++;
5816                 env->imbalance -= load;
5817 
5818 #ifdef CONFIG_PREEMPT
5819                 /*
5820                  * NEWIDLE balancing is a source of latency, so preemptible
5821                  * kernels will stop after the first task is detached to minimize
5822                  * the critical section.
5823                  */
5824                 if (env->idle == CPU_NEWLY_IDLE)
5825                         break;
5826 #endif
5827 
5828                 /*
5829                  * We only want to steal up to the prescribed amount of
5830                  * weighted load.
5831                  */
5832                 if (env->imbalance <= 0)
5833                         break;
5834 
5835                 continue;
5836 next:
5837                 list_move_tail(&p->se.group_node, tasks);
5838         }
5839 
5840         /*
5841          * Right now, this is one of only two places we collect this stat
5842          * so we can safely collect detach_one_task() stats here rather
5843          * than inside detach_one_task().
5844          */
5845         schedstat_add(env->sd, lb_gained[env->idle], detached);
5846 
5847         return detached;
5848 }
5849 
5850 /*
5851  * attach_task() -- attach the task detached by detach_task() to its new rq.
5852  */
5853 static void attach_task(struct rq *rq, struct task_struct *p)
5854 {
5855         lockdep_assert_held(&rq->lock);
5856 
5857         BUG_ON(task_rq(p) != rq);
5858         p->on_rq = TASK_ON_RQ_QUEUED;
5859         activate_task(rq, p, 0);
5860         check_preempt_curr(rq, p, 0);
5861 }
5862 
5863 /*
5864  * attach_one_task() -- attaches the task returned from detach_one_task() to
5865  * its new rq.
5866  */
5867 static void attach_one_task(struct rq *rq, struct task_struct *p)
5868 {
5869         raw_spin_lock(&rq->lock);
5870         attach_task(rq, p);
5871         raw_spin_unlock(&rq->lock);
5872 }
5873 
5874 /*
5875  * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
5876  * new rq.
5877  */
5878 static void attach_tasks(struct lb_env *env)
5879 {
5880         struct list_head *tasks = &env->tasks;
5881         struct task_struct *p;
5882 
5883         raw_spin_lock(&env->dst_rq->lock);
5884 
5885         while (!list_empty(tasks)) {
5886                 p = list_first_entry(tasks, struct task_struct, se.group_node);
5887                 list_del_init(&p->se.group_node);
5888 
5889                 attach_task(env->dst_rq, p);
5890         }
5891 
5892         raw_spin_unlock(&env->dst_rq->lock);
5893 }
5894 
5895 #ifdef CONFIG_FAIR_GROUP_SCHED
5896 static void update_blocked_averages(int cpu)
5897 {
5898         struct rq *rq = cpu_rq(cpu);
5899         struct cfs_rq *cfs_rq;
5900         unsigned long flags;
5901 
5902         raw_spin_lock_irqsave(&rq->lock, flags);
5903         update_rq_clock(rq);
5904 
5905         /*
5906          * Iterates the task_group tree in a bottom up fashion, see
5907          * list_add_leaf_cfs_rq() for details.
5908          */
5909         for_each_leaf_cfs_rq(rq, cfs_rq) {
5910                 /* throttled entities do not contribute to load */
5911                 if (throttled_hierarchy(cfs_rq))
5912                         continue;
5913 
5914                 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
5915                         update_tg_load_avg(cfs_rq, 0);
5916         }
5917         raw_spin_unlock_irqrestore(&rq->lock, flags);
5918 }
5919 
5920 /*
5921  * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5922  * This needs to be done in a top-down fashion because the load of a child
5923  * group is a fraction of its parents load.
5924  */
5925 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5926 {
5927         struct rq *rq = rq_of(cfs_rq);
5928         struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5929         unsigned long now = jiffies;
5930         unsigned long load;
5931 
5932         if (cfs_rq->last_h_load_update == now)
5933                 return;
5934 
5935         cfs_rq->h_load_next = NULL;
5936         for_each_sched_entity(se) {
5937                 cfs_rq = cfs_rq_of(se);
5938                 cfs_rq->h_load_next = se;
5939                 if (cfs_rq->last_h_load_update == now)
5940                         break;
5941         }
5942 
5943         if (!se) {
5944                 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
5945                 cfs_rq->last_h_load_update = now;
5946         }
5947 
5948         while ((se = cfs_rq->h_load_next) != NULL) {
5949                 load = cfs_rq->h_load;
5950                 load = div64_ul(load * se->avg.load_avg,
5951                         cfs_rq_load_avg(cfs_rq) + 1);
5952                 cfs_rq = group_cfs_rq(se);
5953                 cfs_rq->h_load = load;
5954                 cfs_rq->last_h_load_update = now;
5955         }
5956 }
5957 
5958 static unsigned long task_h_load(struct task_struct *p)
5959 {
5960         struct cfs_rq *cfs_rq = task_cfs_rq(p);
5961 
5962         update_cfs_rq_h_load(cfs_rq);
5963         return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
5964                         cfs_rq_load_avg(cfs_rq) + 1);
5965 }
5966 #else
5967 static inline void update_blocked_averages(int cpu)
5968 {
5969         struct rq *rq = cpu_rq(cpu);
5970         struct cfs_rq *cfs_rq = &rq->cfs;
5971         unsigned long flags;
5972 
5973         raw_spin_lock_irqsave(&rq->lock, flags);
5974         update_rq_clock(rq);
5975         update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
5976         raw_spin_unlock_irqrestore(&rq->lock, flags);
5977 }
5978 
5979 static unsigned long task_h_load(struct task_struct *p)
5980 {
5981         return p->se.avg.load_avg;
5982 }
5983 #endif
5984 
5985 /********** Helpers for find_busiest_group ************************/
5986 
5987 enum group_type {
5988         group_other = 0,
5989         group_imbalanced,
5990         group_overloaded,
5991 };
5992 
5993 /*
5994  * sg_lb_stats - stats of a sched_group required for load_balancing
5995  */
5996 struct sg_lb_stats {
5997         unsigned long avg_load; /*Avg load across the CPUs of the group */
5998         unsigned long group_load; /* Total load over the CPUs of the group */
5999         unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6000         unsigned long load_per_task;
6001         unsigned long group_capacity;
6002         unsigned long group_util; /* Total utilization of the group */
6003         unsigned int sum_nr_running; /* Nr tasks running in the group */
6004         unsigned int idle_cpus;
6005         unsigned int group_weight;
6006         enum group_type group_type;
6007         int group_no_capacity;
6008 #ifdef CONFIG_NUMA_BALANCING
6009         unsigned int nr_numa_running;
6010         unsigned int nr_preferred_running;
6011 #endif
6012 };
6013 
6014 /*
6015  * sd_lb_stats - Structure to store the statistics of a sched_domain
6016  *               during load balancing.
6017  */
6018 struct sd_lb_stats {
6019         struct sched_group *busiest;    /* Busiest group in this sd */
6020         struct sched_group *local;      /* Local group in this sd */
6021         unsigned long total_load;       /* Total load of all groups in sd */
6022         unsigned long total_capacity;   /* Total capacity of all groups in sd */
6023         unsigned long avg_load; /* Average load across all groups in sd */
6024 
6025         struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6026         struct sg_lb_stats local_stat;  /* Statistics of the local group */
6027 };
6028 
6029 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6030 {
6031         /*
6032          * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6033          * local_stat because update_sg_lb_stats() does a full clear/assignment.
6034          * We must however clear busiest_stat::avg_load because
6035          * update_sd_pick_busiest() reads this before assignment.
6036          */
6037         *sds = (struct sd_lb_stats){
6038                 .busiest = NULL,
6039                 .local = NULL,
6040                 .total_load = 0UL,
6041                 .total_capacity = 0UL,
6042                 .busiest_stat = {
6043                         .avg_load = 0UL,
6044                         .sum_nr_running = 0,
6045                         .group_type = group_other,
6046                 },
6047         };
6048 }
6049 
6050 /**
6051  * get_sd_load_idx - Obtain the load index for a given sched domain.
6052  * @sd: The sched_domain whose load_idx is to be obtained.
6053  * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6054  *
6055  * Return: The load index.
6056  */
6057 static inline int get_sd_load_idx(struct sched_domain *sd,
6058                                         enum cpu_idle_type idle)
6059 {
6060         int load_idx;
6061 
6062         switch (idle) {
6063         case CPU_NOT_IDLE:
6064                 load_idx = sd->busy_idx;
6065                 break;
6066 
6067         case CPU_NEWLY_IDLE:
6068                 load_idx = sd->newidle_idx;
6069                 break;
6070         default:
6071                 load_idx = sd->idle_idx;
6072                 break;
6073         }
6074 
6075         return load_idx;
6076 }
6077 
6078 static unsigned long scale_rt_capacity(int cpu)
6079 {
6080         struct rq *rq = cpu_rq(cpu);
6081         u64 total, used, age_stamp, avg;
6082         s64 delta;
6083 
6084         /*
6085          * Since we're reading these variables without serialization make sure
6086          * we read them once before doing sanity checks on them.
6087          */
6088         age_stamp = READ_ONCE(rq->age_stamp);
6089         avg = READ_ONCE(rq->rt_avg);
6090         delta = __rq_clock_broken(rq) - age_stamp;
6091 
6092         if (unlikely(delta < 0))
6093                 delta = 0;
6094 
6095         total = sched_avg_period() + delta;
6096 
6097         used = div_u64(avg, total);
6098 
6099         if (likely(used < SCHED_CAPACITY_SCALE))
6100                 return SCHED_CAPACITY_SCALE - used;
6101 
6102         return 1;
6103 }
6104 
6105 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6106 {
6107         unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6108         struct sched_group *sdg = sd->groups;
6109 
6110         cpu_rq(cpu)->cpu_capacity_orig = capacity;
6111 
6112         capacity *= scale_rt_capacity(cpu);
6113         capacity >>= SCHED_CAPACITY_SHIFT;
6114 
6115         if (!capacity)
6116                 capacity = 1;
6117 
6118         cpu_rq(cpu)->cpu_capacity = capacity;
6119         sdg->sgc->capacity = capacity;
6120 }
6121 
6122 void update_group_capacity(struct sched_domain *sd, int cpu)
6123 {
6124         struct sched_domain *child = sd->child;
6125         struct sched_group *group, *sdg = sd->groups;
6126         unsigned long capacity;
6127         unsigned long interval;
6128 
6129         interval = msecs_to_jiffies(sd->balance_interval);
6130         interval = clamp(interval, 1UL, max_load_balance_interval);
6131         sdg->sgc->next_update = jiffies + interval;
6132 
6133         if (!child) {
6134                 update_cpu_capacity(sd, cpu);
6135                 return;
6136         }
6137 
6138         capacity = 0;
6139 
6140         if (child->flags & SD_OVERLAP) {
6141                 /*
6142                  * SD_OVERLAP domains cannot assume that child groups
6143                  * span the current group.
