Version:  2.0.40 2.2.26 2.4.37 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 3.18 3.19 4.0 4.1

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/cpuidle.h>
 27 #include <linux/slab.h>
 28 #include <linux/profile.h>
 29 #include <linux/interrupt.h>
 30 #include <linux/mempolicy.h>
 31 #include <linux/migrate.h>
 32 #include <linux/task_work.h>
 33 
 34 #include <trace/events/sched.h>
 35 
 36 #include "sched.h"
 37 
 38 /*
 39  * Targeted preemption latency for CPU-bound tasks:
 40  * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
 41  *
 42  * NOTE: this latency value is not the same as the concept of
 43  * 'timeslice length' - timeslices in CFS are of variable length
 44  * and have no persistent notion like in traditional, time-slice
 45  * based scheduling concepts.
 46  *
 47  * (to see the precise effective timeslice length of your workload,
 48  *  run vmstat and monitor the context-switches (cs) field)
 49  */
 50 unsigned int sysctl_sched_latency = 6000000ULL;
 51 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
 52 
 53 /*
 54  * The initial- and re-scaling of tunables is configurable
 55  * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
 56  *
 57  * Options are:
 58  * SCHED_TUNABLESCALING_NONE - unscaled, always *1
 59  * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
 60  * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
 61  */
 62 enum sched_tunable_scaling sysctl_sched_tunable_scaling
 63         = SCHED_TUNABLESCALING_LOG;
 64 
 65 /*
 66  * Minimal preemption granularity for CPU-bound tasks:
 67  * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
 68  */
 69 unsigned int sysctl_sched_min_granularity = 750000ULL;
 70 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
 71 
 72 /*
 73  * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
 74  */
 75 static unsigned int sched_nr_latency = 8;
 76 
 77 /*
 78  * After fork, child runs first. If set to 0 (default) then
 79  * parent will (try to) run first.
 80  */
 81 unsigned int sysctl_sched_child_runs_first __read_mostly;
 82 
 83 /*
 84  * SCHED_OTHER wake-up granularity.
 85  * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
 86  *
 87  * This option delays the preemption effects of decoupled workloads
 88  * and reduces their over-scheduling. Synchronous workloads will still
 89  * have immediate wakeup/sleep latencies.
 90  */
 91 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
 92 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
 93 
 94 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
 95 
 96 /*
 97  * The exponential sliding  window over which load is averaged for shares
 98  * distribution.
 99  * (default: 10msec)
100  */
101 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
102 
103 #ifdef CONFIG_CFS_BANDWIDTH
104 /*
105  * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
106  * each time a cfs_rq requests quota.
107  *
108  * Note: in the case that the slice exceeds the runtime remaining (either due
109  * to consumption or the quota being specified to be smaller than the slice)
110  * we will always only issue the remaining available time.
111  *
112  * default: 5 msec, units: microseconds
113   */
114 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
115 #endif
116 
117 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
118 {
119         lw->weight += inc;
120         lw->inv_weight = 0;
121 }
122 
123 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
124 {
125         lw->weight -= dec;
126         lw->inv_weight = 0;
127 }
128 
129 static inline void update_load_set(struct load_weight *lw, unsigned long w)
130 {
131         lw->weight = w;
132         lw->inv_weight = 0;
133 }
134 
135 /*
136  * Increase the granularity value when there are more CPUs,
137  * because with more CPUs the 'effective latency' as visible
138  * to users decreases. But the relationship is not linear,
139  * so pick a second-best guess by going with the log2 of the
140  * number of CPUs.
141  *
142  * This idea comes from the SD scheduler of Con Kolivas:
143  */
144 static int get_update_sysctl_factor(void)
145 {
146         unsigned int cpus = min_t(int, num_online_cpus(), 8);
147         unsigned int factor;
148 
149         switch (sysctl_sched_tunable_scaling) {
150         case SCHED_TUNABLESCALING_NONE:
151                 factor = 1;
152                 break;
153         case SCHED_TUNABLESCALING_LINEAR:
154                 factor = cpus;
155                 break;
156         case SCHED_TUNABLESCALING_LOG:
157         default:
158                 factor = 1 + ilog2(cpus);
159                 break;
160         }
161 
162         return factor;
163 }
164 
165 static void update_sysctl(void)
166 {
167         unsigned int factor = get_update_sysctl_factor();
168 
169 #define SET_SYSCTL(name) \
170         (sysctl_##name = (factor) * normalized_sysctl_##name)
171         SET_SYSCTL(sched_min_granularity);
172         SET_SYSCTL(sched_latency);
173         SET_SYSCTL(sched_wakeup_granularity);
174 #undef SET_SYSCTL
175 }
176 
177 void sched_init_granularity(void)
178 {
179         update_sysctl();
180 }
181 
182 #define WMULT_CONST     (~0U)
183 #define WMULT_SHIFT     32
184 
185 static void __update_inv_weight(struct load_weight *lw)
186 {
187         unsigned long w;
188 
189         if (likely(lw->inv_weight))
190                 return;
191 
192         w = scale_load_down(lw->weight);
193 
194         if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
195                 lw->inv_weight = 1;
196         else if (unlikely(!w))
197                 lw->inv_weight = WMULT_CONST;
198         else
199                 lw->inv_weight = WMULT_CONST / w;
200 }
201 
202 /*
203  * delta_exec * weight / lw.weight
204  *   OR
205  * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
206  *
207  * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
208  * we're guaranteed shift stays positive because inv_weight is guaranteed to
209  * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
210  *
211  * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
212  * weight/lw.weight <= 1, and therefore our shift will also be positive.
