Version:  2.0.40 2.2.26 2.4.37 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 3.10 3.11 3.12 3.13 3.14 3.15 3.16 3.17

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 <pzijlstr@redhat.com>
 21  */
 22 
 23 #include <linux/latencytop.h>
 24 #include <linux/sched.h>
 25 #include <linux/cpumask.h>
 26 #include <linux/slab.h>
 27 #include <linux/profile.h>
 28 #include <linux/interrupt.h>
 29 #include <linux/mempolicy.h>
 30 #include <linux/migrate.h>
 31 #include <linux/task_work.h>
 32 
 33 #include <trace/events/sched.h>
 34 
 35 #include "sched.h"
 36 
 37 /*
 38  * Targeted preemption latency for CPU-bound tasks:
 39  * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
 40  *
 41  * NOTE: this latency value is not the same as the concept of
 42  * 'timeslice length' - timeslices in CFS are of variable length
 43  * and have no persistent notion like in traditional, time-slice
 44  * based scheduling concepts.
 45  *
 46  * (to see the precise effective timeslice length of your workload,
 47  *  run vmstat and monitor the context-switches (cs) field)
 48  */
 49 unsigned int sysctl_sched_latency = 6000000ULL;
 50 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
 51 
 52 /*
 53  * The initial- and re-scaling of tunables is configurable
 54  * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
 55  *
 56  * Options are:
 57  * SCHED_TUNABLESCALING_NONE - unscaled, always *1
 58  * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
 59  * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
 60  */
 61 enum sched_tunable_scaling sysctl_sched_tunable_scaling
 62         = SCHED_TUNABLESCALING_LOG;
 63 
 64 /*
 65  * Minimal preemption granularity for CPU-bound tasks:
 66  * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
 67  */
 68 unsigned int sysctl_sched_min_granularity = 750000ULL;
 69 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
 70 
 71 /*
 72  * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
 73  */
 74 static unsigned int sched_nr_latency = 8;
 75 
 76 /*
 77  * After fork, child runs first. If set to 0 (default) then
 78  * parent will (try to) run first.
 79  */
 80 unsigned int sysctl_sched_child_runs_first __read_mostly;
 81 
 82 /*
 83  * SCHED_OTHER wake-up granularity.
 84  * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
 85  *
 86  * This option delays the preemption effects of decoupled workloads
 87  * and reduces their over-scheduling. Synchronous workloads will still
 88  * have immediate wakeup/sleep latencies.
 89  */
 90 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
 91 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
 92 
 93 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
 94 
 95 /*
 96  * The exponential sliding  window over which load is averaged for shares
 97  * distribution.
 98  * (default: 10msec)
 99  */
100 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
101 
102 #ifdef CONFIG_CFS_BANDWIDTH
103 /*
104  * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
105  * each time a cfs_rq requests quota.
106  *
107  * Note: in the case that the slice exceeds the runtime remaining (either due
108  * to consumption or the quota being specified to be smaller than the slice)
109  * we will always only issue the remaining available time.
110  *
111  * default: 5 msec, units: microseconds
112   */
113 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
114 #endif
115 
116 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
117 {
118         lw->weight += inc;
119         lw->inv_weight = 0;
120 }
121 
122 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
123 {
124         lw->weight -= dec;
125         lw->inv_weight = 0;
126 }
127 
128 static inline void update_load_set(struct load_weight *lw, unsigned long w)
129 {
130         lw->weight = w;
131         lw->inv_weight = 0;
132 }
133 
134 /*
135  * Increase the granularity value when there are more CPUs,
136  * because with more CPUs the 'effective latency' as visible
137  * to users decreases. But the relationship is not linear,
138  * so pick a second-best guess by going with the log2 of the
139  * number of CPUs.
140  *
141  * This idea comes from the SD scheduler of Con Kolivas:
142  */
143 static int get_update_sysctl_factor(void)
144 {
145         unsigned int cpus = min_t(int, num_online_cpus(), 8);
146         unsigned int factor;
147 
148         switch (sysctl_sched_tunable_scaling) {
149         case SCHED_TUNABLESCALING_NONE:
150                 factor = 1;
151                 break;
152         case SCHED_TUNABLESCALING_LINEAR:
153                 factor = cpus;
154                 break;
155         case SCHED_TUNABLESCALING_LOG:
156         default:
157                 factor = 1 + ilog2(cpus);
158                 break;
159         }
160 
161         return factor;
162 }
163 
164 static void update_sysctl(void)
165 {
166         unsigned int factor = get_update_sysctl_factor();
167 
168 #define SET_SYSCTL(name) \
169         (sysctl_##name = (factor) * normalized_sysctl_##name)
170         SET_SYSCTL(sched_min_granularity);
171         SET_SYSCTL(sched_latency);
172         SET_SYSCTL(sched_wakeup_granularity);
173 #undef SET_SYSCTL
174 }
175 
176 void sched_init_granularity(void)
177 {
178         update_sysctl();
179 }
180 
181 #define WMULT_CONST     (~0U)
182 #define WMULT_SHIFT     32
183 
184 static void __update_inv_weight(struct load_weight *lw)
185 {
186         unsigned long w;
187 
188         if (likely(lw->inv_weight))
189                 return;
190 
191         w = scale_load_down(lw->weight);
192 
193         if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
194                 lw->inv_weight = 1;
195         else if (unlikely(!w))
196                 lw->inv_weight = WMULT_CONST;
197         else
198                 lw->inv_weight = WMULT_CONST / w;
199 }
200 
201 /*
202  * delta_exec * weight / lw.weight
203  *   OR
204  * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
205  *
206  * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
207  * we're guaranteed shift stays positive because inv_weight is guaranteed to
208  * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
209  *
210  * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
211  * weight/lw.weight <= 1, and therefore our shift will also be positive.
212  */
213 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
214 {
215         u64 fact = scale_load_down(weight);
216         int shift = WMULT_SHIFT;
217 
218         __update_inv_weight(lw);
219 
220         if (unlikely(fact >> 32)) {
221                 while (fact >> 32) {
222                         fact >>= 1;
223                         shift--;
224                 }
225         }
226 
227         /* hint to use a 32x32->64 mul */
228         fact = (u64)(u32)fact * lw->inv_weight;
229 
230         while (fact >> 32) {
231                 fact >>= 1;
232                 shift--;
233         }
234 
235         return mul_u64_u32_shr(delta_exec, fact, shift);
236 }
237 
238 
239 const struct sched_class fair_sched_class;
240 
241 /**************************************************************
242  * CFS operations on generic schedulable entities:
243  */
244 
245 #ifdef CONFIG_FAIR_GROUP_SCHED
246 
247 /* cpu runqueue to which this cfs_rq is attached */
248 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
249 {
250         return cfs_rq->rq;
251 }
252 
253 /* An entity is a task if it doesn't "own" a runqueue */
254 #define entity_is_task(se)      (!se->my_q)
255 
256 static inline struct task_struct *task_of(struct sched_entity *se)
257 {
258 #ifdef CONFIG_SCHED_DEBUG
259         WARN_ON_ONCE(!entity_is_task(se));
260 #endif
261         return container_of(se, struct task_struct, se);
262 }
263 
264 /* Walk up scheduling entities hierarchy */
265 #define for_each_sched_entity(se) \
266                 for (; se; se = se->parent)
267 
268 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
269 {
270         return p->se.cfs_rq;
271 }
272 
273 /* runqueue on which this entity is (to be) queued */
274 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
275 {
276         return se->cfs_rq;
277 }
278 
279 /* runqueue "owned" by this group */
280 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
281 {
282         return grp->my_q;
283 }
284 
285 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
286                                        int force_update);
287 
288 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
289 {
290         if (!cfs_rq->on_list) {
291                 /*
292                  * Ensure we either appear before our parent (if already
293                  * enqueued) or force our parent to appear after us when it is
294                  * enqueued.  The fact that we always enqueue bottom-up
295                  * reduces this to two cases.
296                  */
297                 if (cfs_rq->tg->parent &&
298                     cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
299                         list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
300                                 &rq_of(cfs_rq)->leaf_cfs_rq_list);
301                 } else {
302                         list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
303                                 &rq_of(cfs_rq)->leaf_cfs_rq_list);
304                 }
305 
306                 cfs_rq->on_list = 1;
307                 /* We should have no load, but we need to update last_decay. */
308                 update_cfs_rq_blocked_load(cfs_rq, 0);
309         }
310 }
311 
312 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
313 {
314         if (cfs_rq->on_list) {
315                 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
316                 cfs_rq->on_list = 0;
317         }
318 }
319 
320 /* Iterate thr' all leaf cfs_rq's on a runqueue */
321 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
322         list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
323 
324 /* Do the two (enqueued) entities belong to the same group ? */
325 static inline struct cfs_rq *
326 is_same_group(struct sched_entity *se, struct sched_entity *pse)
327 {
328         if (se->cfs_rq == pse->cfs_rq)
329                 return se->cfs_rq;
330 
331         return NULL;
332 }
333 
334 static inline struct sched_entity *parent_entity(struct sched_entity *se)
335 {
336         return se->parent;
337 }
338 
339 static void
340 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
341 {
342         int se_depth, pse_depth;
343 
344         /*
345          * preemption test can be made between sibling entities who are in the
346          * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
347          * both tasks until we find their ancestors who are siblings of common
348          * parent.
349          */
350 
351         /* First walk up until both entities are at same depth */
352         se_depth = (*se)->depth;
353         pse_depth = (*pse)->depth;
354 
355         while (se_depth > pse_depth) {
356                 se_depth--;
357                 *se = parent_entity(*se);
358         }
359 
360         while (pse_depth > se_depth) {
361                 pse_depth--;
362                 *pse = parent_entity(*pse);
363         }
364 
365         while (!is_same_group(*se, *pse)) {
366                 *se = parent_entity(*se);
367                 *pse = parent_entity(*pse);
368         }
369 }
370 
371 #else   /* !CONFIG_FAIR_GROUP_SCHED */
372 
373 static inline struct task_struct *task_of(struct sched_entity *se)
374 {
375         return container_of(se, struct task_struct, se);
376 }
377 
378 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
379 {
380         return container_of(cfs_rq, struct rq, cfs);
381 }
382 
383 #define entity_is_task(se)      1
384 
385 #define for_each_sched_entity(se) \
386                 for (; se; se = NULL)
387 
388 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
389 {
390         return &task_rq(p)->cfs;
391 }
392 
393 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
394 {
395         struct task_struct *p = task_of(se);
396         struct rq *rq = task_rq(p);
397 
398         return &rq->cfs;
399 }
400 
401 /* runqueue "owned" by this group */
402 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
403 {
404         return NULL;
405 }
406 
407 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
408 {
409 }
410 
411 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
412 {
413 }
414 
415 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
416                 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
417 
418 static inline struct sched_entity *parent_entity(struct sched_entity *se)
419 {
420         return NULL;
421 }
422 
423 static inline void
424 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
425 {
426 }
427 
428 #endif  /* CONFIG_FAIR_GROUP_SCHED */
429 
430 static __always_inline
431 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
432 
433 /**************************************************************
434  * Scheduling class tree data structure manipulation methods:
435  */
436 
437 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
438 {
439         s64 delta = (s64)(vruntime - max_vruntime);
440         if (delta > 0)
441                 max_vruntime = vruntime;
442 
443         return max_vruntime;
444 }
445 
446 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
447 {
448         s64 delta = (s64)(vruntime - min_vruntime);
449         if (delta < 0)
450                 min_vruntime = vruntime;
451 
452         return min_vruntime;
453 }
454 
455 static inline int entity_before(struct sched_entity *a,
456                                 struct sched_entity *b)
457 {
458         return (s64)(a->vruntime - b->vruntime) < 0;
459 }
460 
461 static void update_min_vruntime(struct cfs_rq *cfs_rq)
462 {
463         u64 vruntime = cfs_rq->min_vruntime;
464 
465         if (cfs_rq->curr)
466                 vruntime = cfs_rq->curr->vruntime;
467 
468         if (cfs_rq->rb_leftmost) {
469                 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
470                                                    struct sched_entity,
471                                                    run_node);
472 
473                 if (!cfs_rq->curr)
474                         vruntime = se->vruntime;
475                 else
476                         vruntime = min_vruntime(vruntime, se->vruntime);
477         }
478 
479         /* ensure we never gain time by being placed backwards. */
480         cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
481 #ifndef CONFIG_64BIT
482         smp_wmb();
483         cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
484 #endif
485 }
486 
487 /*
488  * Enqueue an entity into the rb-tree:
489  */
490 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
491 {
492         struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
493         struct rb_node *parent = NULL;
494         struct sched_entity *entry;
495         int leftmost = 1;
496 
497         /*
498          * Find the right place in the rbtree:
499          */
500         while (*link) {
501                 parent = *link;
502                 entry = rb_entry(parent, struct sched_entity, run_node);
503                 /*
504                  * We dont care about collisions. Nodes with
505                  * the same key stay together.
506                  */
507                 if (entity_before(se, entry)) {
508                         link = &parent->rb_left;
509                 } else {
510                         link = &parent->rb_right;
511                         leftmost = 0;
512                 }
513         }
514 
515         /*
516          * Maintain a cache of leftmost tree entries (it is frequently
517          * used):
518          */
519         if (leftmost)
520                 cfs_rq->rb_leftmost = &se->run_node;
521 
522         rb_link_node(&se->run_node, parent, link);
523         rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
524 }
525 
526 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
527 {
528         if (cfs_rq->rb_leftmost == &se->run_node) {
529                 struct rb_node *next_node;
530 
531                 next_node = rb_next(&se->run_node);
532                 cfs_rq->rb_leftmost = next_node;
533         }
534 
535         rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
536 }
537 
538 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
539 {
540         struct rb_node *left = cfs_rq->rb_leftmost;
541 
542         if (!left)
543                 return NULL;
544 
545         return rb_entry(left, struct sched_entity, run_node);
546 }
547 
548 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
549 {
550         struct rb_node *next = rb_next(&se->run_node);
551 
552         if (!next)
553                 return NULL;
554 
555         return rb_entry(next, struct sched_entity, run_node);
556 }
557 
558 #ifdef CONFIG_SCHED_DEBUG
559 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
560 {
561         struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
562 
563         if (!last)
564                 return NULL;
565 
566         return rb_entry(last, struct sched_entity, run_node);
567 }
568 
569 /**************************************************************
570  * Scheduling class statistics methods:
571  */
572 
573 int sched_proc_update_handler(struct ctl_table *table, int write,
574                 void __user *buffer, size_t *lenp,
575                 loff_t *ppos)
576 {
577         int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
578         int factor = get_update_sysctl_factor();
579 
580         if (ret || !write)
581                 return ret;
582 
583         sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
584                                         sysctl_sched_min_granularity);
585 
586 #define WRT_SYSCTL(name) \
587         (normalized_sysctl_##name = sysctl_##name / (factor))
588         WRT_SYSCTL(sched_min_granularity);
589         WRT_SYSCTL(sched_latency);
590         WRT_SYSCTL(sched_wakeup_granularity);
591 #undef WRT_SYSCTL
592 
593         return 0;
594 }
595 #endif
596 
597 /*
598  * delta /= w
599  */
600 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
601 {
602         if (unlikely(se->load.weight != NICE_0_LOAD))
603                 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
604 
605         return delta;
606 }
607 
608 /*
609  * The idea is to set a period in which each task runs once.
610  *
611  * When there are too many tasks (sched_nr_latency) we have to stretch
612  * this period because otherwise the slices get too small.
613  *
614  * p = (nr <= nl) ? l : l*nr/nl
615  */
616 static u64 __sched_period(unsigned long nr_running)
617 {
618         u64 period = sysctl_sched_latency;
619         unsigned long nr_latency = sched_nr_latency;
620 
621         if (unlikely(nr_running > nr_latency)) {
622                 period = sysctl_sched_min_granularity;
623                 period *= nr_running;
624         }
625 
626         return period;
627 }
628 
629 /*
630  * We calculate the wall-time slice from the period by taking a part
631  * proportional to the weight.
632  *
633  * s = p*P[w/rw]
634  */
635 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
636 {
637         u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
638 
639         for_each_sched_entity(se) {
640                 struct load_weight *load;
641                 struct load_weight lw;
642 
643                 cfs_rq = cfs_rq_of(se);
644                 load = &cfs_rq->load;
645 
646                 if (unlikely(!se->on_rq)) {
647                         lw = cfs_rq->load;
648 
649                         update_load_add(&lw, se->load.weight);
650                         load = &lw;
651                 }
652                 slice = __calc_delta(slice, se->load.weight, load);
653         }
654         return slice;
655 }
656 
657 /*
658  * We calculate the vruntime slice of a to-be-inserted task.
659  *
660  * vs = s/w
661  */
662 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
663 {
664         return calc_delta_fair(sched_slice(cfs_rq, se), se);
665 }
666 
667 #ifdef CONFIG_SMP
668 static unsigned long task_h_load(struct task_struct *p);
669 
670 static inline void __update_task_entity_contrib(struct sched_entity *se);
671 
672 /* Give new task start runnable values to heavy its load in infant time */
673 void init_task_runnable_average(struct task_struct *p)
674 {
675         u32 slice;
676 
677         p->se.avg.decay_count = 0;
678         slice = sched_slice(task_cfs_rq(p), &p->se) >> 10;
679         p->se.avg.runnable_avg_sum = slice;
680         p->se.avg.runnable_avg_period = slice;
681         __update_task_entity_contrib(&p->se);
682 }
683 #else
684 void init_task_runnable_average(struct task_struct *p)
685 {
686 }
687 #endif
688 
689 /*
690  * Update the current task's runtime statistics.
691  */
692 static void update_curr(struct cfs_rq *cfs_rq)
693 {
694         struct sched_entity *curr = cfs_rq->curr;
695         u64 now = rq_clock_task(rq_of(cfs_rq));
696         u64 delta_exec;
697 
698         if (unlikely(!curr))
699                 return;
700 
701         delta_exec = now - curr->exec_start;
702         if (unlikely((s64)delta_exec <= 0))
703                 return;
704 
705         curr->exec_start = now;
706 
707         schedstat_set(curr->statistics.exec_max,
708                       max(delta_exec, curr->statistics.exec_max));
709 
710         curr->sum_exec_runtime += delta_exec;
711         schedstat_add(cfs_rq, exec_clock, delta_exec);
712 
713         curr->vruntime += calc_delta_fair(delta_exec, curr);
714         update_min_vruntime(cfs_rq);
715 
716         if (entity_is_task(curr)) {
717                 struct task_struct *curtask = task_of(curr);
718 
719                 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
720                 cpuacct_charge(curtask, delta_exec);
721                 account_group_exec_runtime(curtask, delta_exec);
722         }
723 
724         account_cfs_rq_runtime(cfs_rq, delta_exec);
725 }
726 
727 static inline void
728 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
729 {
730         schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
731 }
732 
733 /*
734  * Task is being enqueued - update stats:
735  */
736 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
737 {
738         /*
739          * Are we enqueueing a waiting task? (for current tasks
740          * a dequeue/enqueue event is a NOP)
741          */
742         if (se != cfs_rq->curr)
743                 update_stats_wait_start(cfs_rq, se);
744 }
745 
746 static void
747 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
748 {
749         schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
750                         rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
751         schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
752         schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
753                         rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
754 #ifdef CONFIG_SCHEDSTATS
755         if (entity_is_task(se)) {
756                 trace_sched_stat_wait(task_of(se),
757                         rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
758         }
759 #endif
760         schedstat_set(se->statistics.wait_start, 0);
761 }
762 
763 static inline void
764 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
765 {
766         /*
767          * Mark the end of the wait period if dequeueing a
768          * waiting task:
769          */
770         if (se != cfs_rq->curr)
771                 update_stats_wait_end(cfs_rq, se);
772 }
773 
774 /*
775  * We are picking a new current task - update its stats:
776  */
777 static inline void
778 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
779 {
780         /*
781          * We are starting a new run period:
782          */
783         se->exec_start = rq_clock_task(rq_of(cfs_rq));
784 }
785 
786 /**************************************************
787  * Scheduling class queueing methods:
788  */
789 
790 #ifdef CONFIG_NUMA_BALANCING
791 /*
792  * Approximate time to scan a full NUMA task in ms. The task scan period is
793  * calculated based on the tasks virtual memory size and
794  * numa_balancing_scan_size.
795  */
796 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
797 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
798 
799 /* Portion of address space to scan in MB */
800 unsigned int sysctl_numa_balancing_scan_size = 256;
801 
802 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
803 unsigned int sysctl_numa_balancing_scan_delay = 1000;
804 
805 static unsigned int task_nr_scan_windows(struct task_struct *p)
806 {
807         unsigned long rss = 0;
808         unsigned long nr_scan_pages;
809 
810         /*
811          * Calculations based on RSS as non-present and empty pages are skipped
812          * by the PTE scanner and NUMA hinting faults should be trapped based
813          * on resident pages
814          */
815         nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
816         rss = get_mm_rss(p->mm);
817         if (!rss)
818                 rss = nr_scan_pages;
819 
820         rss = round_up(rss, nr_scan_pages);
821         return rss / nr_scan_pages;
822 }
823 
824 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
825 #define MAX_SCAN_WINDOW 2560
826 
827 static unsigned int task_scan_min(struct task_struct *p)
828 {
829         unsigned int scan, floor;
830         unsigned int windows = 1;
831 
832         if (sysctl_numa_balancing_scan_size < MAX_SCAN_WINDOW)
833                 windows = MAX_SCAN_WINDOW / sysctl_numa_balancing_scan_size;
834         floor = 1000 / windows;
835 
836         scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
837         return max_t(unsigned int, floor, scan);
838 }
839 
840 static unsigned int task_scan_max(struct task_struct *p)
841 {
842         unsigned int smin = task_scan_min(p);
843         unsigned int smax;
844 
845         /* Watch for min being lower than max due to floor calculations */
846         smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
847         return max(smin, smax);
848 }
849 
850 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
851 {
852         rq->nr_numa_running += (p->numa_preferred_nid != -1);
853         rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
854 }
855 
856 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
857 {
858         rq->nr_numa_running -= (p->numa_preferred_nid != -1);
859         rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
860 }
861 
862 struct numa_group {
863         atomic_t refcount;
864 
865         spinlock_t lock; /* nr_tasks, tasks */
866         int nr_tasks;
867         pid_t gid;
868         struct list_head task_list;
869 
870         struct rcu_head rcu;
871         nodemask_t active_nodes;
872         unsigned long total_faults;
873         /*
874          * Faults_cpu is used to decide whether memory should move
875          * towards the CPU. As a consequence, these stats are weighted
876          * more by CPU use than by memory faults.
877          */
878         unsigned long *faults_cpu;
879         unsigned long faults[0];
880 };
881 
882 /* Shared or private faults. */
883 #define NR_NUMA_HINT_FAULT_TYPES 2
884 
885 /* Memory and CPU locality */
886 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
887 
888 /* Averaged statistics, and temporary buffers. */
889 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
890 
891 pid_t task_numa_group_id(struct task_struct *p)
892 {
893         return p->numa_group ? p->numa_group->gid : 0;
894 }
895 
896 static inline int task_faults_idx(int nid, int priv)
897 {
898         return NR_NUMA_HINT_FAULT_TYPES * nid + priv;
899 }
900 
901 static inline unsigned long task_faults(struct task_struct *p, int nid)
902 {
903         if (!p->numa_faults_memory)
904                 return 0;
905 
906         return p->numa_faults_memory[task_faults_idx(nid, 0)] +
907                 p->numa_faults_memory[task_faults_idx(nid, 1)];
908 }
909 
910 static inline unsigned long group_faults(struct task_struct *p, int nid)
911 {
912         if (!p->numa_group)
913                 return 0;
914 
915         return p->numa_group->faults[task_faults_idx(nid, 0)] +
916                 p->numa_group->faults[task_faults_idx(nid, 1)];
917 }
918 
919 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
920 {
921         return group->faults_cpu[task_faults_idx(nid, 0)] +
922                 group->faults_cpu[task_faults_idx(nid, 1)];
923 }
924 
925 /*
926  * These return the fraction of accesses done by a particular task, or
927  * task group, on a particular numa node.  The group weight is given a
928  * larger multiplier, in order to group tasks together that are almost
929  * evenly spread out between numa nodes.
930  */
931 static inline unsigned long task_weight(struct task_struct *p, int nid)
932 {
933         unsigned long total_faults;
934 
935         if (!p->numa_faults_memory)
936                 return 0;
937 
938         total_faults = p->total_numa_faults;
939 
940         if (!total_faults)
941                 return 0;
942 
943         return 1000 * task_faults(p, nid) / total_faults;
944 }
945 
946 static inline unsigned long group_weight(struct task_struct *p, int nid)
947 {
948         if (!p->numa_group || !p->numa_group->total_faults)
949                 return 0;
950 
951         return 1000 * group_faults(p, nid) / p->numa_group->total_faults;
952 }
953 
954 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
955                                 int src_nid, int dst_cpu)
956 {
957         struct numa_group *ng = p->numa_group;
958         int dst_nid = cpu_to_node(dst_cpu);
959         int last_cpupid, this_cpupid;
960 
961         this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
962 
963         /*
964          * Multi-stage node selection is used in conjunction with a periodic
965          * migration fault to build a temporal task<->page relation. By using
966          * a two-stage filter we remove short/unlikely relations.
967          *
968          * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
969          * a task's usage of a particular page (n_p) per total usage of this
970          * page (n_t) (in a given time-span) to a probability.
971          *
972          * Our periodic faults will sample this probability and getting the
973          * same result twice in a row, given these samples are fully
974          * independent, is then given by P(n)^2, provided our sample period
975          * is sufficiently short compared to the usage pattern.
976          *
977          * This quadric squishes small probabilities, making it less likely we
978          * act on an unlikely task<->page relation.
979          */
980         last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
981         if (!cpupid_pid_unset(last_cpupid) &&
982                                 cpupid_to_nid(last_cpupid) != dst_nid)
983                 return false;
984 
985         /* Always allow migrate on private faults */
986         if (cpupid_match_pid(p, last_cpupid))
987                 return true;
988 
989         /* A shared fault, but p->numa_group has not been set up yet. */
990         if (!ng)
991                 return true;
992 
993         /*
994          * Do not migrate if the destination is not a node that
995          * is actively used by this numa group.
996          */
997         if (!node_isset(dst_nid, ng->active_nodes))
998                 return false;
999 
1000         /*
1001          * Source is a node that is not actively used by this
1002          * numa group, while the destination is. Migrate.
1003          */
1004         if (!node_isset(src_nid, ng->active_nodes))
1005                 return true;
1006 
1007         /*
1008          * Both source and destination are nodes in active
1009          * use by this numa group. Maximize memory bandwidth
1010          * by migrating from more heavily used groups, to less
1011          * heavily used ones, spreading the load around.
1012          * Use a 1/4 hysteresis to avoid spurious page movement.
1013          */
1014         return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1015 }
1016 
1017 static unsigned long weighted_cpuload(const int cpu);
1018 static unsigned long source_load(int cpu, int type);
1019 static unsigned long target_load(int cpu, int type);
1020 static unsigned long capacity_of(int cpu);
1021 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1022 
1023 /* Cached statistics for all CPUs within a node */
1024 struct numa_stats {
1025         unsigned long nr_running;
1026         unsigned long load;
1027 
1028         /* Total compute capacity of CPUs on a node */
1029         unsigned long compute_capacity;
1030 
1031         /* Approximate capacity in terms of runnable tasks on a node */
1032         unsigned long task_capacity;
1033         int has_free_capacity;
1034 };
1035 
1036 /*
1037  * XXX borrowed from update_sg_lb_stats
1038  */
1039 static void update_numa_stats(struct numa_stats *ns, int nid)
1040 {
1041         int cpu, cpus = 0;
1042 
1043         memset(ns, 0, sizeof(*ns));
1044         for_each_cpu(cpu, cpumask_of_node(nid)) {
1045                 struct rq *rq = cpu_rq(cpu);
1046 
1047                 ns->nr_running += rq->nr_running;
1048                 ns->load += weighted_cpuload(cpu);
1049                 ns->compute_capacity += capacity_of(cpu);
1050 
1051                 cpus++;
1052         }
1053 
1054         /*
1055          * If we raced with hotplug and there are no CPUs left in our mask
1056          * the @ns structure is NULL'ed and task_numa_compare() will
1057          * not find this node attractive.
1058          *
1059          * We'll either bail at !has_free_capacity, or we'll detect a huge
1060          * imbalance and bail there.
1061          */
1062         if (!cpus)
1063                 return;
1064 
1065         ns->task_capacity =
1066                 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE);
1067         ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1068 }
1069 
1070 struct task_numa_env {
1071         struct task_struct *p;
1072 
1073         int src_cpu, src_nid;
1074         int dst_cpu, dst_nid;
1075 
1076         struct numa_stats src_stats, dst_stats;
1077 
1078         int imbalance_pct;
1079 
1080         struct task_struct *best_task;
1081         long best_imp;
1082         int best_cpu;
1083 };
1084 
1085 static void task_numa_assign(struct task_numa_env *env,
1086                              struct task_struct *p, long imp)
1087 {
1088         if (env->best_task)
1089                 put_task_struct(env->best_task);
1090         if (p)
1091                 get_task_struct(p);
1092 
1093         env->best_task = p;
1094         env->best_imp = imp;
1095         env->best_cpu = env->dst_cpu;
1096 }
1097 
1098 static bool load_too_imbalanced(long src_load, long dst_load,
1099                                 struct task_numa_env *env)
1100 {
1101         long imb, old_imb;
1102         long orig_src_load, orig_dst_load;
1103         long src_capacity, dst_capacity;
1104 
1105         /*
1106          * The load is corrected for the CPU capacity available on each node.
1107          *
1108          * src_load        dst_load
1109          * ------------ vs ---------
1110          * src_capacity    dst_capacity
1111          */
1112         src_capacity = env->src_stats.compute_capacity;
1113         dst_capacity = env->dst_stats.compute_capacity;
1114 
1115         /* We care about the slope of the imbalance, not the direction. */
1116         if (dst_load < src_load)
1117                 swap(dst_load, src_load);
1118 
1119         /* Is the difference below the threshold? */
1120         imb = dst_load * src_capacity * 100 -
1121               src_load * dst_capacity * env->imbalance_pct;
1122         if (imb <= 0)
1123                 return false;
1124 
1125         /*
1126          * The imbalance is above the allowed threshold.
1127          * Compare it with the old imbalance.