6144                  */
6145 
6146                 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6147                         struct sched_group_capacity *sgc;
6148                         struct rq *rq = cpu_rq(cpu);
6149 
6150                         /*
6151                          * build_sched_domains() -> init_sched_groups_capacity()
6152                          * gets here before we've attached the domains to the
6153                          * runqueues.
6154                          *
6155                          * Use capacity_of(), which is set irrespective of domains
6156                          * in update_cpu_capacity().
6157                          *
6158                          * This avoids capacity from being 0 and
6159                          * causing divide-by-zero issues on boot.
6160                          */
6161                         if (unlikely(!rq->sd)) {
6162                                 capacity += capacity_of(cpu);
6163                                 continue;
6164                         }
6165 
6166                         sgc = rq->sd->groups->sgc;
6167                         capacity += sgc->capacity;
6168                 }
6169         } else  {
6170                 /*
6171                  * !SD_OVERLAP domains can assume that child groups
6172                  * span the current group.
6173                  */ 
6174 
6175                 group = child->groups;
6176                 do {
6177                         capacity += group->sgc->capacity;
6178                         group = group->next;
6179                 } while (group != child->groups);
6180         }
6181 
6182         sdg->sgc->capacity = capacity;
6183 }
6184 
6185 /*
6186  * Check whether the capacity of the rq has been noticeably reduced by side
6187  * activity. The imbalance_pct is used for the threshold.
6188  * Return true is the capacity is reduced
6189  */
6190 static inline int
6191 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6192 {
6193         return ((rq->cpu_capacity * sd->imbalance_pct) <
6194                                 (rq->cpu_capacity_orig * 100));
6195 }
6196 
6197 /*
6198  * Group imbalance indicates (and tries to solve) the problem where balancing
6199  * groups is inadequate due to tsk_cpus_allowed() constraints.
6200  *
6201  * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6202  * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6203  * Something like:
6204  *
6205  *      { 0 1 2 3 } { 4 5 6 7 }
6206  *              *     * * *
6207  *
6208  * If we were to balance group-wise we'd place two tasks in the first group and
6209  * two tasks in the second group. Clearly this is undesired as it will overload
6210  * cpu 3 and leave one of the cpus in the second group unused.
6211  *
6212  * The current solution to this issue is detecting the skew in the first group
6213  * by noticing the lower domain failed to reach balance and had difficulty
6214  * moving tasks due to affinity constraints.
6215  *
6216  * When this is so detected; this group becomes a candidate for busiest; see
6217  * update_sd_pick_busiest(). And calculate_imbalance() and
6218  * find_busiest_group() avoid some of the usual balance conditions to allow it
6219  * to create an effective group imbalance.
6220  *
6221  * This is a somewhat tricky proposition since the next run might not find the
6222  * group imbalance and decide the groups need to be balanced again. A most
6223  * subtle and fragile situation.
6224  */
6225 
6226 static inline int sg_imbalanced(struct sched_group *group)
6227 {
6228         return group->sgc->imbalance;
6229 }
6230 
6231 /*
6232  * group_has_capacity returns true if the group has spare capacity that could
6233  * be used by some tasks.
6234  * We consider that a group has spare capacity if the  * number of task is
6235  * smaller than the number of CPUs or if the utilization is lower than the
6236  * available capacity for CFS tasks.
6237  * For the latter, we use a threshold to stabilize the state, to take into
6238  * account the variance of the tasks' load and to return true if the available
6239  * capacity in meaningful for the load balancer.
6240  * As an example, an available capacity of 1% can appear but it doesn't make
6241  * any benefit for the load balance.
6242  */
6243 static inline bool
6244 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6245 {
6246         if (sgs->sum_nr_running < sgs->group_weight)
6247                 return true;
6248 
6249         if ((sgs->group_capacity * 100) >
6250                         (sgs->group_util * env->sd->imbalance_pct))
6251                 return true;
6252 
6253         return false;
6254 }
6255 
6256 /*
6257  *  group_is_overloaded returns true if the group has more tasks than it can
6258  *  handle.
6259  *  group_is_overloaded is not equals to !group_has_capacity because a group
6260  *  with the exact right number of tasks, has no more spare capacity but is not
6261  *  overloaded so both group_has_capacity and group_is_overloaded return
6262  *  false.
6263  */
6264 static inline bool
6265 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6266 {
6267         if (sgs->sum_nr_running <= sgs->group_weight)
6268                 return false;
6269 
6270         if ((sgs->group_capacity * 100) <
6271                         (sgs->group_util * env->sd->imbalance_pct))
6272                 return true;
6273 
6274         return false;
6275 }
6276 
6277 static inline enum
6278 group_type group_classify(struct sched_group *group,
6279                           struct sg_lb_stats *sgs)
6280 {
6281         if (sgs->group_no_capacity)
6282                 return group_overloaded;
6283 
6284         if (sg_imbalanced(group))
6285                 return group_imbalanced;
6286 
6287         return group_other;
6288 }
6289 
6290 /**
6291  * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6292  * @env: The load balancing environment.
6293  * @group: sched_group whose statistics are to be updated.
6294  * @load_idx: Load index of sched_domain of this_cpu for load calc.
6295  * @local_group: Does group contain this_cpu.
6296  * @sgs: variable to hold the statistics for this group.
6297  * @overload: Indicate more than one runnable task for any CPU.
6298  */
6299 static inline void update_sg_lb_stats(struct lb_env *env,
6300                         struct sched_group *group, int load_idx,
6301                         int local_group, struct sg_lb_stats *sgs,
6302                         bool *overload)
6303 {
6304         unsigned long load;
6305         int i;
6306 
6307         memset(sgs, 0, sizeof(*sgs));
6308 
6309         for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6310                 struct rq *rq = cpu_rq(i);
6311 
6312                 /* Bias balancing toward cpus of our domain */
6313                 if (local_group)
6314                         load = target_load(i, load_idx);
6315                 else
6316                         load = source_load(i, load_idx);
6317 
6318                 sgs->group_load += load;
6319                 sgs->group_util += cpu_util(i);
6320                 sgs->sum_nr_running += rq->cfs.h_nr_running;
6321 
6322                 if (rq->nr_running > 1)
6323                         *overload = true;
6324 
6325 #ifdef CONFIG_NUMA_BALANCING
6326                 sgs->nr_numa_running += rq->nr_numa_running;
6327                 sgs->nr_preferred_running += rq->nr_preferred_running;
6328 #endif
6329                 sgs->sum_weighted_load += weighted_cpuload(i);
6330                 if (idle_cpu(i))
6331                         sgs->idle_cpus++;
6332         }
6333 
6334         /* Adjust by relative CPU capacity of the group */
6335         sgs->group_capacity = group->sgc->capacity;
6336         sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6337 
6338         if (sgs->sum_nr_running)
6339                 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6340 
6341         sgs->group_weight = group->group_weight;
6342 
6343         sgs->group_no_capacity = group_is_overloaded(env, sgs);
6344         sgs->group_type = group_classify(group, sgs);
6345 }
6346 
6347 /**
6348  * update_sd_pick_busiest - return 1 on busiest group
6349  * @env: The load balancing environment.
6350  * @sds: sched_domain statistics
6351  * @sg: sched_group candidate to be checked for being the busiest
6352  * @sgs: sched_group statistics
6353  *
6354  * Determine if @sg is a busier group than the previously selected
6355  * busiest group.
6356  *
6357  * Return: %true if @sg is a busier group than the previously selected
6358  * busiest group. %false otherwise.
6359  */
6360 static bool update_sd_pick_busiest(struct lb_env *env,
6361                                    struct sd_lb_stats *sds,
6362                                    struct sched_group *sg,
6363                                    struct sg_lb_stats *sgs)
6364 {
6365         struct sg_lb_stats *busiest = &sds->busiest_stat;
6366 
6367         if (sgs->group_type > busiest->group_type)
6368                 return true;
6369 
6370         if (sgs->group_type < busiest->group_type)
6371                 return false;
6372 
6373         if (sgs->avg_load <= busiest->avg_load)
6374                 return false;
6375 
6376         /* This is the busiest node in its class. */
6377         if (!(env->sd->flags & SD_ASYM_PACKING))
6378                 return true;
6379 
6380         /*
6381          * ASYM_PACKING needs to move all the work to the lowest
6382          * numbered CPUs in the group, therefore mark all groups
6383          * higher than ourself as busy.
6384          */
6385         if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6386                 if (!sds->busiest)
6387                         return true;
6388 
6389                 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
6390                         return true;
6391         }
6392 
6393         return false;
6394 }
6395 
6396 #ifdef CONFIG_NUMA_BALANCING
6397 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6398 {
6399         if (sgs->sum_nr_running > sgs->nr_numa_running)
6400                 return regular;
6401         if (sgs->sum_nr_running > sgs->nr_preferred_running)
6402                 return remote;
6403         return all;
6404 }
6405 
6406 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6407 {
6408         if (rq->nr_running > rq->nr_numa_running)
6409                 return regular;
6410         if (rq->nr_running > rq->nr_preferred_running)
6411                 return remote;
6412         return all;
6413 }
6414 #else
6415 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6416 {
6417         return all;
6418 }
6419 
6420 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6421 {
6422         return regular;
6423 }
6424 #endif /* CONFIG_NUMA_BALANCING */
6425 
6426 /**
6427  * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6428  * @env: The load balancing environment.
6429  * @sds: variable to hold the statistics for this sched_domain.