213  */
214 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
215 {
216         u64 fact = scale_load_down(weight);
217         int shift = WMULT_SHIFT;
218 
219         __update_inv_weight(lw);
220 
221         if (unlikely(fact >> 32)) {
222                 while (fact >> 32) {
223                         fact >>= 1;
224                         shift--;
225                 }
226         }
227 
228         /* hint to use a 32x32->64 mul */
229         fact = (u64)(u32)fact * lw->inv_weight;
230 
231         while (fact >> 32) {
232                 fact >>= 1;
233                 shift--;
234         }
235 
236         return mul_u64_u32_shr(delta_exec, fact, shift);
237 }
238 
239 
240 const struct sched_class fair_sched_class;
241 
242 /**************************************************************
243  * CFS operations on generic schedulable entities:
244  */
245 
246 #ifdef CONFIG_FAIR_GROUP_SCHED
247 
248 /* cpu runqueue to which this cfs_rq is attached */
249 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
250 {
251         return cfs_rq->rq;
252 }
253 
254 /* An entity is a task if it doesn't "own" a runqueue */
255 #define entity_is_task(se)      (!se->my_q)
256 
257 static inline struct task_struct *task_of(struct sched_entity *se)
258 {
259 #ifdef CONFIG_SCHED_DEBUG
260         WARN_ON_ONCE(!entity_is_task(se));
261 #endif
262         return container_of(se, struct task_struct, se);
263 }
264 
265 /* Walk up scheduling entities hierarchy */
266 #define for_each_sched_entity(se) \
267                 for (; se; se = se->parent)
268 
269 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
270 {
271         return p->se.cfs_rq;
272 }
273 
274 /* runqueue on which this entity is (to be) queued */
275 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
276 {
277         return se->cfs_rq;
278 }
279 
280 /* runqueue "owned" by this group */
281 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
282 {
283         return grp->my_q;
284 }
285 
286 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
287                                        int force_update);
288 
289 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
290 {
291         if (!cfs_rq->on_list) {
292                 /*
293                  * Ensure we either appear before our parent (if already
294                  * enqueued) or force our parent to appear after us when it is
295                  * enqueued.  The fact that we always enqueue bottom-up
296                  * reduces this to two cases.
297                  */
298                 if (cfs_rq->tg->parent &&
299                     cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
300                         list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
301                                 &rq_of(cfs_rq)->leaf_cfs_rq_list);
302                 } else {
303                         list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
304                                 &rq_of(cfs_rq)->leaf_cfs_rq_list);
305                 }
306 
307                 cfs_rq->on_list = 1;
308                 /* We should have no load, but we need to update last_decay. */
309                 update_cfs_rq_blocked_load(cfs_rq, 0);
310         }
311 }
312 
313 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
314 {
315         if (cfs_rq->on_list) {
316                 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
317                 cfs_rq->on_list = 0;
318         }
319 }
320 
321 /* Iterate thr' all leaf cfs_rq's on a runqueue */
322 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
323         list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
324 
325 /* Do the two (enqueued) entities belong to the same group ? */
326 static inline struct cfs_rq *
327 is_same_group(struct sched_entity *se, struct sched_entity *pse)
328 {
329         if (se->cfs_rq == pse->cfs_rq)
330                 return se->cfs_rq;
331 
332         return NULL;
333 }
334 
335 static inline struct sched_entity *parent_entity(struct sched_entity *se)
336 {
337         return se->parent;
338 }
339 
340 static void
341 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
342 {
343         int se_depth, pse_depth;
344 
345         /*
346          * preemption test can be made between sibling entities who are in the
347          * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
348          * both tasks until we find their ancestors who are siblings of common
349          * parent.