1128          */
1129         orig_src_load = env->src_stats.load;
1130         orig_dst_load = env->dst_stats.load;
1131 
1132         if (orig_dst_load < orig_src_load)
1133                 swap(orig_dst_load, orig_src_load);
1134 
1135         old_imb = orig_dst_load * src_capacity * 100 -
1136                   orig_src_load * dst_capacity * env->imbalance_pct;
1137 
1138         /* Would this change make things worse? */
1139         return (imb > old_imb);
1140 }
1141 
1142 /*
1143  * This checks if the overall compute and NUMA accesses of the system would
1144  * be improved if the source tasks was migrated to the target dst_cpu taking
1145  * into account that it might be best if task running on the dst_cpu should
1146  * be exchanged with the source task
1147  */
1148 static void task_numa_compare(struct task_numa_env *env,
1149                               long taskimp, long groupimp)
1150 {
1151         struct rq *src_rq = cpu_rq(env->src_cpu);
1152         struct rq *dst_rq = cpu_rq(env->dst_cpu);
1153         struct task_struct *cur;
1154         long src_load, dst_load;
1155         long load;
1156         long imp = env->p->numa_group ? groupimp : taskimp;
1157         long moveimp = imp;
1158 
1159         rcu_read_lock();
1160         cur = ACCESS_ONCE(dst_rq->curr);
1161         if (cur->pid == 0) /* idle */
1162                 cur = NULL;
1163 
1164         /*
1165          * "imp" is the fault differential for the source task between the
1166          * source and destination node. Calculate the total differential for
1167          * the source task and potential destination task. The more negative
1168          * the value is, the more rmeote accesses that would be expected to
1169          * be incurred if the tasks were swapped.
1170          */
1171         if (cur) {
1172                 /* Skip this swap candidate if cannot move to the source cpu */
1173                 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1174                         goto unlock;
1175 
1176                 /*
1177                  * If dst and source tasks are in the same NUMA group, or not
1178                  * in any group then look only at task weights.
1179                  */
1180                 if (cur->numa_group == env->p->numa_group) {
1181                         imp = taskimp + task_weight(cur, env->src_nid) -
1182                               task_weight(cur, env->dst_nid);
1183                         /*
1184                          * Add some hysteresis to prevent swapping the
1185                          * tasks within a group over tiny differences.
1186                          */
1187                         if (cur->numa_group)
1188                                 imp -= imp/16;
1189                 } else {
1190                         /*
1191                          * Compare the group weights. If a task is all by
1192                          * itself (not part of a group), use the task weight
1193                          * instead.
1194                          */
1195                         if (cur->numa_group)
1196                                 imp += group_weight(cur, env->src_nid) -
1197                                        group_weight(cur, env->dst_nid);
1198                         else
1199                                 imp += task_weight(cur, env->src_nid) -
1200                                        task_weight(cur, env->dst_nid);
1201                 }
1202         }
1203 
1204         if (imp <= env->best_imp && moveimp <= env->best_imp)
1205                 goto unlock;
1206 
1207         if (!cur) {
1208                 /* Is there capacity at our destination? */
1209                 if (env->src_stats.has_free_capacity &&
1210                     !env->dst_stats.has_free_capacity)
1211                         goto unlock;
1212 
1213                 goto balance;
1214         }
1215 
1216         /* Balance doesn't matter much if we're running a task per cpu */
1217         if (imp > env->best_imp && src_rq->nr_running == 1 &&
1218                         dst_rq->nr_running == 1)
1219                 goto assign;
1220 
1221         /*
1222          * In the overloaded case, try and keep the load balanced.
1223          */
1224 balance:
1225         load = task_h_load(env->p);
1226         dst_load = env->dst_stats.load + load;
1227         src_load = env->src_stats.load - load;
1228 
1229         if (moveimp > imp && moveimp > env->best_imp) {
1230                 /*
1231                  * If the improvement from just moving env->p direction is
1232                  * better than swapping tasks around, check if a move is
1233                  * possible. Store a slightly smaller score than moveimp,
1234                  * so an actually idle CPU will win.
1235                  */
1236                 if (!load_too_imbalanced(src_load, dst_load, env)) {
1237                         imp = moveimp - 1;
1238                         cur = NULL;
1239                         goto assign;
1240                 }
1241         }
1242 
1243         if (imp <= env->best_imp)
1244                 goto unlock;
1245 
1246         if (cur) {
1247                 load = task_h_load(cur);
1248                 dst_load -= load;
1249                 src_load += load;
1250         }
1251 
1252         if (load_too_imbalanced(src_load, dst_load, env))
1253                 goto unlock;
1254 
1255 assign:
1256         task_numa_assign(env, cur, imp);
1257 unlock:
1258         rcu_read_unlock();
1259 }
1260 
1261 static void task_numa_find_cpu(struct task_numa_env *env,
1262                                 long taskimp, long groupimp)
1263 {
1264         int cpu;
1265 
1266         for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1267                 /* Skip this CPU if the source task cannot migrate */
1268                 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1269                         continue;
1270 
1271                 env->dst_cpu = cpu;
1272                 task_numa_compare(env, taskimp, groupimp);
1273         }
1274 }
1275 
1276 static int task_numa_migrate(struct task_struct *p)
1277 {
1278         struct task_numa_env env = {
1279                 .p = p,
1280 
1281                 .src_cpu = task_cpu(p),
1282                 .src_nid = task_node(p),
1283 
1284                 .imbalance_pct = 112,
1285 
1286                 .best_task = NULL,
1287                 .best_imp = 0,
1288                 .best_cpu = -1
1289         };
1290         struct sched_domain *sd;
1291         unsigned long taskweight, groupweight;
1292         int nid, ret;
1293         long taskimp, groupimp;
1294 
1295         /*
1296          * Pick the lowest SD_NUMA domain, as that would have the smallest
1297          * imbalance and would be the first to start moving tasks about.
1298          *
1299          * And we want to avoid any moving of tasks about, as that would create
1300          * random movement of tasks -- counter the numa conditions we're trying
1301          * to satisfy here.
1302          */
1303         rcu_read_lock();
1304         sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1305         if (sd)
1306                 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1307         rcu_read_unlock();
1308 
1309         /*
1310          * Cpusets can break the scheduler domain tree into smaller
1311          * balance domains, some of which do not cross NUMA boundaries.
1312          * Tasks that are "trapped" in such domains cannot be migrated
1313          * elsewhere, so there is no point in (re)trying.
1314          */
1315         if (unlikely(!sd)) {
1316                 p->numa_preferred_nid = task_node(p);
1317                 return -EINVAL;
1318         }
1319 
1320         taskweight = task_weight(p, env.src_nid);
1321         groupweight = group_weight(p, env.src_nid);
1322         update_numa_stats(&env.src_stats, env.src_nid);
1323         env.dst_nid = p->numa_preferred_nid;
1324         taskimp = task_weight(p, env.dst_nid) - taskweight;
1325         groupimp = group_weight(p, env.dst_nid) - groupweight;
1326         update_numa_stats(&env.dst_stats, env.dst_nid);
1327 
1328         /* Try to find a spot on the preferred nid. */
1329         task_numa_find_cpu(&env, taskimp, groupimp);
1330 
1331         /* No space available on the preferred nid. Look elsewhere. */
1332         if (env.best_cpu == -1) {
1333                 for_each_online_node(nid) {
1334                         if (nid == env.src_nid || nid == p->numa_preferred_nid)
1335                                 continue;
1336 
1337                         /* Only consider nodes where both task and groups benefit */
1338                         taskimp = task_weight(p, nid) - taskweight;
1339                         groupimp = group_weight(p, nid) - groupweight;
1340                         if (taskimp < 0 && groupimp < 0)
1341                                 continue;
1342 
1343                         env.dst_nid = nid;
1344                         update_numa_stats(&env.dst_stats, env.dst_nid);
1345                         task_numa_find_cpu(&env, taskimp, groupimp);
1346                 }
1347         }
1348 
1349         /*
1350          * If the task is part of a workload that spans multiple NUMA nodes,
1351          * and is migrating into one of the workload's active nodes, remember
1352          * this node as the task's preferred numa node, so the workload can
1353          * settle down.
1354          * A task that migrated to a second choice node will be better off
1355          * trying for a better one later. Do not set the preferred node here.
1356          */
1357         if (p->numa_group) {
1358                 if (env.best_cpu == -1)
1359                         nid = env.src_nid;
1360                 else
1361                         nid = env.dst_nid;
1362 
1363                 if (node_isset(nid, p->numa_group->active_nodes))
1364                         sched_setnuma(p, env.dst_nid);
1365         }
1366 
1367         /* No better CPU than the current one was found. */
1368         if (env.best_cpu == -1)
1369                 return -EAGAIN;
1370 
1371         /*
1372          * Reset the scan period if the task is being rescheduled on an
1373          * alternative node to recheck if the tasks is now properly placed.
1374          */
1375         p->numa_scan_period = task_scan_min(p);
1376 
1377         if (env.best_task == NULL) {
1378                 ret = migrate_task_to(p, env.best_cpu);
1379                 if (ret != 0)
1380                         trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1381                 return ret;
1382         }
1383 
1384         ret = migrate_swap(p, env.best_task);
1385         if (ret != 0)
1386                 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1387         put_task_struct(env.best_task);
1388         return ret;
1389 }
1390 
1391 /* Attempt to migrate a task to a CPU on the preferred node. */
1392 static void numa_migrate_preferred(struct task_struct *p)
1393 {
1394         unsigned long interval = HZ;
1395 
1396         /* This task has no NUMA fault statistics yet */
1397         if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults_memory))
1398                 return;
1399 
1400         /* Periodically retry migrating the task to the preferred node */
1401         interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1402         p->numa_migrate_retry = jiffies + interval;
1403 
1404         /* Success if task is already running on preferred CPU */
1405         if (task_node(p) == p->numa_preferred_nid)
1406                 return;
1407 
1408         /* Otherwise, try migrate to a CPU on the preferred node */
1409         task_numa_migrate(p);
1410 }
1411 
1412 /*
1413  * Find the nodes on which the workload is actively running. We do this by
1414  * tracking the nodes from which NUMA hinting faults are triggered. This can
1415  * be different from the set of nodes where the workload's memory is currently
1416  * located.
1417  *
1418  * The bitmask is used to make smarter decisions on when to do NUMA page
1419  * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1420  * are added when they cause over 6/16 of the maximum number of faults, but
1421  * only removed when they drop below 3/16.
1422  */
1423 static void update_numa_active_node_mask(struct numa_group *numa_group)
1424 {
1425         unsigned long faults, max_faults = 0;
1426         int nid;
1427 
1428         for_each_online_node(nid) {
1429                 faults = group_faults_cpu(numa_group, nid);
1430                 if (faults > max_faults)
1431                         max_faults = faults;
1432         }
1433 
1434         for_each_online_node(nid) {
1435                 faults = group_faults_cpu(numa_group, nid);
1436                 if (!node_isset(nid, numa_group->active_nodes)) {
1437                         if (faults > max_faults * 6 / 16)
1438                                 node_set(nid, numa_group->active_nodes);
1439                 } else if (faults < max_faults * 3 / 16)
1440                         node_clear(nid, numa_group->active_nodes);
1441         }
1442 }
1443 
1444 /*
1445  * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1446  * increments. The more local the fault statistics are, the higher the scan
1447  * period will be for the next scan window. If local/(local+remote) ratio is
1448  * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1449  * the scan period will decrease. Aim for 70% local accesses.
1450  */
1451 #define NUMA_PERIOD_SLOTS 10
1452 #define NUMA_PERIOD_THRESHOLD 7
1453 
1454 /*
1455  * Increase the scan period (slow down scanning) if the majority of
1456  * our memory is already on our local node, or if the majority of
1457  * the page accesses are shared with other processes.
1458  * Otherwise, decrease the scan period.
1459  */
1460 static void update_task_scan_period(struct task_struct *p,
1461                         unsigned long shared, unsigned long private)
1462 {
1463         unsigned int period_slot;
1464         int ratio;
1465         int diff;
1466 
1467         unsigned long remote = p->numa_faults_locality[0];
1468         unsigned long local = p->numa_faults_locality[1];
1469 
1470         /*
1471          * If there were no record hinting faults then either the task is
1472          * completely idle or all activity is areas that are not of interest
1473          * to automatic numa balancing. Scan slower
1474          */
1475         if (local + shared == 0) {
1476                 p->numa_scan_period = min(p->numa_scan_period_max,
1477                         p->numa_scan_period << 1);
1478 
1479                 p->mm->numa_next_scan = jiffies +
1480                         msecs_to_jiffies(p->numa_scan_period);
1481 
1482                 return;
1483         }
1484 
1485         /*
1486          * Prepare to scale scan period relative to the current period.
1487          *       == NUMA_PERIOD_THRESHOLD scan period stays the same
1488          *       <  NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1489          *       >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1490          */
1491         period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1492         ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1493         if (ratio >= NUMA_PERIOD_THRESHOLD) {
1494                 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1495                 if (!slot)
1496                         slot = 1;
1497                 diff = slot * period_slot;
1498         } else {
1499                 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1500 
1501                 /*
1502                  * Scale scan rate increases based on sharing. There is an
1503                  * inverse relationship between the degree of sharing and
1504                  * the adjustment made to the scanning period. Broadly
1505                  * speaking the intent is that there is little point
1506                  * scanning faster if shared accesses dominate as it may
1507                  * simply bounce migrations uselessly
1508                  */
1509                 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared));
1510                 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1511         }
1512 
1513         p->numa_scan_period = clamp(p->numa_scan_period + diff,
1514                         task_scan_min(p), task_scan_max(p));
1515         memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1516 }
1517 
1518 /*
1519  * Get the fraction of time the task has been running since the last
1520  * NUMA placement cycle. The scheduler keeps similar statistics, but
1521  * decays those on a 32ms period, which is orders of magnitude off
1522  * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1523  * stats only if the task is so new there are no NUMA statistics yet.
1524  */
1525 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1526 {
1527         u64 runtime, delta, now;
1528         /* Use the start of this time slice to avoid calculations. */
1529         now = p->se.exec_start;
1530         runtime = p->se.sum_exec_runtime;
1531 
1532         if (p->last_task_numa_placement) {
1533                 delta = runtime - p->last_sum_exec_runtime;
1534                 *period = now - p->last_task_numa_placement;
1535         } else {
1536                 delta = p->se.avg.runnable_avg_sum;
1537                 *period = p->se.avg.runnable_avg_period;
1538         }
1539 
1540         p->last_sum_exec_runtime = runtime;
1541         p->last_task_numa_placement = now;
1542 
1543         return delta;
1544 }
1545 
1546 static void task_numa_placement(struct task_struct *p)
1547 {
1548         int seq, nid, max_nid = -1, max_group_nid = -1;
1549         unsigned long max_faults = 0, max_group_faults = 0;
1550         unsigned long fault_types[2] = { 0, 0 };
1551         unsigned long total_faults;
1552         u64 runtime, period;
1553         spinlock_t *group_lock = NULL;
1554 
1555         seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1556         if (p->numa_scan_seq == seq)
1557                 return;
1558         p->numa_scan_seq = seq;
1559         p->numa_scan_period_max = task_scan_max(p);
1560 
1561         total_faults = p->numa_faults_locality[0] +
1562                        p->numa_faults_locality[1];
1563         runtime = numa_get_avg_runtime(p, &period);
1564 
1565         /* If the task is part of a group prevent parallel updates to group stats */
1566         if (p->numa_group) {
1567                 group_lock = &p->numa_group->lock;
1568                 spin_lock_irq(group_lock);
1569         }
1570 
1571         /* Find the node with the highest number of faults */
1572         for_each_online_node(nid) {
1573                 unsigned long faults = 0, group_faults = 0;
1574                 int priv, i;
1575 
1576                 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1577                         long diff, f_diff, f_weight;
1578 
1579                         i = task_faults_idx(nid, priv);
1580 
1581                         /* Decay existing window, copy faults since last scan */
1582                         diff = p->numa_faults_buffer_memory[i] - p->numa_faults_memory[i] / 2;
1583                         fault_types[priv] += p->numa_faults_buffer_memory[i];
1584                         p->numa_faults_buffer_memory[i] = 0;
1585 
1586                         /*
1587                          * Normalize the faults_from, so all tasks in a group
1588                          * count according to CPU use, instead of by the raw
1589                          * number of faults. Tasks with little runtime have
1590                          * little over-all impact on throughput, and thus their
1591                          * faults are less important.
1592                          */
1593                         f_weight = div64_u64(runtime << 16, period + 1);
1594                         f_weight = (f_weight * p->numa_faults_buffer_cpu[i]) /
1595                                    (total_faults + 1);
1596                         f_diff = f_weight - p->numa_faults_cpu[i] / 2;
1597                         p->numa_faults_buffer_cpu[i] = 0;
1598 
1599                         p->numa_faults_memory[i] += diff;
1600                         p->numa_faults_cpu[i] += f_diff;
1601                         faults += p->numa_faults_memory[i];
1602                         p->total_numa_faults += diff;
1603                         if (p->numa_group) {
1604                                 /* safe because we can only change our own group */
1605                                 p->numa_group->faults[i] += diff;
1606                                 p->numa_group->faults_cpu[i] += f_diff;
1607                                 p->numa_group->total_faults += diff;
1608                                 group_faults += p->numa_group->faults[i];
1609                         }
1610                 }
1611 
1612                 if (faults > max_faults) {
1613                         max_faults = faults;
1614                         max_nid = nid;
1615                 }
1616 
1617                 if (group_faults > max_group_faults) {
1618                         max_group_faults = group_faults;
1619                         max_group_nid = nid;
1620                 }
1621         }
1622 
1623         update_task_scan_period(p, fault_types[0], fault_types[1]);
1624 
1625         if (p->numa_group) {
1626                 update_numa_active_node_mask(p->numa_group);
1627                 spin_unlock_irq(group_lock);
1628                 max_nid = max_group_nid;
1629         }
1630 
1631         if (max_faults) {
1632                 /* Set the new preferred node */
1633                 if (max_nid != p->numa_preferred_nid)
1634                         sched_setnuma(p, max_nid);
1635 
1636                 if (task_node(p) != p->numa_preferred_nid)
1637                         numa_migrate_preferred(p);
1638         }
1639 }
1640 
1641 static inline int get_numa_group(struct numa_group *grp)
1642 {
1643         return atomic_inc_not_zero(&grp->refcount);
1644 }
1645 
1646 static inline void put_numa_group(struct numa_group *grp)
1647 {
1648         if (atomic_dec_and_test(&grp->refcount))
1649                 kfree_rcu(grp, rcu);
1650 }
1651 
1652 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1653                         int *priv)
1654 {
1655         struct numa_group *grp, *my_grp;
1656         struct task_struct *tsk;
1657         bool join = false;
1658         int cpu = cpupid_to_cpu(cpupid);
1659         int i;
1660 
1661         if (unlikely(!p->numa_group)) {
1662                 unsigned int size = sizeof(struct numa_group) +
1663                                     4*nr_node_ids*sizeof(unsigned long);
1664 
1665                 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1666                 if (!grp)
1667                         return;
1668 
1669                 atomic_set(&grp->refcount, 1);
1670                 spin_lock_init(&grp->lock);
1671                 INIT_LIST_HEAD(&grp->task_list);
1672                 grp->gid = p->pid;
1673                 /* Second half of the array tracks nids where faults happen */
1674                 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1675                                                 nr_node_ids;
1676 
1677                 node_set(task_node(current), grp->active_nodes);
1678 
1679                 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1680                         grp->faults[i] = p->numa_faults_memory[i];
1681 
1682                 grp->total_faults = p->total_numa_faults;
1683 
1684                 list_add(&p->numa_entry, &grp->task_list);
1685                 grp->nr_tasks++;
1686                 rcu_assign_pointer(p->numa_group, grp);
1687         }
1688 
1689         rcu_read_lock();
1690         tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1691 
1692         if (!cpupid_match_pid(tsk, cpupid))
1693                 goto no_join;
1694 
1695         grp = rcu_dereference(tsk->numa_group);
1696         if (!grp)
1697                 goto no_join;
1698 
1699         my_grp = p->numa_group;
1700         if (grp == my_grp)
1701                 goto no_join;
1702 
1703         /*
1704          * Only join the other group if its bigger; if we're the bigger group,
1705          * the other task will join us.
1706          */
1707         if (my_grp->nr_tasks > grp->nr_tasks)
1708                 goto no_join;
1709 
1710         /*
1711          * Tie-break on the grp address.
1712          */
1713         if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1714                 goto no_join;
1715 
1716         /* Always join threads in the same process. */
1717         if (tsk->mm == current->mm)
1718                 join = true;
1719 
1720         /* Simple filter to avoid false positives due to PID collisions */
1721         if (flags & TNF_SHARED)
1722                 join = true;
1723 
1724         /* Update priv based on whether false sharing was detected */
1725         *priv = !join;
1726 
1727         if (join && !get_numa_group(grp))
1728                 goto no_join;
1729 
1730         rcu_read_unlock();
1731 
1732         if (!join)
1733                 return;
1734 
1735         BUG_ON(irqs_disabled());
1736         double_lock_irq(&my_grp->lock, &grp->lock);
1737 
1738         for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
1739                 my_grp->faults[i] -= p->numa_faults_memory[i];
1740                 grp->faults[i] += p->numa_faults_memory[i];
1741         }
1742         my_grp->total_faults -= p->total_numa_faults;
1743         grp->total_faults += p->total_numa_faults;
1744 
1745         list_move(&p->numa_entry, &grp->task_list);
1746         my_grp->nr_tasks--;
1747         grp->nr_tasks++;
1748 
1749         spin_unlock(&my_grp->lock);
1750         spin_unlock_irq(&grp->lock);
1751 
1752         rcu_assign_pointer(p->numa_group, grp);
1753 
1754         put_numa_group(my_grp);
1755         return;
1756 
1757 no_join:
1758         rcu_read_unlock();
1759         return;
1760 }
1761 
1762 void task_numa_free(struct task_struct *p)
1763 {
1764         struct numa_group *grp = p->numa_group;
1765         void *numa_faults = p->numa_faults_memory;
1766         unsigned long flags;
1767         int i;
1768 
1769         if (grp) {
1770                 spin_lock_irqsave(&grp->lock, flags);
1771                 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1772                         grp->faults[i] -= p->numa_faults_memory[i];
1773                 grp->total_faults -= p->total_numa_faults;
1774 
1775                 list_del(&p->numa_entry);
1776                 grp->nr_tasks--;
1777                 spin_unlock_irqrestore(&grp->lock, flags);
1778                 rcu_assign_pointer(p->numa_group, NULL);
1779                 put_numa_group(grp);
1780         }
1781 
1782         p->numa_faults_memory = NULL;
1783         p->numa_faults_buffer_memory = NULL;
1784         p->numa_faults_cpu= NULL;
1785         p->numa_faults_buffer_cpu = NULL;
1786         kfree(numa_faults);
1787 }
1788 
1789 /*
1790  * Got a PROT_NONE fault for a page on @node.
1791  */
1792 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
1793 {
1794         struct task_struct *p = current;
1795         bool migrated = flags & TNF_MIGRATED;
1796         int cpu_node = task_node(current);
1797         int local = !!(flags & TNF_FAULT_LOCAL);
1798         int priv;
1799 
1800         if (!numabalancing_enabled)
1801                 return;
1802 
1803         /* for example, ksmd faulting in a user's mm */
1804         if (!p->mm)
1805                 return;
1806 
1807         /* Do not worry about placement if exiting */
1808         if (p->state == TASK_DEAD)
1809                 return;
1810 
1811         /* Allocate buffer to track faults on a per-node basis */
1812         if (unlikely(!p->numa_faults_memory)) {
1813                 int size = sizeof(*p->numa_faults_memory) *
1814                            NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
1815 
1816                 p->numa_faults_memory = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
1817                 if (!p->numa_faults_memory)
1818                         return;
1819 
1820                 BUG_ON(p->numa_faults_buffer_memory);
1821                 /*
1822                  * The averaged statistics, shared & private, memory & cpu,
1823                  * occupy the first half of the array. The second half of the
1824                  * array is for current counters, which are averaged into the
1825                  * first set by task_numa_placement.
1826                  */
1827                 p->numa_faults_cpu = p->numa_faults_memory + (2 * nr_node_ids);
1828                 p->numa_faults_buffer_memory = p->numa_faults_memory + (4 * nr_node_ids);
1829                 p->numa_faults_buffer_cpu = p->numa_faults_memory + (6 * nr_node_ids);
1830                 p->total_numa_faults = 0;
1831                 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1832         }
1833 
1834         /*
1835          * First accesses are treated as private, otherwise consider accesses
1836          * to be private if the accessing pid has not changed
1837          */
1838         if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
1839                 priv = 1;
1840         } else {
1841                 priv = cpupid_match_pid(p, last_cpupid);
1842                 if (!priv && !(flags & TNF_NO_GROUP))
1843                         task_numa_group(p, last_cpupid, flags, &priv);
1844         }
1845 
1846         /*
1847          * If a workload spans multiple NUMA nodes, a shared fault that
1848          * occurs wholly within the set of nodes that the workload is
1849          * actively using should be counted as local. This allows the
1850          * scan rate to slow down when a workload has settled down.
1851          */
1852         if (!priv && !local && p->numa_group &&
1853                         node_isset(cpu_node, p->numa_group->active_nodes) &&
1854                         node_isset(mem_node, p->numa_group->active_nodes))
1855                 local = 1;
1856 
1857         task_numa_placement(p);
1858 
1859         /*
1860          * Retry task to preferred node migration periodically, in case it
1861          * case it previously failed, or the scheduler moved us.
1862          */
1863         if (time_after(jiffies, p->numa_migrate_retry))
1864                 numa_migrate_preferred(p);
1865 
1866         if (migrated)
1867                 p->numa_pages_migrated += pages;
1868 
1869         p->numa_faults_buffer_memory[task_faults_idx(mem_node, priv)] += pages;
1870         p->numa_faults_buffer_cpu[task_faults_idx(cpu_node, priv)] += pages;
1871         p->numa_faults_locality[local] += pages;
1872 }
1873 
1874 static void reset_ptenuma_scan(struct task_struct *p)
1875 {
1876         ACCESS_ONCE(p->mm->numa_scan_seq)++;
1877         p->mm->numa_scan_offset = 0;
1878 }
1879 
1880 /*
1881  * The expensive part of numa migration is done from task_work context.
1882  * Triggered from task_tick_numa().
1883  */
1884 void task_numa_work(struct callback_head *work)
1885 {
1886         unsigned long migrate, next_scan, now = jiffies;
1887         struct task_struct *p = current;
1888         struct mm_struct *mm = p->mm;
1889         struct vm_area_struct *vma;
1890         unsigned long start, end;
1891         unsigned long nr_pte_updates = 0;
1892         long pages;
1893 
1894         WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1895 
1896         work->next = work; /* protect against double add */
1897         /*
1898          * Who cares about NUMA placement when they're dying.
1899          *
1900          * NOTE: make sure not to dereference p->mm before this check,
1901          * exit_task_work() happens _after_ exit_mm() so we could be called
1902          * without p->mm even though we still had it when we enqueued this
1903          * work.
1904          */
1905         if (p->flags & PF_EXITING)
1906                 return;
1907 
1908         if (!mm->numa_next_scan) {
1909                 mm->numa_next_scan = now +
1910                         msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1911         }
1912 
1913         /*
1914          * Enforce maximal scan/migration frequency..
1915          */
1916         migrate = mm->numa_next_scan;
1917         if (time_before(now, migrate))
1918                 return;
1919 
1920         if (p->numa_scan_period == 0) {
1921                 p->numa_scan_period_max = task_scan_max(p);
1922                 p->numa_scan_period = task_scan_min(p);
1923         }
1924 
1925         next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1926         if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1927                 return;
1928 
1929         /*
1930          * Delay this task enough that another task of this mm will likely win
1931          * the next time around.
1932          */
1933         p->node_stamp += 2 * TICK_NSEC;
1934 
1935         start = mm->numa_scan_offset;
1936         pages = sysctl_numa_balancing_scan_size;
1937         pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1938         if (!pages)
1939                 return;
1940 
1941         down_read(&mm->mmap_sem);
1942         vma = find_vma(mm, start);
1943         if (!vma) {
1944                 reset_ptenuma_scan(p);
1945                 start = 0;
1946                 vma = mm->mmap;
1947         }
1948         for (; vma; vma = vma->vm_next) {
1949                 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1950                         continue;
1951 
1952                 /*
1953                  * Shared library pages mapped by multiple processes are not
1954                  * migrated as it is expected they are cache replicated. Avoid
1955                  * hinting faults in read-only file-backed mappings or the vdso
1956                  * as migrating the pages will be of marginal benefit.
1957                  */
1958                 if (!vma->vm_mm ||
1959                     (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1960                         continue;
1961 
1962                 /*
1963                  * Skip inaccessible VMAs to avoid any confusion between
1964                  * PROT_NONE and NUMA hinting ptes
1965                  */
1966                 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
1967                         continue;
1968 
1969                 do {
1970                         start = max(start, vma->vm_start);
1971                         end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1972                         end = min(end, vma->vm_end);
1973                         nr_pte_updates += change_prot_numa(vma, start, end);
1974 
1975                         /*
1976                          * Scan sysctl_numa_balancing_scan_size but ensure that
1977                          * at least one PTE is updated so that unused virtual
1978                          * address space is quickly skipped.
1979                          */
1980                         if (nr_pte_updates)
1981                                 pages -= (end - start) >> PAGE_SHIFT;
1982 
1983                         start = end;
1984                         if (pages <= 0)
1985                                 goto out;
1986 
1987                         cond_resched();
1988                 } while (end != vma->vm_end);
1989         }
1990 
1991 out:
1992         /*
1993          * It is possible to reach the end of the VMA list but the last few
1994          * VMAs are not guaranteed to the vma_migratable. If they are not, we
1995          * would find the !migratable VMA on the next scan but not reset the
1996          * scanner to the start so check it now.
1997          */
1998         if (vma)
1999                 mm->numa_scan_offset = start;
2000         else
2001                 reset_ptenuma_scan(p);
2002         up_read(&mm->mmap_sem);
2003 }
2004 
2005 /*
2006  * Drive the periodic memory faults..
2007  */
2008 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2009 {
2010         struct callback_head *work = &curr->numa_work;
2011         u64 period, now;
2012 
2013         /*
2014          * We don't care about NUMA placement if we don't have memory.
2015          */
2016         if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2017                 return;
2018 
2019         /*
2020          * Using runtime rather than walltime has the dual advantage that
2021          * we (mostly) drive the selection from busy threads and that the
2022          * task needs to have done some actual work before we bother with
2023          * NUMA placement.