6430  */
6431 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6432 {
6433         struct sched_domain *child = env->sd->child;
6434         struct sched_group *sg = env->sd->groups;
6435         struct sg_lb_stats tmp_sgs;
6436         int load_idx, prefer_sibling = 0;
6437         bool overload = false;
6438 
6439         if (child && child->flags & SD_PREFER_SIBLING)
6440                 prefer_sibling = 1;
6441 
6442         load_idx = get_sd_load_idx(env->sd, env->idle);
6443 
6444         do {
6445                 struct sg_lb_stats *sgs = &tmp_sgs;
6446                 int local_group;
6447 
6448                 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6449                 if (local_group) {
6450                         sds->local = sg;
6451                         sgs = &sds->local_stat;
6452 
6453                         if (env->idle != CPU_NEWLY_IDLE ||
6454                             time_after_eq(jiffies, sg->sgc->next_update))
6455                                 update_group_capacity(env->sd, env->dst_cpu);
6456                 }
6457 
6458                 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6459                                                 &overload);
6460 
6461                 if (local_group)
6462                         goto next_group;
6463 
6464                 /*
6465                  * In case the child domain prefers tasks go to siblings
6466                  * first, lower the sg capacity so that we'll try
6467                  * and move all the excess tasks away. We lower the capacity
6468                  * of a group only if the local group has the capacity to fit
6469                  * these excess tasks. The extra check prevents the case where
6470                  * you always pull from the heaviest group when it is already
6471                  * under-utilized (possible with a large weight task outweighs
6472                  * the tasks on the system).
6473                  */
6474                 if (prefer_sibling && sds->local &&
6475                     group_has_capacity(env, &sds->local_stat) &&
6476                     (sgs->sum_nr_running > 1)) {
6477                         sgs->group_no_capacity = 1;
6478                         sgs->group_type = group_classify(sg, sgs);
6479                 }
6480 
6481                 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6482                         sds->busiest = sg;
6483                         sds->busiest_stat = *sgs;
6484                 }
6485 
6486 next_group:
6487                 /* Now, start updating sd_lb_stats */
6488                 sds->total_load += sgs->group_load;
6489                 sds->total_capacity += sgs->group_capacity;
6490 
6491                 sg = sg->next;
6492         } while (sg != env->sd->groups);
6493 
6494         if (env->sd->flags & SD_NUMA)
6495                 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6496 
6497         if (!env->sd->parent) {
6498                 /* update overload indicator if we are at root domain */
6499                 if (env->dst_rq->rd->overload != overload)
6500                         env->dst_rq->rd->overload = overload;
6501         }
6502 
6503 }
6504 
6505 /**
6506  * check_asym_packing - Check to see if the group is packed into the
6507  *                      sched doman.
6508  *
6509  * This is primarily intended to used at the sibling level.  Some
6510  * cores like POWER7 prefer to use lower numbered SMT threads.  In the
6511  * case of POWER7, it can move to lower SMT modes only when higher
6512  * threads are idle.  When in lower SMT modes, the threads will
6513  * perform better since they share less core resources.  Hence when we
6514  * have idle threads, we want them to be the higher ones.
6515  *
6516  * This packing function is run on idle threads.  It checks to see if
6517  * the busiest CPU in this domain (core in the P7 case) has a higher
6518  * CPU number than the packing function is being run on.  Here we are
6519  * assuming lower CPU number will be equivalent to lower a SMT thread
6520  * number.
6521  *
6522  * Return: 1 when packing is required and a task should be moved to
6523  * this CPU.  The amount of the imbalance is returned in *imbalance.
6524  *
6525  * @env: The load balancing environment.
6526  * @sds: Statistics of the sched_domain which is to be packed
6527  */
6528 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6529 {
6530         int busiest_cpu;
6531 
6532         if (!(env->sd->flags & SD_ASYM_PACKING))
6533                 return 0;
6534 
6535         if (!sds->busiest)
6536                 return 0;
6537 
6538         busiest_cpu = group_first_cpu(sds->busiest);
6539         if (env->dst_cpu > busiest_cpu)
6540                 return 0;
6541 
6542         env->imbalance = DIV_ROUND_CLOSEST(
6543                 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6544                 SCHED_CAPACITY_SCALE);
6545 
6546         return 1;
6547 }
6548 
6549 /**
6550  * fix_small_imbalance - Calculate the minor imbalance that exists
6551  *                      amongst the groups of a sched_domain, during
6552  *                      load balancing.
6553  * @env: The load balancing environment.
6554  * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6555  */
6556 static inline
6557 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6558 {
6559         unsigned long tmp, capa_now = 0, capa_move = 0;
6560         unsigned int imbn = 2;
6561         unsigned long scaled_busy_load_per_task;
6562         struct sg_lb_stats *local, *busiest;
6563 
6564         local = &sds->local_stat;
6565         busiest = &sds->busiest_stat;
6566 
6567         if (!local->sum_nr_running)
6568                 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6569         else if (busiest->load_per_task > local->load_per_task)
6570                 imbn = 1;
6571 
6572         scaled_busy_load_per_task =
6573                 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6574                 busiest->group_capacity;
6575 
6576         if (busiest->avg_load + scaled_busy_load_per_task >=
6577             local->avg_load + (scaled_busy_load_per_task * imbn)) {
6578                 env->imbalance = busiest->load_per_task;
6579                 return;
6580         }
6581 
6582         /*
6583          * OK, we don't have enough imbalance to justify moving tasks,
6584          * however we may be able to increase total CPU capacity used by
6585          * moving them.
6586          */
6587 
6588         capa_now += busiest->group_capacity *
6589                         min(busiest->load_per_task, busiest->avg_load);
6590         capa_now += local->group_capacity *
6591                         min(local->load_per_task, local->avg_load);
6592         capa_now /= SCHED_CAPACITY_SCALE;
6593 
6594         /* Amount of load we'd subtract */
6595         if (busiest->avg_load > scaled_busy_load_per_task) {
6596                 capa_move += busiest->group_capacity *
6597                             min(busiest->load_per_task,
6598                                 busiest->avg_load - scaled_busy_load_per_task);
6599         }
6600 
6601         /* Amount of load we'd add */
6602         if (busiest->avg_load * busiest->group_capacity <
6603             busiest->load_per_task * SCHED_CAPACITY_SCALE) {
6604                 tmp = (busiest->avg_load * busiest->group_capacity) /
6605                       local->group_capacity;
6606         } else {
6607                 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6608                       local->group_capacity;
6609         }
6610         capa_move += local->group_capacity *
6611                     min(local->load_per_task, local->avg_load + tmp);
6612         capa_move /= SCHED_CAPACITY_SCALE;
6613 
6614         /* Move if we gain throughput */
6615         if (capa_move > capa_now)
6616                 env->imbalance = busiest->load_per_task;
6617 }
6618 
6619 /**
6620  * calculate_imbalance - Calculate the amount of imbalance present within the
6621  *                       groups of a given sched_domain during load balance.
6622  * @env: load balance environment
6623  * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6624  */
6625 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6626 {
6627         unsigned long max_pull, load_above_capacity = ~0UL;
6628         struct sg_lb_stats *local, *busiest;
6629 
6630         local = &sds->local_stat;
6631         busiest = &sds->busiest_stat;
6632 
6633         if (busiest->group_type == group_imbalanced) {
6634                 /*
6635                  * In the group_imb case we cannot rely on group-wide averages
6636                  * to ensure cpu-load equilibrium, look at wider averages. XXX
6637                  */
6638                 busiest->load_per_task =
6639                         min(busiest->load_per_task, sds->avg_load);
6640         }
6641 
6642         /*
6643          * In the presence of smp nice balancing, certain scenarios can have
6644          * max load less than avg load(as we skip the groups at or below
6645          * its cpu_capacity, while calculating max_load..)
6646          */
6647         if (busiest->avg_load <= sds->avg_load ||
6648             local->avg_load >= sds->avg_load) {
6649                 env->imbalance = 0;
6650                 return fix_small_imbalance(env, sds);
6651         }
6652 
6653         /*
6654          * If there aren't any idle cpus, avoid creating some.
6655          */
6656         if (busiest->group_type == group_overloaded &&
6657             local->group_type   == group_overloaded) {
6658                 load_above_capacity = busiest->sum_nr_running *
6659                                         SCHED_LOAD_SCALE;
6660                 if (load_above_capacity > busiest->group_capacity)
6661                         load_above_capacity -= busiest->group_capacity;
6662                 else
6663                         load_above_capacity = ~0UL;
6664         }
6665 
6666         /*
6667          * We're trying to get all the cpus to the average_load, so we don't
6668          * want to push ourselves above the average load, nor do we wish to
6669          * reduce the max loaded cpu below the average load. At the same time,
6670          * we also don't want to reduce the group load below the group capacity
6671          * (so that we can implement power-savings policies etc). Thus we look
6672          * for the minimum possible imbalance.
6673          */
6674         max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6675 
6676         /* How much load to actually move to equalise the imbalance */
6677         env->imbalance = min(
6678                 max_pull * busiest->group_capacity,
6679                 (sds->avg_load - local->avg_load) * local->group_capacity
6680         ) / SCHED_CAPACITY_SCALE;
6681 
6682         /*
6683          * if *imbalance is less than the average load per runnable task
6684          * there is no guarantee that any tasks will be moved so we'll have
6685          * a think about bumping its value to force at least one task to be
6686          * moved
6687          */
6688         if (env->imbalance < busiest->load_per_task)
6689                 return fix_small_imbalance(env, sds);
6690 }
6691 
6692 /******* find_busiest_group() helpers end here *********************/
6693 
6694 /**
6695  * find_busiest_group - Returns the busiest group within the sched_domain
6696  * if there is an imbalance. If there isn't an imbalance, and
6697  * the user has opted for power-savings, it returns a group whose
6698  * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6699  * such a group exists.
6700  *
6701  * Also calculates the amount of weighted load which should be moved
6702  * to restore balance.
6703  *
6704  * @env: The load balancing environment.
6705  *
6706  * Return:      - The busiest group if imbalance exists.
6707  *              - If no imbalance and user has opted for power-savings balance,
6708  *                 return the least loaded group whose CPUs can be
6709  *                 put to idle by rebalancing its tasks onto our group.
6710  */
6711 static struct sched_group *find_busiest_group(struct lb_env *env)
6712 {
6713         struct sg_lb_stats *local, *busiest;
6714         struct sd_lb_stats sds;
6715 
6716         init_sd_lb_stats(&sds);
6717 
6718         /*
6719          * Compute the various statistics relavent for load balancing at
6720          * this level.