350          */
351 
352         /* First walk up until both entities are at same depth */
353         se_depth = (*se)->depth;
354         pse_depth = (*pse)->depth;
355 
356         while (se_depth > pse_depth) {
357                 se_depth--;
358                 *se = parent_entity(*se);
359         }
360 
361         while (pse_depth > se_depth) {
362                 pse_depth--;
363                 *pse = parent_entity(*pse);
364         }
365 
366         while (!is_same_group(*se, *pse)) {
367                 *se = parent_entity(*se);
368                 *pse = parent_entity(*pse);
369         }
370 }
371 
372 #else   /* !CONFIG_FAIR_GROUP_SCHED */
373 
374 static inline struct task_struct *task_of(struct sched_entity *se)
375 {
376         return container_of(se, struct task_struct, se);
377 }
378 
379 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
380 {
381         return container_of(cfs_rq, struct rq, cfs);
382 }
383 
384 #define entity_is_task(se)      1
385 
386 #define for_each_sched_entity(se) \
387                 for (; se; se = NULL)
388 
389 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
390 {
391         return &task_rq(p)->cfs;
392 }
393 
394 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
395 {
396         struct task_struct *p = task_of(se);
397         struct rq *rq = task_rq(p);
398 
399         return &rq->cfs;
400 }
401 
402 /* runqueue "owned" by this group */
403 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
404 {
405         return NULL;
406 }
407 
408 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
409 {
410 }
411 
412 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
413 {
414 }
415 
416 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
417                 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
418 
419 static inline struct sched_entity *parent_entity(struct sched_entity *se)
420 {
421         return NULL;
422 }
423 
424 static inline void
425 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
426 {
427 }
428 
429 #endif  /* CONFIG_FAIR_GROUP_SCHED */
430 
431 static __always_inline
432 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
433 
434 /**************************************************************
435  * Scheduling class tree data structure manipulation methods:
436  */
437 
438 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
439 {
440         s64 delta = (s64)(vruntime - max_vruntime);
441         if (delta > 0)
442                 max_vruntime = vruntime;
443 
444         return max_vruntime;
445 }
446 
447 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
448 {
449         s64 delta = (s64)(vruntime - min_vruntime);
450         if (delta < 0)
451                 min_vruntime = vruntime;
452 
453         return min_vruntime;
454 }
455 
456 static inline int entity_before(struct sched_entity *a,
457                                 struct sched_entity *b)
458 {
459         return (s64)(a->vruntime - b->vruntime) < 0;
460 }
461 
462 static void update_min_vruntime(struct cfs_rq *cfs_rq)
463 {
464         u64 vruntime = cfs_rq->min_vruntime;
465 
466         if (cfs_rq->curr)
467                 vruntime = cfs_rq->curr->vruntime;
468 
469         if (cfs_rq->rb_leftmost) {
470                 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
471                                                    struct sched_entity,
472                                                    run_node);
473 
474                 if (!cfs_rq->curr)
475                         vruntime = se->vruntime;
476                 else
477                         vruntime = min_vruntime(vruntime, se->vruntime);
478         }
479 
480         /* ensure we never gain time by being placed backwards. */
481         cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
482 #ifndef CONFIG_64BIT
483         smp_wmb();
484         cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
485 #endif
486 }
487 
488 /*
489  * Enqueue an entity into the rb-tree:
490  */
491 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
492 {
493         struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
494         struct rb_node *parent = NULL;
495         struct sched_entity *entry;
496         int leftmost = 1;
497 
498         /*
499          * Find the right place in the rbtree:
500          */
501         while (*link) {
502                 parent = *link;
503                 entry = rb_entry(parent, struct sched_entity, run_node);
504                 /*
505                  * We dont care about collisions. Nodes with
506                  * the same key stay together.
507                  */
508                 if (entity_before(se, entry)) {
509                         link = &parent->rb_left;
510                 } else {
511                         link = &parent->rb_right;
512                         leftmost = 0;
513                 }
514         }
515 
516         /*
517          * Maintain a cache of leftmost tree entries (it is frequently
518          * used):
519          */
520         if (leftmost)
521                 cfs_rq->rb_leftmost = &se->run_node;
522 
523         rb_link_node(&se->run_node, parent, link);
524         rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
525 }
526 
527 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
528 {
529         if (cfs_rq->rb_leftmost == &se->run_node) {
530                 struct rb_node *next_node;
531 
532                 next_node = rb_next(&se->run_node);
533                 cfs_rq->rb_leftmost = next_node;
534         }
535 
536         rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
537 }
538 
539 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
540 {
541         struct rb_node *left = cfs_rq->rb_leftmost;
542 
543         if (!left)
544                 return NULL;
545 
546         return rb_entry(left, struct sched_entity, run_node);
547 }
548 
549 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
550 {
551         struct rb_node *next = rb_next(&se->run_node);
552 
553         if (!next)
554                 return NULL;
555 
556         return rb_entry(next, struct sched_entity, run_node);
557 }
558 
559 #ifdef CONFIG_SCHED_DEBUG
560 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
561 {
562         struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
563 
564         if (!last)
565                 return NULL;
566 
567         return rb_entry(last, struct sched_entity, run_node);
568 }
569 
570 /**************************************************************
571  * Scheduling class statistics methods:
572  */
573 
574 int sched_proc_update_handler(struct ctl_table *table, int write,
575                 void __user *buffer, size_t *lenp,
576                 loff_t *ppos)
577 {
578         int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
579         int factor = get_update_sysctl_factor();
580 
581         if (ret || !write)
582                 return ret;
583 
584         sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
585                                         sysctl_sched_min_granularity);
586 
587 #define WRT_SYSCTL(name) \
588         (normalized_sysctl_##name = sysctl_##name / (factor))
589         WRT_SYSCTL(sched_min_granularity);
590         WRT_SYSCTL(sched_latency);
591         WRT_SYSCTL(sched_wakeup_granularity);
592 #undef WRT_SYSCTL
593 
594         return 0;
595 }
596 #endif
597 
598 /*
599  * delta /= w
600  */
601 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
602 {
603         if (unlikely(se->load.weight != NICE_0_LOAD))
604                 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
605 
606         return delta;
607 }
608 
609 /*
610  * The idea is to set a period in which each task runs once.
611  *
612  * When there are too many tasks (sched_nr_latency) we have to stretch
613  * this period because otherwise the slices get too small.
614  *
615  * p = (nr <= nl) ? l : l*nr/nl
616  */
617 static u64 __sched_period(unsigned long nr_running)
618 {
619         u64 period = sysctl_sched_latency;
620         unsigned long nr_latency = sched_nr_latency;
621 
622         if (unlikely(nr_running > nr_latency)) {
623                 period = sysctl_sched_min_granularity;
624                 period *= nr_running;
625         }
626 
627         return period;
628 }
629 
630 /*
631  * We calculate the wall-time slice from the period by taking a part
632  * proportional to the weight.