2024          */
2025         now = curr->se.sum_exec_runtime;
2026         period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2027 
2028         if (now - curr->node_stamp > period) {
2029                 if (!curr->node_stamp)
2030                         curr->numa_scan_period = task_scan_min(curr);
2031                 curr->node_stamp += period;
2032 
2033                 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2034                         init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2035                         task_work_add(curr, work, true);
2036                 }
2037         }
2038 }
2039 #else
2040 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2041 {
2042 }
2043 
2044 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2045 {
2046 }
2047 
2048 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2049 {
2050 }
2051 #endif /* CONFIG_NUMA_BALANCING */
2052 
2053 static void
2054 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2055 {
2056         update_load_add(&cfs_rq->load, se->load.weight);
2057         if (!parent_entity(se))
2058                 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2059 #ifdef CONFIG_SMP
2060         if (entity_is_task(se)) {
2061                 struct rq *rq = rq_of(cfs_rq);
2062 
2063                 account_numa_enqueue(rq, task_of(se));
2064                 list_add(&se->group_node, &rq->cfs_tasks);
2065         }
2066 #endif
2067         cfs_rq->nr_running++;
2068 }
2069 
2070 static void
2071 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2072 {
2073         update_load_sub(&cfs_rq->load, se->load.weight);
2074         if (!parent_entity(se))
2075                 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2076         if (entity_is_task(se)) {
2077                 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2078                 list_del_init(&se->group_node);
2079         }
2080         cfs_rq->nr_running--;
2081 }
2082 
2083 #ifdef CONFIG_FAIR_GROUP_SCHED
2084 # ifdef CONFIG_SMP
2085 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2086 {
2087         long tg_weight;
2088 
2089         /*
2090          * Use this CPU's actual weight instead of the last load_contribution
2091          * to gain a more accurate current total weight. See
2092          * update_cfs_rq_load_contribution().
2093          */
2094         tg_weight = atomic_long_read(&tg->load_avg);
2095         tg_weight -= cfs_rq->tg_load_contrib;
2096         tg_weight += cfs_rq->load.weight;
2097 
2098         return tg_weight;
2099 }
2100 
2101 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2102 {
2103         long tg_weight, load, shares;
2104 
2105         tg_weight = calc_tg_weight(tg, cfs_rq);
2106         load = cfs_rq->load.weight;
2107 
2108         shares = (tg->shares * load);
2109         if (tg_weight)
2110                 shares /= tg_weight;
2111 
2112         if (shares < MIN_SHARES)
2113                 shares = MIN_SHARES;
2114         if (shares > tg->shares)
2115                 shares = tg->shares;
2116 
2117         return shares;
2118 }
2119 # else /* CONFIG_SMP */
2120 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2121 {
2122         return tg->shares;
2123 }
2124 # endif /* CONFIG_SMP */
2125 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2126                             unsigned long weight)
2127 {
2128         if (se->on_rq) {
2129                 /* commit outstanding execution time */
2130                 if (cfs_rq->curr == se)
2131                         update_curr(cfs_rq);
2132                 account_entity_dequeue(cfs_rq, se);
2133         }
2134 
2135         update_load_set(&se->load, weight);
2136 
2137         if (se->on_rq)
2138                 account_entity_enqueue(cfs_rq, se);
2139 }
2140 
2141 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2142 
2143 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2144 {
2145         struct task_group *tg;
2146         struct sched_entity *se;
2147         long shares;
2148 
2149         tg = cfs_rq->tg;
2150         se = tg->se[cpu_of(rq_of(cfs_rq))];
2151         if (!se || throttled_hierarchy(cfs_rq))
2152                 return;
2153 #ifndef CONFIG_SMP
2154         if (likely(se->load.weight == tg->shares))
2155                 return;
2156 #endif
2157         shares = calc_cfs_shares(cfs_rq, tg);
2158 
2159         reweight_entity(cfs_rq_of(se), se, shares);
2160 }
2161 #else /* CONFIG_FAIR_GROUP_SCHED */
2162 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2163 {
2164 }
2165 #endif /* CONFIG_FAIR_GROUP_SCHED */
2166 
2167 #ifdef CONFIG_SMP
2168 /*
2169  * We choose a half-life close to 1 scheduling period.
2170  * Note: The tables below are dependent on this value.
2171  */
2172 #define LOAD_AVG_PERIOD 32
2173 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2174 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2175 
2176 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2177 static const u32 runnable_avg_yN_inv[] = {
2178         0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2179         0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2180         0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2181         0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2182         0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2183         0x85aac367, 0x82cd8698,
2184 };
2185 
2186 /*
2187  * Precomputed \Sum y^k { 1<=k<=n }.  These are floor(true_value) to prevent
2188  * over-estimates when re-combining.
2189  */
2190 static const u32 runnable_avg_yN_sum[] = {
2191             0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2192          9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2193         17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2194 };
2195 
2196 /*
2197  * Approximate:
2198  *   val * y^n,    where y^32 ~= 0.5 (~1 scheduling period)
2199  */
2200 static __always_inline u64 decay_load(u64 val, u64 n)
2201 {
2202         unsigned int local_n;
2203 
2204         if (!n)
2205                 return val;
2206         else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2207                 return 0;
2208 
2209         /* after bounds checking we can collapse to 32-bit */
2210         local_n = n;
2211 
2212         /*
2213          * As y^PERIOD = 1/2, we can combine
2214          *    y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
2215          * With a look-up table which covers k^n (n<PERIOD)
2216          *
2217          * To achieve constant time decay_load.
2218          */
2219         if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2220                 val >>= local_n / LOAD_AVG_PERIOD;
2221                 local_n %= LOAD_AVG_PERIOD;
2222         }
2223 
2224         val *= runnable_avg_yN_inv[local_n];
2225         /* We don't use SRR here since we always want to round down. */
2226         return val >> 32;
2227 }
2228 
2229 /*
2230  * For updates fully spanning n periods, the contribution to runnable
2231  * average will be: \Sum 1024*y^n
2232  *
2233  * We can compute this reasonably efficiently by combining:
2234  *   y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for  n <PERIOD}
2235  */
2236 static u32 __compute_runnable_contrib(u64 n)
2237 {
2238         u32 contrib = 0;
2239 
2240         if (likely(n <= LOAD_AVG_PERIOD))
2241                 return runnable_avg_yN_sum[n];
2242         else if (unlikely(n >= LOAD_AVG_MAX_N))
2243                 return LOAD_AVG_MAX;
2244 
2245         /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2246         do {
2247                 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2248                 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2249 
2250                 n -= LOAD_AVG_PERIOD;
2251         } while (n > LOAD_AVG_PERIOD);
2252 
2253         contrib = decay_load(contrib, n);
2254         return contrib + runnable_avg_yN_sum[n];
2255 }
2256 
2257 /*
2258  * We can represent the historical contribution to runnable average as the
2259  * coefficients of a geometric series.  To do this we sub-divide our runnable
2260  * history into segments of approximately 1ms (1024us); label the segment that
2261  * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2262  *
2263  * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2264  *      p0            p1           p2
2265  *     (now)       (~1ms ago)  (~2ms ago)
2266  *
2267  * Let u_i denote the fraction of p_i that the entity was runnable.
2268  *
2269  * We then designate the fractions u_i as our co-efficients, yielding the
2270  * following representation of historical load:
2271  *   u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2272  *
2273  * We choose y based on the with of a reasonably scheduling period, fixing:
2274  *   y^32 = 0.5
2275  *
2276  * This means that the contribution to load ~32ms ago (u_32) will be weighted
2277  * approximately half as much as the contribution to load within the last ms
2278  * (u_0).
2279  *
2280  * When a period "rolls over" and we have new u_0`, multiplying the previous
2281  * sum again by y is sufficient to update:
2282  *   load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2283  *            = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2284  */
2285 static __always_inline int __update_entity_runnable_avg(u64 now,
2286                                                         struct sched_avg *sa,
2287                                                         int runnable)
2288 {
2289         u64 delta, periods;
2290         u32 runnable_contrib;
2291         int delta_w, decayed = 0;
2292 
2293         delta = now - sa->last_runnable_update;
2294         /*
2295          * This should only happen when time goes backwards, which it
2296          * unfortunately does during sched clock init when we swap over to TSC.
2297          */
2298         if ((s64)delta < 0) {
2299                 sa->last_runnable_update = now;
2300                 return 0;
2301         }
2302 
2303         /*
2304          * Use 1024ns as the unit of measurement since it's a reasonable
2305          * approximation of 1us and fast to compute.
2306          */
2307         delta >>= 10;
2308         if (!delta)
2309                 return 0;
2310         sa->last_runnable_update = now;
2311 
2312         /* delta_w is the amount already accumulated against our next period */
2313         delta_w = sa->runnable_avg_period % 1024;
2314         if (delta + delta_w >= 1024) {
2315                 /* period roll-over */
2316                 decayed = 1;
2317 
2318                 /*
2319                  * Now that we know we're crossing a period boundary, figure
2320                  * out how much from delta we need to complete the current
2321                  * period and accrue it.
2322                  */
2323                 delta_w = 1024 - delta_w;
2324                 if (runnable)
2325                         sa->runnable_avg_sum += delta_w;
2326                 sa->runnable_avg_period += delta_w;
2327 
2328                 delta -= delta_w;
2329 
2330                 /* Figure out how many additional periods this update spans */
2331                 periods = delta / 1024;
2332                 delta %= 1024;
2333 
2334                 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2335                                                   periods + 1);
2336                 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
2337                                                      periods + 1);
2338 
2339                 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2340                 runnable_contrib = __compute_runnable_contrib(periods);
2341                 if (runnable)
2342                         sa->runnable_avg_sum += runnable_contrib;
2343                 sa->runnable_avg_period += runnable_contrib;
2344         }
2345 
2346         /* Remainder of delta accrued against u_0` */
2347         if (runnable)
2348                 sa->runnable_avg_sum += delta;
2349         sa->runnable_avg_period += delta;
2350 
2351         return decayed;
2352 }
2353 
2354 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2355 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2356 {
2357         struct cfs_rq *cfs_rq = cfs_rq_of(se);
2358         u64 decays = atomic64_read(&cfs_rq->decay_counter);
2359 
2360         decays -= se->avg.decay_count;
2361         if (!decays)
2362                 return 0;
2363 
2364         se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2365         se->avg.decay_count = 0;
2366 
2367         return decays;
2368 }
2369 
2370 #ifdef CONFIG_FAIR_GROUP_SCHED
2371 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2372                                                  int force_update)
2373 {
2374         struct task_group *tg = cfs_rq->tg;
2375         long tg_contrib;
2376 
2377         tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2378         tg_contrib -= cfs_rq->tg_load_contrib;
2379 
2380         if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2381                 atomic_long_add(tg_contrib, &tg->load_avg);
2382                 cfs_rq->tg_load_contrib += tg_contrib;
2383         }
2384 }
2385 
2386 /*
2387  * Aggregate cfs_rq runnable averages into an equivalent task_group
2388  * representation for computing load contributions.
2389  */
2390 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2391                                                   struct cfs_rq *cfs_rq)
2392 {
2393         struct task_group *tg = cfs_rq->tg;
2394         long contrib;
2395 
2396         /* The fraction of a cpu used by this cfs_rq */
2397         contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT,
2398                           sa->runnable_avg_period + 1);
2399         contrib -= cfs_rq->tg_runnable_contrib;
2400 
2401         if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2402                 atomic_add(contrib, &tg->runnable_avg);
2403                 cfs_rq->tg_runnable_contrib += contrib;
2404         }
2405 }
2406 
2407 static inline void __update_group_entity_contrib(struct sched_entity *se)
2408 {
2409         struct cfs_rq *cfs_rq = group_cfs_rq(se);
2410         struct task_group *tg = cfs_rq->tg;
2411         int runnable_avg;
2412 
2413         u64 contrib;
2414 
2415         contrib = cfs_rq->tg_load_contrib * tg->shares;
2416         se->avg.load_avg_contrib = div_u64(contrib,
2417                                      atomic_long_read(&tg->load_avg) + 1);
2418 
2419         /*
2420          * For group entities we need to compute a correction term in the case
2421          * that they are consuming <1 cpu so that we would contribute the same
2422          * load as a task of equal weight.
2423          *
2424          * Explicitly co-ordinating this measurement would be expensive, but
2425          * fortunately the sum of each cpus contribution forms a usable
2426          * lower-bound on the true value.
2427          *
2428          * Consider the aggregate of 2 contributions.  Either they are disjoint
2429          * (and the sum represents true value) or they are disjoint and we are
2430          * understating by the aggregate of their overlap.
2431          *
2432          * Extending this to N cpus, for a given overlap, the maximum amount we
2433          * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2434          * cpus that overlap for this interval and w_i is the interval width.
2435          *
2436          * On a small machine; the first term is well-bounded which bounds the
2437          * total error since w_i is a subset of the period.  Whereas on a
2438          * larger machine, while this first term can be larger, if w_i is the
2439          * of consequential size guaranteed to see n_i*w_i quickly converge to
2440          * our upper bound of 1-cpu.
2441          */
2442         runnable_avg = atomic_read(&tg->runnable_avg);
2443         if (runnable_avg < NICE_0_LOAD) {
2444                 se->avg.load_avg_contrib *= runnable_avg;
2445                 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2446         }
2447 }
2448 
2449 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2450 {
2451         __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
2452         __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2453 }
2454 #else /* CONFIG_FAIR_GROUP_SCHED */
2455 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2456                                                  int force_update) {}
2457 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2458                                                   struct cfs_rq *cfs_rq) {}
2459 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2460 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2461 #endif /* CONFIG_FAIR_GROUP_SCHED */
2462 
2463 static inline void __update_task_entity_contrib(struct sched_entity *se)
2464 {
2465         u32 contrib;
2466 
2467         /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2468         contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2469         contrib /= (se->avg.runnable_avg_period + 1);
2470         se->avg.load_avg_contrib = scale_load(contrib);
2471 }
2472 
2473 /* Compute the current contribution to load_avg by se, return any delta */
2474 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2475 {
2476         long old_contrib = se->avg.load_avg_contrib;
2477 
2478         if (entity_is_task(se)) {
2479                 __update_task_entity_contrib(se);
2480         } else {
2481                 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2482                 __update_group_entity_contrib(se);
2483         }
2484 
2485         return se->avg.load_avg_contrib - old_contrib;
2486 }
2487 
2488 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2489                                                  long load_contrib)
2490 {
2491         if (likely(load_contrib < cfs_rq->blocked_load_avg))
2492                 cfs_rq->blocked_load_avg -= load_contrib;
2493         else
2494                 cfs_rq->blocked_load_avg = 0;
2495 }
2496 
2497 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2498 
2499 /* Update a sched_entity's runnable average */
2500 static inline void update_entity_load_avg(struct sched_entity *se,
2501                                           int update_cfs_rq)
2502 {
2503         struct cfs_rq *cfs_rq = cfs_rq_of(se);
2504         long contrib_delta;
2505         u64 now;
2506 
2507         /*
2508          * For a group entity we need to use their owned cfs_rq_clock_task() in
2509          * case they are the parent of a throttled hierarchy.
2510          */
2511         if (entity_is_task(se))
2512                 now = cfs_rq_clock_task(cfs_rq);
2513         else
2514                 now = cfs_rq_clock_task(group_cfs_rq(se));
2515 
2516         if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
2517                 return;
2518 
2519         contrib_delta = __update_entity_load_avg_contrib(se);
2520 
2521         if (!update_cfs_rq)
2522                 return;
2523 
2524         if (se->on_rq)
2525                 cfs_rq->runnable_load_avg += contrib_delta;
2526         else
2527                 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2528 }
2529 
2530 /*
2531  * Decay the load contributed by all blocked children and account this so that
2532  * their contribution may appropriately discounted when they wake up.
2533  */
2534 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2535 {
2536         u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2537         u64 decays;
2538 
2539         decays = now - cfs_rq->last_decay;
2540         if (!decays && !force_update)
2541                 return;
2542 
2543         if (atomic_long_read(&cfs_rq->removed_load)) {
2544                 unsigned long removed_load;
2545                 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2546                 subtract_blocked_load_contrib(cfs_rq, removed_load);
2547         }
2548 
2549         if (decays) {
2550                 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2551                                                       decays);
2552                 atomic64_add(decays, &cfs_rq->decay_counter);
2553                 cfs_rq->last_decay = now;
2554         }
2555 
2556         __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2557 }
2558 
2559 /* Add the load generated by se into cfs_rq's child load-average */
2560 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2561                                                   struct sched_entity *se,
2562                                                   int wakeup)
2563 {
2564         /*
2565          * We track migrations using entity decay_count <= 0, on a wake-up
2566          * migration we use a negative decay count to track the remote decays
2567          * accumulated while sleeping.
2568          *
2569          * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2570          * are seen by enqueue_entity_load_avg() as a migration with an already
2571          * constructed load_avg_contrib.
2572          */
2573         if (unlikely(se->avg.decay_count <= 0)) {
2574                 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2575                 if (se->avg.decay_count) {
2576                         /*
2577                          * In a wake-up migration we have to approximate the
2578                          * time sleeping.  This is because we can't synchronize
2579                          * clock_task between the two cpus, and it is not
2580                          * guaranteed to be read-safe.  Instead, we can
2581                          * approximate this using our carried decays, which are
2582                          * explicitly atomically readable.
2583                          */
2584                         se->avg.last_runnable_update -= (-se->avg.decay_count)
2585                                                         << 20;
2586                         update_entity_load_avg(se, 0);
2587                         /* Indicate that we're now synchronized and on-rq */
2588                         se->avg.decay_count = 0;
2589                 }
2590                 wakeup = 0;
2591         } else {
2592                 __synchronize_entity_decay(se);
2593         }
2594 
2595         /* migrated tasks did not contribute to our blocked load */
2596         if (wakeup) {
2597                 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2598                 update_entity_load_avg(se, 0);
2599         }
2600 
2601         cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2602         /* we force update consideration on load-balancer moves */
2603         update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2604 }
2605 
2606 /*
2607  * Remove se's load from this cfs_rq child load-average, if the entity is
2608  * transitioning to a blocked state we track its projected decay using
2609  * blocked_load_avg.
2610  */
2611 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2612                                                   struct sched_entity *se,
2613                                                   int sleep)
2614 {
2615         update_entity_load_avg(se, 1);
2616         /* we force update consideration on load-balancer moves */
2617         update_cfs_rq_blocked_load(cfs_rq, !sleep);
2618 
2619         cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2620         if (sleep) {
2621                 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2622                 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2623         } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2624 }
2625 
2626 /*
2627  * Update the rq's load with the elapsed running time before entering
2628  * idle. if the last scheduled task is not a CFS task, idle_enter will
2629  * be the only way to update the runnable statistic.
2630  */
2631 void idle_enter_fair(struct rq *this_rq)
2632 {
2633         update_rq_runnable_avg(this_rq, 1);
2634 }
2635 
2636 /*
2637  * Update the rq's load with the elapsed idle time before a task is
2638  * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2639  * be the only way to update the runnable statistic.
2640  */
2641 void idle_exit_fair(struct rq *this_rq)
2642 {
2643         update_rq_runnable_avg(this_rq, 0);
2644 }
2645 
2646 static int idle_balance(struct rq *this_rq);
2647 
2648 #else /* CONFIG_SMP */
2649 
2650 static inline void update_entity_load_avg(struct sched_entity *se,
2651                                           int update_cfs_rq) {}
2652 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2653 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2654                                            struct sched_entity *se,
2655                                            int wakeup) {}
2656 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2657                                            struct sched_entity *se,
2658                                            int sleep) {}
2659 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2660                                               int force_update) {}
2661 
2662 static inline int idle_balance(struct rq *rq)
2663 {
2664         return 0;
2665 }
2666 
2667 #endif /* CONFIG_SMP */
2668 
2669 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2670 {
2671 #ifdef CONFIG_SCHEDSTATS
2672         struct task_struct *tsk = NULL;
2673 
2674         if (entity_is_task(se))
2675                 tsk = task_of(se);
2676 
2677         if (se->statistics.sleep_start) {
2678                 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2679 
2680                 if ((s64)delta < 0)
2681                         delta = 0;
2682 
2683                 if (unlikely(delta > se->statistics.sleep_max))
2684                         se->statistics.sleep_max = delta;
2685 
2686                 se->statistics.sleep_start = 0;
2687                 se->statistics.sum_sleep_runtime += delta;
2688 
2689                 if (tsk) {
2690                         account_scheduler_latency(tsk, delta >> 10, 1);
2691                         trace_sched_stat_sleep(tsk, delta);
2692                 }
2693         }
2694         if (se->statistics.block_start) {
2695                 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2696 
2697                 if ((s64)delta < 0)
2698                         delta = 0;
2699 
2700                 if (unlikely(delta > se->statistics.block_max))
2701                         se->statistics.block_max = delta;
2702 
2703                 se->statistics.block_start = 0;
2704                 se->statistics.sum_sleep_runtime += delta;
2705 
2706                 if (tsk) {
2707                         if (tsk->in_iowait) {
2708                                 se->statistics.iowait_sum += delta;
2709                                 se->statistics.iowait_count++;
2710                                 trace_sched_stat_iowait(tsk, delta);
2711                         }
2712 
2713                         trace_sched_stat_blocked(tsk, delta);
2714 
2715                         /*
2716                          * Blocking time is in units of nanosecs, so shift by
2717                          * 20 to get a milliseconds-range estimation of the
2718                          * amount of time that the task spent sleeping:
2719                          */
2720                         if (unlikely(prof_on == SLEEP_PROFILING)) {
2721                                 profile_hits(SLEEP_PROFILING,
2722                                                 (void *)get_wchan(tsk),
2723                                                 delta >> 20);
2724                         }
2725                         account_scheduler_latency(tsk, delta >> 10, 0);
2726                 }
2727         }
2728 #endif
2729 }
2730 
2731 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2732 {
2733 #ifdef CONFIG_SCHED_DEBUG
2734         s64 d = se->vruntime - cfs_rq->min_vruntime;
2735 
2736         if (d < 0)
2737                 d = -d;
2738 
2739         if (d > 3*sysctl_sched_latency)
2740                 schedstat_inc(cfs_rq, nr_spread_over);
2741 #endif
2742 }
2743 
2744 static void
2745 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2746 {
2747         u64 vruntime = cfs_rq->min_vruntime;
2748 
2749         /*
2750          * The 'current' period is already promised to the current tasks,
2751          * however the extra weight of the new task will slow them down a
2752          * little, place the new task so that it fits in the slot that
2753          * stays open at the end.
2754          */
2755         if (initial && sched_feat(START_DEBIT))
2756                 vruntime += sched_vslice(cfs_rq, se);
2757 
2758         /* sleeps up to a single latency don't count. */
2759         if (!initial) {
2760                 unsigned long thresh = sysctl_sched_latency;
2761 
2762                 /*
2763                  * Halve their sleep time's effect, to allow
2764                  * for a gentler effect of sleepers:
2765                  */
2766                 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2767                         thresh >>= 1;
2768 
2769                 vruntime -= thresh;
2770         }
2771 
2772         /* ensure we never gain time by being placed backwards. */
2773         se->vruntime = max_vruntime(se->vruntime, vruntime);
2774 }
2775 
2776 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2777 
2778 static void
2779 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2780 {
2781         /*
2782          * Update the normalized vruntime before updating min_vruntime
2783          * through calling update_curr().
2784          */
2785         if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2786                 se->vruntime += cfs_rq->min_vruntime;
2787 
2788         /*
2789          * Update run-time statistics of the 'current'.
2790          */
2791         update_curr(cfs_rq);
2792         enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2793         account_entity_enqueue(cfs_rq, se);
2794         update_cfs_shares(cfs_rq);
2795 
2796         if (flags & ENQUEUE_WAKEUP) {
2797                 place_entity(cfs_rq, se, 0);
2798                 enqueue_sleeper(cfs_rq, se);
2799         }
2800 
2801         update_stats_enqueue(cfs_rq, se);
2802         check_spread(cfs_rq, se);
2803         if (se != cfs_rq->curr)
2804                 __enqueue_entity(cfs_rq, se);
2805         se->on_rq = 1;
2806 
2807         if (cfs_rq->nr_running == 1) {
2808                 list_add_leaf_cfs_rq(cfs_rq);
2809                 check_enqueue_throttle(cfs_rq);
2810         }
2811 }
2812 
2813 static void __clear_buddies_last(struct sched_entity *se)
2814 {
2815         for_each_sched_entity(se) {
2816                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2817                 if (cfs_rq->last != se)
2818                         break;
2819 
2820                 cfs_rq->last = NULL;
2821         }
2822 }
2823 
2824 static void __clear_buddies_next(struct sched_entity *se)
2825 {
2826         for_each_sched_entity(se) {
2827                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2828                 if (cfs_rq->next != se)
2829                         break;
2830 
2831                 cfs_rq->next = NULL;
2832         }
2833 }
2834 
2835 static void __clear_buddies_skip(struct sched_entity *se)
2836 {
2837         for_each_sched_entity(se) {
2838                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2839                 if (cfs_rq->skip != se)
2840                         break;
2841 
2842                 cfs_rq->skip = NULL;
2843         }
2844 }
2845 
2846 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2847 {
2848         if (cfs_rq->last == se)
2849                 __clear_buddies_last(se);
2850 
2851         if (cfs_rq->next == se)
2852                 __clear_buddies_next(se);
2853 
2854         if (cfs_rq->skip == se)
2855                 __clear_buddies_skip(se);
2856 }
2857 
2858 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2859 
2860 static void
2861 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2862 {
2863         /*
2864          * Update run-time statistics of the 'current'.
2865          */
2866         update_curr(cfs_rq);
2867         dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2868 
2869         update_stats_dequeue(cfs_rq, se);
2870         if (flags & DEQUEUE_SLEEP) {
2871 #ifdef CONFIG_SCHEDSTATS
2872                 if (entity_is_task(se)) {
2873                         struct task_struct *tsk = task_of(se);
2874 
2875                         if (tsk->state & TASK_INTERRUPTIBLE)
2876                                 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2877                         if (tsk->state & TASK_UNINTERRUPTIBLE)
2878                                 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2879                 }
2880 #endif
2881         }
2882 
2883         clear_buddies(cfs_rq, se);
2884 
2885         if (se != cfs_rq->curr)
2886                 __dequeue_entity(cfs_rq, se);
2887         se->on_rq = 0;
2888         account_entity_dequeue(cfs_rq, se);
2889 
2890         /*
2891          * Normalize the entity after updating the min_vruntime because the
2892          * update can refer to the ->curr item and we need to reflect this
2893          * movement in our normalized position.
2894          */
2895         if (!(flags & DEQUEUE_SLEEP))
2896                 se->vruntime -= cfs_rq->min_vruntime;
2897 
2898         /* return excess runtime on last dequeue */
2899         return_cfs_rq_runtime(cfs_rq);
2900 
2901         update_min_vruntime(cfs_rq);
2902         update_cfs_shares(cfs_rq);
2903 }
2904 
2905 /*
2906  * Preempt the current task with a newly woken task if needed:
2907  */
2908 static void
2909 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2910 {
2911         unsigned long ideal_runtime, delta_exec;
2912         struct sched_entity *se;
2913         s64 delta;
2914 
2915         ideal_runtime = sched_slice(cfs_rq, curr);
2916         delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2917         if (delta_exec > ideal_runtime) {
2918                 resched_curr(rq_of(cfs_rq));
2919                 /*
2920                  * The current task ran long enough, ensure it doesn't get
2921                  * re-elected due to buddy favours.
2922                  */
2923                 clear_buddies(cfs_rq, curr);
2924                 return;
2925         }
2926 
2927         /*
2928          * Ensure that a task that missed wakeup preemption by a
2929          * narrow margin doesn't have to wait for a full slice.
2930          * This also mitigates buddy induced latencies under load.
2931          */
2932         if (delta_exec < sysctl_sched_min_granularity)
2933                 return;
2934 
2935         se = __pick_first_entity(cfs_rq);
2936         delta = curr->vruntime - se->vruntime;
2937 
2938         if (delta < 0)
2939                 return;
2940 
2941         if (delta > ideal_runtime)
2942                 resched_curr(rq_of(cfs_rq));
2943 }
2944 
2945 static void
2946 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2947 {
2948         /* 'current' is not kept within the tree. */
2949         if (se->on_rq) {
2950                 /*
2951                  * Any task has to be enqueued before it get to execute on
2952                  * a CPU. So account for the time it spent waiting on the
2953                  * runqueue.
2954                  */
2955                 update_stats_wait_end(cfs_rq, se);
2956                 __dequeue_entity(cfs_rq, se);
2957         }
2958 
2959         update_stats_curr_start(cfs_rq, se);
2960         cfs_rq->curr = se;
2961 #ifdef CONFIG_SCHEDSTATS
2962         /*
2963          * Track our maximum slice length, if the CPU's load is at
2964          * least twice that of our own weight (i.e. dont track it
2965          * when there are only lesser-weight tasks around):
2966          */
2967         if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2968                 se->statistics.slice_max = max(se->statistics.slice_max,
2969                         se->sum_exec_runtime - se->prev_sum_exec_runtime);
2970         }
2971 #endif
2972         se->prev_sum_exec_runtime = se->sum_exec_runtime;
2973 }
2974 
2975 static int
2976 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2977 
2978 /*
2979  * Pick the next process, keeping these things in mind, in this order:
2980  * 1) keep things fair between processes/task groups
2981  * 2) pick the "next" process, since someone really wants that to run
2982  * 3) pick the "last" process, for cache locality
2983  * 4) do not run the "skip" process, if something else is available
2984  */
2985 static struct sched_entity *
2986 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2987 {
2988         struct sched_entity *left = __pick_first_entity(cfs_rq);
2989         struct sched_entity *se;
2990 
2991         /*
2992          * If curr is set we have to see if its left of the leftmost entity
2993          * still in the tree, provided there was anything in the tree at all.
2994          */
2995         if (!left || (curr && entity_before(curr, left)))
2996                 left = curr;
2997 
2998         se = left; /* ideally we run the leftmost entity */
2999 
3000         /*
3001          * Avoid running the skip buddy, if running something else can
3002          * be done without getting too unfair.
3003          */
3004         if (cfs_rq->skip == se) {
3005                 struct sched_entity *second;
3006 
3007                 if (se == curr) {
3008                         second = __pick_first_entity(cfs_rq);
3009                 } else {
3010                         second = __pick_next_entity(se);
3011                         if (!second || (curr && entity_before(curr, second)))
3012                                 second = curr;
3013                 }
3014 
3015                 if (second && wakeup_preempt_entity(second, left) < 1)
3016                         se = second;
3017         }
3018 
3019         /*
3020          * Prefer last buddy, try to return the CPU to a preempted task.
3021          */
3022         if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3023                 se = cfs_rq->last;
3024 
3025         /*
3026          * Someone really wants this to run. If it's not unfair, run it.
3027          */
3028         if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3029                 se = cfs_rq->next;
3030 
3031         clear_buddies(cfs_rq, se);
3032 
3033         return se;
3034 }
3035 
3036 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3037 
3038 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3039 {
3040         /*
3041          * If still on the runqueue then deactivate_task()
3042          * was not called and update_curr() has to be done:
3043          */
3044         if (prev->on_rq)
3045                 update_curr(cfs_rq);
3046 
3047         /* throttle cfs_rqs exceeding runtime */
3048         check_cfs_rq_runtime(cfs_rq);
3049 
3050         check_spread(cfs_rq, prev);
3051         if (prev->on_rq) {
3052                 update_stats_wait_start(cfs_rq, prev);
3053                 /* Put 'current' back into the tree. */
3054                 __enqueue_entity(cfs_rq, prev);
3055                 /* in !on_rq case, update occurred at dequeue */
3056                 update_entity_load_avg(prev, 1);
3057         }
3058         cfs_rq->curr = NULL;
3059 }
3060 
3061 static void
3062 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3063 {
3064         /*
3065          * Update run-time statistics of the 'current'.