6721          */
6722         update_sd_lb_stats(env, &sds);
6723         local = &sds.local_stat;
6724         busiest = &sds.busiest_stat;
6725 
6726         /* ASYM feature bypasses nice load balance check */
6727         if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6728             check_asym_packing(env, &sds))
6729                 return sds.busiest;
6730 
6731         /* There is no busy sibling group to pull tasks from */
6732         if (!sds.busiest || busiest->sum_nr_running == 0)
6733                 goto out_balanced;
6734 
6735         sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
6736                                                 / sds.total_capacity;
6737 
6738         /*
6739          * If the busiest group is imbalanced the below checks don't
6740          * work because they assume all things are equal, which typically
6741          * isn't true due to cpus_allowed constraints and the like.
6742          */
6743         if (busiest->group_type == group_imbalanced)
6744                 goto force_balance;
6745 
6746         /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6747         if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
6748             busiest->group_no_capacity)
6749                 goto force_balance;
6750 
6751         /*
6752          * If the local group is busier than the selected busiest group
6753          * don't try and pull any tasks.
6754          */
6755         if (local->avg_load >= busiest->avg_load)
6756                 goto out_balanced;
6757 
6758         /*
6759          * Don't pull any tasks if this group is already above the domain
6760          * average load.
6761          */
6762         if (local->avg_load >= sds.avg_load)
6763                 goto out_balanced;
6764 
6765         if (env->idle == CPU_IDLE) {
6766                 /*
6767                  * This cpu is idle. If the busiest group is not overloaded
6768                  * and there is no imbalance between this and busiest group
6769                  * wrt idle cpus, it is balanced. The imbalance becomes
6770                  * significant if the diff is greater than 1 otherwise we
6771                  * might end up to just move the imbalance on another group
6772                  */
6773                 if ((busiest->group_type != group_overloaded) &&
6774                                 (local->idle_cpus <= (busiest->idle_cpus + 1)))
6775                         goto out_balanced;
6776         } else {
6777                 /*
6778                  * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6779                  * imbalance_pct to be conservative.
6780                  */
6781                 if (100 * busiest->avg_load <=
6782                                 env->sd->imbalance_pct * local->avg_load)
6783                         goto out_balanced;
6784         }
6785 
6786 force_balance:
6787         /* Looks like there is an imbalance. Compute it */
6788         calculate_imbalance(env, &sds);
6789         return sds.busiest;
6790 
6791 out_balanced:
6792         env->imbalance = 0;
6793         return NULL;
6794 }
6795 
6796 /*
6797  * find_busiest_queue - find the busiest runqueue among the cpus in group.
6798  */
6799 static struct rq *find_busiest_queue(struct lb_env *env,
6800                                      struct sched_group *group)
6801 {
6802         struct rq *busiest = NULL, *rq;
6803         unsigned long busiest_load = 0, busiest_capacity = 1;
6804         int i;
6805 
6806         for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6807                 unsigned long capacity, wl;
6808                 enum fbq_type rt;
6809 
6810                 rq = cpu_rq(i);
6811                 rt = fbq_classify_rq(rq);
6812 
6813                 /*
6814                  * We classify groups/runqueues into three groups:
6815                  *  - regular: there are !numa tasks
6816                  *  - remote:  there are numa tasks that run on the 'wrong' node
6817                  *  - all:     there is no distinction
6818                  *
6819                  * In order to avoid migrating ideally placed numa tasks,
6820                  * ignore those when there's better options.
6821                  *
6822                  * If we ignore the actual busiest queue to migrate another
6823                  * task, the next balance pass can still reduce the busiest
6824                  * queue by moving tasks around inside the node.
6825                  *
6826                  * If we cannot move enough load due to this classification
6827                  * the next pass will adjust the group classification and
6828                  * allow migration of more tasks.
6829                  *
6830                  * Both cases only affect the total convergence complexity.
6831                  */
6832                 if (rt > env->fbq_type)
6833                         continue;
6834 
6835                 capacity = capacity_of(i);
6836 
6837                 wl = weighted_cpuload(i);
6838 
6839                 /*
6840                  * When comparing with imbalance, use weighted_cpuload()
6841                  * which is not scaled with the cpu capacity.
6842                  */
6843 
6844                 if (rq->nr_running == 1 && wl > env->imbalance &&
6845                     !check_cpu_capacity(rq, env->sd))
6846                         continue;
6847 
6848                 /*
6849                  * For the load comparisons with the other cpu's, consider
6850                  * the weighted_cpuload() scaled with the cpu capacity, so
6851                  * that the load can be moved away from the cpu that is
6852                  * potentially running at a lower capacity.
6853                  *
6854                  * Thus we're looking for max(wl_i / capacity_i), crosswise
6855                  * multiplication to rid ourselves of the division works out
6856                  * to: wl_i * capacity_j > wl_j * capacity_i;  where j is
6857                  * our previous maximum.
6858                  */
6859                 if (wl * busiest_capacity > busiest_load * capacity) {
6860                         busiest_load = wl;
6861                         busiest_capacity = capacity;
6862                         busiest = rq;
6863                 }
6864         }
6865 
6866         return busiest;
6867 }
6868 
6869 /*
6870  * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6871  * so long as it is large enough.
6872  */
6873 #define MAX_PINNED_INTERVAL     512
6874 
6875 /* Working cpumask for load_balance and load_balance_newidle. */
6876 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6877 
6878 static int need_active_balance(struct lb_env *env)
6879 {
6880         struct sched_domain *sd = env->sd;
6881 
6882         if (env->idle == CPU_NEWLY_IDLE) {
6883 
6884                 /*
6885                  * ASYM_PACKING needs to force migrate tasks from busy but
6886                  * higher numbered CPUs in order to pack all tasks in the
6887                  * lowest numbered CPUs.
6888                  */
6889                 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6890                         return 1;
6891         }
6892 
6893         /*
6894          * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
6895          * It's worth migrating the task if the src_cpu's capacity is reduced
6896          * because of other sched_class or IRQs if more capacity stays
6897          * available on dst_cpu.
6898          */
6899         if ((env->idle != CPU_NOT_IDLE) &&
6900             (env->src_rq->cfs.h_nr_running == 1)) {
6901                 if ((check_cpu_capacity(env->src_rq, sd)) &&
6902                     (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
6903                         return 1;
6904         }
6905 
6906         return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6907 }
6908 
6909 static int active_load_balance_cpu_stop(void *data);
6910 
6911 static int should_we_balance(struct lb_env *env)
6912 {
6913         struct sched_group *sg = env->sd->groups;
6914         struct cpumask *sg_cpus, *sg_mask;
6915         int cpu, balance_cpu = -1;
6916 
6917         /*
6918          * In the newly idle case, we will allow all the cpu's
6919          * to do the newly idle load balance.
6920          */
6921         if (env->idle == CPU_NEWLY_IDLE)
6922                 return 1;
6923 
6924         sg_cpus = sched_group_cpus(sg);
6925         sg_mask = sched_group_mask(sg);
6926         /* Try to find first idle cpu */
6927         for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6928                 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6929                         continue;
6930 
6931                 balance_cpu = cpu;
6932                 break;
6933         }
6934 
6935         if (balance_cpu == -1)
6936                 balance_cpu = group_balance_cpu(sg);
6937 
6938         /*
6939          * First idle cpu or the first cpu(busiest) in this sched group
6940          * is eligible for doing load balancing at this and above domains.
6941          */
6942         return balance_cpu == env->dst_cpu;
6943 }
6944 
6945 /*
6946  * Check this_cpu to ensure it is balanced within domain. Attempt to move
6947  * tasks if there is an imbalance.
6948  */
6949 static int load_balance(int this_cpu, struct rq *this_rq,
6950                         struct sched_domain *sd, enum cpu_idle_type idle,
6951                         int *continue_balancing)
6952 {
6953         int ld_moved, cur_ld_moved, active_balance = 0;
6954         struct sched_domain *sd_parent = sd->parent;
6955         struct sched_group *group;
6956         struct rq *busiest;
6957         unsigned long flags;
6958         struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
6959 
6960         struct lb_env env = {
6961                 .sd             = sd,
6962                 .dst_cpu        = this_cpu,
6963                 .dst_rq         = this_rq,
6964                 .dst_grpmask    = sched_group_cpus(sd->groups),
6965                 .idle           = idle,
6966                 .loop_break     = sched_nr_migrate_break,
6967                 .cpus           = cpus,
6968                 .fbq_type       = all,
6969                 .tasks          = LIST_HEAD_INIT(env.tasks),
6970         };
6971 
6972         /*
6973          * For NEWLY_IDLE load_balancing, we don't need to consider
6974          * other cpus in our group
6975          */
6976         if (idle == CPU_NEWLY_IDLE)
6977                 env.dst_grpmask = NULL;
6978 
6979         cpumask_copy(cpus, cpu_active_mask);
6980 
6981         schedstat_inc(sd, lb_count[idle]);
6982 
6983 redo:
6984         if (!should_we_balance(&env)) {
6985                 *continue_balancing = 0;
6986                 goto out_balanced;
6987         }
6988 
6989         group = find_busiest_group(&env);
6990         if (!group) {
6991                 schedstat_inc(sd, lb_nobusyg[idle]);
6992                 goto out_balanced;
6993         }
6994 
6995         busiest = find_busiest_queue(&env, group);
6996         if (!busiest) {
6997                 schedstat_inc(sd, lb_nobusyq[idle]);
6998                 goto out_balanced;
6999         }
7000 
7001         BUG_ON(busiest == env.dst_rq);
7002 
7003         schedstat_add(sd, lb_imbalance[idle], env.imbalance);
7004 
7005         env.src_cpu = busiest->cpu;
7006         env.src_rq = busiest;
7007 
7008         ld_moved = 0;
7009         if (busiest->nr_running > 1) {
7010                 /*
7011                  * Attempt to move tasks. If find_busiest_group has found
7012                  * an imbalance but busiest->nr_running <= 1, the group is
7013                  * still unbalanced. ld_moved simply stays zero, so it is
7014                  * correctly treated as an imbalance.
7015                  */
7016                 env.flags |= LBF_ALL_PINNED;
7017                 env.loop_max  = min(sysctl_sched_nr_migrate, busiest->nr_running);
7018 
7019 more_balance:
7020                 raw_spin_lock_irqsave(&busiest->lock, flags);
7021 
7022                 /*
7023                  * cur_ld_moved - load moved in current iteration
7024                  * ld_moved     - cumulative load moved across iterations
7025                  */
7026                 cur_ld_moved = detach_tasks(&env);
7027 
7028                 /*
7029                  * We've detached some tasks from busiest_rq. Every
7030                  * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7031                  * unlock busiest->lock, and we are able to be sure
7032                  * that nobody can manipulate the tasks in parallel.