633  *
634  * s = p*P[w/rw]
635  */
636 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
637 {
638         u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
639 
640         for_each_sched_entity(se) {
641                 struct load_weight *load;
642                 struct load_weight lw;
643 
644                 cfs_rq = cfs_rq_of(se);
645                 load = &cfs_rq->load;
646 
647                 if (unlikely(!se->on_rq)) {
648                         lw = cfs_rq->load;
649 
650                         update_load_add(&lw, se->load.weight);
651                         load = &lw;
652                 }
653                 slice = __calc_delta(slice, se->load.weight, load);
654         }
655         return slice;
656 }
657 
658 /*
659  * We calculate the vruntime slice of a to-be-inserted task.
660  *
661  * vs = s/w
662  */
663 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
664 {
665         return calc_delta_fair(sched_slice(cfs_rq, se), se);
666 }
667 
668 #ifdef CONFIG_SMP
669 static int select_idle_sibling(struct task_struct *p, int cpu);
670 static unsigned long task_h_load(struct task_struct *p);
671 
672 static inline void __update_task_entity_contrib(struct sched_entity *se);
673 static inline void __update_task_entity_utilization(struct sched_entity *se);
674 
675 /* Give new task start runnable values to heavy its load in infant time */
676 void init_task_runnable_average(struct task_struct *p)
677 {
678         u32 slice;
679 
680         slice = sched_slice(task_cfs_rq(p), &p->se) >> 10;
681         p->se.avg.runnable_avg_sum = p->se.avg.running_avg_sum = slice;
682         p->se.avg.avg_period = slice;
683         __update_task_entity_contrib(&p->se);
684         __update_task_entity_utilization(&p->se);
685 }
686 #else
687 void init_task_runnable_average(struct task_struct *p)
688 {
689 }
690 #endif
691 
692 /*
693  * Update the current task's runtime statistics.
694  */
695 static void update_curr(struct cfs_rq *cfs_rq)
696 {
697         struct sched_entity *curr = cfs_rq->curr;
698         u64 now = rq_clock_task(rq_of(cfs_rq));
699         u64 delta_exec;
700 
701         if (unlikely(!curr))
702                 return;
703 
704         delta_exec = now - curr->exec_start;
705         if (unlikely((s64)delta_exec <= 0))
706                 return;
707 
708         curr->exec_start = now;
709 
710         schedstat_set(curr->statistics.exec_max,
711                       max(delta_exec, curr->statistics.exec_max));
712 
713         curr->sum_exec_runtime += delta_exec;
714         schedstat_add(cfs_rq, exec_clock, delta_exec);
715 
716         curr->vruntime += calc_delta_fair(delta_exec, curr);
717         update_min_vruntime(cfs_rq);
718 
719         if (entity_is_task(curr)) {
720                 struct task_struct *curtask = task_of(curr);
721 
722                 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
723                 cpuacct_charge(curtask, delta_exec);
724                 account_group_exec_runtime(curtask, delta_exec);
725         }
726 
727         account_cfs_rq_runtime(cfs_rq, delta_exec);
728 }
729 
730 static void update_curr_fair(struct rq *rq)
731 {
732         update_curr(cfs_rq_of(&rq->curr->se));
733 }
734 
735 static inline void
736 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
737 {
738         schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
739 }
740 
741 /*
742  * Task is being enqueued - update stats:
743  */
744 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
745 {
746         /*
747          * Are we enqueueing a waiting task? (for current tasks
748          * a dequeue/enqueue event is a NOP)
749          */
750         if (se != cfs_rq->curr)
751                 update_stats_wait_start(cfs_rq, se);
752 }
753 
754 static void
755 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
756 {
757         schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
758                         rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
759         schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
760         schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
761                         rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
762 #ifdef CONFIG_SCHEDSTATS
763         if (entity_is_task(se)) {
764                 trace_sched_stat_wait(task_of(se),
765                         rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
766         }
767 #endif
768         schedstat_set(se->statistics.wait_start, 0);
769 }
770 
771 static inline void
772 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
773 {
774         /*
775          * Mark the end of the wait period if dequeueing a
776          * waiting task:
777          */
778         if (se != cfs_rq->curr)
779                 update_stats_wait_end(cfs_rq, se);
780 }
781 
782 /*
783  * We are picking a new current task - update its stats:
784  */
785 static inline void
786 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
787 {
788         /*
789          * We are starting a new run period:
790          */
791         se->exec_start = rq_clock_task(rq_of(cfs_rq));
792 }
793 
794 /**************************************************
795  * Scheduling class queueing methods:
796  */
797 
798 #ifdef CONFIG_NUMA_BALANCING
799 /*
800  * Approximate time to scan a full NUMA task in ms. The task scan period is
801  * calculated based on the tasks virtual memory size and
802  * numa_balancing_scan_size.