3066          */
3067         update_curr(cfs_rq);
3068 
3069         /*
3070          * Ensure that runnable average is periodically updated.
3071          */
3072         update_entity_load_avg(curr, 1);
3073         update_cfs_rq_blocked_load(cfs_rq, 1);
3074         update_cfs_shares(cfs_rq);
3075 
3076 #ifdef CONFIG_SCHED_HRTICK
3077         /*
3078          * queued ticks are scheduled to match the slice, so don't bother
3079          * validating it and just reschedule.
3080          */
3081         if (queued) {
3082                 resched_curr(rq_of(cfs_rq));
3083                 return;
3084         }
3085         /*
3086          * don't let the period tick interfere with the hrtick preemption
3087          */
3088         if (!sched_feat(DOUBLE_TICK) &&
3089                         hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3090                 return;
3091 #endif
3092 
3093         if (cfs_rq->nr_running > 1)
3094                 check_preempt_tick(cfs_rq, curr);
3095 }
3096 
3097 
3098 /**************************************************
3099  * CFS bandwidth control machinery
3100  */
3101 
3102 #ifdef CONFIG_CFS_BANDWIDTH
3103 
3104 #ifdef HAVE_JUMP_LABEL
3105 static struct static_key __cfs_bandwidth_used;
3106 
3107 static inline bool cfs_bandwidth_used(void)
3108 {
3109         return static_key_false(&__cfs_bandwidth_used);
3110 }
3111 
3112 void cfs_bandwidth_usage_inc(void)
3113 {
3114         static_key_slow_inc(&__cfs_bandwidth_used);
3115 }
3116 
3117 void cfs_bandwidth_usage_dec(void)
3118 {
3119         static_key_slow_dec(&__cfs_bandwidth_used);
3120 }
3121 #else /* HAVE_JUMP_LABEL */
3122 static bool cfs_bandwidth_used(void)
3123 {
3124         return true;
3125 }
3126 
3127 void cfs_bandwidth_usage_inc(void) {}
3128 void cfs_bandwidth_usage_dec(void) {}
3129 #endif /* HAVE_JUMP_LABEL */
3130 
3131 /*
3132  * default period for cfs group bandwidth.
3133  * default: 0.1s, units: nanoseconds
3134  */
3135 static inline u64 default_cfs_period(void)
3136 {
3137         return 100000000ULL;
3138 }
3139 
3140 static inline u64 sched_cfs_bandwidth_slice(void)
3141 {
3142         return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3143 }
3144 
3145 /*
3146  * Replenish runtime according to assigned quota and update expiration time.
3147  * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3148  * additional synchronization around rq->lock.
3149  *
3150  * requires cfs_b->lock
3151  */
3152 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3153 {
3154         u64 now;
3155 
3156         if (cfs_b->quota == RUNTIME_INF)
3157                 return;
3158 
3159         now = sched_clock_cpu(smp_processor_id());
3160         cfs_b->runtime = cfs_b->quota;
3161         cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3162 }
3163 
3164 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3165 {
3166         return &tg->cfs_bandwidth;
3167 }
3168 
3169 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3170 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3171 {
3172         if (unlikely(cfs_rq->throttle_count))
3173                 return cfs_rq->throttled_clock_task;
3174 
3175         return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3176 }
3177 
3178 /* returns 0 on failure to allocate runtime */
3179 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3180 {
3181         struct task_group *tg = cfs_rq->tg;
3182         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3183         u64 amount = 0, min_amount, expires;
3184 
3185         /* note: this is a positive sum as runtime_remaining <= 0 */
3186         min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3187 
3188         raw_spin_lock(&cfs_b->lock);
3189         if (cfs_b->quota == RUNTIME_INF)
3190                 amount = min_amount;
3191         else {
3192                 /*
3193                  * If the bandwidth pool has become inactive, then at least one
3194                  * period must have elapsed since the last consumption.
3195                  * Refresh the global state and ensure bandwidth timer becomes
3196                  * active.
3197                  */
3198                 if (!cfs_b->timer_active) {
3199                         __refill_cfs_bandwidth_runtime(cfs_b);
3200                         __start_cfs_bandwidth(cfs_b, false);
3201                 }
3202 
3203                 if (cfs_b->runtime > 0) {
3204                         amount = min(cfs_b->runtime, min_amount);
3205                         cfs_b->runtime -= amount;
3206                         cfs_b->idle = 0;
3207                 }
3208         }
3209         expires = cfs_b->runtime_expires;
3210         raw_spin_unlock(&cfs_b->lock);
3211 
3212         cfs_rq->runtime_remaining += amount;
3213         /*
3214          * we may have advanced our local expiration to account for allowed
3215          * spread between our sched_clock and the one on which runtime was
3216          * issued.
3217          */
3218         if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3219                 cfs_rq->runtime_expires = expires;
3220 
3221         return cfs_rq->runtime_remaining > 0;
3222 }
3223 
3224 /*
3225  * Note: This depends on the synchronization provided by sched_clock and the
3226  * fact that rq->clock snapshots this value.
3227  */
3228 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3229 {
3230         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3231 
3232         /* if the deadline is ahead of our clock, nothing to do */
3233         if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3234                 return;
3235 
3236         if (cfs_rq->runtime_remaining < 0)
3237                 return;
3238 
3239         /*
3240          * If the local deadline has passed we have to consider the
3241          * possibility that our sched_clock is 'fast' and the global deadline
3242          * has not truly expired.
3243          *
3244          * Fortunately we can check determine whether this the case by checking
3245          * whether the global deadline has advanced. It is valid to compare
3246          * cfs_b->runtime_expires without any locks since we only care about
3247          * exact equality, so a partial write will still work.
3248          */
3249 
3250         if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3251                 /* extend local deadline, drift is bounded above by 2 ticks */
3252                 cfs_rq->runtime_expires += TICK_NSEC;
3253         } else {
3254                 /* global deadline is ahead, expiration has passed */
3255                 cfs_rq->runtime_remaining = 0;
3256         }
3257 }
3258 
3259 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3260 {
3261         /* dock delta_exec before expiring quota (as it could span periods) */
3262         cfs_rq->runtime_remaining -= delta_exec;
3263         expire_cfs_rq_runtime(cfs_rq);
3264 
3265         if (likely(cfs_rq->runtime_remaining > 0))
3266                 return;
3267 
3268         /*
3269          * if we're unable to extend our runtime we resched so that the active
3270          * hierarchy can be throttled
3271          */
3272         if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3273                 resched_curr(rq_of(cfs_rq));
3274 }
3275 
3276 static __always_inline
3277 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3278 {
3279         if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3280                 return;
3281 
3282         __account_cfs_rq_runtime(cfs_rq, delta_exec);
3283 }
3284 
3285 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3286 {
3287         return cfs_bandwidth_used() && cfs_rq->throttled;
3288 }
3289 
3290 /* check whether cfs_rq, or any parent, is throttled */
3291 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3292 {
3293         return cfs_bandwidth_used() && cfs_rq->throttle_count;
3294 }
3295 
3296 /*
3297  * Ensure that neither of the group entities corresponding to src_cpu or
3298  * dest_cpu are members of a throttled hierarchy when performing group
3299  * load-balance operations.
3300  */
3301 static inline int throttled_lb_pair(struct task_group *tg,
3302                                     int src_cpu, int dest_cpu)
3303 {
3304         struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3305 
3306         src_cfs_rq = tg->cfs_rq[src_cpu];
3307         dest_cfs_rq = tg->cfs_rq[dest_cpu];
3308 
3309         return throttled_hierarchy(src_cfs_rq) ||
3310                throttled_hierarchy(dest_cfs_rq);
3311 }
3312 
3313 /* updated child weight may affect parent so we have to do this bottom up */
3314 static int tg_unthrottle_up(struct task_group *tg, void *data)
3315 {
3316         struct rq *rq = data;
3317         struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3318 
3319         cfs_rq->throttle_count--;
3320 #ifdef CONFIG_SMP
3321         if (!cfs_rq->throttle_count) {
3322                 /* adjust cfs_rq_clock_task() */
3323                 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3324                                              cfs_rq->throttled_clock_task;
3325         }
3326 #endif
3327 
3328         return 0;
3329 }
3330 
3331 static int tg_throttle_down(struct task_group *tg, void *data)
3332 {
3333         struct rq *rq = data;
3334         struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3335 
3336         /* group is entering throttled state, stop time */
3337         if (!cfs_rq->throttle_count)
3338                 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3339         cfs_rq->throttle_count++;
3340 
3341         return 0;
3342 }
3343 
3344 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3345 {
3346         struct rq *rq = rq_of(cfs_rq);
3347         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3348         struct sched_entity *se;
3349         long task_delta, dequeue = 1;
3350 
3351         se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3352 
3353         /* freeze hierarchy runnable averages while throttled */
3354         rcu_read_lock();
3355         walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3356         rcu_read_unlock();
3357 
3358         task_delta = cfs_rq->h_nr_running;
3359         for_each_sched_entity(se) {
3360                 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3361                 /* throttled entity or throttle-on-deactivate */
3362                 if (!se->on_rq)
3363                         break;
3364 
3365                 if (dequeue)
3366                         dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3367                 qcfs_rq->h_nr_running -= task_delta;
3368 
3369                 if (qcfs_rq->load.weight)
3370                         dequeue = 0;
3371         }
3372 
3373         if (!se)
3374                 sub_nr_running(rq, task_delta);
3375 
3376         cfs_rq->throttled = 1;
3377         cfs_rq->throttled_clock = rq_clock(rq);
3378         raw_spin_lock(&cfs_b->lock);
3379         /*
3380          * Add to the _head_ of the list, so that an already-started
3381          * distribute_cfs_runtime will not see us
3382          */
3383         list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3384         if (!cfs_b->timer_active)
3385                 __start_cfs_bandwidth(cfs_b, false);
3386         raw_spin_unlock(&cfs_b->lock);
3387 }
3388 
3389 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3390 {
3391         struct rq *rq = rq_of(cfs_rq);
3392         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3393         struct sched_entity *se;
3394         int enqueue = 1;
3395         long task_delta;
3396 
3397         se = cfs_rq->tg->se[cpu_of(rq)];
3398 
3399         cfs_rq->throttled = 0;
3400 
3401         update_rq_clock(rq);
3402 
3403         raw_spin_lock(&cfs_b->lock);
3404         cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3405         list_del_rcu(&cfs_rq->throttled_list);
3406         raw_spin_unlock(&cfs_b->lock);
3407 
3408         /* update hierarchical throttle state */
3409         walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3410 
3411         if (!cfs_rq->load.weight)
3412                 return;
3413 
3414         task_delta = cfs_rq->h_nr_running;
3415         for_each_sched_entity(se) {
3416                 if (se->on_rq)
3417                         enqueue = 0;
3418 
3419                 cfs_rq = cfs_rq_of(se);
3420                 if (enqueue)
3421                         enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3422                 cfs_rq->h_nr_running += task_delta;
3423 
3424                 if (cfs_rq_throttled(cfs_rq))
3425                         break;
3426         }
3427 
3428         if (!se)
3429                 add_nr_running(rq, task_delta);
3430 
3431         /* determine whether we need to wake up potentially idle cpu */
3432         if (rq->curr == rq->idle && rq->cfs.nr_running)
3433                 resched_curr(rq);
3434 }
3435 
3436 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3437                 u64 remaining, u64 expires)
3438 {
3439         struct cfs_rq *cfs_rq;
3440         u64 runtime;
3441         u64 starting_runtime = remaining;
3442 
3443         rcu_read_lock();
3444         list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3445                                 throttled_list) {
3446                 struct rq *rq = rq_of(cfs_rq);
3447 
3448                 raw_spin_lock(&rq->lock);
3449                 if (!cfs_rq_throttled(cfs_rq))
3450                         goto next;
3451 
3452                 runtime = -cfs_rq->runtime_remaining + 1;
3453                 if (runtime > remaining)
3454                         runtime = remaining;
3455                 remaining -= runtime;
3456 
3457                 cfs_rq->runtime_remaining += runtime;
3458                 cfs_rq->runtime_expires = expires;
3459 
3460                 /* we check whether we're throttled above */
3461                 if (cfs_rq->runtime_remaining > 0)
3462                         unthrottle_cfs_rq(cfs_rq);
3463 
3464 next:
3465                 raw_spin_unlock(&rq->lock);
3466 
3467                 if (!remaining)
3468                         break;
3469         }
3470         rcu_read_unlock();
3471 
3472         return starting_runtime - remaining;
3473 }
3474 
3475 /*
3476  * Responsible for refilling a task_group's bandwidth and unthrottling its
3477  * cfs_rqs as appropriate. If there has been no activity within the last
3478  * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3479  * used to track this state.
3480  */
3481 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3482 {
3483         u64 runtime, runtime_expires;
3484         int throttled;
3485 
3486         /* no need to continue the timer with no bandwidth constraint */
3487         if (cfs_b->quota == RUNTIME_INF)
3488                 goto out_deactivate;
3489 
3490         throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3491         cfs_b->nr_periods += overrun;
3492 
3493         /*
3494          * idle depends on !throttled (for the case of a large deficit), and if
3495          * we're going inactive then everything else can be deferred
3496          */
3497         if (cfs_b->idle && !throttled)
3498                 goto out_deactivate;
3499 
3500         /*
3501          * if we have relooped after returning idle once, we need to update our
3502          * status as actually running, so that other cpus doing
3503          * __start_cfs_bandwidth will stop trying to cancel us.
3504          */
3505         cfs_b->timer_active = 1;
3506 
3507         __refill_cfs_bandwidth_runtime(cfs_b);
3508 
3509         if (!throttled) {
3510                 /* mark as potentially idle for the upcoming period */
3511                 cfs_b->idle = 1;
3512                 return 0;
3513         }
3514 
3515         /* account preceding periods in which throttling occurred */
3516         cfs_b->nr_throttled += overrun;
3517 
3518         runtime_expires = cfs_b->runtime_expires;
3519 
3520         /*
3521          * This check is repeated as we are holding onto the new bandwidth while
3522          * we unthrottle. This can potentially race with an unthrottled group
3523          * trying to acquire new bandwidth from the global pool. This can result
3524          * in us over-using our runtime if it is all used during this loop, but
3525          * only by limited amounts in that extreme case.
3526          */
3527         while (throttled && cfs_b->runtime > 0) {
3528                 runtime = cfs_b->runtime;
3529                 raw_spin_unlock(&cfs_b->lock);
3530                 /* we can't nest cfs_b->lock while distributing bandwidth */
3531                 runtime = distribute_cfs_runtime(cfs_b, runtime,
3532                                                  runtime_expires);
3533                 raw_spin_lock(&cfs_b->lock);
3534 
3535                 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3536 
3537                 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3538         }
3539 
3540         /*
3541          * While we are ensured activity in the period following an
3542          * unthrottle, this also covers the case in which the new bandwidth is
3543          * insufficient to cover the existing bandwidth deficit.  (Forcing the
3544          * timer to remain active while there are any throttled entities.)
3545          */
3546         cfs_b->idle = 0;
3547 
3548         return 0;
3549 
3550 out_deactivate:
3551         cfs_b->timer_active = 0;
3552         return 1;
3553 }
3554 
3555 /* a cfs_rq won't donate quota below this amount */
3556 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3557 /* minimum remaining period time to redistribute slack quota */
3558 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3559 /* how long we wait to gather additional slack before distributing */
3560 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3561 
3562 /*
3563  * Are we near the end of the current quota period?
3564  *
3565  * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3566  * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3567  * migrate_hrtimers, base is never cleared, so we are fine.
3568  */
3569 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3570 {
3571         struct hrtimer *refresh_timer = &cfs_b->period_timer;
3572         u64 remaining;
3573 
3574         /* if the call-back is running a quota refresh is already occurring */
3575         if (hrtimer_callback_running(refresh_timer))
3576                 return 1;
3577 
3578         /* is a quota refresh about to occur? */
3579         remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3580         if (remaining < min_expire)
3581                 return 1;
3582 
3583         return 0;
3584 }
3585 
3586 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3587 {
3588         u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3589 
3590         /* if there's a quota refresh soon don't bother with slack */
3591         if (runtime_refresh_within(cfs_b, min_left))
3592                 return;
3593 
3594         start_bandwidth_timer(&cfs_b->slack_timer,
3595                                 ns_to_ktime(cfs_bandwidth_slack_period));
3596 }
3597 
3598 /* we know any runtime found here is valid as update_curr() precedes return */
3599 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3600 {
3601         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3602         s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3603 
3604         if (slack_runtime <= 0)
3605                 return;
3606 
3607         raw_spin_lock(&cfs_b->lock);
3608         if (cfs_b->quota != RUNTIME_INF &&
3609             cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3610                 cfs_b->runtime += slack_runtime;
3611 
3612                 /* we are under rq->lock, defer unthrottling using a timer */
3613                 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3614                     !list_empty(&cfs_b->throttled_cfs_rq))
3615                         start_cfs_slack_bandwidth(cfs_b);
3616         }
3617         raw_spin_unlock(&cfs_b->lock);
3618 
3619         /* even if it's not valid for return we don't want to try again */
3620         cfs_rq->runtime_remaining -= slack_runtime;
3621 }
3622 
3623 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3624 {
3625         if (!cfs_bandwidth_used())
3626                 return;
3627 
3628         if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3629                 return;
3630 
3631         __return_cfs_rq_runtime(cfs_rq);
3632 }
3633 
3634 /*
3635  * This is done with a timer (instead of inline with bandwidth return) since
3636  * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3637  */
3638 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3639 {
3640         u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3641         u64 expires;
3642 
3643         /* confirm we're still not at a refresh boundary */
3644         raw_spin_lock(&cfs_b->lock);
3645         if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3646                 raw_spin_unlock(&cfs_b->lock);
3647                 return;
3648         }
3649 
3650         if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3651                 runtime = cfs_b->runtime;
3652 
3653         expires = cfs_b->runtime_expires;
3654         raw_spin_unlock(&cfs_b->lock);
3655 
3656         if (!runtime)
3657                 return;
3658 
3659         runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3660 
3661         raw_spin_lock(&cfs_b->lock);
3662         if (expires == cfs_b->runtime_expires)
3663                 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3664         raw_spin_unlock(&cfs_b->lock);
3665 }
3666 
3667 /*
3668  * When a group wakes up we want to make sure that its quota is not already
3669  * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3670  * runtime as update_curr() throttling can not not trigger until it's on-rq.
3671  */
3672 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3673 {
3674         if (!cfs_bandwidth_used())
3675                 return;
3676 
3677         /* an active group must be handled by the update_curr()->put() path */
3678         if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3679                 return;
3680 
3681         /* ensure the group is not already throttled */
3682         if (cfs_rq_throttled(cfs_rq))
3683                 return;
3684 
3685         /* update runtime allocation */
3686         account_cfs_rq_runtime(cfs_rq, 0);
3687         if (cfs_rq->runtime_remaining <= 0)
3688                 throttle_cfs_rq(cfs_rq);
3689 }
3690 
3691 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3692 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3693 {
3694         if (!cfs_bandwidth_used())
3695                 return false;
3696 
3697         if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3698                 return false;
3699 
3700         /*
3701          * it's possible for a throttled entity to be forced into a running
3702          * state (e.g. set_curr_task), in this case we're finished.
3703          */
3704         if (cfs_rq_throttled(cfs_rq))
3705                 return true;
3706 
3707         throttle_cfs_rq(cfs_rq);
3708         return true;
3709 }
3710 
3711 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3712 {
3713         struct cfs_bandwidth *cfs_b =
3714                 container_of(timer, struct cfs_bandwidth, slack_timer);
3715         do_sched_cfs_slack_timer(cfs_b);
3716 
3717         return HRTIMER_NORESTART;
3718 }
3719 
3720 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3721 {
3722         struct cfs_bandwidth *cfs_b =
3723                 container_of(timer, struct cfs_bandwidth, period_timer);
3724         ktime_t now;
3725         int overrun;
3726         int idle = 0;
3727 
3728         raw_spin_lock(&cfs_b->lock);
3729         for (;;) {
3730                 now = hrtimer_cb_get_time(timer);
3731                 overrun = hrtimer_forward(timer, now, cfs_b->period);
3732 
3733                 if (!overrun)
3734                         break;
3735 
3736                 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3737         }
3738         raw_spin_unlock(&cfs_b->lock);
3739 
3740         return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3741 }
3742 
3743 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3744 {
3745         raw_spin_lock_init(&cfs_b->lock);
3746         cfs_b->runtime = 0;
3747         cfs_b->quota = RUNTIME_INF;
3748         cfs_b->period = ns_to_ktime(default_cfs_period());
3749 
3750         INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3751         hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3752         cfs_b->period_timer.function = sched_cfs_period_timer;
3753         hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3754         cfs_b->slack_timer.function = sched_cfs_slack_timer;
3755 }
3756 
3757 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3758 {
3759         cfs_rq->runtime_enabled = 0;
3760         INIT_LIST_HEAD(&cfs_rq->throttled_list);
3761 }
3762 
3763 /* requires cfs_b->lock, may release to reprogram timer */
3764 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b, bool force)
3765 {
3766         /*
3767          * The timer may be active because we're trying to set a new bandwidth
3768          * period or because we're racing with the tear-down path
3769          * (timer_active==0 becomes visible before the hrtimer call-back
3770          * terminates).  In either case we ensure that it's re-programmed
3771          */
3772         while (unlikely(hrtimer_active(&cfs_b->period_timer)) &&
3773                hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) {
3774                 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3775                 raw_spin_unlock(&cfs_b->lock);
3776                 cpu_relax();
3777                 raw_spin_lock(&cfs_b->lock);
3778                 /* if someone else restarted the timer then we're done */
3779                 if (!force && cfs_b->timer_active)
3780                         return;
3781         }
3782 
3783         cfs_b->timer_active = 1;
3784         start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3785 }
3786 
3787 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3788 {
3789         hrtimer_cancel(&cfs_b->period_timer);
3790         hrtimer_cancel(&cfs_b->slack_timer);
3791 }
3792 
3793 static void __maybe_unused update_runtime_enabled(struct rq *rq)
3794 {
3795         struct cfs_rq *cfs_rq;
3796 
3797         for_each_leaf_cfs_rq(rq, cfs_rq) {
3798                 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
3799 
3800                 raw_spin_lock(&cfs_b->lock);
3801                 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
3802                 raw_spin_unlock(&cfs_b->lock);
3803         }
3804 }
3805 
3806 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3807 {
3808         struct cfs_rq *cfs_rq;
3809 
3810         for_each_leaf_cfs_rq(rq, cfs_rq) {
3811                 if (!cfs_rq->runtime_enabled)
3812                         continue;
3813 
3814                 /*
3815                  * clock_task is not advancing so we just need to make sure
3816                  * there's some valid quota amount
3817                  */
3818                 cfs_rq->runtime_remaining = 1;
3819                 /*
3820                  * Offline rq is schedulable till cpu is completely disabled
3821                  * in take_cpu_down(), so we prevent new cfs throttling here.
3822                  */
3823                 cfs_rq->runtime_enabled = 0;
3824 
3825                 if (cfs_rq_throttled(cfs_rq))
3826                         unthrottle_cfs_rq(cfs_rq);
3827         }
3828 }
3829 
3830 #else /* CONFIG_CFS_BANDWIDTH */
3831 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3832 {
3833         return rq_clock_task(rq_of(cfs_rq));
3834 }
3835 
3836 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
3837 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
3838 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3839 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3840 
3841 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3842 {
3843         return 0;
3844 }
3845 
3846 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3847 {
3848         return 0;
3849 }
3850 
3851 static inline int throttled_lb_pair(struct task_group *tg,
3852                                     int src_cpu, int dest_cpu)
3853 {
3854         return 0;
3855 }
3856 
3857 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3858 
3859 #ifdef CONFIG_FAIR_GROUP_SCHED
3860 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3861 #endif
3862 
3863 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3864 {
3865         return NULL;
3866 }
3867 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3868 static inline void update_runtime_enabled(struct rq *rq) {}
3869 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3870 
3871 #endif /* CONFIG_CFS_BANDWIDTH */
3872 
3873 /**************************************************
3874  * CFS operations on tasks:
3875  */
3876 
3877 #ifdef CONFIG_SCHED_HRTICK
3878 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3879 {
3880         struct sched_entity *se = &p->se;
3881         struct cfs_rq *cfs_rq = cfs_rq_of(se);
3882 
3883         WARN_ON(task_rq(p) != rq);
3884 
3885         if (cfs_rq->nr_running > 1) {
3886                 u64 slice = sched_slice(cfs_rq, se);
3887                 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3888                 s64 delta = slice - ran;
3889 
3890                 if (delta < 0) {
3891                         if (rq->curr == p)
3892                                 resched_curr(rq);
3893                         return;
3894                 }
3895 
3896                 /*
3897                  * Don't schedule slices shorter than 10000ns, that just
3898                  * doesn't make sense. Rely on vruntime for fairness.
3899                  */
3900                 if (rq->curr != p)
3901                         delta = max_t(s64, 10000LL, delta);
3902 
3903                 hrtick_start(rq, delta);
3904         }
3905 }
3906 
3907 /*
3908  * called from enqueue/dequeue and updates the hrtick when the
3909  * current task is from our class and nr_running is low enough
3910  * to matter.
3911  */
3912 static void hrtick_update(struct rq *rq)
3913 {
3914         struct task_struct *curr = rq->curr;
3915 
3916         if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3917                 return;
3918 
3919         if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3920                 hrtick_start_fair(rq, curr);
3921 }
3922 #else /* !CONFIG_SCHED_HRTICK */
3923 static inline void
3924 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3925 {
3926 }
3927 
3928 static inline void hrtick_update(struct rq *rq)
3929 {
3930 }
3931 #endif
3932 
3933 /*
3934  * The enqueue_task method is called before nr_running is
3935  * increased. Here we update the fair scheduling stats and
3936  * then put the task into the rbtree:
3937  */
3938 static void
3939 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3940 {
3941         struct cfs_rq *cfs_rq;
3942         struct sched_entity *se = &p->se;
3943 
3944         for_each_sched_entity(se) {
3945                 if (se->on_rq)
3946                         break;
3947                 cfs_rq = cfs_rq_of(se);
3948                 enqueue_entity(cfs_rq, se, flags);
3949 
3950                 /*
3951                  * end evaluation on encountering a throttled cfs_rq
3952                  *
3953                  * note: in the case of encountering a throttled cfs_rq we will
3954                  * post the final h_nr_running increment below.
3955                 */
3956                 if (cfs_rq_throttled(cfs_rq))
3957                         break;
3958                 cfs_rq->h_nr_running++;
3959 
3960                 flags = ENQUEUE_WAKEUP;
3961         }
3962 
3963         for_each_sched_entity(se) {
3964                 cfs_rq = cfs_rq_of(se);
3965                 cfs_rq->h_nr_running++;
3966 
3967                 if (cfs_rq_throttled(cfs_rq))
3968                         break;
3969 
3970                 update_cfs_shares(cfs_rq);
3971                 update_entity_load_avg(se, 1);
3972         }
3973 
3974         if (!se) {
3975                 update_rq_runnable_avg(rq, rq->nr_running);
3976                 add_nr_running(rq, 1);
3977         }
3978         hrtick_update(rq);
3979 }
3980 
3981 static void set_next_buddy(struct sched_entity *se);
3982 
3983 /*
3984  * The dequeue_task method is called before nr_running is
3985  * decreased. We remove the task from the rbtree and
3986  * update the fair scheduling stats:
3987  */
3988 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3989 {
3990         struct cfs_rq *cfs_rq;
3991         struct sched_entity *se = &p->se;
3992         int task_sleep = flags & DEQUEUE_SLEEP;
3993 
3994         for_each_sched_entity(se) {
3995                 cfs_rq = cfs_rq_of(se);
3996                 dequeue_entity(cfs_rq, se, flags);
3997 
3998                 /*
3999                  * end evaluation on encountering a throttled cfs_rq
4000                  *
4001                  * note: in the case of encountering a throttled cfs_rq we will
4002                  * post the final h_nr_running decrement below.
4003                 */
4004                 if (cfs_rq_throttled(cfs_rq))
4005                         break;
4006                 cfs_rq->h_nr_running--;
4007 
4008                 /* Don't dequeue parent if it has other entities besides us */
4009                 if (cfs_rq->load.weight) {
4010                         /*
4011                          * Bias pick_next to pick a task from this cfs_rq, as
4012                          * p is sleeping when it is within its sched_slice.
4013                          */
4014                         if (task_sleep && parent_entity(se))
4015                                 set_next_buddy(parent_entity(se));
4016 
4017                         /* avoid re-evaluating load for this entity */
4018                         se = parent_entity(se);
4019                         break;
4020                 }
4021                 flags |= DEQUEUE_SLEEP;
4022         }
4023 
4024         for_each_sched_entity(se) {
4025                 cfs_rq = cfs_rq_of(se);
4026                 cfs_rq->h_nr_running--;
4027 
4028                 if (cfs_rq_throttled(cfs_rq))
4029                         break;
4030 
4031                 update_cfs_shares(cfs_rq);
4032                 update_entity_load_avg(se, 1);
4033         }
4034 
4035         if (!se) {
4036                 sub_nr_running(rq, 1);
4037                 update_rq_runnable_avg(rq, 1);
4038         }
4039         hrtick_update(rq);
4040 }
4041 
4042 #ifdef CONFIG_SMP
4043 /* Used instead of source_load when we know the type == 0 */
4044 static unsigned long weighted_cpuload(const int cpu)
4045 {
4046         return cpu_rq(cpu)->cfs.runnable_load_avg;
4047 }
4048 
4049 /*
4050  * Return a low guess at the load of a migration-source cpu weighted
4051  * according to the scheduling class and "nice" value.
4052  *
4053  * We want to under-estimate the load of migration sources, to
4054  * balance conservatively.
4055  */
4056 static unsigned long source_load(int cpu, int type)
4057 {
4058         struct rq *rq = cpu_rq(cpu);
4059         unsigned long total = weighted_cpuload(cpu);
4060 
4061         if (type == 0 || !sched_feat(LB_BIAS))
4062                 return total;
4063 
4064         return min(rq->cpu_load[type-1], total);
4065 }
4066 
4067 /*
4068  * Return a high guess at the load of a migration-target cpu weighted
4069  * according to the scheduling class and "nice" value.