7033                  * See task_rq_lock() family for the details.
7034                  */
7035 
7036                 raw_spin_unlock(&busiest->lock);
7037 
7038                 if (cur_ld_moved) {
7039                         attach_tasks(&env);
7040                         ld_moved += cur_ld_moved;
7041                 }
7042 
7043                 local_irq_restore(flags);
7044 
7045                 if (env.flags & LBF_NEED_BREAK) {
7046                         env.flags &= ~LBF_NEED_BREAK;
7047                         goto more_balance;
7048                 }
7049 
7050                 /*
7051                  * Revisit (affine) tasks on src_cpu that couldn't be moved to
7052                  * us and move them to an alternate dst_cpu in our sched_group
7053                  * where they can run. The upper limit on how many times we
7054                  * iterate on same src_cpu is dependent on number of cpus in our
7055                  * sched_group.
7056                  *
7057                  * This changes load balance semantics a bit on who can move
7058                  * load to a given_cpu. In addition to the given_cpu itself
7059                  * (or a ilb_cpu acting on its behalf where given_cpu is
7060                  * nohz-idle), we now have balance_cpu in a position to move
7061                  * load to given_cpu. In rare situations, this may cause
7062                  * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7063                  * _independently_ and at _same_ time to move some load to
7064                  * given_cpu) causing exceess load to be moved to given_cpu.
7065                  * This however should not happen so much in practice and
7066                  * moreover subsequent load balance cycles should correct the
7067                  * excess load moved.
7068                  */
7069                 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7070 
7071                         /* Prevent to re-select dst_cpu via env's cpus */
7072                         cpumask_clear_cpu(env.dst_cpu, env.cpus);
7073 
7074                         env.dst_rq       = cpu_rq(env.new_dst_cpu);
7075                         env.dst_cpu      = env.new_dst_cpu;
7076                         env.flags       &= ~LBF_DST_PINNED;
7077                         env.loop         = 0;
7078                         env.loop_break   = sched_nr_migrate_break;
7079 
7080                         /*
7081                          * Go back to "more_balance" rather than "redo" since we
7082                          * need to continue with same src_cpu.
7083                          */
7084                         goto more_balance;
7085                 }
7086 
7087                 /*
7088                  * We failed to reach balance because of affinity.
7089                  */
7090                 if (sd_parent) {
7091                         int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7092 
7093                         if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7094                                 *group_imbalance = 1;
7095                 }
7096 
7097                 /* All tasks on this runqueue were pinned by CPU affinity */
7098                 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7099                         cpumask_clear_cpu(cpu_of(busiest), cpus);
7100                         if (!cpumask_empty(cpus)) {
7101                                 env.loop = 0;
7102                                 env.loop_break = sched_nr_migrate_break;
7103                                 goto redo;
7104                         }
7105                         goto out_all_pinned;
7106                 }
7107         }
7108 
7109         if (!ld_moved) {
7110                 schedstat_inc(sd, lb_failed[idle]);
7111                 /*
7112                  * Increment the failure counter only on periodic balance.
7113                  * We do not want newidle balance, which can be very
7114                  * frequent, pollute the failure counter causing
7115                  * excessive cache_hot migrations and active balances.
7116                  */
7117                 if (idle != CPU_NEWLY_IDLE)
7118                         sd->nr_balance_failed++;
7119 
7120                 if (need_active_balance(&env)) {
7121                         raw_spin_lock_irqsave(&busiest->lock, flags);
7122 
7123                         /* don't kick the active_load_balance_cpu_stop,
7124                          * if the curr task on busiest cpu can't be
7125                          * moved to this_cpu
7126                          */
7127                         if (!cpumask_test_cpu(this_cpu,
7128                                         tsk_cpus_allowed(busiest->curr))) {
7129                                 raw_spin_unlock_irqrestore(&busiest->lock,
7130                                                             flags);
7131                                 env.flags |= LBF_ALL_PINNED;
7132                                 goto out_one_pinned;
7133                         }
7134 
7135                         /*
7136                          * ->active_balance synchronizes accesses to
7137                          * ->active_balance_work.  Once set, it's cleared
7138                          * only after active load balance is finished.
7139                          */
7140                         if (!busiest->active_balance) {
7141                                 busiest->active_balance = 1;
7142                                 busiest->push_cpu = this_cpu;
7143                                 active_balance = 1;
7144                         }
7145                         raw_spin_unlock_irqrestore(&busiest->lock, flags);
7146 
7147                         if (active_balance) {
7148                                 stop_one_cpu_nowait(cpu_of(busiest),
7149                                         active_load_balance_cpu_stop, busiest,
7150                                         &busiest->active_balance_work);
7151                         }
7152 
7153                         /*
7154                          * We've kicked active balancing, reset the failure
7155                          * counter.
7156                          */
7157                         sd->nr_balance_failed = sd->cache_nice_tries+1;
7158                 }
7159         } else
7160                 sd->nr_balance_failed = 0;
7161 
7162         if (likely(!active_balance)) {
7163                 /* We were unbalanced, so reset the balancing interval */
7164                 sd->balance_interval = sd->min_interval;
7165         } else {
7166                 /*
7167                  * If we've begun active balancing, start to back off. This
7168                  * case may not be covered by the all_pinned logic if there
7169                  * is only 1 task on the busy runqueue (because we don't call
7170                  * detach_tasks).
7171                  */
7172                 if (sd->balance_interval < sd->max_interval)
7173                         sd->balance_interval *= 2;
7174         }
7175 
7176         goto out;
7177 
7178 out_balanced:
7179         /*
7180          * We reach balance although we may have faced some affinity
7181          * constraints. Clear the imbalance flag if it was set.
7182          */
7183         if (sd_parent) {
7184                 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7185 
7186                 if (*group_imbalance)
7187                         *group_imbalance = 0;
7188         }
7189 
7190 out_all_pinned:
7191         /*
7192          * We reach balance because all tasks are pinned at this level so
7193          * we can't migrate them. Let the imbalance flag set so parent level
7194          * can try to migrate them.
7195          */
7196         schedstat_inc(sd, lb_balanced[idle]);
7197 
7198         sd->nr_balance_failed = 0;
7199 
7200 out_one_pinned:
7201         /* tune up the balancing interval */
7202         if (((env.flags & LBF_ALL_PINNED) &&
7203                         sd->balance_interval < MAX_PINNED_INTERVAL) ||
7204                         (sd->balance_interval < sd->max_interval))
7205                 sd->balance_interval *= 2;
7206 
7207         ld_moved = 0;
7208 out:
7209         return ld_moved;
7210 }
7211 
7212 static inline unsigned long
7213 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7214 {
7215         unsigned long interval = sd->balance_interval;
7216 
7217         if (cpu_busy)
7218                 interval *= sd->busy_factor;
7219 
7220         /* scale ms to jiffies */
7221         interval = msecs_to_jiffies(interval);
7222         interval = clamp(interval, 1UL, max_load_balance_interval);
7223 
7224         return interval;
7225 }
7226 
7227 static inline void
7228 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
7229 {
7230         unsigned long interval, next;
7231 
7232         interval = get_sd_balance_interval(sd, cpu_busy);
7233         next = sd->last_balance + interval;
7234 
7235         if (time_after(*next_balance, next))
7236                 *next_balance = next;
7237 }
7238 
7239 /*
7240  * idle_balance is called by schedule() if this_cpu is about to become
7241  * idle. Attempts to pull tasks from other CPUs.
7242  */
7243 static int idle_balance(struct rq *this_rq)
7244 {
7245         unsigned long next_balance = jiffies + HZ;
7246         int this_cpu = this_rq->cpu;
7247         struct sched_domain *sd;
7248         int pulled_task = 0;
7249         u64 curr_cost = 0;
7250 
7251         idle_enter_fair(this_rq);
7252 
7253         /*
7254          * We must set idle_stamp _before_ calling idle_balance(), such that we
7255          * measure the duration of idle_balance() as idle time.
7256          */
7257         this_rq->idle_stamp = rq_clock(this_rq);
7258 
7259         if (this_rq->avg_idle < sysctl_sched_migration_cost ||
7260             !this_rq->rd->overload) {
7261                 rcu_read_lock();
7262                 sd = rcu_dereference_check_sched_domain(this_rq->sd);
7263                 if (sd)
7264                         update_next_balance(sd, 0, &next_balance);
7265                 rcu_read_unlock();
7266 
7267                 goto out;
7268         }
7269 
7270         raw_spin_unlock(&this_rq->lock);
7271 
7272         update_blocked_averages(this_cpu);
7273         rcu_read_lock();
7274         for_each_domain(this_cpu, sd) {
7275                 int continue_balancing = 1;
7276                 u64 t0, domain_cost;
7277 
7278                 if (!(sd->flags & SD_LOAD_BALANCE))
7279                         continue;
7280 
7281                 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7282                         update_next_balance(sd, 0, &next_balance);
7283                         break;
7284                 }
7285 
7286                 if (sd->flags & SD_BALANCE_NEWIDLE) {
7287                         t0 = sched_clock_cpu(this_cpu);
7288 
7289                         pulled_task = load_balance(this_cpu, this_rq,
7290                                                    sd, CPU_NEWLY_IDLE,
7291                                                    &continue_balancing);
7292 
7293                         domain_cost = sched_clock_cpu(this_cpu) - t0;
7294                         if (domain_cost > sd->max_newidle_lb_cost)
7295                                 sd->max_newidle_lb_cost = domain_cost;
7296 
7297                         curr_cost += domain_cost;
7298                 }
7299 
7300                 update_next_balance(sd, 0, &next_balance);
7301 
7302                 /*
7303                  * Stop searching for tasks to pull if there are
7304                  * now runnable tasks on this rq.
7305                  */
7306                 if (pulled_task || this_rq->nr_running > 0)
7307                         break;
7308         }
7309         rcu_read_unlock();
7310 
7311         raw_spin_lock(&this_rq->lock);
7312 
7313         if (curr_cost > this_rq->max_idle_balance_cost)
7314                 this_rq->max_idle_balance_cost = curr_cost;
7315 
7316         /*
7317          * While browsing the domains, we released the rq lock, a task could
7318          * have been enqueued in the meantime. Since we're not going idle,
7319          * pretend we pulled a task.