803  */
804 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
805 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
806 
807 /* Portion of address space to scan in MB */
808 unsigned int sysctl_numa_balancing_scan_size = 256;
809 
810 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
811 unsigned int sysctl_numa_balancing_scan_delay = 1000;
812 
813 static unsigned int task_nr_scan_windows(struct task_struct *p)
814 {
815         unsigned long rss = 0;
816         unsigned long nr_scan_pages;
817 
818         /*
819          * Calculations based on RSS as non-present and empty pages are skipped
820          * by the PTE scanner and NUMA hinting faults should be trapped based
821          * on resident pages
822          */
823         nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
824         rss = get_mm_rss(p->mm);
825         if (!rss)
826                 rss = nr_scan_pages;
827 
828         rss = round_up(rss, nr_scan_pages);
829         return rss / nr_scan_pages;
830 }
831 
832 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
833 #define MAX_SCAN_WINDOW 2560
834 
835 static unsigned int task_scan_min(struct task_struct *p)
836 {
837         unsigned int scan_size = ACCESS_ONCE(sysctl_numa_balancing_scan_size);
838         unsigned int scan, floor;
839         unsigned int windows = 1;
840 
841         if (scan_size < MAX_SCAN_WINDOW)
842                 windows = MAX_SCAN_WINDOW / scan_size;
843         floor = 1000 / windows;
844 
845         scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
846         return max_t(unsigned int, floor, scan);
847 }
848 
849 static unsigned int task_scan_max(struct task_struct *p)
850 {
851         unsigned int smin = task_scan_min(p);
852         unsigned int smax;
853 
854         /* Watch for min being lower than max due to floor calculations */
855         smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
856         return max(smin, smax);
857 }
858 
859 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
860 {
861         rq->nr_numa_running += (p->numa_preferred_nid != -1);
862         rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
863 }
864 
865 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
866 {
867         rq->nr_numa_running -= (p->numa_preferred_nid != -1);
868         rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
869 }
870 
871 struct numa_group {
872         atomic_t refcount;
873 
874         spinlock_t lock; /* nr_tasks, tasks */
875         int nr_tasks;
876         pid_t gid;
877 
878         struct rcu_head rcu;
879         nodemask_t active_nodes;
880         unsigned long total_faults;
881         /*
882          * Faults_cpu is used to decide whether memory should move
883          * towards the CPU. As a consequence, these stats are weighted
884          * more by CPU use than by memory faults.
885          */
886         unsigned long *faults_cpu;
887         unsigned long faults[0];
888 };
889 
890 /* Shared or private faults. */
891 #define NR_NUMA_HINT_FAULT_TYPES 2
892 
893 /* Memory and CPU locality */
894 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
895 
896 /* Averaged statistics, and temporary buffers. */
897 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
898 
899 pid_t task_numa_group_id(struct task_struct *p)
900 {
901         return p->numa_group ? p->numa_group->gid : 0;
902 }
903 
904 /*
905  * The averaged statistics, shared & private, memory & cpu,
906  * occupy the first half of the array. The second half of the
907  * array is for current counters, which are averaged into the
908  * first set by task_numa_placement.
909  */
910 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
911 {
912         return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
913 }
914 
915 static inline unsigned long task_faults(struct task_struct *p, int nid)
916 {
917         if (!p->numa_faults)
918                 return 0;
919 
920         return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
921                 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
922 }
923 
924 static inline unsigned long group_faults(struct task_struct *p, int nid)
925 {
926         if (!p->numa_group)
927                 return 0;
928 
929         return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
930                 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
931 }
932 
933 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
934 {
935         return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
936                 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
937 }
938 
939 /* Handle placement on systems where not all nodes are directly connected. */
940 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
941                                         int maxdist, bool task)
942 {
943         unsigned long score = 0;
944         int node;
945 
946         /*
947          * All nodes are directly connected, and the same distance
948          * from each other. No need for fancy placement algorithms.
949          */
950         if (sched_numa_topology_type == NUMA_DIRECT)
951                 return 0;
952 
953         /*
954          * This code is called for each node, introducing N^2 complexity,
955          * which should be ok given the number of nodes rarely exceeds 8.
956          */
957         for_each_online_node(node) {
958                 unsigned long faults;
959                 int dist = node_distance(nid, node);
960 
961                 /*
962                  * The furthest away nodes in the system are not interesting
963                  * for placement; nid was already counted.
964                  */
965                 if (dist == sched_max_numa_distance || node == nid)
966                         continue;
967 
968                 /*
969                  * On systems with a backplane NUMA topology, compare groups
970                  * of nodes, and move tasks towards the group with the most
971                  * memory accesses. When comparing two nodes at distance
972                  * "hoplimit", only nodes closer by than "hoplimit" are part
973                  * of each group. Skip other nodes.
974                  */
975                 if (sched_numa_topology_type == NUMA_BACKPLANE &&
976                                         dist > maxdist)
977                         continue;
978 
979                 /* Add up the faults from nearby nodes. */
980                 if (task)
981                         faults = task_faults(p, node);
982                 else
983                         faults = group_faults(p, node);
984 
985                 /*
986                  * On systems with a glueless mesh NUMA topology, there are
987                  * no fixed "groups of nodes". Instead, nodes that are not
988                  * directly connected bounce traffic through intermediate
989                  * nodes; a numa_group can occupy any set of nodes.
990                  * The further away a node is, the less the faults count.
991                  * This seems to result in good task placement.
992                  */
993                 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
994                         faults *= (sched_max_numa_distance - dist);
995                         faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
996                 }
997 
998                 score += faults;
999         }
1000 
1001         return score;
1002 }
1003 
1004 /*
1005  * These return the fraction of accesses done by a particular task, or
1006  * task group, on a particular numa node.  The group weight is given a
1007  * larger multiplier, in order to group tasks together that are almost
1008  * evenly spread out between numa nodes.