4070  */
4071 static unsigned long target_load(int cpu, int type)
4072 {
4073         struct rq *rq = cpu_rq(cpu);
4074         unsigned long total = weighted_cpuload(cpu);
4075 
4076         if (type == 0 || !sched_feat(LB_BIAS))
4077                 return total;
4078 
4079         return max(rq->cpu_load[type-1], total);
4080 }
4081 
4082 static unsigned long capacity_of(int cpu)
4083 {
4084         return cpu_rq(cpu)->cpu_capacity;
4085 }
4086 
4087 static unsigned long cpu_avg_load_per_task(int cpu)
4088 {
4089         struct rq *rq = cpu_rq(cpu);
4090         unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
4091         unsigned long load_avg = rq->cfs.runnable_load_avg;
4092 
4093         if (nr_running)
4094                 return load_avg / nr_running;
4095 
4096         return 0;
4097 }
4098 
4099 static void record_wakee(struct task_struct *p)
4100 {
4101         /*
4102          * Rough decay (wiping) for cost saving, don't worry
4103          * about the boundary, really active task won't care
4104          * about the loss.
4105          */
4106         if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4107                 current->wakee_flips >>= 1;
4108                 current->wakee_flip_decay_ts = jiffies;
4109         }
4110 
4111         if (current->last_wakee != p) {
4112                 current->last_wakee = p;
4113                 current->wakee_flips++;
4114         }
4115 }
4116 
4117 static void task_waking_fair(struct task_struct *p)
4118 {
4119         struct sched_entity *se = &p->se;
4120         struct cfs_rq *cfs_rq = cfs_rq_of(se);
4121         u64 min_vruntime;
4122 
4123 #ifndef CONFIG_64BIT
4124         u64 min_vruntime_copy;
4125 
4126         do {
4127                 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4128                 smp_rmb();
4129                 min_vruntime = cfs_rq->min_vruntime;
4130         } while (min_vruntime != min_vruntime_copy);
4131 #else
4132         min_vruntime = cfs_rq->min_vruntime;
4133 #endif
4134 
4135         se->vruntime -= min_vruntime;
4136         record_wakee(p);
4137 }
4138 
4139 #ifdef CONFIG_FAIR_GROUP_SCHED
4140 /*
4141  * effective_load() calculates the load change as seen from the root_task_group
4142  *
4143  * Adding load to a group doesn't make a group heavier, but can cause movement
4144  * of group shares between cpus. Assuming the shares were perfectly aligned one
4145  * can calculate the shift in shares.
4146  *
4147  * Calculate the effective load difference if @wl is added (subtracted) to @tg
4148  * on this @cpu and results in a total addition (subtraction) of @wg to the
4149  * total group weight.
4150  *
4151  * Given a runqueue weight distribution (rw_i) we can compute a shares
4152  * distribution (s_i) using:
4153  *
4154  *   s_i = rw_i / \Sum rw_j                                             (1)
4155  *
4156  * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4157  * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4158  * shares distribution (s_i):
4159  *
4160  *   rw_i = {   2,   4,   1,   0 }
4161  *   s_i  = { 2/7, 4/7, 1/7,   0 }
4162  *
4163  * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4164  * task used to run on and the CPU the waker is running on), we need to
4165  * compute the effect of waking a task on either CPU and, in case of a sync
4166  * wakeup, compute the effect of the current task going to sleep.
4167  *
4168  * So for a change of @wl to the local @cpu with an overall group weight change
4169  * of @wl we can compute the new shares distribution (s'_i) using:
4170  *
4171  *   s'_i = (rw_i + @wl) / (@wg + \Sum rw_j)                            (2)
4172  *
4173  * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4174  * differences in waking a task to CPU 0. The additional task changes the
4175  * weight and shares distributions like:
4176  *
4177  *   rw'_i = {   3,   4,   1,   0 }
4178  *   s'_i  = { 3/8, 4/8, 1/8,   0 }
4179  *
4180  * We can then compute the difference in effective weight by using:
4181  *
4182  *   dw_i = S * (s'_i - s_i)                                            (3)
4183  *
4184  * Where 'S' is the group weight as seen by its parent.
4185  *
4186  * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4187  * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4188  * 4/7) times the weight of the group.
4189  */
4190 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4191 {
4192         struct sched_entity *se = tg->se[cpu];
4193 
4194         if (!tg->parent)        /* the trivial, non-cgroup case */
4195                 return wl;
4196 
4197         for_each_sched_entity(se) {
4198                 long w, W;
4199 
4200                 tg = se->my_q->tg;
4201 
4202                 /*
4203                  * W = @wg + \Sum rw_j
4204                  */
4205                 W = wg + calc_tg_weight(tg, se->my_q);
4206 
4207                 /*
4208                  * w = rw_i + @wl
4209                  */
4210                 w = se->my_q->load.weight + wl;
4211 
4212                 /*
4213                  * wl = S * s'_i; see (2)
4214                  */
4215                 if (W > 0 && w < W)
4216                         wl = (w * tg->shares) / W;
4217                 else
4218                         wl = tg->shares;
4219 
4220                 /*
4221                  * Per the above, wl is the new se->load.weight value; since
4222                  * those are clipped to [MIN_SHARES, ...) do so now. See
4223                  * calc_cfs_shares().
4224                  */
4225                 if (wl < MIN_SHARES)
4226                         wl = MIN_SHARES;
4227 
4228                 /*
4229                  * wl = dw_i = S * (s'_i - s_i); see (3)
4230                  */
4231                 wl -= se->load.weight;
4232 
4233                 /*
4234                  * Recursively apply this logic to all parent groups to compute
4235                  * the final effective load change on the root group. Since
4236                  * only the @tg group gets extra weight, all parent groups can
4237                  * only redistribute existing shares. @wl is the shift in shares
4238                  * resulting from this level per the above.
4239                  */
4240                 wg = 0;
4241         }
4242 
4243         return wl;
4244 }
4245 #else
4246 
4247 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4248 {
4249         return wl;
4250 }
4251 
4252 #endif
4253 
4254 static int wake_wide(struct task_struct *p)
4255 {
4256         int factor = this_cpu_read(sd_llc_size);
4257 
4258         /*
4259          * Yeah, it's the switching-frequency, could means many wakee or
4260          * rapidly switch, use factor here will just help to automatically
4261          * adjust the loose-degree, so bigger node will lead to more pull.
4262          */
4263         if (p->wakee_flips > factor) {
4264                 /*
4265                  * wakee is somewhat hot, it needs certain amount of cpu
4266                  * resource, so if waker is far more hot, prefer to leave
4267                  * it alone.
4268                  */
4269                 if (current->wakee_flips > (factor * p->wakee_flips))
4270                         return 1;
4271         }
4272 
4273         return 0;
4274 }
4275 
4276 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4277 {
4278         s64 this_load, load;
4279         int idx, this_cpu, prev_cpu;
4280         unsigned long tl_per_task;
4281         struct task_group *tg;
4282         unsigned long weight;
4283         int balanced;
4284 
4285         /*
4286          * If we wake multiple tasks be careful to not bounce
4287          * ourselves around too much.
4288          */
4289         if (wake_wide(p))
4290                 return 0;
4291 
4292         idx       = sd->wake_idx;
4293         this_cpu  = smp_processor_id();
4294         prev_cpu  = task_cpu(p);
4295         load      = source_load(prev_cpu, idx);
4296         this_load = target_load(this_cpu, idx);
4297 
4298         /*
4299          * If sync wakeup then subtract the (maximum possible)
4300          * effect of the currently running task from the load
4301          * of the current CPU:
4302          */
4303         if (sync) {
4304                 tg = task_group(current);
4305                 weight = current->se.load.weight;
4306 
4307                 this_load += effective_load(tg, this_cpu, -weight, -weight);
4308                 load += effective_load(tg, prev_cpu, 0, -weight);
4309         }
4310 
4311         tg = task_group(p);
4312         weight = p->se.load.weight;
4313 
4314         /*
4315          * In low-load situations, where prev_cpu is idle and this_cpu is idle
4316          * due to the sync cause above having dropped this_load to 0, we'll
4317          * always have an imbalance, but there's really nothing you can do
4318          * about that, so that's good too.
4319          *
4320          * Otherwise check if either cpus are near enough in load to allow this
4321          * task to be woken on this_cpu.
4322          */
4323         if (this_load > 0) {
4324                 s64 this_eff_load, prev_eff_load;
4325 
4326                 this_eff_load = 100;
4327                 this_eff_load *= capacity_of(prev_cpu);
4328                 this_eff_load *= this_load +
4329                         effective_load(tg, this_cpu, weight, weight);
4330 
4331                 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4332                 prev_eff_load *= capacity_of(this_cpu);
4333                 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4334 
4335                 balanced = this_eff_load <= prev_eff_load;
4336         } else
4337                 balanced = true;
4338 
4339         /*
4340          * If the currently running task will sleep within
4341          * a reasonable amount of time then attract this newly
4342          * woken task:
4343          */
4344         if (sync && balanced)
4345                 return 1;
4346 
4347         schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4348         tl_per_task = cpu_avg_load_per_task(this_cpu);
4349 
4350         if (balanced ||
4351             (this_load <= load &&
4352              this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
4353                 /*
4354                  * This domain has SD_WAKE_AFFINE and
4355                  * p is cache cold in this domain, and
4356                  * there is no bad imbalance.
4357                  */
4358                 schedstat_inc(sd, ttwu_move_affine);
4359                 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4360 
4361                 return 1;
4362         }
4363         return 0;
4364 }
4365 
4366 /*
4367  * find_idlest_group finds and returns the least busy CPU group within the
4368  * domain.
4369  */
4370 static struct sched_group *
4371 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4372                   int this_cpu, int sd_flag)
4373 {
4374         struct sched_group *idlest = NULL, *group = sd->groups;
4375         unsigned long min_load = ULONG_MAX, this_load = 0;
4376         int load_idx = sd->forkexec_idx;
4377         int imbalance = 100 + (sd->imbalance_pct-100)/2;
4378 
4379         if (sd_flag & SD_BALANCE_WAKE)
4380                 load_idx = sd->wake_idx;
4381 
4382         do {
4383                 unsigned long load, avg_load;
4384                 int local_group;
4385                 int i;
4386 
4387                 /* Skip over this group if it has no CPUs allowed */
4388                 if (!cpumask_intersects(sched_group_cpus(group),
4389                                         tsk_cpus_allowed(p)))
4390                         continue;
4391 
4392                 local_group = cpumask_test_cpu(this_cpu,
4393                                                sched_group_cpus(group));
4394 
4395                 /* Tally up the load of all CPUs in the group */
4396                 avg_load = 0;
4397 
4398                 for_each_cpu(i, sched_group_cpus(group)) {
4399                         /* Bias balancing toward cpus of our domain */
4400                         if (local_group)
4401                                 load = source_load(i, load_idx);
4402                         else
4403                                 load = target_load(i, load_idx);
4404 
4405                         avg_load += load;
4406                 }
4407 
4408                 /* Adjust by relative CPU capacity of the group */
4409                 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
4410 
4411                 if (local_group) {
4412                         this_load = avg_load;
4413                 } else if (avg_load < min_load) {
4414                         min_load = avg_load;
4415                         idlest = group;
4416                 }
4417         } while (group = group->next, group != sd->groups);
4418 
4419         if (!idlest || 100*this_load < imbalance*min_load)
4420                 return NULL;
4421         return idlest;
4422 }
4423 
4424 /*
4425  * find_idlest_cpu - find the idlest cpu among the cpus in group.
4426  */
4427 static int
4428 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4429 {
4430         unsigned long load, min_load = ULONG_MAX;
4431         int idlest = -1;
4432         int i;
4433 
4434         /* Traverse only the allowed CPUs */
4435         for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4436                 load = weighted_cpuload(i);
4437 
4438                 if (load < min_load || (load == min_load && i == this_cpu)) {
4439                         min_load = load;
4440                         idlest = i;
4441                 }
4442         }
4443 
4444         return idlest;
4445 }
4446 
4447 /*
4448  * Try and locate an idle CPU in the sched_domain.
4449  */
4450 static int select_idle_sibling(struct task_struct *p, int target)
4451 {
4452         struct sched_domain *sd;
4453         struct sched_group *sg;
4454         int i = task_cpu(p);
4455 
4456         if (idle_cpu(target))
4457                 return target;
4458 
4459         /*
4460          * If the prevous cpu is cache affine and idle, don't be stupid.
4461          */
4462         if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4463                 return i;
4464 
4465         /*
4466          * Otherwise, iterate the domains and find an elegible idle cpu.
4467          */
4468         sd = rcu_dereference(per_cpu(sd_llc, target));
4469         for_each_lower_domain(sd) {
4470                 sg = sd->groups;
4471                 do {
4472                         if (!cpumask_intersects(sched_group_cpus(sg),
4473                                                 tsk_cpus_allowed(p)))
4474                                 goto next;
4475 
4476                         for_each_cpu(i, sched_group_cpus(sg)) {
4477                                 if (i == target || !idle_cpu(i))
4478                                         goto next;
4479                         }
4480 
4481                         target = cpumask_first_and(sched_group_cpus(sg),
4482                                         tsk_cpus_allowed(p));
4483                         goto done;
4484 next:
4485                         sg = sg->next;
4486                 } while (sg != sd->groups);
4487         }
4488 done:
4489         return target;
4490 }
4491 
4492 /*
4493  * select_task_rq_fair: Select target runqueue for the waking task in domains
4494  * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4495  * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4496  *
4497  * Balances load by selecting the idlest cpu in the idlest group, or under
4498  * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4499  *
4500  * Returns the target cpu number.
4501  *
4502  * preempt must be disabled.
4503  */
4504 static int
4505 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4506 {
4507         struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4508         int cpu = smp_processor_id();
4509         int new_cpu = cpu;
4510         int want_affine = 0;
4511         int sync = wake_flags & WF_SYNC;
4512 
4513         if (p->nr_cpus_allowed == 1)
4514                 return prev_cpu;
4515 
4516         if (sd_flag & SD_BALANCE_WAKE) {
4517                 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
4518                         want_affine = 1;
4519                 new_cpu = prev_cpu;
4520         }
4521 
4522         rcu_read_lock();
4523         for_each_domain(cpu, tmp) {
4524                 if (!(tmp->flags & SD_LOAD_BALANCE))
4525                         continue;
4526 
4527                 /*
4528                  * If both cpu and prev_cpu are part of this domain,
4529                  * cpu is a valid SD_WAKE_AFFINE target.
4530                  */
4531                 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4532                     cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4533                         affine_sd = tmp;
4534                         break;
4535                 }
4536 
4537                 if (tmp->flags & sd_flag)
4538                         sd = tmp;
4539         }
4540 
4541         if (affine_sd && cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4542                 prev_cpu = cpu;
4543 
4544         if (sd_flag & SD_BALANCE_WAKE) {
4545                 new_cpu = select_idle_sibling(p, prev_cpu);
4546                 goto unlock;
4547         }
4548 
4549         while (sd) {
4550                 struct sched_group *group;
4551                 int weight;
4552 
4553                 if (!(sd->flags & sd_flag)) {
4554                         sd = sd->child;
4555                         continue;
4556                 }
4557 
4558                 group = find_idlest_group(sd, p, cpu, sd_flag);
4559                 if (!group) {
4560                         sd = sd->child;
4561                         continue;
4562                 }
4563 
4564                 new_cpu = find_idlest_cpu(group, p, cpu);
4565                 if (new_cpu == -1 || new_cpu == cpu) {
4566                         /* Now try balancing at a lower domain level of cpu */
4567                         sd = sd->child;
4568                         continue;
4569                 }
4570 
4571                 /* Now try balancing at a lower domain level of new_cpu */
4572                 cpu = new_cpu;
4573                 weight = sd->span_weight;
4574                 sd = NULL;
4575                 for_each_domain(cpu, tmp) {
4576                         if (weight <= tmp->span_weight)
4577                                 break;
4578                         if (tmp->flags & sd_flag)
4579                                 sd = tmp;
4580                 }
4581                 /* while loop will break here if sd == NULL */
4582         }
4583 unlock:
4584         rcu_read_unlock();
4585 
4586         return new_cpu;
4587 }
4588 
4589 /*
4590  * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4591  * cfs_rq_of(p) references at time of call are still valid and identify the
4592  * previous cpu.  However, the caller only guarantees p->pi_lock is held; no
4593  * other assumptions, including the state of rq->lock, should be made.
4594  */
4595 static void
4596 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4597 {
4598         struct sched_entity *se = &p->se;
4599         struct cfs_rq *cfs_rq = cfs_rq_of(se);
4600 
4601         /*
4602          * Load tracking: accumulate removed load so that it can be processed
4603          * when we next update owning cfs_rq under rq->lock.  Tasks contribute
4604          * to blocked load iff they have a positive decay-count.  It can never
4605          * be negative here since on-rq tasks have decay-count == 0.
4606          */
4607         if (se->avg.decay_count) {
4608                 se->avg.decay_count = -__synchronize_entity_decay(se);
4609                 atomic_long_add(se->avg.load_avg_contrib,
4610                                                 &cfs_rq->removed_load);
4611         }
4612 
4613         /* We have migrated, no longer consider this task hot */
4614         se->exec_start = 0;
4615 }
4616 #endif /* CONFIG_SMP */
4617 
4618 static unsigned long
4619 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4620 {
4621         unsigned long gran = sysctl_sched_wakeup_granularity;
4622 
4623         /*
4624          * Since its curr running now, convert the gran from real-time
4625          * to virtual-time in his units.
4626          *
4627          * By using 'se' instead of 'curr' we penalize light tasks, so
4628          * they get preempted easier. That is, if 'se' < 'curr' then
4629          * the resulting gran will be larger, therefore penalizing the
4630          * lighter, if otoh 'se' > 'curr' then the resulting gran will
4631          * be smaller, again penalizing the lighter task.
4632          *
4633          * This is especially important for buddies when the leftmost
4634          * task is higher priority than the buddy.
4635          */
4636         return calc_delta_fair(gran, se);
4637 }
4638 
4639 /*
4640  * Should 'se' preempt 'curr'.
4641  *
4642  *             |s1
4643  *        |s2
4644  *   |s3
4645  *         g
4646  *      |<--->|c
4647  *
4648  *  w(c, s1) = -1
4649  *  w(c, s2) =  0
4650  *  w(c, s3) =  1
4651  *
4652  */
4653 static int
4654 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4655 {
4656         s64 gran, vdiff = curr->vruntime - se->vruntime;
4657 
4658         if (vdiff <= 0)
4659                 return -1;
4660 
4661         gran = wakeup_gran(curr, se);
4662         if (vdiff > gran)
4663                 return 1;
4664 
4665         return 0;
4666 }
4667 
4668 static void set_last_buddy(struct sched_entity *se)
4669 {
4670         if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4671                 return;
4672 
4673         for_each_sched_entity(se)
4674                 cfs_rq_of(se)->last = se;
4675 }
4676 
4677 static void set_next_buddy(struct sched_entity *se)
4678 {
4679         if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4680                 return;
4681 
4682         for_each_sched_entity(se)
4683                 cfs_rq_of(se)->next = se;
4684 }
4685 
4686 static void set_skip_buddy(struct sched_entity *se)
4687 {
4688         for_each_sched_entity(se)
4689                 cfs_rq_of(se)->skip = se;
4690 }
4691 
4692 /*
4693  * Preempt the current task with a newly woken task if needed:
4694  */
4695 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4696 {
4697         struct task_struct *curr = rq->curr;
4698         struct sched_entity *se = &curr->se, *pse = &p->se;
4699         struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4700         int scale = cfs_rq->nr_running >= sched_nr_latency;
4701         int next_buddy_marked = 0;
4702 
4703         if (unlikely(se == pse))
4704                 return;
4705 
4706         /*
4707          * This is possible from callers such as move_task(), in which we
4708          * unconditionally check_prempt_curr() after an enqueue (which may have
4709          * lead to a throttle).  This both saves work and prevents false
4710          * next-buddy nomination below.
4711          */
4712         if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4713                 return;
4714 
4715         if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4716                 set_next_buddy(pse);
4717                 next_buddy_marked = 1;
4718         }
4719 
4720         /*
4721          * We can come here with TIF_NEED_RESCHED already set from new task
4722          * wake up path.
4723          *
4724          * Note: this also catches the edge-case of curr being in a throttled
4725          * group (e.g. via set_curr_task), since update_curr() (in the
4726          * enqueue of curr) will have resulted in resched being set.  This
4727          * prevents us from potentially nominating it as a false LAST_BUDDY
4728          * below.
4729          */
4730         if (test_tsk_need_resched(curr))
4731                 return;
4732 
4733         /* Idle tasks are by definition preempted by non-idle tasks. */
4734         if (unlikely(curr->policy == SCHED_IDLE) &&
4735             likely(p->policy != SCHED_IDLE))
4736                 goto preempt;
4737 
4738         /*
4739          * Batch and idle tasks do not preempt non-idle tasks (their preemption
4740          * is driven by the tick):
4741          */
4742         if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4743                 return;
4744 
4745         find_matching_se(&se, &pse);
4746         update_curr(cfs_rq_of(se));
4747         BUG_ON(!pse);
4748         if (wakeup_preempt_entity(se, pse) == 1) {
4749                 /*
4750                  * Bias pick_next to pick the sched entity that is
4751                  * triggering this preemption.
4752                  */
4753                 if (!next_buddy_marked)
4754                         set_next_buddy(pse);
4755                 goto preempt;
4756         }
4757 
4758         return;
4759 
4760 preempt:
4761         resched_curr(rq);
4762         /*
4763          * Only set the backward buddy when the current task is still
4764          * on the rq. This can happen when a wakeup gets interleaved
4765          * with schedule on the ->pre_schedule() or idle_balance()
4766          * point, either of which can * drop the rq lock.
4767          *
4768          * Also, during early boot the idle thread is in the fair class,
4769          * for obvious reasons its a bad idea to schedule back to it.
4770          */
4771         if (unlikely(!se->on_rq || curr == rq->idle))
4772                 return;
4773 
4774         if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4775                 set_last_buddy(se);
4776 }
4777 
4778 static struct task_struct *
4779 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
4780 {
4781         struct cfs_rq *cfs_rq = &rq->cfs;
4782         struct sched_entity *se;
4783         struct task_struct *p;
4784         int new_tasks;
4785 
4786 again:
4787 #ifdef CONFIG_FAIR_GROUP_SCHED
4788         if (!cfs_rq->nr_running)
4789                 goto idle;
4790 
4791         if (prev->sched_class != &fair_sched_class)
4792                 goto simple;
4793 
4794         /*
4795          * Because of the set_next_buddy() in dequeue_task_fair() it is rather
4796          * likely that a next task is from the same cgroup as the current.
4797          *
4798          * Therefore attempt to avoid putting and setting the entire cgroup
4799          * hierarchy, only change the part that actually changes.
4800          */
4801 
4802         do {
4803                 struct sched_entity *curr = cfs_rq->curr;
4804 
4805                 /*
4806                  * Since we got here without doing put_prev_entity() we also
4807                  * have to consider cfs_rq->curr. If it is still a runnable
4808                  * entity, update_curr() will update its vruntime, otherwise
4809                  * forget we've ever seen it.
4810                  */
4811                 if (curr && curr->on_rq)
4812                         update_curr(cfs_rq);
4813                 else
4814                         curr = NULL;
4815 
4816                 /*
4817                  * This call to check_cfs_rq_runtime() will do the throttle and
4818                  * dequeue its entity in the parent(s). Therefore the 'simple'
4819                  * nr_running test will indeed be correct.
4820                  */
4821                 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
4822                         goto simple;
4823 
4824                 se = pick_next_entity(cfs_rq, curr);
4825                 cfs_rq = group_cfs_rq(se);
4826         } while (cfs_rq);
4827 
4828         p = task_of(se);
4829 
4830         /*
4831          * Since we haven't yet done put_prev_entity and if the selected task
4832          * is a different task than we started out with, try and touch the
4833          * least amount of cfs_rqs.
4834          */
4835         if (prev != p) {
4836                 struct sched_entity *pse = &prev->se;
4837 
4838                 while (!(cfs_rq = is_same_group(se, pse))) {
4839                         int se_depth = se->depth;
4840                         int pse_depth = pse->depth;
4841 
4842                         if (se_depth <= pse_depth) {
4843                                 put_prev_entity(cfs_rq_of(pse), pse);
4844                                 pse = parent_entity(pse);
4845                         }
4846                         if (se_depth >= pse_depth) {
4847                                 set_next_entity(cfs_rq_of(se), se);
4848                                 se = parent_entity(se);
4849                         }
4850                 }
4851 
4852                 put_prev_entity(cfs_rq, pse);
4853                 set_next_entity(cfs_rq, se);
4854         }
4855 
4856         if (hrtick_enabled(rq))
4857                 hrtick_start_fair(rq, p);
4858 
4859         return p;
4860 simple:
4861         cfs_rq = &rq->cfs;
4862 #endif
4863 
4864         if (!cfs_rq->nr_running)
4865                 goto idle;
4866 
4867         put_prev_task(rq, prev);
4868 
4869         do {
4870                 se = pick_next_entity(cfs_rq, NULL);
4871                 set_next_entity(cfs_rq, se);
4872                 cfs_rq = group_cfs_rq(se);
4873         } while (cfs_rq);
4874 
4875         p = task_of(se);
4876 
4877         if (hrtick_enabled(rq))
4878                 hrtick_start_fair(rq, p);
4879 
4880         return p;
4881 
4882 idle:
4883         new_tasks = idle_balance(rq);
4884         /*
4885          * Because idle_balance() releases (and re-acquires) rq->lock, it is
4886          * possible for any higher priority task to appear. In that case we
4887          * must re-start the pick_next_entity() loop.
4888          */
4889         if (new_tasks < 0)
4890                 return RETRY_TASK;
4891 
4892         if (new_tasks > 0)
4893                 goto again;
4894 
4895         return NULL;
4896 }
4897 
4898 /*
4899  * Account for a descheduled task:
4900  */
4901 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4902 {
4903         struct sched_entity *se = &prev->se;
4904         struct cfs_rq *cfs_rq;
4905 
4906         for_each_sched_entity(se) {
4907                 cfs_rq = cfs_rq_of(se);
4908                 put_prev_entity(cfs_rq, se);
4909         }
4910 }
4911 
4912 /*
4913  * sched_yield() is very simple
4914  *
4915  * The magic of dealing with the ->skip buddy is in pick_next_entity.
4916  */
4917 static void yield_task_fair(struct rq *rq)
4918 {
4919         struct task_struct *curr = rq->curr;
4920         struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4921         struct sched_entity *se = &curr->se;
4922 
4923         /*
4924          * Are we the only task in the tree?
4925          */
4926         if (unlikely(rq->nr_running == 1))
4927                 return;
4928 
4929         clear_buddies(cfs_rq, se);
4930 
4931         if (curr->policy != SCHED_BATCH) {
4932                 update_rq_clock(rq);
4933                 /*
4934                  * Update run-time statistics of the 'current'.
4935                  */
4936                 update_curr(cfs_rq);
4937                 /*
4938                  * Tell update_rq_clock() that we've just updated,
4939                  * so we don't do microscopic update in schedule()
4940                  * and double the fastpath cost.
4941                  */
4942                  rq->skip_clock_update = 1;
4943         }
4944 
4945         set_skip_buddy(se);
4946 }
4947 
4948 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4949 {
4950         struct sched_entity *se = &p->se;
4951 
4952         /* throttled hierarchies are not runnable */
4953         if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4954                 return false;
4955 
4956         /* Tell the scheduler that we'd really like pse to run next. */
4957         set_next_buddy(se);
4958 
4959         yield_task_fair(rq);
4960 
4961         return true;
4962 }
4963 
4964 #ifdef CONFIG_SMP
4965 /**************************************************
4966  * Fair scheduling class load-balancing methods.
4967  *
4968  * BASICS
4969  *
4970  * The purpose of load-balancing is to achieve the same basic fairness the
4971  * per-cpu scheduler provides, namely provide a proportional amount of compute
4972  * time to each task. This is expressed in the following equation:
4973  *
4974  *   W_i,n/P_i == W_j,n/P_j for all i,j                               (1)
4975  *
4976  * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4977  * W_i,0 is defined as:
4978  *
4979  *   W_i,0 = \Sum_j w_i,j                                             (2)
4980  *
4981  * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4982  * is derived from the nice value as per prio_to_weight[].
4983  *
4984  * The weight average is an exponential decay average of the instantaneous
4985  * weight:
4986  *
4987  *   W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0               (3)
4988  *
4989  * C_i is the compute capacity of cpu i, typically it is the
4990  * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4991  * can also include other factors [XXX].
4992  *
4993  * To achieve this balance we define a measure of imbalance which follows
4994  * directly from (1):
4995  *
4996  *   imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j }    (4)
4997  *
4998  * We them move tasks around to minimize the imbalance. In the continuous
4999  * function space it is obvious this converges, in the discrete case we get
5000  * a few fun cases generally called infeasible weight scenarios.
5001  *
5002  * [XXX expand on:
5003  *     - infeasible weights;
5004  *     - local vs global optima in the discrete case. ]
5005  *
5006  *
5007  * SCHED DOMAINS
5008  *
5009  * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5010  * for all i,j solution, we create a tree of cpus that follows the hardware
5011  * topology where each level pairs two lower groups (or better). This results
5012  * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5013  * tree to only the first of the previous level and we decrease the frequency
5014  * of load-balance at each level inv. proportional to the number of cpus in
5015  * the groups.
5016  *
5017  * This yields:
5018  *
5019  *     log_2 n     1     n
5020  *   \Sum       { --- * --- * 2^i } = O(n)                            (5)
5021  *     i = 0      2^i   2^i
5022  *                               `- size of each group
5023  *         |         |     `- number of cpus doing load-balance
5024  *         |         `- freq
5025  *         `- sum over all levels
5026  *
5027  * Coupled with a limit on how many tasks we can migrate every balance pass,
5028  * this makes (5) the runtime complexity of the balancer.
5029  *
5030  * An important property here is that each CPU is still (indirectly) connected
5031  * to every other cpu in at most O(log n) steps:
5032  *
5033  * The adjacency matrix of the resulting graph is given by:
5034  *
5035  *             log_2 n     
5036  *   A_i,j = \Union     (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1)  (6)
5037  *             k = 0
5038  *
5039  * And you'll find that:
5040  *
5041  *   A^(log_2 n)_i,j != 0  for all i,j                                (7)
5042  *
5043  * Showing there's indeed a path between every cpu in at most O(log n) steps.
5044  * The task movement gives a factor of O(m), giving a convergence complexity
5045  * of:
5046  *
5047  *   O(nm log n),  n := nr_cpus, m := nr_tasks                        (8)
5048  *
5049  *
5050  * WORK CONSERVING
5051  *
5052  * In order to avoid CPUs going idle while there's still work to do, new idle
5053  * balancing is more aggressive and has the newly idle cpu iterate up the domain
5054  * tree itself instead of relying on other CPUs to bring it work.
5055  *
5056  * This adds some complexity to both (5) and (8) but it reduces the total idle
5057  * time.
5058  *
5059  * [XXX more?]
5060  *
5061  *
5062  * CGROUPS
5063  *
5064  * Cgroups make a horror show out of (2), instead of a simple sum we get:
5065  *
5066  *                                s_k,i
5067  *   W_i,0 = \Sum_j \Prod_k w_k * -----                               (9)
5068  *                                 S_k
5069  *
5070  * Where
5071  *
5072  *   s_k,i = \Sum_j w_i,j,k  and  S_k = \Sum_i s_k,i                 (10)
5073  *
5074  * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5075  *
5076  * The big problem is S_k, its a global sum needed to compute a local (W_i)
5077  * property.