7320          */
7321         if (this_rq->cfs.h_nr_running && !pulled_task)
7322                 pulled_task = 1;
7323 
7324 out:
7325         /* Move the next balance forward */
7326         if (time_after(this_rq->next_balance, next_balance))
7327                 this_rq->next_balance = next_balance;
7328 
7329         /* Is there a task of a high priority class? */
7330         if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7331                 pulled_task = -1;
7332 
7333         if (pulled_task) {
7334                 idle_exit_fair(this_rq);
7335                 this_rq->idle_stamp = 0;
7336         }
7337 
7338         return pulled_task;
7339 }
7340 
7341 /*
7342  * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7343  * running tasks off the busiest CPU onto idle CPUs. It requires at
7344  * least 1 task to be running on each physical CPU where possible, and
7345  * avoids physical / logical imbalances.
7346  */
7347 static int active_load_balance_cpu_stop(void *data)
7348 {
7349         struct rq *busiest_rq = data;
7350         int busiest_cpu = cpu_of(busiest_rq);
7351         int target_cpu = busiest_rq->push_cpu;
7352         struct rq *target_rq = cpu_rq(target_cpu);
7353         struct sched_domain *sd;
7354         struct task_struct *p = NULL;
7355 
7356         raw_spin_lock_irq(&busiest_rq->lock);
7357 
7358         /* make sure the requested cpu hasn't gone down in the meantime */
7359         if (unlikely(busiest_cpu != smp_processor_id() ||
7360                      !busiest_rq->active_balance))
7361                 goto out_unlock;
7362 
7363         /* Is there any task to move? */
7364         if (busiest_rq->nr_running <= 1)
7365                 goto out_unlock;
7366 
7367         /*
7368          * This condition is "impossible", if it occurs
7369          * we need to fix it. Originally reported by
7370          * Bjorn Helgaas on a 128-cpu setup.
7371          */
7372         BUG_ON(busiest_rq == target_rq);
7373 
7374         /* Search for an sd spanning us and the target CPU. */
7375         rcu_read_lock();
7376         for_each_domain(target_cpu, sd) {
7377                 if ((sd->flags & SD_LOAD_BALANCE) &&
7378                     cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7379                                 break;
7380         }
7381 
7382         if (likely(sd)) {
7383                 struct lb_env env = {
7384                         .sd             = sd,
7385                         .dst_cpu        = target_cpu,
7386                         .dst_rq         = target_rq,
7387                         .src_cpu        = busiest_rq->cpu,
7388                         .src_rq         = busiest_rq,
7389                         .idle           = CPU_IDLE,
7390                 };
7391 
7392                 schedstat_inc(sd, alb_count);
7393 
7394                 p = detach_one_task(&env);
7395                 if (p)
7396                         schedstat_inc(sd, alb_pushed);
7397                 else
7398                         schedstat_inc(sd, alb_failed);
7399         }
7400         rcu_read_unlock();
7401 out_unlock:
7402         busiest_rq->active_balance = 0;
7403         raw_spin_unlock(&busiest_rq->lock);
7404 
7405         if (p)
7406                 attach_one_task(target_rq, p);
7407 
7408         local_irq_enable();
7409 
7410         return 0;
7411 }
7412 
7413 static inline int on_null_domain(struct rq *rq)
7414 {
7415         return unlikely(!rcu_dereference_sched(rq->sd));
7416 }
7417 
7418 #ifdef CONFIG_NO_HZ_COMMON
7419 /*
7420  * idle load balancing details
7421  * - When one of the busy CPUs notice that there may be an idle rebalancing
7422  *   needed, they will kick the idle load balancer, which then does idle
7423  *   load balancing for all the idle CPUs.
7424  */
7425 static struct {
7426         cpumask_var_t idle_cpus_mask;
7427         atomic_t nr_cpus;
7428         unsigned long next_balance;     /* in jiffy units */
7429 } nohz ____cacheline_aligned;
7430 
7431 static inline int find_new_ilb(void)
7432 {
7433         int ilb = cpumask_first(nohz.idle_cpus_mask);
7434 
7435         if (ilb < nr_cpu_ids && idle_cpu(ilb))
7436                 return ilb;
7437 
7438         return nr_cpu_ids;
7439 }
7440 
7441 /*
7442  * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7443  * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7444  * CPU (if there is one).
7445  */
7446 static void nohz_balancer_kick(void)
7447 {
7448         int ilb_cpu;
7449 
7450         nohz.next_balance++;
7451 
7452         ilb_cpu = find_new_ilb();
7453 
7454         if (ilb_cpu >= nr_cpu_ids)
7455                 return;
7456 
7457         if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7458                 return;
7459         /*
7460          * Use smp_send_reschedule() instead of resched_cpu().
7461          * This way we generate a sched IPI on the target cpu which
7462          * is idle. And the softirq performing nohz idle load balance
7463          * will be run before returning from the IPI.
7464          */
7465         smp_send_reschedule(ilb_cpu);
7466         return;
7467 }
7468 
7469 static inline void nohz_balance_exit_idle(int cpu)
7470 {
7471         if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7472                 /*
7473                  * Completely isolated CPUs don't ever set, so we must test.
7474                  */
7475                 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7476                         cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7477                         atomic_dec(&nohz.nr_cpus);
7478                 }
7479                 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7480         }
7481 }
7482 
7483 static inline void set_cpu_sd_state_busy(void)
7484 {
7485         struct sched_domain *sd;
7486         int cpu = smp_processor_id();
7487 
7488         rcu_read_lock();
7489         sd = rcu_dereference(per_cpu(sd_busy, cpu));
7490 
7491         if (!sd || !sd->nohz_idle)
7492                 goto unlock;
7493         sd->nohz_idle = 0;
7494 
7495         atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7496 unlock:
7497         rcu_read_unlock();
7498 }
7499 
7500 void set_cpu_sd_state_idle(void)
7501 {
7502         struct sched_domain *sd;
7503         int cpu = smp_processor_id();
7504 
7505         rcu_read_lock();
7506         sd = rcu_dereference(per_cpu(sd_busy, cpu));
7507 
7508         if (!sd || sd->nohz_idle)
7509                 goto unlock;
7510         sd->nohz_idle = 1;
7511 
7512         atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7513 unlock:
7514         rcu_read_unlock();
7515 }
7516 
7517 /*
7518  * This routine will record that the cpu is going idle with tick stopped.
7519  * This info will be used in performing idle load balancing in the future.
7520  */
7521 void nohz_balance_enter_idle(int cpu)
7522 {
7523         /*
7524          * If this cpu is going down, then nothing needs to be done.
7525          */
7526         if (!cpu_active(cpu))
7527                 return;
7528 
7529         if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7530                 return;
7531 
7532         /*
7533          * If we're a completely isolated CPU, we don't play.
7534          */
7535         if (on_null_domain(cpu_rq(cpu)))
7536                 return;
7537 
7538         cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7539         atomic_inc(&nohz.nr_cpus);
7540         set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7541 }
7542 
7543 static int sched_ilb_notifier(struct notifier_block *nfb,
7544                                         unsigned long action, void *hcpu)
7545 {
7546         switch (action & ~CPU_TASKS_FROZEN) {
7547         case CPU_DYING:
7548                 nohz_balance_exit_idle(smp_processor_id());
7549                 return NOTIFY_OK;
7550         default:
7551                 return NOTIFY_DONE;
7552         }
7553 }
7554 #endif
7555 
7556 static DEFINE_SPINLOCK(balancing);
7557 
7558 /*
7559  * Scale the max load_balance interval with the number of CPUs in the system.
7560  * This trades load-balance latency on larger machines for less cross talk.
7561  */
7562 void update_max_interval(void)
7563 {
7564         max_load_balance_interval = HZ*num_online_cpus()/10;
7565 }
7566 
7567 /*
7568  * It checks each scheduling domain to see if it is due to be balanced,
7569  * and initiates a balancing operation if so.
7570  *
7571  * Balancing parameters are set up in init_sched_domains.
7572  */
7573 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7574 {
7575         int continue_balancing = 1;
7576         int cpu = rq->cpu;
7577         unsigned long interval;
7578         struct sched_domain *sd;
7579         /* Earliest time when we have to do rebalance again */
7580         unsigned long next_balance = jiffies + 60*HZ;
7581         int update_next_balance = 0;
7582         int need_serialize, need_decay = 0;
7583         u64 max_cost = 0;
7584 
7585         update_blocked_averages(cpu);
7586 
7587         rcu_read_lock();
7588         for_each_domain(cpu, sd) {
7589                 /*
7590                  * Decay the newidle max times here because this is a regular
7591                  * visit to all the domains. Decay ~1% per second.
7592                  */
7593                 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7594                         sd->max_newidle_lb_cost =
7595                                 (sd->max_newidle_lb_cost * 253) / 256;
7596                         sd->next_decay_max_lb_cost = jiffies + HZ;
7597                         need_decay = 1;
7598                 }
7599                 max_cost += sd->max_newidle_lb_cost;
7600 
7601                 if (!(sd->flags & SD_LOAD_BALANCE))
7602                         continue;
7603 
7604                 /*
7605                  * Stop the load balance at this level. There is another
7606                  * CPU in our sched group which is doing load balancing more
7607                  * actively.
7608                  */
7609                 if (!continue_balancing) {
7610                         if (need_decay)
7611                                 continue;
7612                         break;
7613                 }
7614 
7615                 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7616 
7617                 need_serialize = sd->flags & SD_SERIALIZE;
7618                 if (need_serialize) {
7619                         if (!spin_trylock(&balancing))
7620                                 goto out;
7621                 }
7622 
7623                 if (time_after_eq(jiffies, sd->last_balance + interval)) {
7624                         if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7625                                 /*
7626                                  * The LBF_DST_PINNED logic could have changed
7627                                  * env->dst_cpu, so we can't know our idle
7628                                  * state even if we migrated tasks. Update it.
7629                                  */
7630                                 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7631                         }
7632                         sd->last_balance = jiffies;
7633                         interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7634                 }
7635                 if (need_serialize)
7636                         spin_unlock(&balancing);
7637 out:
7638                 if (time_after(next_balance, sd->last_balance + interval)) {
7639                         next_balance = sd->last_balance + interval;
7640                         update_next_balance = 1;
7641                 }
7642         }
7643         if (need_decay) {
7644                 /*
7645                  * Ensure the rq-wide value also decays but keep it at a
7646                  * reasonable floor to avoid funnies with rq->avg_idle.