1009  */
1010 static inline unsigned long task_weight(struct task_struct *p, int nid,
1011                                         int dist)
1012 {
1013         unsigned long faults, total_faults;
1014 
1015         if (!p->numa_faults)
1016                 return 0;
1017 
1018         total_faults = p->total_numa_faults;
1019 
1020         if (!total_faults)
1021                 return 0;
1022 
1023         faults = task_faults(p, nid);
1024         faults += score_nearby_nodes(p, nid, dist, true);
1025 
1026         return 1000 * faults / total_faults;
1027 }
1028 
1029 static inline unsigned long group_weight(struct task_struct *p, int nid,
1030                                          int dist)
1031 {
1032         unsigned long faults, total_faults;
1033 
1034         if (!p->numa_group)
1035                 return 0;
1036 
1037         total_faults = p->numa_group->total_faults;
1038 
1039         if (!total_faults)
1040                 return 0;
1041 
1042         faults = group_faults(p, nid);
1043         faults += score_nearby_nodes(p, nid, dist, false);
1044 
1045         return 1000 * faults / total_faults;
1046 }
1047 
1048 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1049                                 int src_nid, int dst_cpu)
1050 {
1051         struct numa_group *ng = p->numa_group;
1052         int dst_nid = cpu_to_node(dst_cpu);
1053         int last_cpupid, this_cpupid;
1054 
1055         this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1056 
1057         /*
1058          * Multi-stage node selection is used in conjunction with a periodic
1059          * migration fault to build a temporal task<->page relation. By using
1060          * a two-stage filter we remove short/unlikely relations.
1061          *
1062          * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1063          * a task's usage of a particular page (n_p) per total usage of this
1064          * page (n_t) (in a given time-span) to a probability.
1065          *
1066          * Our periodic faults will sample this probability and getting the
1067          * same result twice in a row, given these samples are fully
1068          * independent, is then given by P(n)^2, provided our sample period
1069          * is sufficiently short compared to the usage pattern.
1070          *
1071          * This quadric squishes small probabilities, making it less likely we
1072          * act on an unlikely task<->page relation.
1073          */
1074         last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1075         if (!cpupid_pid_unset(last_cpupid) &&
1076                                 cpupid_to_nid(last_cpupid) != dst_nid)
1077                 return false;
1078 
1079         /* Always allow migrate on private faults */
1080         if (cpupid_match_pid(p, last_cpupid))
1081                 return true;
1082 
1083         /* A shared fault, but p->numa_group has not been set up yet. */
1084         if (!ng)
1085                 return true;
1086 
1087         /*
1088          * Do not migrate if the destination is not a node that
1089          * is actively used by this numa group.
1090          */
1091         if (!node_isset(dst_nid, ng->active_nodes))
1092                 return false;
1093 
1094         /*
1095          * Source is a node that is not actively used by this
1096          * numa group, while the destination is. Migrate.
1097          */
1098         if (!node_isset(src_nid, ng->active_nodes))
1099                 return true;
1100 
1101         /*
1102          * Both source and destination are nodes in active
1103          * use by this numa group. Maximize memory bandwidth
1104          * by migrating from more heavily used groups, to less
1105          * heavily used ones, spreading the load around.
1106          * Use a 1/4 hysteresis to avoid spurious page movement.
1107          */
1108         return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1109 }
1110 
1111 static unsigned long weighted_cpuload(const int cpu);
1112 static unsigned long source_load(int cpu, int type);
1113 static unsigned long target_load(int cpu, int type);
1114 static unsigned long capacity_of(int cpu);
1115 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1116 
1117 /* Cached statistics for all CPUs within a node */
1118 struct numa_stats {
1119         unsigned long nr_running;
1120         unsigned long load;
1121 
1122         /* Total compute capacity of CPUs on a node */
1123         unsigned long compute_capacity;
1124 
1125         /* Approximate capacity in terms of runnable tasks on a node */
1126         unsigned long task_capacity;
1127         int has_free_capacity;
1128 };
1129 
1130 /*
1131  * XXX borrowed from update_sg_lb_stats
1132  */
1133 static void update_numa_stats(struct numa_stats *ns, int nid)
1134 {
1135         int smt, cpu, cpus = 0;
1136         unsigned long capacity;
1137 
1138         memset(ns, 0, sizeof(*ns));
1139         for_each_cpu(cpu, cpumask_of_node(nid)) {
1140                 struct rq *rq = cpu_rq(cpu);
1141 
1142                 ns->nr_running += rq->nr_running;
1143                 ns->load += weighted_cpuload(cpu);
1144                 ns->compute_capacity += capacity_of(cpu);
1145 
1146                 cpus++;
1147         }
1148 
1149         /*
1150          * If we raced with hotplug and there are no CPUs left in our mask
1151          * the @ns structure is NULL'ed and task_numa_compare() will
1152          * not find this node attractive.
1153          *
1154          * We'll either bail at !has_free_capacity, or we'll detect a huge
1155          * imbalance and bail there.