5078  *
5079  * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5080  *      rewrite all of this once again.]
5081  */ 
5082 
5083 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5084 
5085 enum fbq_type { regular, remote, all };
5086 
5087 #define LBF_ALL_PINNED  0x01
5088 #define LBF_NEED_BREAK  0x02
5089 #define LBF_DST_PINNED  0x04
5090 #define LBF_SOME_PINNED 0x08
5091 
5092 struct lb_env {
5093         struct sched_domain     *sd;
5094 
5095         struct rq               *src_rq;
5096         int                     src_cpu;
5097 
5098         int                     dst_cpu;
5099         struct rq               *dst_rq;
5100 
5101         struct cpumask          *dst_grpmask;
5102         int                     new_dst_cpu;
5103         enum cpu_idle_type      idle;
5104         long                    imbalance;
5105         /* The set of CPUs under consideration for load-balancing */
5106         struct cpumask          *cpus;
5107 
5108         unsigned int            flags;
5109 
5110         unsigned int            loop;
5111         unsigned int            loop_break;
5112         unsigned int            loop_max;
5113 
5114         enum fbq_type           fbq_type;
5115 };
5116 
5117 /*
5118  * move_task - move a task from one runqueue to another runqueue.
5119  * Both runqueues must be locked.
5120  */
5121 static void move_task(struct task_struct *p, struct lb_env *env)
5122 {
5123         deactivate_task(env->src_rq, p, 0);
5124         set_task_cpu(p, env->dst_cpu);
5125         activate_task(env->dst_rq, p, 0);
5126         check_preempt_curr(env->dst_rq, p, 0);
5127 }
5128 
5129 /*
5130  * Is this task likely cache-hot:
5131  */
5132 static int task_hot(struct task_struct *p, struct lb_env *env)
5133 {
5134         s64 delta;
5135 
5136         if (p->sched_class != &fair_sched_class)
5137                 return 0;
5138 
5139         if (unlikely(p->policy == SCHED_IDLE))
5140                 return 0;
5141 
5142         /*
5143          * Buddy candidates are cache hot:
5144          */
5145         if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
5146                         (&p->se == cfs_rq_of(&p->se)->next ||
5147                          &p->se == cfs_rq_of(&p->se)->last))
5148                 return 1;
5149 
5150         if (sysctl_sched_migration_cost == -1)
5151                 return 1;
5152         if (sysctl_sched_migration_cost == 0)
5153                 return 0;
5154 
5155         delta = rq_clock_task(env->src_rq) - p->se.exec_start;
5156 
5157         return delta < (s64)sysctl_sched_migration_cost;
5158 }
5159 
5160 #ifdef CONFIG_NUMA_BALANCING
5161 /* Returns true if the destination node has incurred more faults */
5162 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
5163 {
5164         struct numa_group *numa_group = rcu_dereference(p->numa_group);
5165         int src_nid, dst_nid;
5166 
5167         if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults_memory ||
5168             !(env->sd->flags & SD_NUMA)) {
5169                 return false;
5170         }
5171 
5172         src_nid = cpu_to_node(env->src_cpu);
5173         dst_nid = cpu_to_node(env->dst_cpu);
5174 
5175         if (src_nid == dst_nid)
5176                 return false;
5177 
5178         if (numa_group) {
5179                 /* Task is already in the group's interleave set. */
5180                 if (node_isset(src_nid, numa_group->active_nodes))
5181                         return false;
5182 
5183                 /* Task is moving into the group's interleave set. */
5184                 if (node_isset(dst_nid, numa_group->active_nodes))
5185                         return true;
5186 
5187                 return group_faults(p, dst_nid) > group_faults(p, src_nid);
5188         }
5189 
5190         /* Encourage migration to the preferred node. */
5191         if (dst_nid == p->numa_preferred_nid)
5192                 return true;
5193 
5194         return task_faults(p, dst_nid) > task_faults(p, src_nid);
5195 }
5196 
5197 
5198 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5199 {
5200         struct numa_group *numa_group = rcu_dereference(p->numa_group);
5201         int src_nid, dst_nid;
5202 
5203         if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
5204                 return false;
5205 
5206         if (!p->numa_faults_memory || !(env->sd->flags & SD_NUMA))
5207                 return false;
5208 
5209         src_nid = cpu_to_node(env->src_cpu);
5210         dst_nid = cpu_to_node(env->dst_cpu);
5211 
5212         if (src_nid == dst_nid)
5213                 return false;
5214 
5215         if (numa_group) {
5216                 /* Task is moving within/into the group's interleave set. */
5217                 if (node_isset(dst_nid, numa_group->active_nodes))
5218                         return false;
5219 
5220                 /* Task is moving out of the group's interleave set. */
5221                 if (node_isset(src_nid, numa_group->active_nodes))
5222                         return true;
5223 
5224                 return group_faults(p, dst_nid) < group_faults(p, src_nid);
5225         }
5226 
5227         /* Migrating away from the preferred node is always bad. */
5228         if (src_nid == p->numa_preferred_nid)
5229                 return true;
5230 
5231         return task_faults(p, dst_nid) < task_faults(p, src_nid);
5232 }
5233 
5234 #else
5235 static inline bool migrate_improves_locality(struct task_struct *p,
5236                                              struct lb_env *env)
5237 {
5238         return false;
5239 }
5240 
5241 static inline bool migrate_degrades_locality(struct task_struct *p,
5242                                              struct lb_env *env)
5243 {
5244         return false;
5245 }
5246 #endif
5247 
5248 /*
5249  * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5250  */
5251 static
5252 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5253 {
5254         int tsk_cache_hot = 0;
5255         /*
5256          * We do not migrate tasks that are:
5257          * 1) throttled_lb_pair, or
5258          * 2) cannot be migrated to this CPU due to cpus_allowed, or
5259          * 3) running (obviously), or
5260          * 4) are cache-hot on their current CPU.
5261          */
5262         if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5263                 return 0;
5264 
5265         if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5266                 int cpu;
5267 
5268                 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5269 
5270                 env->flags |= LBF_SOME_PINNED;
5271 
5272                 /*
5273                  * Remember if this task can be migrated to any other cpu in
5274                  * our sched_group. We may want to revisit it if we couldn't
5275                  * meet load balance goals by pulling other tasks on src_cpu.
5276                  *
5277                  * Also avoid computing new_dst_cpu if we have already computed
5278                  * one in current iteration.
5279                  */
5280                 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5281                         return 0;
5282 
5283                 /* Prevent to re-select dst_cpu via env's cpus */
5284                 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5285                         if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5286                                 env->flags |= LBF_DST_PINNED;
5287                                 env->new_dst_cpu = cpu;
5288                                 break;
5289                         }
5290                 }
5291 
5292                 return 0;
5293         }
5294 
5295         /* Record that we found atleast one task that could run on dst_cpu */
5296         env->flags &= ~LBF_ALL_PINNED;
5297 
5298         if (task_running(env->src_rq, p)) {
5299                 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5300                 return 0;
5301         }
5302 
5303         /*
5304          * Aggressive migration if:
5305          * 1) destination numa is preferred
5306          * 2) task is cache cold, or
5307          * 3) too many balance attempts have failed.
5308          */
5309         tsk_cache_hot = task_hot(p, env);
5310         if (!tsk_cache_hot)
5311                 tsk_cache_hot = migrate_degrades_locality(p, env);
5312 
5313         if (migrate_improves_locality(p, env)) {
5314 #ifdef CONFIG_SCHEDSTATS
5315                 if (tsk_cache_hot) {
5316                         schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5317                         schedstat_inc(p, se.statistics.nr_forced_migrations);
5318                 }
5319 #endif
5320                 return 1;
5321         }
5322 
5323         if (!tsk_cache_hot ||
5324                 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5325 
5326                 if (tsk_cache_hot) {
5327                         schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5328                         schedstat_inc(p, se.statistics.nr_forced_migrations);
5329                 }
5330 
5331                 return 1;
5332         }
5333 
5334         schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5335         return 0;
5336 }
5337 
5338 /*
5339  * move_one_task tries to move exactly one task from busiest to this_rq, as
5340  * part of active balancing operations within "domain".
5341  * Returns 1 if successful and 0 otherwise.
5342  *
5343  * Called with both runqueues locked.
5344  */
5345 static int move_one_task(struct lb_env *env)
5346 {
5347         struct task_struct *p, *n;
5348 
5349         list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5350                 if (!can_migrate_task(p, env))
5351                         continue;
5352 
5353                 move_task(p, env);
5354                 /*
5355                  * Right now, this is only the second place move_task()
5356                  * is called, so we can safely collect move_task()
5357                  * stats here rather than inside move_task().
5358                  */
5359                 schedstat_inc(env->sd, lb_gained[env->idle]);
5360                 return 1;
5361         }
5362         return 0;
5363 }
5364 
5365 static const unsigned int sched_nr_migrate_break = 32;
5366 
5367 /*
5368  * move_tasks tries to move up to imbalance weighted load from busiest to
5369  * this_rq, as part of a balancing operation within domain "sd".
5370  * Returns 1 if successful and 0 otherwise.
5371  *
5372  * Called with both runqueues locked.
5373  */
5374 static int move_tasks(struct lb_env *env)
5375 {
5376         struct list_head *tasks = &env->src_rq->cfs_tasks;
5377         struct task_struct *p;
5378         unsigned long load;
5379         int pulled = 0;
5380 
5381         if (env->imbalance <= 0)
5382                 return 0;
5383 
5384         while (!list_empty(tasks)) {
5385                 p = list_first_entry(tasks, struct task_struct, se.group_node);
5386 
5387                 env->loop++;
5388                 /* We've more or less seen every task there is, call it quits */
5389                 if (env->loop > env->loop_max)
5390                         break;
5391 
5392                 /* take a breather every nr_migrate tasks */
5393                 if (env->loop > env->loop_break) {
5394                         env->loop_break += sched_nr_migrate_break;
5395                         env->flags |= LBF_NEED_BREAK;
5396                         break;
5397                 }
5398 
5399                 if (!can_migrate_task(p, env))
5400                         goto next;
5401 
5402                 load = task_h_load(p);
5403 
5404                 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5405                         goto next;
5406 
5407                 if ((load / 2) > env->imbalance)
5408                         goto next;
5409 
5410                 move_task(p, env);
5411                 pulled++;
5412                 env->imbalance -= load;
5413 
5414 #ifdef CONFIG_PREEMPT
5415                 /*
5416                  * NEWIDLE balancing is a source of latency, so preemptible
5417                  * kernels will stop after the first task is pulled to minimize
5418                  * the critical section.
5419                  */
5420                 if (env->idle == CPU_NEWLY_IDLE)
5421                         break;
5422 #endif
5423 
5424                 /*
5425                  * We only want to steal up to the prescribed amount of
5426                  * weighted load.
5427                  */
5428                 if (env->imbalance <= 0)
5429                         break;
5430 
5431                 continue;
5432 next:
5433                 list_move_tail(&p->se.group_node, tasks);
5434         }
5435 
5436         /*
5437          * Right now, this is one of only two places move_task() is called,
5438          * so we can safely collect move_task() stats here rather than
5439          * inside move_task().
5440          */
5441         schedstat_add(env->sd, lb_gained[env->idle], pulled);
5442 
5443         return pulled;
5444 }
5445 
5446 #ifdef CONFIG_FAIR_GROUP_SCHED
5447 /*
5448  * update tg->load_weight by folding this cpu's load_avg
5449  */
5450 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5451 {
5452         struct sched_entity *se = tg->se[cpu];
5453         struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5454 
5455         /* throttled entities do not contribute to load */
5456         if (throttled_hierarchy(cfs_rq))
5457                 return;
5458 
5459         update_cfs_rq_blocked_load(cfs_rq, 1);
5460 
5461         if (se) {
5462                 update_entity_load_avg(se, 1);
5463                 /*
5464                  * We pivot on our runnable average having decayed to zero for
5465                  * list removal.  This generally implies that all our children
5466                  * have also been removed (modulo rounding error or bandwidth
5467                  * control); however, such cases are rare and we can fix these
5468                  * at enqueue.
5469                  *
5470                  * TODO: fix up out-of-order children on enqueue.
5471                  */
5472                 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5473                         list_del_leaf_cfs_rq(cfs_rq);
5474         } else {
5475                 struct rq *rq = rq_of(cfs_rq);
5476                 update_rq_runnable_avg(rq, rq->nr_running);
5477         }
5478 }
5479 
5480 static void update_blocked_averages(int cpu)
5481 {
5482         struct rq *rq = cpu_rq(cpu);
5483         struct cfs_rq *cfs_rq;
5484         unsigned long flags;
5485 
5486         raw_spin_lock_irqsave(&rq->lock, flags);
5487         update_rq_clock(rq);
5488         /*
5489          * Iterates the task_group tree in a bottom up fashion, see
5490          * list_add_leaf_cfs_rq() for details.
5491          */
5492         for_each_leaf_cfs_rq(rq, cfs_rq) {
5493                 /*
5494                  * Note: We may want to consider periodically releasing
5495                  * rq->lock about these updates so that creating many task
5496                  * groups does not result in continually extending hold time.
5497                  */
5498                 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5499         }
5500 
5501         raw_spin_unlock_irqrestore(&rq->lock, flags);
5502 }
5503 
5504 /*
5505  * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5506  * This needs to be done in a top-down fashion because the load of a child
5507  * group is a fraction of its parents load.
5508  */
5509 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5510 {
5511         struct rq *rq = rq_of(cfs_rq);
5512         struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5513         unsigned long now = jiffies;
5514         unsigned long load;
5515 
5516         if (cfs_rq->last_h_load_update == now)
5517                 return;
5518 
5519         cfs_rq->h_load_next = NULL;
5520         for_each_sched_entity(se) {
5521                 cfs_rq = cfs_rq_of(se);
5522                 cfs_rq->h_load_next = se;
5523                 if (cfs_rq->last_h_load_update == now)
5524                         break;
5525         }
5526 
5527         if (!se) {
5528                 cfs_rq->h_load = cfs_rq->runnable_load_avg;
5529                 cfs_rq->last_h_load_update = now;
5530         }
5531 
5532         while ((se = cfs_rq->h_load_next) != NULL) {
5533                 load = cfs_rq->h_load;
5534                 load = div64_ul(load * se->avg.load_avg_contrib,
5535                                 cfs_rq->runnable_load_avg + 1);
5536                 cfs_rq = group_cfs_rq(se);
5537                 cfs_rq->h_load = load;
5538                 cfs_rq->last_h_load_update = now;
5539         }
5540 }
5541 
5542 static unsigned long task_h_load(struct task_struct *p)
5543 {
5544         struct cfs_rq *cfs_rq = task_cfs_rq(p);
5545 
5546         update_cfs_rq_h_load(cfs_rq);
5547         return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
5548                         cfs_rq->runnable_load_avg + 1);
5549 }
5550 #else
5551 static inline void update_blocked_averages(int cpu)
5552 {
5553 }
5554 
5555 static unsigned long task_h_load(struct task_struct *p)
5556 {
5557         return p->se.avg.load_avg_contrib;
5558 }
5559 #endif
5560 
5561 /********** Helpers for find_busiest_group ************************/
5562 /*
5563  * sg_lb_stats - stats of a sched_group required for load_balancing
5564  */
5565 struct sg_lb_stats {
5566         unsigned long avg_load; /*Avg load across the CPUs of the group */
5567         unsigned long group_load; /* Total load over the CPUs of the group */
5568         unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5569         unsigned long load_per_task;
5570         unsigned long group_capacity;
5571         unsigned int sum_nr_running; /* Nr tasks running in the group */
5572         unsigned int group_capacity_factor;
5573         unsigned int idle_cpus;
5574         unsigned int group_weight;
5575         int group_imb; /* Is there an imbalance in the group ? */
5576         int group_has_free_capacity;
5577 #ifdef CONFIG_NUMA_BALANCING
5578         unsigned int nr_numa_running;
5579         unsigned int nr_preferred_running;
5580 #endif
5581 };
5582 
5583 /*
5584  * sd_lb_stats - Structure to store the statistics of a sched_domain
5585  *               during load balancing.
5586  */
5587 struct sd_lb_stats {
5588         struct sched_group *busiest;    /* Busiest group in this sd */
5589         struct sched_group *local;      /* Local group in this sd */
5590         unsigned long total_load;       /* Total load of all groups in sd */
5591         unsigned long total_capacity;   /* Total capacity of all groups in sd */
5592         unsigned long avg_load; /* Average load across all groups in sd */
5593 
5594         struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5595         struct sg_lb_stats local_stat;  /* Statistics of the local group */
5596 };
5597 
5598 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5599 {
5600         /*
5601          * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5602          * local_stat because update_sg_lb_stats() does a full clear/assignment.
5603          * We must however clear busiest_stat::avg_load because
5604          * update_sd_pick_busiest() reads this before assignment.
5605          */
5606         *sds = (struct sd_lb_stats){
5607                 .busiest = NULL,
5608                 .local = NULL,
5609                 .total_load = 0UL,
5610                 .total_capacity = 0UL,
5611                 .busiest_stat = {
5612                         .avg_load = 0UL,
5613                 },
5614         };
5615 }
5616 
5617 /**
5618  * get_sd_load_idx - Obtain the load index for a given sched domain.
5619  * @sd: The sched_domain whose load_idx is to be obtained.
5620  * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5621  *
5622  * Return: The load index.
5623  */
5624 static inline int get_sd_load_idx(struct sched_domain *sd,
5625                                         enum cpu_idle_type idle)
5626 {
5627         int load_idx;
5628 
5629         switch (idle) {
5630         case CPU_NOT_IDLE:
5631                 load_idx = sd->busy_idx;
5632                 break;
5633 
5634         case CPU_NEWLY_IDLE:
5635                 load_idx = sd->newidle_idx;
5636                 break;
5637         default:
5638                 load_idx = sd->idle_idx;
5639                 break;
5640         }
5641 
5642         return load_idx;
5643 }
5644 
5645 static unsigned long default_scale_capacity(struct sched_domain *sd, int cpu)
5646 {
5647         return SCHED_CAPACITY_SCALE;
5648 }
5649 
5650 unsigned long __weak arch_scale_freq_capacity(struct sched_domain *sd, int cpu)
5651 {
5652         return default_scale_capacity(sd, cpu);
5653 }
5654 
5655 static unsigned long default_scale_smt_capacity(struct sched_domain *sd, int cpu)
5656 {
5657         unsigned long weight = sd->span_weight;
5658         unsigned long smt_gain = sd->smt_gain;
5659 
5660         smt_gain /= weight;
5661 
5662         return smt_gain;
5663 }
5664 
5665 unsigned long __weak arch_scale_smt_capacity(struct sched_domain *sd, int cpu)
5666 {
5667         return default_scale_smt_capacity(sd, cpu);
5668 }
5669 
5670 static unsigned long scale_rt_capacity(int cpu)
5671 {
5672         struct rq *rq = cpu_rq(cpu);
5673         u64 total, available, age_stamp, avg;
5674         s64 delta;
5675 
5676         /*
5677          * Since we're reading these variables without serialization make sure
5678          * we read them once before doing sanity checks on them.
5679          */
5680         age_stamp = ACCESS_ONCE(rq->age_stamp);
5681         avg = ACCESS_ONCE(rq->rt_avg);
5682 
5683         delta = rq_clock(rq) - age_stamp;
5684         if (unlikely(delta < 0))
5685                 delta = 0;
5686 
5687         total = sched_avg_period() + delta;
5688 
5689         if (unlikely(total < avg)) {
5690                 /* Ensures that capacity won't end up being negative */
5691                 available = 0;
5692         } else {
5693                 available = total - avg;
5694         }
5695 
5696         if (unlikely((s64)total < SCHED_CAPACITY_SCALE))
5697                 total = SCHED_CAPACITY_SCALE;
5698 
5699         total >>= SCHED_CAPACITY_SHIFT;
5700 
5701         return div_u64(available, total);
5702 }
5703 
5704 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
5705 {
5706         unsigned long weight = sd->span_weight;
5707         unsigned long capacity = SCHED_CAPACITY_SCALE;
5708         struct sched_group *sdg = sd->groups;
5709 
5710         if ((sd->flags & SD_SHARE_CPUCAPACITY) && weight > 1) {
5711                 if (sched_feat(ARCH_CAPACITY))
5712                         capacity *= arch_scale_smt_capacity(sd, cpu);
5713                 else
5714                         capacity *= default_scale_smt_capacity(sd, cpu);
5715 
5716                 capacity >>= SCHED_CAPACITY_SHIFT;
5717         }
5718 
5719         sdg->sgc->capacity_orig = capacity;
5720 
5721         if (sched_feat(ARCH_CAPACITY))
5722                 capacity *= arch_scale_freq_capacity(sd, cpu);
5723         else
5724                 capacity *= default_scale_capacity(sd, cpu);
5725 
5726         capacity >>= SCHED_CAPACITY_SHIFT;
5727 
5728         capacity *= scale_rt_capacity(cpu);
5729         capacity >>= SCHED_CAPACITY_SHIFT;
5730 
5731         if (!capacity)
5732                 capacity = 1;
5733 
5734         cpu_rq(cpu)->cpu_capacity = capacity;
5735         sdg->sgc->capacity = capacity;
5736 }
5737 
5738 void update_group_capacity(struct sched_domain *sd, int cpu)
5739 {
5740         struct sched_domain *child = sd->child;
5741         struct sched_group *group, *sdg = sd->groups;
5742         unsigned long capacity, capacity_orig;
5743         unsigned long interval;
5744 
5745         interval = msecs_to_jiffies(sd->balance_interval);
5746         interval = clamp(interval, 1UL, max_load_balance_interval);
5747         sdg->sgc->next_update = jiffies + interval;
5748 
5749         if (!child) {
5750                 update_cpu_capacity(sd, cpu);
5751                 return;
5752         }
5753 
5754         capacity_orig = capacity = 0;
5755 
5756         if (child->flags & SD_OVERLAP) {
5757                 /*
5758                  * SD_OVERLAP domains cannot assume that child groups
5759                  * span the current group.
5760                  */
5761 
5762                 for_each_cpu(cpu, sched_group_cpus(sdg)) {
5763                         struct sched_group_capacity *sgc;
5764                         struct rq *rq = cpu_rq(cpu);
5765 
5766                         /*
5767                          * build_sched_domains() -> init_sched_groups_capacity()
5768                          * gets here before we've attached the domains to the
5769                          * runqueues.
5770                          *
5771                          * Use capacity_of(), which is set irrespective of domains
5772                          * in update_cpu_capacity().
5773                          *
5774                          * This avoids capacity/capacity_orig from being 0 and
5775                          * causing divide-by-zero issues on boot.
5776                          *
5777                          * Runtime updates will correct capacity_orig.
5778                          */
5779                         if (unlikely(!rq->sd)) {
5780                                 capacity_orig += capacity_of(cpu);
5781                                 capacity += capacity_of(cpu);
5782                                 continue;
5783                         }
5784 
5785                         sgc = rq->sd->groups->sgc;
5786                         capacity_orig += sgc->capacity_orig;
5787                         capacity += sgc->capacity;
5788                 }
5789         } else  {
5790                 /*
5791                  * !SD_OVERLAP domains can assume that child groups
5792                  * span the current group.
5793                  */ 
5794 
5795                 group = child->groups;
5796                 do {
5797                         capacity_orig += group->sgc->capacity_orig;
5798                         capacity += group->sgc->capacity;
5799                         group = group->next;
5800                 } while (group != child->groups);
5801         }
5802 
5803         sdg->sgc->capacity_orig = capacity_orig;
5804         sdg->sgc->capacity = capacity;
5805 }
5806 
5807 /*
5808  * Try and fix up capacity for tiny siblings, this is needed when
5809  * things like SD_ASYM_PACKING need f_b_g to select another sibling
5810  * which on its own isn't powerful enough.
5811  *
5812  * See update_sd_pick_busiest() and check_asym_packing().
5813  */
5814 static inline int
5815 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5816 {
5817         /*
5818          * Only siblings can have significantly less than SCHED_CAPACITY_SCALE
5819          */
5820         if (!(sd->flags & SD_SHARE_CPUCAPACITY))
5821                 return 0;
5822 
5823         /*
5824          * If ~90% of the cpu_capacity is still there, we're good.
5825          */
5826         if (group->sgc->capacity * 32 > group->sgc->capacity_orig * 29)
5827                 return 1;
5828 
5829         return 0;
5830 }
5831 
5832 /*
5833  * Group imbalance indicates (and tries to solve) the problem where balancing
5834  * groups is inadequate due to tsk_cpus_allowed() constraints.
5835  *
5836  * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5837  * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5838  * Something like:
5839  *
5840  *      { 0 1 2 3 } { 4 5 6 7 }
5841  *              *     * * *
5842  *
5843  * If we were to balance group-wise we'd place two tasks in the first group and
5844  * two tasks in the second group. Clearly this is undesired as it will overload
5845  * cpu 3 and leave one of the cpus in the second group unused.
5846  *
5847  * The current solution to this issue is detecting the skew in the first group
5848  * by noticing the lower domain failed to reach balance and had difficulty
5849  * moving tasks due to affinity constraints.
5850  *
5851  * When this is so detected; this group becomes a candidate for busiest; see
5852  * update_sd_pick_busiest(). And calculate_imbalance() and
5853  * find_busiest_group() avoid some of the usual balance conditions to allow it
5854  * to create an effective group imbalance.
5855  *
5856  * This is a somewhat tricky proposition since the next run might not find the
5857  * group imbalance and decide the groups need to be balanced again. A most
5858  * subtle and fragile situation.
5859  */
5860 
5861 static inline int sg_imbalanced(struct sched_group *group)
5862 {
5863         return group->sgc->imbalance;
5864 }
5865 
5866 /*
5867  * Compute the group capacity factor.
5868  *
5869  * Avoid the issue where N*frac(smt_capacity) >= 1 creates 'phantom' cores by
5870  * first dividing out the smt factor and computing the actual number of cores
5871  * and limit unit capacity with that.
5872  */
5873 static inline int sg_capacity_factor(struct lb_env *env, struct sched_group *group)
5874 {
5875         unsigned int capacity_factor, smt, cpus;
5876         unsigned int capacity, capacity_orig;
5877 
5878         capacity = group->sgc->capacity;
5879         capacity_orig = group->sgc->capacity_orig;
5880         cpus = group->group_weight;
5881 
5882         /* smt := ceil(cpus / capacity), assumes: 1 < smt_capacity < 2 */
5883         smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, capacity_orig);
5884         capacity_factor = cpus / smt; /* cores */
5885 
5886         capacity_factor = min_t(unsigned,
5887                 capacity_factor, DIV_ROUND_CLOSEST(capacity, SCHED_CAPACITY_SCALE));
5888         if (!capacity_factor)
5889                 capacity_factor = fix_small_capacity(env->sd, group);
5890 
5891         return capacity_factor;
5892 }
5893 
5894 /**
5895  * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5896  * @env: The load balancing environment.
5897  * @group: sched_group whose statistics are to be updated.
5898  * @load_idx: Load index of sched_domain of this_cpu for load calc.
5899  * @local_group: Does group contain this_cpu.
5900  * @sgs: variable to hold the statistics for this group.
5901  * @overload: Indicate more than one runnable task for any CPU.
5902  */
5903 static inline void update_sg_lb_stats(struct lb_env *env,
5904                         struct sched_group *group, int load_idx,
5905                         int local_group, struct sg_lb_stats *sgs,
5906                         bool *overload)
5907 {
5908         unsigned long load;
5909         int i;
5910 
5911         memset(sgs, 0, sizeof(*sgs));
5912 
5913         for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5914                 struct rq *rq = cpu_rq(i);
5915 
5916                 /* Bias balancing toward cpus of our domain */
5917                 if (local_group)
5918                         load = target_load(i, load_idx);
5919                 else
5920                         load = source_load(i, load_idx);
5921 
5922                 sgs->group_load += load;
5923                 sgs->sum_nr_running += rq->nr_running;
5924 
5925                 if (rq->nr_running > 1)
5926                         *overload = true;
5927 
5928 #ifdef CONFIG_NUMA_BALANCING
5929                 sgs->nr_numa_running += rq->nr_numa_running;
5930                 sgs->nr_preferred_running += rq->nr_preferred_running;
5931 #endif
5932                 sgs->sum_weighted_load += weighted_cpuload(i);
5933                 if (idle_cpu(i))
5934                         sgs->idle_cpus++;
5935         }
5936 
5937         /* Adjust by relative CPU capacity of the group */
5938         sgs->group_capacity = group->sgc->capacity;
5939         sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
5940 
5941         if (sgs->sum_nr_running)
5942                 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
5943 
5944         sgs->group_weight = group->group_weight;
5945 
5946         sgs->group_imb = sg_imbalanced(group);
5947         sgs->group_capacity_factor = sg_capacity_factor(env, group);
5948 
5949         if (sgs->group_capacity_factor > sgs->sum_nr_running)
5950                 sgs->group_has_free_capacity = 1;
5951 }
5952 
5953 /**
5954  * update_sd_pick_busiest - return 1 on busiest group
5955  * @env: The load balancing environment.
5956  * @sds: sched_domain statistics
5957  * @sg: sched_group candidate to be checked for being the busiest
5958  * @sgs: sched_group statistics
5959  *
5960  * Determine if @sg is a busier group than the previously selected
5961  * busiest group.
5962  *
5963  * Return: %true if @sg is a busier group than the previously selected
5964  * busiest group. %false otherwise.
5965  */
5966 static bool update_sd_pick_busiest(struct lb_env *env,
5967                                    struct sd_lb_stats *sds,
5968                                    struct sched_group *sg,
5969                                    struct sg_lb_stats *sgs)
5970 {
5971         if (sgs->avg_load <= sds->busiest_stat.avg_load)
5972                 return false;
5973 
5974         if (sgs->sum_nr_running > sgs->group_capacity_factor)
5975                 return true;
5976 
5977         if (sgs->group_imb)
5978                 return true;
5979 
5980         /*
5981          * ASYM_PACKING needs to move all the work to the lowest
5982          * numbered CPUs in the group, therefore mark all groups
5983          * higher than ourself as busy.
5984          */
5985         if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5986             env->dst_cpu < group_first_cpu(sg)) {
5987                 if (!sds->busiest)
5988                         return true;
5989 
5990                 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5991                         return true;
5992         }
5993 
5994         return false;
5995 }
5996 
5997 #ifdef CONFIG_NUMA_BALANCING
5998 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5999 {
6000         if (sgs->sum_nr_running > sgs->nr_numa_running)
6001                 return regular;
6002         if (sgs->sum_nr_running > sgs->nr_preferred_running)
6003                 return remote;
6004         return all;
6005 }
6006 
6007 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6008 {
6009         if (rq->nr_running > rq->nr_numa_running)
6010                 return regular;
6011         if (rq->nr_running > rq->nr_preferred_running)
6012                 return remote;
6013         return all;
6014 }
6015 #else
6016 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6017 {
6018         return all;
6019 }
6020 
6021 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6022 {
6023         return regular;
6024 }
6025 #endif /* CONFIG_NUMA_BALANCING */
6026 
6027 /**
6028  * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6029  * @env: The load balancing environment.