7647                  */
7648                 rq->max_idle_balance_cost =
7649                         max((u64)sysctl_sched_migration_cost, max_cost);
7650         }
7651         rcu_read_unlock();
7652 
7653         /*
7654          * next_balance will be updated only when there is a need.
7655          * When the cpu is attached to null domain for ex, it will not be
7656          * updated.
7657          */
7658         if (likely(update_next_balance)) {
7659                 rq->next_balance = next_balance;
7660 
7661 #ifdef CONFIG_NO_HZ_COMMON
7662                 /*
7663                  * If this CPU has been elected to perform the nohz idle
7664                  * balance. Other idle CPUs have already rebalanced with
7665                  * nohz_idle_balance() and nohz.next_balance has been
7666                  * updated accordingly. This CPU is now running the idle load
7667                  * balance for itself and we need to update the
7668                  * nohz.next_balance accordingly.
7669                  */
7670                 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
7671                         nohz.next_balance = rq->next_balance;
7672 #endif
7673         }
7674 }
7675 
7676 #ifdef CONFIG_NO_HZ_COMMON
7677 /*
7678  * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7679  * rebalancing for all the cpus for whom scheduler ticks are stopped.
7680  */
7681 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7682 {
7683         int this_cpu = this_rq->cpu;
7684         struct rq *rq;
7685         int balance_cpu;
7686         /* Earliest time when we have to do rebalance again */
7687         unsigned long next_balance = jiffies + 60*HZ;
7688         int update_next_balance = 0;
7689 
7690         if (idle != CPU_IDLE ||
7691             !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7692                 goto end;
7693 
7694         for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7695                 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7696                         continue;
7697 
7698                 /*
7699                  * If this cpu gets work to do, stop the load balancing
7700                  * work being done for other cpus. Next load
7701                  * balancing owner will pick it up.
7702                  */
7703                 if (need_resched())
7704                         break;
7705 
7706                 rq = cpu_rq(balance_cpu);
7707 
7708                 /*
7709                  * If time for next balance is due,
7710                  * do the balance.
7711                  */
7712                 if (time_after_eq(jiffies, rq->next_balance)) {
7713                         raw_spin_lock_irq(&rq->lock);
7714                         update_rq_clock(rq);
7715                         update_idle_cpu_load(rq);
7716                         raw_spin_unlock_irq(&rq->lock);
7717                         rebalance_domains(rq, CPU_IDLE);
7718                 }
7719 
7720                 if (time_after(next_balance, rq->next_balance)) {
7721                         next_balance = rq->next_balance;
7722                         update_next_balance = 1;
7723                 }
7724         }
7725 
7726         /*
7727          * next_balance will be updated only when there is a need.
7728          * When the CPU is attached to null domain for ex, it will not be
7729          * updated.
7730          */
7731         if (likely(update_next_balance))
7732                 nohz.next_balance = next_balance;
7733 end:
7734         clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7735 }
7736 
7737 /*
7738  * Current heuristic for kicking the idle load balancer in the presence
7739  * of an idle cpu in the system.
7740  *   - This rq has more than one task.
7741  *   - This rq has at least one CFS task and the capacity of the CPU is
7742  *     significantly reduced because of RT tasks or IRQs.
7743  *   - At parent of LLC scheduler domain level, this cpu's scheduler group has
7744  *     multiple busy cpu.
7745  *   - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7746  *     domain span are idle.
7747  */
7748 static inline bool nohz_kick_needed(struct rq *rq)
7749 {
7750         unsigned long now = jiffies;
7751         struct sched_domain *sd;
7752         struct sched_group_capacity *sgc;
7753         int nr_busy, cpu = rq->cpu;
7754         bool kick = false;
7755 
7756         if (unlikely(rq->idle_balance))
7757                 return false;
7758 
7759        /*
7760         * We may be recently in ticked or tickless idle mode. At the first
7761         * busy tick after returning from idle, we will update the busy stats.
7762         */
7763         set_cpu_sd_state_busy();
7764         nohz_balance_exit_idle(cpu);
7765 
7766         /*
7767          * None are in tickless mode and hence no need for NOHZ idle load
7768          * balancing.
7769          */
7770         if (likely(!atomic_read(&nohz.nr_cpus)))
7771                 return false;
7772 
7773         if (time_before(now, nohz.next_balance))
7774                 return false;
7775 
7776         if (rq->nr_running >= 2)
7777                 return true;
7778 
7779         rcu_read_lock();
7780         sd = rcu_dereference(per_cpu(sd_busy, cpu));
7781         if (sd) {
7782                 sgc = sd->groups->sgc;
7783                 nr_busy = atomic_read(&sgc->nr_busy_cpus);
7784 
7785                 if (nr_busy > 1) {
7786                         kick = true;
7787                         goto unlock;
7788                 }
7789 
7790         }
7791 
7792         sd = rcu_dereference(rq->sd);
7793         if (sd) {
7794                 if ((rq->cfs.h_nr_running >= 1) &&
7795                                 check_cpu_capacity(rq, sd)) {
7796                         kick = true;
7797                         goto unlock;
7798                 }
7799         }
7800 
7801         sd = rcu_dereference(per_cpu(sd_asym, cpu));
7802         if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7803                                   sched_domain_span(sd)) < cpu)) {
7804                 kick = true;
7805                 goto unlock;
7806         }
7807 
7808 unlock:
7809         rcu_read_unlock();
7810         return kick;
7811 }
7812 #else
7813 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7814 #endif
7815 
7816 /*
7817  * run_rebalance_domains is triggered when needed from the scheduler tick.
7818  * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7819  */
7820 static void run_rebalance_domains(struct softirq_action *h)
7821 {
7822         struct rq *this_rq = this_rq();
7823         enum cpu_idle_type idle = this_rq->idle_balance ?
7824                                                 CPU_IDLE : CPU_NOT_IDLE;
7825 
7826         /*
7827          * If this cpu has a pending nohz_balance_kick, then do the
7828          * balancing on behalf of the other idle cpus whose ticks are
7829          * stopped. Do nohz_idle_balance *before* rebalance_domains to
7830          * give the idle cpus a chance to load balance. Else we may
7831          * load balance only within the local sched_domain hierarchy
7832          * and abort nohz_idle_balance altogether if we pull some load.
7833          */
7834         nohz_idle_balance(this_rq, idle);
7835         rebalance_domains(this_rq, idle);
7836 }
7837 
7838 /*
7839  * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7840  */
7841 void trigger_load_balance(struct rq *rq)
7842 {
7843         /* Don't need to rebalance while attached to NULL domain */
7844         if (unlikely(on_null_domain(rq)))
7845                 return;
7846 
7847         if (time_after_eq(jiffies, rq->next_balance))
7848                 raise_softirq(SCHED_SOFTIRQ);
7849 #ifdef CONFIG_NO_HZ_COMMON
7850         if (nohz_kick_needed(rq))
7851                 nohz_balancer_kick();
7852 #endif
7853 }
7854 
7855 static void rq_online_fair(struct rq *rq)
7856 {
7857         update_sysctl();
7858 
7859         update_runtime_enabled(rq);
7860 }
7861 
7862 static void rq_offline_fair(struct rq *rq)
7863 {
7864         update_sysctl();
7865 
7866         /* Ensure any throttled groups are reachable by pick_next_task */
7867         unthrottle_offline_cfs_rqs(rq);
7868 }
7869 
7870 #endif /* CONFIG_SMP */
7871 
7872 /*
7873  * scheduler tick hitting a task of our scheduling class:
7874  */
7875 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7876 {
7877         struct cfs_rq *cfs_rq;
7878         struct sched_entity *se = &curr->se;
7879 
7880         for_each_sched_entity(se) {
7881                 cfs_rq = cfs_rq_of(se);
7882                 entity_tick(cfs_rq, se, queued);
7883         }
7884 
7885         if (static_branch_unlikely(&sched_numa_balancing))
7886                 task_tick_numa(rq, curr);
7887 }
7888 
7889 /*
7890  * called on fork with the child task as argument from the parent's context
7891  *  - child not yet on the tasklist
7892  *  - preemption disabled
7893  */
7894 static void task_fork_fair(struct task_struct *p)
7895 {
7896         struct cfs_rq *cfs_rq;
7897         struct sched_entity *se = &p->se, *curr;
7898         int this_cpu = smp_processor_id();
7899         struct rq *rq = this_rq();
7900         unsigned long flags;
7901 
7902         raw_spin_lock_irqsave(&rq->lock, flags);
7903 
7904         update_rq_clock(rq);
7905 
7906         cfs_rq = task_cfs_rq(current);
7907         curr = cfs_rq->curr;
7908 
7909         /*
7910          * Not only the cpu but also the task_group of the parent might have
7911          * been changed after parent->se.parent,cfs_rq were copied to
7912          * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7913          * of child point to valid ones.
7914          */
7915         rcu_read_lock();
7916         __set_task_cpu(p, this_cpu);
7917         rcu_read_unlock();
7918 
7919         update_curr(cfs_rq);
7920 
7921         if (curr)
7922                 se->vruntime = curr->vruntime;
7923         place_entity(cfs_rq, se, 1);
7924 
7925         if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
7926                 /*
7927                  * Upon rescheduling, sched_class::put_prev_task() will place
7928                  * 'current' within the tree based on its new key value.
7929                  */
7930                 swap(curr->vruntime, se->vruntime);
7931                 resched_curr(rq);
7932         }
7933 
7934         se->vruntime -= cfs_rq->min_vruntime;
7935 
7936         raw_spin_unlock_irqrestore(&rq->lock, flags);
7937 }
7938 
7939 /*
7940  * Priority of the task has changed. Check to see if we preempt
7941  * the current task.
7942  */
7943 static void
7944 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
7945 {
7946         if (!task_on_rq_queued(p))
7947                 return;
7948 
7949         /*
7950          * Reschedule if we are currently running on this runqueue and
7951          * our priority decreased, or if we are not currently running on
7952          * this runqueue and our priority is higher than the current's
7953          */
7954         if (rq->curr == p) {
7955                 if (p->prio > oldprio)
7956                         resched_curr(rq);
7957         } else
7958                 check_preempt_curr(rq, p, 0);
7959 }
7960 
7961 static inline bool vruntime_normalized(struct task_struct *p)
7962 {
7963         struct sched_entity *se = &p->se;
7964 
7965         /*
7966          * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
7967          * the dequeue_entity(.flags=0) will already have normalized the
7968          * vruntime.