1156          */
1157         if (!cpus)
1158                 return;
1159 
1160         /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1161         smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1162         capacity = cpus / smt; /* cores */
1163 
1164         ns->task_capacity = min_t(unsigned, capacity,
1165                 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1166         ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1167 }
1168 
1169 struct task_numa_env {
1170         struct task_struct *p;
1171 
1172         int src_cpu, src_nid;
1173         int dst_cpu, dst_nid;
1174 
1175         struct numa_stats src_stats, dst_stats;
1176 
1177         int imbalance_pct;
1178         int dist;
1179 
1180         struct task_struct *best_task;
1181         long best_imp;
1182         int best_cpu;
1183 };
1184 
1185 static void task_numa_assign(struct task_numa_env *env,
1186                              struct task_struct *p, long imp)
1187 {
1188         if (env->best_task)
1189                 put_task_struct(env->best_task);
1190         if (p)
1191                 get_task_struct(p);
1192 
1193         env->best_task = p;
1194         env->best_imp = imp;
1195         env->best_cpu = env->dst_cpu;
1196 }
1197 
1198 static bool load_too_imbalanced(long src_load, long dst_load,
1199                                 struct task_numa_env *env)
1200 {
1201         long src_capacity, dst_capacity;
1202         long orig_src_load;
1203         long load_a, load_b;
1204         long moved_load;
1205         long imb;
1206 
1207         /*
1208          * The load is corrected for the CPU capacity available on each node.
1209          *
1210          * src_load        dst_load
1211          * ------------ vs ---------
1212          * src_capacity    dst_capacity
1213          */
1214         src_capacity = env->src_stats.compute_capacity;
1215         dst_capacity = env->dst_stats.compute_capacity;
1216 
1217         /* We care about the slope of the imbalance, not the direction. */
1218         load_a = dst_load;
1219         load_b = src_load;
1220         if (load_a < load_b)
1221                 swap(load_a, load_b);
1222 
1223         /* Is the difference below the threshold? */
1224         imb = load_a * src_capacity * 100 -
1225                 load_b * dst_capacity * env->imbalance_pct;
1226         if (imb <= 0)
1227                 return false;
1228 
1229         /*
1230          * The imbalance is above the allowed threshold.
1231          * Allow a move that brings us closer to a balanced situation,
1232          * without moving things past the point of balance.
1233          */
1234         orig_src_load = env->src_stats.load;
1235 
1236         /*
1237          * In a task swap, there will be one load moving from src to dst,
1238          * and another moving back. This is the net sum of both moves.
1239          * A simple task move will always have a positive value.
1240          * Allow the move if it brings the system closer to a balanced
1241          * situation, without crossing over the balance point.
1242          */
1243         moved_load = orig_src_load - src_load;
1244 
1245         if (moved_load > 0)
1246                 /* Moving src -> dst. Did we overshoot balance? */
1247                 return src_load * dst_capacity < dst_load * src_capacity;
1248         else
1249                 /* Moving dst -> src. Did we overshoot balance? */
1250                 return dst_load * src_capacity < src_load * dst_capacity;
1251 }
1252 
1253 /*
1254  * This checks if the overall compute and NUMA accesses of the system would
1255  * be improved if the source tasks was migrated to the target dst_cpu taking
1256  * into account that it might be best if task running on the dst_cpu should
1257  * be exchanged with the source task
1258  */
1259 static void task_numa_compare(struct task_numa_env *env,
1260                               long taskimp, long groupimp)
1261 {
1262         struct rq *src_rq = cpu_rq(env->src_cpu);
1263         struct rq *dst_rq = cpu_rq(env->dst_cpu);
1264         struct task_struct *cur;
1265         long src_load, dst_load;
1266         long load;
1267         long imp = env->p->numa_group ? groupimp : taskimp;
1268         long moveimp = imp;
1269         int dist = env->dist;
1270 
1271         rcu_read_lock();
1272 
1273         raw_spin_lock_irq(&dst_rq->lock);
1274         cur = dst_rq->curr;
1275         /*
1276          * No need to move the exiting task, and this ensures that ->curr
1277          * wasn't reaped and thus get_task_struct() in task_numa_assign()
1278          * is safe under RCU read lock.
1279          * Note that rcu_read_lock() itself can't protect from the final
1280          * put_task_struct() after the last schedule().
1281          */
1282         if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1283                 cur = NULL;
1284         raw_spin_unlock_irq(&dst_rq->lock);
1285 
1286         /*
1287          * Because we have preemption enabled we can get migrated around and
1288          * end try selecting ourselves (current == env->p) as a swap candidate.
1289          */
1290         if (cur == env->p)
1291                 goto unlock;
1292 
1293         /*
1294          * "imp" is the fault differential for the source task between the
1295          * source and destination node. Calculate the total differential for
1296          * the source task and potential destination task. The more negative
1297          * the value is, the more rmeote accesses that would be expected to
1298          * be incurred if the tasks were swapped.
1299          */
1300         if (cur) {
1301                 /* Skip this swap candidate if cannot move to the source cpu */
1302                 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1303                         goto unlock;
1304 
1305                 /*
1306                  * If dst and source tasks are in the same NUMA group, or not
1307                  * in any group then look only at task weights.
1308                  */
1309                 if (cur->numa_group == env->p->numa_group) {
1310                         imp = taskimp + task_weight(cur, env->src_nid, dist) -
1311                               task_weight(cur, env->dst_nid, dist);
1312                         /*
1313                          * Add some hysteresis to prevent swapping the
1314                          * tasks within a group over tiny differences.