6030  * @sds: variable to hold the statistics for this sched_domain.
6031  */
6032 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6033 {
6034         struct sched_domain *child = env->sd->child;
6035         struct sched_group *sg = env->sd->groups;
6036         struct sg_lb_stats tmp_sgs;
6037         int load_idx, prefer_sibling = 0;
6038         bool overload = false;
6039 
6040         if (child && child->flags & SD_PREFER_SIBLING)
6041                 prefer_sibling = 1;
6042 
6043         load_idx = get_sd_load_idx(env->sd, env->idle);
6044 
6045         do {
6046                 struct sg_lb_stats *sgs = &tmp_sgs;
6047                 int local_group;
6048 
6049                 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6050                 if (local_group) {
6051                         sds->local = sg;
6052                         sgs = &sds->local_stat;
6053 
6054                         if (env->idle != CPU_NEWLY_IDLE ||
6055                             time_after_eq(jiffies, sg->sgc->next_update))
6056                                 update_group_capacity(env->sd, env->dst_cpu);
6057                 }
6058 
6059                 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6060                                                 &overload);
6061 
6062                 if (local_group)
6063                         goto next_group;
6064 
6065                 /*
6066                  * In case the child domain prefers tasks go to siblings
6067                  * first, lower the sg capacity factor to one so that we'll try
6068                  * and move all the excess tasks away. We lower the capacity
6069                  * of a group only if the local group has the capacity to fit
6070                  * these excess tasks, i.e. nr_running < group_capacity_factor. The
6071                  * extra check prevents the case where you always pull from the
6072                  * heaviest group when it is already under-utilized (possible
6073                  * with a large weight task outweighs the tasks on the system).
6074                  */
6075                 if (prefer_sibling && sds->local &&
6076                     sds->local_stat.group_has_free_capacity)
6077                         sgs->group_capacity_factor = min(sgs->group_capacity_factor, 1U);
6078 
6079                 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6080                         sds->busiest = sg;
6081                         sds->busiest_stat = *sgs;
6082                 }
6083 
6084 next_group:
6085                 /* Now, start updating sd_lb_stats */
6086                 sds->total_load += sgs->group_load;
6087                 sds->total_capacity += sgs->group_capacity;
6088 
6089                 sg = sg->next;
6090         } while (sg != env->sd->groups);
6091 
6092         if (env->sd->flags & SD_NUMA)
6093                 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6094 
6095         if (!env->sd->parent) {
6096                 /* update overload indicator if we are at root domain */
6097                 if (env->dst_rq->rd->overload != overload)
6098                         env->dst_rq->rd->overload = overload;
6099         }
6100 
6101 }
6102 
6103 /**
6104  * check_asym_packing - Check to see if the group is packed into the
6105  *                      sched doman.
6106  *
6107  * This is primarily intended to used at the sibling level.  Some
6108  * cores like POWER7 prefer to use lower numbered SMT threads.  In the
6109  * case of POWER7, it can move to lower SMT modes only when higher
6110  * threads are idle.  When in lower SMT modes, the threads will
6111  * perform better since they share less core resources.  Hence when we
6112  * have idle threads, we want them to be the higher ones.
6113  *
6114  * This packing function is run on idle threads.  It checks to see if
6115  * the busiest CPU in this domain (core in the P7 case) has a higher
6116  * CPU number than the packing function is being run on.  Here we are
6117  * assuming lower CPU number will be equivalent to lower a SMT thread
6118  * number.
6119  *
6120  * Return: 1 when packing is required and a task should be moved to
6121  * this CPU.  The amount of the imbalance is returned in *imbalance.
6122  *
6123  * @env: The load balancing environment.
6124  * @sds: Statistics of the sched_domain which is to be packed
6125  */
6126 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6127 {
6128         int busiest_cpu;
6129 
6130         if (!(env->sd->flags & SD_ASYM_PACKING))
6131                 return 0;
6132 
6133         if (!sds->busiest)
6134                 return 0;
6135 
6136         busiest_cpu = group_first_cpu(sds->busiest);
6137         if (env->dst_cpu > busiest_cpu)
6138                 return 0;
6139 
6140         env->imbalance = DIV_ROUND_CLOSEST(
6141                 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6142                 SCHED_CAPACITY_SCALE);
6143 
6144         return 1;
6145 }
6146 
6147 /**
6148  * fix_small_imbalance - Calculate the minor imbalance that exists
6149  *                      amongst the groups of a sched_domain, during
6150  *                      load balancing.
6151  * @env: The load balancing environment.
6152  * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6153  */
6154 static inline
6155 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6156 {
6157         unsigned long tmp, capa_now = 0, capa_move = 0;
6158         unsigned int imbn = 2;
6159         unsigned long scaled_busy_load_per_task;
6160         struct sg_lb_stats *local, *busiest;
6161 
6162         local = &sds->local_stat;
6163         busiest = &sds->busiest_stat;
6164 
6165         if (!local->sum_nr_running)
6166                 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6167         else if (busiest->load_per_task > local->load_per_task)
6168                 imbn = 1;
6169 
6170         scaled_busy_load_per_task =
6171                 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6172                 busiest->group_capacity;
6173 
6174         if (busiest->avg_load + scaled_busy_load_per_task >=
6175             local->avg_load + (scaled_busy_load_per_task * imbn)) {
6176                 env->imbalance = busiest->load_per_task;
6177                 return;
6178         }
6179 
6180         /*
6181          * OK, we don't have enough imbalance to justify moving tasks,
6182          * however we may be able to increase total CPU capacity used by
6183          * moving them.
6184          */
6185 
6186         capa_now += busiest->group_capacity *
6187                         min(busiest->load_per_task, busiest->avg_load);
6188         capa_now += local->group_capacity *
6189                         min(local->load_per_task, local->avg_load);
6190         capa_now /= SCHED_CAPACITY_SCALE;
6191 
6192         /* Amount of load we'd subtract */
6193         if (busiest->avg_load > scaled_busy_load_per_task) {
6194                 capa_move += busiest->group_capacity *
6195                             min(busiest->load_per_task,
6196                                 busiest->avg_load - scaled_busy_load_per_task);
6197         }
6198 
6199         /* Amount of load we'd add */
6200         if (busiest->avg_load * busiest->group_capacity <
6201             busiest->load_per_task * SCHED_CAPACITY_SCALE) {
6202                 tmp = (busiest->avg_load * busiest->group_capacity) /
6203                       local->group_capacity;
6204         } else {
6205                 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6206                       local->group_capacity;
6207         }
6208         capa_move += local->group_capacity *
6209                     min(local->load_per_task, local->avg_load + tmp);
6210         capa_move /= SCHED_CAPACITY_SCALE;
6211 
6212         /* Move if we gain throughput */
6213         if (capa_move > capa_now)
6214                 env->imbalance = busiest->load_per_task;
6215 }
6216 
6217 /**
6218  * calculate_imbalance - Calculate the amount of imbalance present within the
6219  *                       groups of a given sched_domain during load balance.
6220  * @env: load balance environment
6221  * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6222  */
6223 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6224 {
6225         unsigned long max_pull, load_above_capacity = ~0UL;
6226         struct sg_lb_stats *local, *busiest;
6227 
6228         local = &sds->local_stat;
6229         busiest = &sds->busiest_stat;
6230 
6231         if (busiest->group_imb) {
6232                 /*
6233                  * In the group_imb case we cannot rely on group-wide averages
6234                  * to ensure cpu-load equilibrium, look at wider averages. XXX
6235                  */
6236                 busiest->load_per_task =
6237                         min(busiest->load_per_task, sds->avg_load);
6238         }
6239 
6240         /*
6241          * In the presence of smp nice balancing, certain scenarios can have
6242          * max load less than avg load(as we skip the groups at or below
6243          * its cpu_capacity, while calculating max_load..)
6244          */
6245         if (busiest->avg_load <= sds->avg_load ||
6246             local->avg_load >= sds->avg_load) {
6247                 env->imbalance = 0;
6248                 return fix_small_imbalance(env, sds);
6249         }
6250 
6251         if (!busiest->group_imb) {
6252                 /*
6253                  * Don't want to pull so many tasks that a group would go idle.
6254                  * Except of course for the group_imb case, since then we might
6255                  * have to drop below capacity to reach cpu-load equilibrium.
6256                  */
6257                 load_above_capacity =
6258                         (busiest->sum_nr_running - busiest->group_capacity_factor);
6259 
6260                 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_CAPACITY_SCALE);
6261                 load_above_capacity /= busiest->group_capacity;
6262         }
6263 
6264         /*
6265          * We're trying to get all the cpus to the average_load, so we don't
6266          * want to push ourselves above the average load, nor do we wish to
6267          * reduce the max loaded cpu below the average load. At the same time,
6268          * we also don't want to reduce the group load below the group capacity
6269          * (so that we can implement power-savings policies etc). Thus we look
6270          * for the minimum possible imbalance.
6271          */
6272         max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6273 
6274         /* How much load to actually move to equalise the imbalance */
6275         env->imbalance = min(
6276                 max_pull * busiest->group_capacity,
6277                 (sds->avg_load - local->avg_load) * local->group_capacity
6278         ) / SCHED_CAPACITY_SCALE;
6279 
6280         /*
6281          * if *imbalance is less than the average load per runnable task
6282          * there is no guarantee that any tasks will be moved so we'll have
6283          * a think about bumping its value to force at least one task to be
6284          * moved
6285          */
6286         if (env->imbalance < busiest->load_per_task)
6287                 return fix_small_imbalance(env, sds);
6288 }
6289 
6290 /******* find_busiest_group() helpers end here *********************/
6291 
6292 /**
6293  * find_busiest_group - Returns the busiest group within the sched_domain
6294  * if there is an imbalance. If there isn't an imbalance, and
6295  * the user has opted for power-savings, it returns a group whose
6296  * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6297  * such a group exists.
6298  *
6299  * Also calculates the amount of weighted load which should be moved
6300  * to restore balance.
6301  *
6302  * @env: The load balancing environment.
6303  *
6304  * Return:      - The busiest group if imbalance exists.
6305  *              - If no imbalance and user has opted for power-savings balance,
6306  *                 return the least loaded group whose CPUs can be
6307  *                 put to idle by rebalancing its tasks onto our group.
6308  */
6309 static struct sched_group *find_busiest_group(struct lb_env *env)
6310 {
6311         struct sg_lb_stats *local, *busiest;
6312         struct sd_lb_stats sds;
6313 
6314         init_sd_lb_stats(&sds);
6315 
6316         /*
6317          * Compute the various statistics relavent for load balancing at
6318          * this level.
6319          */
6320         update_sd_lb_stats(env, &sds);
6321         local = &sds.local_stat;
6322         busiest = &sds.busiest_stat;
6323 
6324         if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6325             check_asym_packing(env, &sds))
6326                 return sds.busiest;
6327 
6328         /* There is no busy sibling group to pull tasks from */
6329         if (!sds.busiest || busiest->sum_nr_running == 0)
6330                 goto out_balanced;
6331 
6332         sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
6333                                                 / sds.total_capacity;
6334 
6335         /*
6336          * If the busiest group is imbalanced the below checks don't
6337          * work because they assume all things are equal, which typically
6338          * isn't true due to cpus_allowed constraints and the like.
6339          */
6340         if (busiest->group_imb)
6341                 goto force_balance;
6342 
6343         /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6344         if (env->idle == CPU_NEWLY_IDLE && local->group_has_free_capacity &&
6345             !busiest->group_has_free_capacity)
6346                 goto force_balance;
6347 
6348         /*
6349          * If the local group is more busy than the selected busiest group
6350          * don't try and pull any tasks.
6351          */
6352         if (local->avg_load >= busiest->avg_load)
6353                 goto out_balanced;
6354 
6355         /*
6356          * Don't pull any tasks if this group is already above the domain
6357          * average load.
6358          */
6359         if (local->avg_load >= sds.avg_load)
6360                 goto out_balanced;
6361 
6362         if (env->idle == CPU_IDLE) {
6363                 /*
6364                  * This cpu is idle. If the busiest group load doesn't
6365                  * have more tasks than the number of available cpu's and
6366                  * there is no imbalance between this and busiest group
6367                  * wrt to idle cpu's, it is balanced.
6368                  */
6369                 if ((local->idle_cpus < busiest->idle_cpus) &&
6370                     busiest->sum_nr_running <= busiest->group_weight)
6371                         goto out_balanced;
6372         } else {
6373                 /*
6374                  * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6375                  * imbalance_pct to be conservative.
6376                  */
6377                 if (100 * busiest->avg_load <=
6378                                 env->sd->imbalance_pct * local->avg_load)
6379                         goto out_balanced;
6380         }
6381 
6382 force_balance:
6383         /* Looks like there is an imbalance. Compute it */
6384         calculate_imbalance(env, &sds);
6385         return sds.busiest;
6386 
6387 out_balanced:
6388         env->imbalance = 0;
6389         return NULL;
6390 }
6391 
6392 /*
6393  * find_busiest_queue - find the busiest runqueue among the cpus in group.
6394  */
6395 static struct rq *find_busiest_queue(struct lb_env *env,
6396                                      struct sched_group *group)
6397 {
6398         struct rq *busiest = NULL, *rq;
6399         unsigned long busiest_load = 0, busiest_capacity = 1;
6400         int i;
6401 
6402         for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6403                 unsigned long capacity, capacity_factor, wl;
6404                 enum fbq_type rt;
6405 
6406                 rq = cpu_rq(i);
6407                 rt = fbq_classify_rq(rq);
6408 
6409                 /*
6410                  * We classify groups/runqueues into three groups:
6411                  *  - regular: there are !numa tasks
6412                  *  - remote:  there are numa tasks that run on the 'wrong' node
6413                  *  - all:     there is no distinction
6414                  *
6415                  * In order to avoid migrating ideally placed numa tasks,
6416                  * ignore those when there's better options.
6417                  *
6418                  * If we ignore the actual busiest queue to migrate another
6419                  * task, the next balance pass can still reduce the busiest
6420                  * queue by moving tasks around inside the node.
6421                  *
6422                  * If we cannot move enough load due to this classification
6423                  * the next pass will adjust the group classification and
6424                  * allow migration of more tasks.
6425                  *
6426                  * Both cases only affect the total convergence complexity.
6427                  */
6428                 if (rt > env->fbq_type)
6429                         continue;
6430 
6431                 capacity = capacity_of(i);
6432                 capacity_factor = DIV_ROUND_CLOSEST(capacity, SCHED_CAPACITY_SCALE);
6433                 if (!capacity_factor)
6434                         capacity_factor = fix_small_capacity(env->sd, group);
6435 
6436                 wl = weighted_cpuload(i);
6437 
6438                 /*
6439                  * When comparing with imbalance, use weighted_cpuload()
6440                  * which is not scaled with the cpu capacity.
6441                  */
6442                 if (capacity_factor && rq->nr_running == 1 && wl > env->imbalance)
6443                         continue;
6444 
6445                 /*
6446                  * For the load comparisons with the other cpu's, consider
6447                  * the weighted_cpuload() scaled with the cpu capacity, so
6448                  * that the load can be moved away from the cpu that is
6449                  * potentially running at a lower capacity.
6450                  *
6451                  * Thus we're looking for max(wl_i / capacity_i), crosswise
6452                  * multiplication to rid ourselves of the division works out
6453                  * to: wl_i * capacity_j > wl_j * capacity_i;  where j is
6454                  * our previous maximum.
6455                  */
6456                 if (wl * busiest_capacity > busiest_load * capacity) {
6457                         busiest_load = wl;
6458                         busiest_capacity = capacity;
6459                         busiest = rq;
6460                 }
6461         }
6462 
6463         return busiest;
6464 }
6465 
6466 /*
6467  * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6468  * so long as it is large enough.
6469  */
6470 #define MAX_PINNED_INTERVAL     512
6471 
6472 /* Working cpumask for load_balance and load_balance_newidle. */
6473 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6474 
6475 static int need_active_balance(struct lb_env *env)
6476 {
6477         struct sched_domain *sd = env->sd;
6478 
6479         if (env->idle == CPU_NEWLY_IDLE) {
6480 
6481                 /*
6482                  * ASYM_PACKING needs to force migrate tasks from busy but
6483                  * higher numbered CPUs in order to pack all tasks in the
6484                  * lowest numbered CPUs.
6485                  */
6486                 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6487                         return 1;
6488         }
6489 
6490         return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6491 }
6492 
6493 static int active_load_balance_cpu_stop(void *data);
6494 
6495 static int should_we_balance(struct lb_env *env)
6496 {
6497         struct sched_group *sg = env->sd->groups;
6498         struct cpumask *sg_cpus, *sg_mask;
6499         int cpu, balance_cpu = -1;
6500 
6501         /*
6502          * In the newly idle case, we will allow all the cpu's
6503          * to do the newly idle load balance.
6504          */
6505         if (env->idle == CPU_NEWLY_IDLE)
6506                 return 1;
6507 
6508         sg_cpus = sched_group_cpus(sg);
6509         sg_mask = sched_group_mask(sg);
6510         /* Try to find first idle cpu */
6511         for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6512                 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6513                         continue;
6514 
6515                 balance_cpu = cpu;
6516                 break;
6517         }
6518 
6519         if (balance_cpu == -1)
6520                 balance_cpu = group_balance_cpu(sg);
6521 
6522         /*
6523          * First idle cpu or the first cpu(busiest) in this sched group
6524          * is eligible for doing load balancing at this and above domains.
6525          */
6526         return balance_cpu == env->dst_cpu;
6527 }
6528 
6529 /*
6530  * Check this_cpu to ensure it is balanced within domain. Attempt to move
6531  * tasks if there is an imbalance.
6532  */
6533 static int load_balance(int this_cpu, struct rq *this_rq,
6534                         struct sched_domain *sd, enum cpu_idle_type idle,
6535                         int *continue_balancing)
6536 {
6537         int ld_moved, cur_ld_moved, active_balance = 0;
6538         struct sched_domain *sd_parent = sd->parent;
6539         struct sched_group *group;
6540         struct rq *busiest;
6541         unsigned long flags;
6542         struct cpumask *cpus = __get_cpu_var(load_balance_mask);
6543 
6544         struct lb_env env = {
6545                 .sd             = sd,
6546                 .dst_cpu        = this_cpu,
6547                 .dst_rq         = this_rq,
6548                 .dst_grpmask    = sched_group_cpus(sd->groups),
6549                 .idle           = idle,
6550                 .loop_break     = sched_nr_migrate_break,
6551                 .cpus           = cpus,
6552                 .fbq_type       = all,
6553         };
6554 
6555         /*
6556          * For NEWLY_IDLE load_balancing, we don't need to consider
6557          * other cpus in our group
6558          */
6559         if (idle == CPU_NEWLY_IDLE)
6560                 env.dst_grpmask = NULL;
6561 
6562         cpumask_copy(cpus, cpu_active_mask);
6563 
6564         schedstat_inc(sd, lb_count[idle]);
6565 
6566 redo:
6567         if (!should_we_balance(&env)) {
6568                 *continue_balancing = 0;
6569                 goto out_balanced;
6570         }
6571 
6572         group = find_busiest_group(&env);
6573         if (!group) {
6574                 schedstat_inc(sd, lb_nobusyg[idle]);
6575                 goto out_balanced;
6576         }
6577 
6578         busiest = find_busiest_queue(&env, group);
6579         if (!busiest) {
6580                 schedstat_inc(sd, lb_nobusyq[idle]);
6581                 goto out_balanced;
6582         }
6583 
6584         BUG_ON(busiest == env.dst_rq);
6585 
6586         schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6587 
6588         ld_moved = 0;
6589         if (busiest->nr_running > 1) {
6590                 /*
6591                  * Attempt to move tasks. If find_busiest_group has found
6592                  * an imbalance but busiest->nr_running <= 1, the group is
6593                  * still unbalanced. ld_moved simply stays zero, so it is
6594                  * correctly treated as an imbalance.
6595                  */
6596                 env.flags |= LBF_ALL_PINNED;
6597                 env.src_cpu   = busiest->cpu;
6598                 env.src_rq    = busiest;
6599                 env.loop_max  = min(sysctl_sched_nr_migrate, busiest->nr_running);
6600 
6601 more_balance:
6602                 local_irq_save(flags);
6603                 double_rq_lock(env.dst_rq, busiest);
6604 
6605                 /*
6606                  * cur_ld_moved - load moved in current iteration
6607                  * ld_moved     - cumulative load moved across iterations
6608                  */
6609                 cur_ld_moved = move_tasks(&env);
6610                 ld_moved += cur_ld_moved;
6611                 double_rq_unlock(env.dst_rq, busiest);
6612                 local_irq_restore(flags);
6613 
6614                 /*
6615                  * some other cpu did the load balance for us.
6616                  */
6617                 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
6618                         resched_cpu(env.dst_cpu);
6619 
6620                 if (env.flags & LBF_NEED_BREAK) {
6621                         env.flags &= ~LBF_NEED_BREAK;
6622                         goto more_balance;
6623                 }
6624 
6625                 /*
6626                  * Revisit (affine) tasks on src_cpu that couldn't be moved to
6627                  * us and move them to an alternate dst_cpu in our sched_group
6628                  * where they can run. The upper limit on how many times we
6629                  * iterate on same src_cpu is dependent on number of cpus in our
6630                  * sched_group.
6631                  *
6632                  * This changes load balance semantics a bit on who can move
6633                  * load to a given_cpu. In addition to the given_cpu itself
6634                  * (or a ilb_cpu acting on its behalf where given_cpu is
6635                  * nohz-idle), we now have balance_cpu in a position to move
6636                  * load to given_cpu. In rare situations, this may cause
6637                  * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6638                  * _independently_ and at _same_ time to move some load to
6639                  * given_cpu) causing exceess load to be moved to given_cpu.
6640                  * This however should not happen so much in practice and
6641                  * moreover subsequent load balance cycles should correct the
6642                  * excess load moved.
6643                  */
6644                 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
6645 
6646                         /* Prevent to re-select dst_cpu via env's cpus */
6647                         cpumask_clear_cpu(env.dst_cpu, env.cpus);
6648 
6649                         env.dst_rq       = cpu_rq(env.new_dst_cpu);
6650                         env.dst_cpu      = env.new_dst_cpu;
6651                         env.flags       &= ~LBF_DST_PINNED;
6652                         env.loop         = 0;
6653                         env.loop_break   = sched_nr_migrate_break;
6654 
6655                         /*
6656                          * Go back to "more_balance" rather than "redo" since we
6657                          * need to continue with same src_cpu.
6658                          */
6659                         goto more_balance;
6660                 }
6661 
6662                 /*
6663                  * We failed to reach balance because of affinity.
6664                  */
6665                 if (sd_parent) {
6666                         int *group_imbalance = &sd_parent->groups->sgc->imbalance;
6667 
6668                         if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
6669                                 *group_imbalance = 1;
6670                         } else if (*group_imbalance)
6671                                 *group_imbalance = 0;
6672                 }
6673 
6674                 /* All tasks on this runqueue were pinned by CPU affinity */
6675                 if (unlikely(env.flags & LBF_ALL_PINNED)) {
6676                         cpumask_clear_cpu(cpu_of(busiest), cpus);
6677                         if (!cpumask_empty(cpus)) {
6678                                 env.loop = 0;
6679                                 env.loop_break = sched_nr_migrate_break;
6680                                 goto redo;
6681                         }
6682                         goto out_balanced;
6683                 }
6684         }
6685 
6686         if (!ld_moved) {
6687                 schedstat_inc(sd, lb_failed[idle]);
6688                 /*
6689                  * Increment the failure counter only on periodic balance.
6690                  * We do not want newidle balance, which can be very
6691                  * frequent, pollute the failure counter causing
6692                  * excessive cache_hot migrations and active balances.
6693                  */
6694                 if (idle != CPU_NEWLY_IDLE)
6695                         sd->nr_balance_failed++;
6696 
6697                 if (need_active_balance(&env)) {
6698                         raw_spin_lock_irqsave(&busiest->lock, flags);
6699 
6700                         /* don't kick the active_load_balance_cpu_stop,
6701                          * if the curr task on busiest cpu can't be
6702                          * moved to this_cpu
6703                          */
6704                         if (!cpumask_test_cpu(this_cpu,
6705                                         tsk_cpus_allowed(busiest->curr))) {
6706                                 raw_spin_unlock_irqrestore(&busiest->lock,
6707                                                             flags);
6708                                 env.flags |= LBF_ALL_PINNED;
6709                                 goto out_one_pinned;
6710                         }
6711 
6712                         /*
6713                          * ->active_balance synchronizes accesses to
6714                          * ->active_balance_work.  Once set, it's cleared
6715                          * only after active load balance is finished.
6716                          */
6717                         if (!busiest->active_balance) {
6718                                 busiest->active_balance = 1;
6719                                 busiest->push_cpu = this_cpu;
6720                                 active_balance = 1;
6721                         }
6722                         raw_spin_unlock_irqrestore(&busiest->lock, flags);
6723 
6724                         if (active_balance) {
6725                                 stop_one_cpu_nowait(cpu_of(busiest),
6726                                         active_load_balance_cpu_stop, busiest,
6727                                         &busiest->active_balance_work);
6728                         }
6729 
6730                         /*
6731                          * We've kicked active balancing, reset the failure
6732                          * counter.
6733                          */
6734                         sd->nr_balance_failed = sd->cache_nice_tries+1;
6735                 }
6736         } else
6737                 sd->nr_balance_failed = 0;
6738 
6739         if (likely(!active_balance)) {
6740                 /* We were unbalanced, so reset the balancing interval */
6741                 sd->balance_interval = sd->min_interval;
6742         } else {
6743                 /*
6744                  * If we've begun active balancing, start to back off. This
6745                  * case may not be covered by the all_pinned logic if there
6746                  * is only 1 task on the busy runqueue (because we don't call
6747                  * move_tasks).
6748                  */
6749                 if (sd->balance_interval < sd->max_interval)
6750                         sd->balance_interval *= 2;
6751         }
6752 
6753         goto out;
6754 
6755 out_balanced:
6756         schedstat_inc(sd, lb_balanced[idle]);
6757 
6758         sd->nr_balance_failed = 0;
6759 
6760 out_one_pinned:
6761         /* tune up the balancing interval */
6762         if (((env.flags & LBF_ALL_PINNED) &&
6763                         sd->balance_interval < MAX_PINNED_INTERVAL) ||
6764                         (sd->balance_interval < sd->max_interval))
6765                 sd->balance_interval *= 2;
6766 
6767         ld_moved = 0;
6768 out:
6769         return ld_moved;
6770 }
6771 
6772 static inline unsigned long
6773 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
6774 {
6775         unsigned long interval = sd->balance_interval;
6776 
6777         if (cpu_busy)
6778                 interval *= sd->busy_factor;
6779 
6780         /* scale ms to jiffies */
6781         interval = msecs_to_jiffies(interval);
6782         interval = clamp(interval, 1UL, max_load_balance_interval);
6783 
6784         return interval;
6785 }
6786 
6787 static inline void
6788 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
6789 {
6790         unsigned long interval, next;
6791 
6792         interval = get_sd_balance_interval(sd, cpu_busy);
6793         next = sd->last_balance + interval;
6794 
6795         if (time_after(*next_balance, next))
6796                 *next_balance = next;
6797 }
6798 
6799 /*
6800  * idle_balance is called by schedule() if this_cpu is about to become
6801  * idle. Attempts to pull tasks from other CPUs.
6802  */
6803 static int idle_balance(struct rq *this_rq)
6804 {
6805         unsigned long next_balance = jiffies + HZ;
6806         int this_cpu = this_rq->cpu;
6807         struct sched_domain *sd;
6808         int pulled_task = 0;
6809         u64 curr_cost = 0;
6810 
6811         idle_enter_fair(this_rq);
6812 
6813         /*
6814          * We must set idle_stamp _before_ calling idle_balance(), such that we
6815          * measure the duration of idle_balance() as idle time.
6816          */
6817         this_rq->idle_stamp = rq_clock(this_rq);
6818 
6819         if (this_rq->avg_idle < sysctl_sched_migration_cost ||
6820             !this_rq->rd->overload) {
6821                 rcu_read_lock();
6822                 sd = rcu_dereference_check_sched_domain(this_rq->sd);
6823                 if (sd)
6824                         update_next_balance(sd, 0, &next_balance);
6825                 rcu_read_unlock();
6826 
6827                 goto out;
6828         }
6829 
6830         /*
6831          * Drop the rq->lock, but keep IRQ/preempt disabled.
6832          */
6833         raw_spin_unlock(&this_rq->lock);
6834 
6835         update_blocked_averages(this_cpu);
6836         rcu_read_lock();
6837         for_each_domain(this_cpu, sd) {
6838                 int continue_balancing = 1;
6839                 u64 t0, domain_cost;
6840 
6841                 if (!(sd->flags & SD_LOAD_BALANCE))
6842                         continue;
6843 
6844                 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
6845                         update_next_balance(sd, 0, &next_balance);
6846                         break;
6847                 }
6848 
6849                 if (sd->flags & SD_BALANCE_NEWIDLE) {
6850                         t0 = sched_clock_cpu(this_cpu);
6851 
6852                         pulled_task = load_balance(this_cpu, this_rq,
6853                                                    sd, CPU_NEWLY_IDLE,
6854                                                    &continue_balancing);
6855 
6856                         domain_cost = sched_clock_cpu(this_cpu) - t0;
6857                         if (domain_cost > sd->max_newidle_lb_cost)
6858                                 sd->max_newidle_lb_cost = domain_cost;
6859 
6860                         curr_cost += domain_cost;
6861                 }
6862 
6863                 update_next_balance(sd, 0, &next_balance);
6864 
6865                 /*
6866                  * Stop searching for tasks to pull if there are
6867                  * now runnable tasks on this rq.
6868                  */
6869                 if (pulled_task || this_rq->nr_running > 0)
6870                         break;
6871         }
6872         rcu_read_unlock();
6873 
6874         raw_spin_lock(&this_rq->lock);
6875 
6876         if (curr_cost > this_rq->max_idle_balance_cost)
6877                 this_rq->max_idle_balance_cost = curr_cost;
6878 
6879         /*
6880          * While browsing the domains, we released the rq lock, a task could
6881          * have been enqueued in the meantime. Since we're not going idle,
6882          * pretend we pulled a task.