7969          */
7970         if (p->on_rq)
7971                 return true;
7972 
7973         /*
7974          * When !on_rq, vruntime of the task has usually NOT been normalized.
7975          * But there are some cases where it has already been normalized:
7976          *
7977          * - A forked child which is waiting for being woken up by
7978          *   wake_up_new_task().
7979          * - A task which has been woken up by try_to_wake_up() and
7980          *   waiting for actually being woken up by sched_ttwu_pending().
7981          */
7982         if (!se->sum_exec_runtime || p->state == TASK_WAKING)
7983                 return true;
7984 
7985         return false;
7986 }
7987 
7988 static void detach_task_cfs_rq(struct task_struct *p)
7989 {
7990         struct sched_entity *se = &p->se;
7991         struct cfs_rq *cfs_rq = cfs_rq_of(se);
7992 
7993         if (!vruntime_normalized(p)) {
7994                 /*
7995                  * Fix up our vruntime so that the current sleep doesn't
7996                  * cause 'unlimited' sleep bonus.
7997                  */
7998                 place_entity(cfs_rq, se, 0);
7999                 se->vruntime -= cfs_rq->min_vruntime;
8000         }
8001 
8002         /* Catch up with the cfs_rq and remove our load when we leave */
8003         detach_entity_load_avg(cfs_rq, se);
8004 }
8005 
8006 static void attach_task_cfs_rq(struct task_struct *p)
8007 {
8008         struct sched_entity *se = &p->se;
8009         struct cfs_rq *cfs_rq = cfs_rq_of(se);
8010 
8011 #ifdef CONFIG_FAIR_GROUP_SCHED
8012         /*
8013          * Since the real-depth could have been changed (only FAIR
8014          * class maintain depth value), reset depth properly.
8015          */
8016         se->depth = se->parent ? se->parent->depth + 1 : 0;
8017 #endif
8018 
8019         /* Synchronize task with its cfs_rq */
8020         attach_entity_load_avg(cfs_rq, se);
8021 
8022         if (!vruntime_normalized(p))
8023                 se->vruntime += cfs_rq->min_vruntime;
8024 }
8025 
8026 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8027 {
8028         detach_task_cfs_rq(p);
8029 }
8030 
8031 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8032 {
8033         attach_task_cfs_rq(p);
8034 
8035         if (task_on_rq_queued(p)) {
8036                 /*
8037                  * We were most likely switched from sched_rt, so
8038                  * kick off the schedule if running, otherwise just see
8039                  * if we can still preempt the current task.
8040                  */
8041                 if (rq->curr == p)
8042                         resched_curr(rq);
8043                 else
8044                         check_preempt_curr(rq, p, 0);
8045         }
8046 }
8047 
8048 /* Account for a task changing its policy or group.
8049  *
8050  * This routine is mostly called to set cfs_rq->curr field when a task
8051  * migrates between groups/classes.
8052  */
8053 static void set_curr_task_fair(struct rq *rq)
8054 {
8055         struct sched_entity *se = &rq->curr->se;
8056 
8057         for_each_sched_entity(se) {
8058                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8059 
8060                 set_next_entity(cfs_rq, se);
8061                 /* ensure bandwidth has been allocated on our new cfs_rq */
8062                 account_cfs_rq_runtime(cfs_rq, 0);
8063         }
8064 }
8065 
8066 void init_cfs_rq(struct cfs_rq *cfs_rq)
8067 {
8068         cfs_rq->tasks_timeline = RB_ROOT;
8069         cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8070 #ifndef CONFIG_64BIT
8071         cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8072 #endif
8073 #ifdef CONFIG_SMP
8074         atomic_long_set(&cfs_rq->removed_load_avg, 0);
8075         atomic_long_set(&cfs_rq->removed_util_avg, 0);
8076 #endif
8077 }
8078 
8079 #ifdef CONFIG_FAIR_GROUP_SCHED
8080 static void task_move_group_fair(struct task_struct *p)
8081 {
8082         detach_task_cfs_rq(p);
8083         set_task_rq(p, task_cpu(p));
8084 
8085 #ifdef CONFIG_SMP
8086         /* Tell se's cfs_rq has been changed -- migrated */
8087         p->se.avg.last_update_time = 0;
8088 #endif
8089         attach_task_cfs_rq(p);
8090 }
8091 
8092 void free_fair_sched_group(struct task_group *tg)
8093 {
8094         int i;
8095 
8096         destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8097 
8098         for_each_possible_cpu(i) {
8099                 if (tg->cfs_rq)
8100                         kfree(tg->cfs_rq[i]);
8101                 if (tg->se) {
8102                         if (tg->se[i])
8103                                 remove_entity_load_avg(tg->se[i]);
8104                         kfree(tg->se[i]);
8105                 }
8106         }
8107 
8108         kfree(tg->cfs_rq);
8109         kfree(tg->se);
8110 }
8111 
8112 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8113 {
8114         struct cfs_rq *cfs_rq;
8115         struct sched_entity *se;
8116         int i;
8117 
8118         tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8119         if (!tg->cfs_rq)
8120                 goto err;
8121         tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8122         if (!tg->se)
8123                 goto err;
8124 
8125         tg->shares = NICE_0_LOAD;
8126 
8127         init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8128 
8129         for_each_possible_cpu(i) {
8130                 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8131                                       GFP_KERNEL, cpu_to_node(i));
8132                 if (!cfs_rq)
8133                         goto err;
8134 
8135                 se = kzalloc_node(sizeof(struct sched_entity),
8136                                   GFP_KERNEL, cpu_to_node(i));
8137                 if (!se)
8138                         goto err_free_rq;
8139 
8140                 init_cfs_rq(cfs_rq);
8141                 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8142                 init_entity_runnable_average(se);
8143         }
8144 
8145         return 1;
8146 
8147 err_free_rq:
8148         kfree(cfs_rq);
8149 err:
8150         return 0;
8151 }
8152 
8153 void unregister_fair_sched_group(struct task_group *tg, int cpu)
8154 {
8155         struct rq *rq = cpu_rq(cpu);
8156         unsigned long flags;
8157 
8158         /*
8159         * Only empty task groups can be destroyed; so we can speculatively
8160         * check on_list without danger of it being re-added.
8161         */
8162         if (!tg->cfs_rq[cpu]->on_list)
8163                 return;
8164 
8165         raw_spin_lock_irqsave(&rq->lock, flags);
8166         list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8167         raw_spin_unlock_irqrestore(&rq->lock, flags);
8168 }
8169 
8170 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8171                         struct sched_entity *se, int cpu,
8172                         struct sched_entity *parent)
8173 {
8174         struct rq *rq = cpu_rq(cpu);
8175 
8176         cfs_rq->tg = tg;
8177         cfs_rq->rq = rq;
8178         init_cfs_rq_runtime(cfs_rq);
8179 
8180         tg->cfs_rq[cpu] = cfs_rq;
8181         tg->se[cpu] = se;
8182 
8183         /* se could be NULL for root_task_group */
8184         if (!se)
8185                 return;
8186 
8187         if (!parent) {
8188                 se->cfs_rq = &rq->cfs;
8189                 se->depth = 0;
8190         } else {
8191                 se->cfs_rq = parent->my_q;
8192                 se->depth = parent->depth + 1;
8193         }
8194 
8195         se->my_q = cfs_rq;
8196         /* guarantee group entities always have weight */
8197         update_load_set(&se->load, NICE_0_LOAD);
8198         se->parent = parent;
8199 }
8200 
8201 static DEFINE_MUTEX(shares_mutex);
8202 
8203 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8204 {
8205         int i;
8206         unsigned long flags;
8207 
8208         /*
8209          * We can't change the weight of the root cgroup.
8210          */
8211         if (!tg->se[0])
8212                 return -EINVAL;
8213 
8214         shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8215 
8216         mutex_lock(&shares_mutex);
8217         if (tg->shares == shares)
8218                 goto done;
8219 
8220         tg->shares = shares;
8221         for_each_possible_cpu(i) {
8222                 struct rq *rq = cpu_rq(i);
8223                 struct sched_entity *se;
8224 
8225                 se = tg->se[i];
8226                 /* Propagate contribution to hierarchy */
8227                 raw_spin_lock_irqsave(&rq->lock, flags);
8228 
8229                 /* Possible calls to update_curr() need rq clock */
8230                 update_rq_clock(rq);
8231                 for_each_sched_entity(se)
8232                         update_cfs_shares(group_cfs_rq(se));
8233                 raw_spin_unlock_irqrestore(&rq->lock, flags);
8234         }
8235 
8236 done:
8237         mutex_unlock(&shares_mutex);
8238         return 0;
8239 }
8240 #else /* CONFIG_FAIR_GROUP_SCHED */
8241 
8242 void free_fair_sched_group(struct task_group *tg) { }
8243 
8244 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8245 {
8246         return 1;
8247 }
8248 
8249 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
8250 
8251 #endif /* CONFIG_FAIR_GROUP_SCHED */
8252 
8253 
8254 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
8255 {
8256         struct sched_entity *se = &task->se;
8257         unsigned int rr_interval = 0;
8258 
8259         /*
8260          * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8261          * idle runqueue:
8262          */
8263         if (rq->cfs.load.weight)
8264                 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
8265 
8266         return rr_interval;
8267 }
8268 
8269 /*
8270  * All the scheduling class methods:
8271  */
8272 const struct sched_class fair_sched_class = {
8273         .next                   = &idle_sched_class,
8274         .enqueue_task           = enqueue_task_fair,
8275         .dequeue_task           = dequeue_task_fair,
8276         .yield_task             = yield_task_fair,
8277         .yield_to_task          = yield_to_task_fair,
8278 
8279         .check_preempt_curr     = check_preempt_wakeup,
8280 
8281         .pick_next_task         = pick_next_task_fair,
8282         .put_prev_task          = put_prev_task_fair,
8283 
8284 #ifdef CONFIG_SMP
8285         .select_task_rq         = select_task_rq_fair,
8286         .migrate_task_rq        = migrate_task_rq_fair,
8287 
8288         .rq_online              = rq_online_fair,
8289         .rq_offline             = rq_offline_fair,
8290 
8291         .task_waking            = task_waking_fair,
8292         .task_dead              = task_dead_fair,
8293         .set_cpus_allowed       = set_cpus_allowed_common,
8294 #endif
8295 
8296         .set_curr_task          = set_curr_task_fair,
8297         .task_tick              = task_tick_fair,
8298         .task_fork              = task_fork_fair,
8299 
8300