6883          */
6884         if (this_rq->cfs.h_nr_running && !pulled_task)
6885                 pulled_task = 1;
6886 
6887 out:
6888         /* Move the next balance forward */
6889         if (time_after(this_rq->next_balance, next_balance))
6890                 this_rq->next_balance = next_balance;
6891 
6892         /* Is there a task of a high priority class? */
6893         if (this_rq->nr_running != this_rq->cfs.h_nr_running)
6894                 pulled_task = -1;
6895 
6896         if (pulled_task) {
6897                 idle_exit_fair(this_rq);
6898                 this_rq->idle_stamp = 0;
6899         }
6900 
6901         return pulled_task;
6902 }
6903 
6904 /*
6905  * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6906  * running tasks off the busiest CPU onto idle CPUs. It requires at
6907  * least 1 task to be running on each physical CPU where possible, and
6908  * avoids physical / logical imbalances.
6909  */
6910 static int active_load_balance_cpu_stop(void *data)
6911 {
6912         struct rq *busiest_rq = data;
6913         int busiest_cpu = cpu_of(busiest_rq);
6914         int target_cpu = busiest_rq->push_cpu;
6915         struct rq *target_rq = cpu_rq(target_cpu);
6916         struct sched_domain *sd;
6917 
6918         raw_spin_lock_irq(&busiest_rq->lock);
6919 
6920         /* make sure the requested cpu hasn't gone down in the meantime */
6921         if (unlikely(busiest_cpu != smp_processor_id() ||
6922                      !busiest_rq->active_balance))
6923                 goto out_unlock;
6924 
6925         /* Is there any task to move? */
6926         if (busiest_rq->nr_running <= 1)
6927                 goto out_unlock;
6928 
6929         /*
6930          * This condition is "impossible", if it occurs
6931          * we need to fix it. Originally reported by
6932          * Bjorn Helgaas on a 128-cpu setup.
6933          */
6934         BUG_ON(busiest_rq == target_rq);
6935 
6936         /* move a task from busiest_rq to target_rq */
6937         double_lock_balance(busiest_rq, target_rq);
6938 
6939         /* Search for an sd spanning us and the target CPU. */
6940         rcu_read_lock();
6941         for_each_domain(target_cpu, sd) {
6942                 if ((sd->flags & SD_LOAD_BALANCE) &&
6943                     cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6944                                 break;
6945         }
6946 
6947         if (likely(sd)) {
6948                 struct lb_env env = {
6949                         .sd             = sd,
6950                         .dst_cpu        = target_cpu,
6951                         .dst_rq         = target_rq,
6952                         .src_cpu        = busiest_rq->cpu,
6953                         .src_rq         = busiest_rq,
6954                         .idle           = CPU_IDLE,
6955                 };
6956 
6957                 schedstat_inc(sd, alb_count);
6958 
6959                 if (move_one_task(&env))
6960                         schedstat_inc(sd, alb_pushed);
6961                 else
6962                         schedstat_inc(sd, alb_failed);
6963         }
6964         rcu_read_unlock();
6965         double_unlock_balance(busiest_rq, target_rq);
6966 out_unlock:
6967         busiest_rq->active_balance = 0;
6968         raw_spin_unlock_irq(&busiest_rq->lock);
6969         return 0;
6970 }
6971 
6972 static inline int on_null_domain(struct rq *rq)
6973 {
6974         return unlikely(!rcu_dereference_sched(rq->sd));
6975 }
6976 
6977 #ifdef CONFIG_NO_HZ_COMMON
6978 /*
6979  * idle load balancing details
6980  * - When one of the busy CPUs notice that there may be an idle rebalancing
6981  *   needed, they will kick the idle load balancer, which then does idle
6982  *   load balancing for all the idle CPUs.
6983  */
6984 static struct {
6985         cpumask_var_t idle_cpus_mask;
6986         atomic_t nr_cpus;
6987         unsigned long next_balance;     /* in jiffy units */
6988 } nohz ____cacheline_aligned;
6989 
6990 static inline int find_new_ilb(void)
6991 {
6992         int ilb = cpumask_first(nohz.idle_cpus_mask);
6993 
6994         if (ilb < nr_cpu_ids && idle_cpu(ilb))
6995                 return ilb;
6996 
6997         return nr_cpu_ids;
6998 }
6999 
7000 /*
7001  * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7002  * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7003  * CPU (if there is one).
7004  */
7005 static void nohz_balancer_kick(void)
7006 {
7007         int ilb_cpu;
7008 
7009         nohz.next_balance++;
7010 
7011         ilb_cpu = find_new_ilb();
7012 
7013         if (ilb_cpu >= nr_cpu_ids)
7014                 return;
7015 
7016         if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7017                 return;
7018         /*
7019          * Use smp_send_reschedule() instead of resched_cpu().
7020          * This way we generate a sched IPI on the target cpu which
7021          * is idle. And the softirq performing nohz idle load balance
7022          * will be run before returning from the IPI.
7023          */
7024         smp_send_reschedule(ilb_cpu);
7025         return;
7026 }
7027 
7028 static inline void nohz_balance_exit_idle(int cpu)
7029 {
7030         if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7031                 /*
7032                  * Completely isolated CPUs don't ever set, so we must test.
7033                  */
7034                 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7035                         cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7036                         atomic_dec(&nohz.nr_cpus);
7037                 }
7038                 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7039         }
7040 }
7041 
7042 static inline void set_cpu_sd_state_busy(void)
7043 {
7044         struct sched_domain *sd;
7045         int cpu = smp_processor_id();
7046 
7047         rcu_read_lock();
7048         sd = rcu_dereference(per_cpu(sd_busy, cpu));
7049 
7050         if (!sd || !sd->nohz_idle)
7051                 goto unlock;
7052         sd->nohz_idle = 0;
7053 
7054         atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7055 unlock:
7056         rcu_read_unlock();
7057 }
7058 
7059 void set_cpu_sd_state_idle(void)
7060 {
7061         struct sched_domain *sd;
7062         int cpu = smp_processor_id();
7063 
7064         rcu_read_lock();
7065         sd = rcu_dereference(per_cpu(sd_busy, cpu));
7066 
7067         if (!sd || sd->nohz_idle)
7068                 goto unlock;
7069         sd->nohz_idle = 1;
7070 
7071         atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7072 unlock:
7073         rcu_read_unlock();
7074 }
7075 
7076 /*
7077  * This routine will record that the cpu is going idle with tick stopped.
7078  * This info will be used in performing idle load balancing in the future.
7079  */
7080 void nohz_balance_enter_idle(int cpu)
7081 {
7082         /*
7083          * If this cpu is going down, then nothing needs to be done.
7084          */
7085         if (!cpu_active(cpu))
7086                 return;
7087 
7088         if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7089                 return;
7090 
7091         /*
7092          * If we're a completely isolated CPU, we don't play.
7093          */
7094         if (on_null_domain(cpu_rq(cpu)))
7095                 return;
7096 
7097         cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7098         atomic_inc(&nohz.nr_cpus);
7099         set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7100 }
7101 
7102 static int sched_ilb_notifier(struct notifier_block *nfb,
7103                                         unsigned long action, void *hcpu)
7104 {
7105         switch (action & ~CPU_TASKS_FROZEN) {
7106         case CPU_DYING:
7107                 nohz_balance_exit_idle(smp_processor_id());
7108                 return NOTIFY_OK;
7109         default:
7110                 return NOTIFY_DONE;
7111         }
7112 }
7113 #endif
7114 
7115 static DEFINE_SPINLOCK(balancing);
7116 
7117 /*
7118  * Scale the max load_balance interval with the number of CPUs in the system.
7119  * This trades load-balance latency on larger machines for less cross talk.
7120  */
7121 void update_max_interval(void)
7122 {
7123         max_load_balance_interval = HZ*num_online_cpus()/10;
7124 }
7125 
7126 /*
7127  * It checks each scheduling domain to see if it is due to be balanced,
7128  * and initiates a balancing operation if so.
7129  *
7130  * Balancing parameters are set up in init_sched_domains.
7131  */
7132 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7133 {
7134         int continue_balancing = 1;
7135         int cpu = rq->cpu;
7136         unsigned long interval;
7137         struct sched_domain *sd;
7138         /* Earliest time when we have to do rebalance again */
7139         unsigned long next_balance = jiffies + 60*HZ;
7140         int update_next_balance = 0;
7141         int need_serialize, need_decay = 0;
7142         u64 max_cost = 0;
7143 
7144         update_blocked_averages(cpu);
7145 
7146         rcu_read_lock();
7147         for_each_domain(cpu, sd) {
7148                 /*
7149                  * Decay the newidle max times here because this is a regular
7150                  * visit to all the domains. Decay ~1% per second.
7151                  */
7152                 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7153                         sd->max_newidle_lb_cost =
7154                                 (sd->max_newidle_lb_cost * 253) / 256;
7155                         sd->next_decay_max_lb_cost = jiffies + HZ;
7156                         need_decay = 1;
7157                 }
7158                 max_cost += sd->max_newidle_lb_cost;
7159 
7160                 if (!(sd->flags & SD_LOAD_BALANCE))
7161                         continue;
7162 
7163                 /*
7164                  * Stop the load balance at this level. There is another
7165                  * CPU in our sched group which is doing load balancing more
7166                  * actively.
7167                  */
7168                 if (!continue_balancing) {
7169                         if (need_decay)
7170                                 continue;
7171                         break;
7172                 }
7173 
7174                 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7175 
7176                 need_serialize = sd->flags & SD_SERIALIZE;
7177                 if (need_serialize) {
7178                         if (!spin_trylock(&balancing))
7179                                 goto out;
7180                 }
7181 
7182                 if (time_after_eq(jiffies, sd->last_balance + interval)) {
7183                         if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7184                                 /*
7185                                  * The LBF_DST_PINNED logic could have changed
7186                                  * env->dst_cpu, so we can't know our idle
7187                                  * state even if we migrated tasks. Update it.
7188                                  */
7189                                 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7190                         }
7191                         sd->last_balance = jiffies;
7192                         interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7193                 }
7194                 if (need_serialize)
7195                         spin_unlock(&balancing);
7196 out:
7197                 if (time_after(next_balance, sd->last_balance + interval)) {
7198                         next_balance = sd->last_balance + interval;
7199                         update_next_balance = 1;
7200                 }
7201         }
7202         if (need_decay) {
7203                 /*
7204                  * Ensure the rq-wide value also decays but keep it at a
7205                  * reasonable floor to avoid funnies with rq->avg_idle.
7206                  */
7207                 rq->max_idle_balance_cost =
7208                         max((u64)sysctl_sched_migration_cost, max_cost);
7209         }
7210         rcu_read_unlock();
7211 
7212         /*
7213          * next_balance will be updated only when there is a need.
7214          * When the cpu is attached to null domain for ex, it will not be
7215          * updated.
7216          */
7217         if (likely(update_next_balance))
7218                 rq->next_balance = next_balance;
7219 }
7220 
7221 #ifdef CONFIG_NO_HZ_COMMON
7222 /*
7223  * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7224  * rebalancing for all the cpus for whom scheduler ticks are stopped.
7225  */
7226 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7227 {
7228         int this_cpu = this_rq->cpu;
7229         struct rq *rq;
7230         int balance_cpu;
7231 
7232         if (idle != CPU_IDLE ||
7233             !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7234                 goto end;
7235 
7236         for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7237                 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7238                         continue;
7239 
7240                 /*
7241                  * If this cpu gets work to do, stop the load balancing
7242                  * work being done for other cpus. Next load
7243                  * balancing owner will pick it up.
7244                  */
7245                 if (need_resched())
7246                         break;
7247 
7248                 rq = cpu_rq(balance_cpu);
7249 
7250                 /*
7251                  * If time for next balance is due,
7252                  * do the balance.
7253                  */
7254                 if (time_after_eq(jiffies, rq->next_balance)) {
7255                         raw_spin_lock_irq(&rq->lock);
7256                         update_rq_clock(rq);
7257                         update_idle_cpu_load(rq);
7258                         raw_spin_unlock_irq(&rq->lock);
7259                         rebalance_domains(rq, CPU_IDLE);
7260                 }
7261 
7262                 if (time_after(this_rq->next_balance, rq->next_balance))
7263                         this_rq->next_balance = rq->next_balance;
7264         }
7265         nohz.next_balance = this_rq->next_balance;
7266 end:
7267         clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7268 }
7269 
7270 /*
7271  * Current heuristic for kicking the idle load balancer in the presence
7272  * of an idle cpu is the system.
7273  *   - This rq has more than one task.
7274  *   - At any scheduler domain level, this cpu's scheduler group has multiple
7275  *     busy cpu's exceeding the group's capacity.
7276  *   - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7277  *     domain span are idle.
7278  */
7279 static inline int nohz_kick_needed(struct rq *rq)
7280 {
7281         unsigned long now = jiffies;
7282         struct sched_domain *sd;
7283         struct sched_group_capacity *sgc;
7284         int nr_busy, cpu = rq->cpu;
7285 
7286         if (unlikely(rq->idle_balance))
7287                 return 0;
7288 
7289        /*
7290         * We may be recently in ticked or tickless idle mode. At the first
7291         * busy tick after returning from idle, we will update the busy stats.
7292         */
7293         set_cpu_sd_state_busy();
7294         nohz_balance_exit_idle(cpu);
7295 
7296         /*
7297          * None are in tickless mode and hence no need for NOHZ idle load
7298          * balancing.
7299          */
7300         if (likely(!atomic_read(&nohz.nr_cpus)))
7301                 return 0;
7302 
7303         if (time_before(now, nohz.next_balance))
7304                 return 0;
7305 
7306         if (rq->nr_running >= 2)
7307                 goto need_kick;
7308 
7309         rcu_read_lock();
7310         sd = rcu_dereference(per_cpu(sd_busy, cpu));
7311 
7312         if (sd) {
7313                 sgc = sd->groups->sgc;
7314                 nr_busy = atomic_read(&sgc->nr_busy_cpus);
7315 
7316                 if (nr_busy > 1)
7317                         goto need_kick_unlock;
7318         }
7319 
7320         sd = rcu_dereference(per_cpu(sd_asym, cpu));
7321 
7322         if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7323                                   sched_domain_span(sd)) < cpu))
7324                 goto need_kick_unlock;
7325 
7326         rcu_read_unlock();
7327         return 0;
7328 
7329 need_kick_unlock:
7330         rcu_read_unlock();
7331 need_kick:
7332         return 1;
7333 }
7334 #else
7335 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7336 #endif
7337 
7338 /*
7339  * run_rebalance_domains is triggered when needed from the scheduler tick.
7340  * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7341  */
7342 static void run_rebalance_domains(struct softirq_action *h)
7343 {
7344         struct rq *this_rq = this_rq();
7345         enum cpu_idle_type idle = this_rq->idle_balance ?
7346                                                 CPU_IDLE : CPU_NOT_IDLE;
7347 
7348         rebalance_domains(this_rq, idle);
7349 
7350         /*
7351          * If this cpu has a pending nohz_balance_kick, then do the
7352          * balancing on behalf of the other idle cpus whose ticks are
7353          * stopped.
7354          */
7355         nohz_idle_balance(this_rq, idle);
7356 }
7357 
7358 /*
7359  * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7360  */
7361 void trigger_load_balance(struct rq *rq)
7362 {
7363         /* Don't need to rebalance while attached to NULL domain */
7364         if (unlikely(on_null_domain(rq)))
7365                 return;
7366 
7367         if (time_after_eq(jiffies, rq->next_balance))
7368                 raise_softirq(SCHED_SOFTIRQ);
7369 #ifdef CONFIG_NO_HZ_COMMON
7370         if (nohz_kick_needed(rq))
7371                 nohz_balancer_kick();
7372 #endif
7373 }
7374 
7375 static void rq_online_fair(struct rq *rq)
7376 {
7377         update_sysctl();
7378 
7379         update_runtime_enabled(rq);
7380 }
7381 
7382 static void rq_offline_fair(struct rq *rq)
7383 {
7384         update_sysctl();
7385 
7386         /* Ensure any throttled groups are reachable by pick_next_task */
7387         unthrottle_offline_cfs_rqs(rq);
7388 }
7389 
7390 #endif /* CONFIG_SMP */
7391 
7392 /*
7393  * scheduler tick hitting a task of our scheduling class:
7394  */
7395 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7396 {
7397         struct cfs_rq *cfs_rq;
7398         struct sched_entity *se = &curr->se;
7399 
7400         for_each_sched_entity(se) {
7401                 cfs_rq = cfs_rq_of(se);
7402                 entity_tick(cfs_rq, se, queued);
7403         }
7404 
7405         if (numabalancing_enabled)
7406                 task_tick_numa(rq, curr);
7407 
7408         update_rq_runnable_avg(rq, 1);
7409 }
7410 
7411 /*
7412  * called on fork with the child task as argument from the parent's context
7413  *  - child not yet on the tasklist
7414  *  - preemption disabled
7415  */
7416 static void task_fork_fair(struct task_struct *p)
7417 {
7418         struct cfs_rq *cfs_rq;
7419         struct sched_entity *se = &p->se, *curr;
7420         int this_cpu = smp_processor_id();
7421         struct rq *rq = this_rq();
7422         unsigned long flags;
7423 
7424         raw_spin_lock_irqsave(&rq->lock, flags);
7425 
7426         update_rq_clock(rq);
7427 
7428         cfs_rq = task_cfs_rq(current);
7429         curr = cfs_rq->curr;
7430 
7431         /*
7432          * Not only the cpu but also the task_group of the parent might have
7433          * been changed after parent->se.parent,cfs_rq were copied to
7434          * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7435          * of child point to valid ones.
7436          */
7437         rcu_read_lock();
7438         __set_task_cpu(p, this_cpu);
7439         rcu_read_unlock();
7440 
7441         update_curr(cfs_rq);
7442 
7443         if (curr)
7444                 se->vruntime = curr->vruntime;
7445         place_entity(cfs_rq, se, 1);
7446 
7447         if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
7448                 /*
7449                  * Upon rescheduling, sched_class::put_prev_task() will place
7450                  * 'current' within the tree based on its new key value.
7451                  */
7452                 swap(curr->vruntime, se->vruntime);
7453                 resched_curr(rq);
7454         }
7455 
7456         se->vruntime -= cfs_rq->min_vruntime;
7457 
7458         raw_spin_unlock_irqrestore(&rq->lock, flags);
7459 }
7460 
7461 /*
7462  * Priority of the task has changed. Check to see if we preempt
7463  * the current task.
7464  */
7465 static void
7466 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
7467 {
7468         if (!p->se.on_rq)
7469                 return;
7470 
7471         /*
7472          * Reschedule if we are currently running on this runqueue and
7473          * our priority decreased, or if we are not currently running on
7474          * this runqueue and our priority is higher than the current's
7475          */
7476         if (rq->curr == p) {
7477                 if (p->prio > oldprio)
7478                         resched_curr(rq);
7479         } else
7480                 check_preempt_curr(rq, p, 0);
7481 }
7482 
7483 static void switched_from_fair(struct rq *rq, struct task_struct *p)
7484 {
7485         struct sched_entity *se = &p->se;
7486         struct cfs_rq *cfs_rq = cfs_rq_of(se);
7487 
7488         /*
7489          * Ensure the task's vruntime is normalized, so that when it's
7490          * switched back to the fair class the enqueue_entity(.flags=0) will
7491          * do the right thing.
7492          *
7493          * If it's on_rq, then the dequeue_entity(.flags=0) will already
7494          * have normalized the vruntime, if it's !on_rq, then only when
7495          * the task is sleeping will it still have non-normalized vruntime.
7496          */
7497         if (!p->on_rq && p->state != TASK_RUNNING) {
7498                 /*
7499                  * Fix up our vruntime so that the current sleep doesn't
7500                  * cause 'unlimited' sleep bonus.
7501                  */
7502                 place_entity(cfs_rq, se, 0);
7503                 se->vruntime -= cfs_rq->min_vruntime;
7504         }
7505 
7506 #ifdef CONFIG_SMP
7507         /*
7508         * Remove our load from contribution when we leave sched_fair
7509         * and ensure we don't carry in an old decay_count if we
7510         * switch back.
7511         */
7512         if (se->avg.decay_count) {
7513                 __synchronize_entity_decay(se);
7514                 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
7515         }
7516 #endif
7517 }
7518 
7519 /*
7520  * We switched to the sched_fair class.
7521  */
7522 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7523 {
7524         struct sched_entity *se = &p->se;
7525 #ifdef CONFIG_FAIR_GROUP_SCHED
7526         /*
7527          * Since the real-depth could have been changed (only FAIR
7528          * class maintain depth value), reset depth properly.
7529          */
7530         se->depth = se->parent ? se->parent->depth + 1 : 0;
7531 #endif
7532         if (!se->on_rq)
7533                 return;
7534 
7535         /*
7536          * We were most likely switched from sched_rt, so
7537          * kick off the schedule if running, otherwise just see
7538          * if we can still preempt the current task.
7539          */
7540         if (rq->curr == p)
7541                 resched_curr(rq);
7542         else
7543                 check_preempt_curr(rq, p, 0);
7544 }
7545 
7546 /* Account for a task changing its policy or group.
7547  *
7548  * This routine is mostly called to set cfs_rq->curr field when a task
7549  * migrates between groups/classes.
7550  */
7551 static void set_curr_task_fair(struct rq *rq)
7552 {
7553         struct sched_entity *se = &rq->curr->se;
7554 
7555         for_each_sched_entity(se) {
7556                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7557 
7558                 set_next_entity(cfs_rq, se);
7559                 /* ensure bandwidth has been allocated on our new cfs_rq */
7560                 account_cfs_rq_runtime(cfs_rq, 0);
7561         }
7562 }
7563 
7564 void init_cfs_rq(struct cfs_rq *cfs_rq)
7565 {
7566         cfs_rq->tasks_timeline = RB_ROOT;
7567         cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7568 #ifndef CONFIG_64BIT
7569         cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7570 #endif
7571 #ifdef CONFIG_SMP
7572         atomic64_set(&cfs_rq->decay_counter, 1);
7573         atomic_long_set(&cfs_rq->removed_load, 0);
7574 #endif
7575 }
7576 
7577 #ifdef CONFIG_FAIR_GROUP_SCHED
7578 static void task_move_group_fair(struct task_struct *p, int on_rq)
7579 {
7580         struct sched_entity *se = &p->se;
7581         struct cfs_rq *cfs_rq;
7582 
7583         /*
7584          * If the task was not on the rq at the time of this cgroup movement
7585          * it must have been asleep, sleeping tasks keep their ->vruntime
7586          * absolute on their old rq until wakeup (needed for the fair sleeper
7587          * bonus in place_entity()).
7588          *
7589          * If it was on the rq, we've just 'preempted' it, which does convert
7590          * ->vruntime to a relative base.
7591          *
7592          * Make sure both cases convert their relative position when migrating
7593          * to another cgroup's rq. This does somewhat interfere with the
7594          * fair sleeper stuff for the first placement, but who cares.
7595          */
7596         /*
7597          * When !on_rq, vruntime of the task has usually NOT been normalized.
7598          * But there are some cases where it has already been normalized:
7599          *
7600          * - Moving a forked child which is waiting for being woken up by
7601          *   wake_up_new_task().
7602          * - Moving a task which has been woken up by try_to_wake_up() and
7603          *   waiting for actually being woken up by sched_ttwu_pending().
7604          *
7605          * To prevent boost or penalty in the new cfs_rq caused by delta
7606          * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7607          */
7608         if (!on_rq && (!se->sum_exec_runtime || p->state == TASK_WAKING))
7609                 on_rq = 1;
7610 
7611         if (!on_rq)
7612                 se->vruntime -= cfs_rq_of(se)->min_vruntime;
7613         set_task_rq(p, task_cpu(p));
7614         se->depth = se->parent ? se->parent->depth + 1 : 0;
7615         if (!on_rq) {
7616                 cfs_rq = cfs_rq_of(se);
7617                 se->vruntime += cfs_rq->min_vruntime;
7618 #ifdef CONFIG_SMP
7619                 /*
7620                  * migrate_task_rq_fair() will have removed our previous
7621                  * contribution, but we must synchronize for ongoing future
7622                  * decay.
7623                  */
7624                 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7625                 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
7626 #endif
7627         }
7628 }
7629 
7630 void free_fair_sched_group(struct task_group *tg)
7631 {
7632         int i;
7633 
7634         destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7635 
7636         for_each_possible_cpu(i) {
7637                 if (tg->cfs_rq)
7638                         kfree(tg->cfs_rq[i]);
7639                 if (tg->se)
7640                         kfree(tg->se[i]);
7641         }
7642 
7643         kfree(tg->cfs_rq);
7644         kfree(tg->se);
7645 }
7646 
7647 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7648 {
7649         struct cfs_rq *cfs_rq;
7650         struct sched_entity *se;
7651         int i;
7652 
7653         tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7654         if (!tg->cfs_rq)
7655                 goto err;
7656         tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7657         if (!tg->se)
7658                 goto err;
7659 
7660         tg->shares = NICE_0_LOAD;
7661 
7662         init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7663 
7664         for_each_possible_cpu(i) {
7665                 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7666                                       GFP_KERNEL, cpu_to_node(i));
7667                 if (!cfs_rq)
7668                         goto err;
7669 
7670                 se = kzalloc_node(sizeof(struct sched_entity),
7671                                   GFP_KERNEL, cpu_to_node(i));
7672                 if (!se)
7673                         goto err_free_rq;
7674 
7675                 init_cfs_rq(cfs_rq);
7676                 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7677         }
7678 
7679         return 1;
7680 
7681 err_free_rq:
7682         kfree(cfs_rq);
7683 err:
7684         return 0;
7685 }
7686 
7687 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7688 {
7689         struct rq *rq = cpu_rq(cpu);
7690         unsigned long flags;
7691 
7692         /*
7693         * Only empty task groups can be destroyed; so we can speculatively
7694         * check on_list without danger of it being re-added.
7695         */
7696         if (!tg->cfs_rq[cpu]->on_list)
7697                 return;
7698 
7699         raw_spin_lock_irqsave(&rq->lock, flags);
7700         list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7701         raw_spin_unlock_irqrestore(&rq->lock, flags);
7702 }
7703 
7704 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7705                         struct sched_entity *se, int cpu,
7706                         struct sched_entity *parent)
7707 {
7708         struct rq *rq = cpu_rq(cpu);
7709 
7710         cfs_rq->tg = tg;
7711         cfs_rq->rq = rq;
7712         init_cfs_rq_runtime(cfs_rq);
7713 
7714         tg->cfs_rq[cpu] = cfs_rq;
7715         tg->se[cpu] = se;
7716 
7717         /* se could be NULL for root_task_group */
7718         if (!se)
7719                 return;
7720 
7721         if (!parent) {
7722                 se->cfs_rq = &rq->cfs;
7723                 se->depth = 0;
7724         } else {
7725                 se->cfs_rq = parent->my_q;
7726                 se->depth = parent->depth + 1;
7727         }
7728 
7729         se->my_q = cfs_rq;
7730         /* guarantee group entities always have weight */
7731         update_load_set(&se->load, NICE_0_LOAD);
7732         se->parent = parent;
7733 }
7734 
7735 static DEFINE_MUTEX(shares_mutex);
7736 
7737 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7738 {
7739         int i;
7740         unsigned long flags;
7741 
7742         /*
7743          * We can't change the weight of the root cgroup.
7744          */
7745         if (!tg->se[0])
7746                 return -EINVAL;
7747 
7748         shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7749 
7750         mutex_lock(&shares_mutex);
7751         if (tg->shares == shares)
7752                 goto done;
7753 
7754         tg->shares = shares;
7755         for_each_possible_cpu(i) {
7756                 struct rq *rq = cpu_rq(i);
7757                 struct sched_entity *se;
7758 
7759                 se = tg->se[i];
7760                 /* Propagate contribution to hierarchy */
7761                 raw_spin_lock_irqsave(&rq->lock, flags);
7762 
7763                 /* Possible calls to update_curr() need rq clock */
7764                 update_rq_clock(rq);
7765                 for_each_sched_entity(se)
7766                         update_cfs_shares(group_cfs_rq(se));
7767                 raw_spin_unlock_irqrestore(&rq->lock, flags);
7768         }
7769 
7770 done:
7771         mutex_unlock(&shares_mutex);
7772         return 0;
7773 }
7774 #else /* CONFIG_FAIR_GROUP_SCHED */
7775 
7776 void free_fair_sched_group(struct task_group *tg) { }
7777 
7778 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7779 {
7780         return 1;
7781 }
7782 
7783 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7784 
7785 #endif /* CONFIG_FAIR_GROUP_SCHED */
7786 
7787 
7788 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
7789 {
7790         struct sched_entity *se = &task->se;
7791         unsigned int rr_interval = 0;
7792 
7793         /*
7794          * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7795          * idle runqueue:
7796          */
7797         if (rq->cfs.load.weight)
7798                 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
7799 
7800         return rr_interval;
7801 }
7802 
7803 /*
7804  * All the scheduling class methods:
7805  */
7806 const struct sched_class fair_sched_class = {
7807         .next                   = &idle_sched_class,
7808         .enqueue_task           = enqueue_task_fair,
7809         .dequeue_task           = dequeue_task_fair,
7810         .yield_task             = yield_task_fair,
7811         .yield_to_task          = yield_to_task_fair,
7812 
7813         .check_preempt_curr     = check_preempt_wakeup,
7814 
7815         .pick_next_task         = pick_next_task_fair,
7816         .put_prev_task          = put_prev_task_fair,
7817 
7818 #ifdef CONFIG_SMP
7819         .select_task_rq         = select_task_rq_fair,
7820         .migrate_task_rq        = migrate_task_rq_fair,
7821 
7822         .rq_online              = rq_online_fair,
7823         .rq_offline             = rq_offline_fair,
7824 
7825         .task_waking            = task_waking_fair,
7826 #endif
7827 
7828         .set_curr_task          = set_curr_task_fair,
7829         .task_tick              = task_tick_fair,
7830         .task_fork              = task_fork_fair,
7831 
7832         .prio_changed           = prio_changed_fair,
7833         .switched_from          = switched_from_fair,
7834         .switched_to            = switched_to_fair,
7835 
7836         .get_rr_interval        = get_rr_interval_fair,
7837 
7838 #ifdef CONFIG_FAIR_GROUP_SCHED
7839         .task_move_group        = task_move_group_fair,
7840 #endif
7841 };
7842 
7843 #ifdef CONFIG_SCHED_DEBUG
7844 void print_cfs_stats(struct seq_file *m, int cpu)
7845 {
7846         struct cfs_rq *cfs_rq;
7847 
7848         rcu_read_lock();
7849         for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7850                 print_cfs_rq(m, cpu, cfs_rq);
7851         rcu_read_unlock();
7852 }
7853 #endif
7854 
7855 __init void init_sched_fair_class(void)
7856 {
7857 #ifdef CONFIG_SMP
7858         open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7859 
7860 #ifdef CONFIG_NO_HZ_COMMON
7861         nohz.next_balance = jiffies;
7862         zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7863         cpu_notifier(sched_ilb_notifier, 0);
7864 #endif
7865 #endif /* SMP */
7866 
7867 }
7868 

This page was automatically generated by LXR 0.3.1 (source).  •  Linux is a registered trademark of Linus Torvalds  •  Contact us