1315                          */
1316                         if (cur->numa_group)
1317                                 imp -= imp/16;
1318                 } else {
1319                         /*
1320                          * Compare the group weights. If a task is all by
1321                          * itself (not part of a group), use the task weight
1322                          * instead.
1323                          */
1324                         if (cur->numa_group)
1325                                 imp += group_weight(cur, env->src_nid, dist) -
1326                                        group_weight(cur, env->dst_nid, dist);
1327                         else
1328                                 imp += task_weight(cur, env->src_nid, dist) -
1329                                        task_weight(cur, env->dst_nid, dist);
1330                 }
1331         }
1332 
1333         if (imp <= env->best_imp && moveimp <= env->best_imp)
1334                 goto unlock;
1335 
1336         if (!cur) {
1337                 /* Is there capacity at our destination? */
1338                 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1339                     !env->dst_stats.has_free_capacity)
1340                         goto unlock;
1341 
1342                 goto balance;
1343         }
1344 
1345         /* Balance doesn't matter much if we're running a task per cpu */
1346         if (imp > env->best_imp && src_rq->nr_running == 1 &&
1347                         dst_rq->nr_running == 1)
1348                 goto assign;
1349 
1350         /*
1351          * In the overloaded case, try and keep the load balanced.
1352          */
1353 balance:
1354         load = task_h_load(env->p);
1355         dst_load = env->dst_stats.load + load;
1356         src_load = env->src_stats.load - load;
1357 
1358         if (moveimp > imp && moveimp > env->best_imp) {
1359                 /*
1360                  * If the improvement from just moving env->p direction is
1361                  * better than swapping tasks around, check if a move is
1362                  * possible. Store a slightly smaller score than moveimp,
1363                  * so an actually idle CPU will win.
1364                  */
1365                 if (!load_too_imbalanced(src_load, dst_load, env)) {
1366                         imp = moveimp - 1;
1367                         cur = NULL;
1368                         goto assign;
1369                 }
1370         }
1371 
1372         if (imp <= env->best_imp)
1373                 goto unlock;
1374 
1375         if (cur) {
1376                 load = task_h_load(cur);
1377                 dst_load -= load;
1378                 src_load += load;
1379         }
1380 
1381         if (load_too_imbalanced(src_load, dst_load, env))
1382                 goto unlock;
1383 
1384         /*
1385          * One idle CPU per node is evaluated for a task numa move.
1386          * Call select_idle_sibling to maybe find a better one.
1387          */
1388         if (!cur)
1389                 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1390 
1391 assign:
1392         task_numa_assign(env, cur, imp);
1393 unlock:
1394         rcu_read_unlock();
1395 }
1396 
1397 static void task_numa_find_cpu(struct task_numa_env *env,
1398                                 long taskimp, long groupimp)
1399 {
1400         int cpu;
1401 
1402         for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1403                 /* Skip this CPU if the source task cannot migrate */
1404                 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1405                         continue;
1406 
1407                 env->dst_cpu = cpu;
1408                 task_numa_compare(env, taskimp, groupimp);
1409         }
1410 }
1411 
1412 static int task_numa_migrate(struct task_struct *p)
1413 {
1414         struct task_numa_env env = {
1415                 .p = p,
1416 
1417                 .src_cpu = task_cpu(p),
1418                 .src_nid = task_node(p),
1419 
1420                 .imbalance_pct = 112,
1421 
1422                 .best_task = NULL,
1423                 .best_imp = 0,
1424                 .best_cpu = -1
1425         };
1426         struct sched_domain *sd;
1427         unsigned long taskweight, groupweight;
1428         int nid, ret, dist;
1429         long taskimp, groupimp;
1430 
1431         /*
1432          * Pick the lowest SD_NUMA domain, as that would have the smallest
1433          * imbalance and would be the first to start moving tasks about.
1434          *
1435          * And we want to avoid any moving of tasks about, as that would create
1436          * random movement of tasks -- counter the numa conditions we're trying
1437          * to satisfy here.
1438          */
1439         rcu_read_lock();
1440         sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1441         if (sd)
1442                 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1443         rcu_read_unlock();
1444 
1445         /*
1446          * Cpusets can break the scheduler domain tree into smaller
1447          * balance domains, some of which do not cross NUMA boundaries.
1448          * Tasks that are "trapped" in such domains cannot be migrated
1449          * elsewhere, so there is no point in (re)trying.
1450          */
1451         if (unlikely(!sd)) {
1452                 p->numa_preferred_nid = task_node(p);
1453                 return -EINVAL;
1454         }
1455 
1456         env.dst_nid = p->numa_preferred_nid;
1457         dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1458         taskweight = task_weight(p, env.src_nid, dist);
1459         groupweight = group_weight(p, env.src_nid, dist);
1460         update_numa_stats(&env.src_stats, env.src_nid);
1461         taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1462         groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1463         update_numa_stats(&env.dst_stats, env.dst_nid);
1464 
1465         /* Try to find a spot on the preferred nid. */
1466         task_numa_find_cpu(&env, taskimp, groupimp);
1467 
1468         /*
1469          * Look at other nodes in these cases:
1470          * - there is no space available on the preferred_nid
1471          * - the task is part of a numa_group that is interleaved across
1472          *   multiple NUMA nodes; in order to better consolidate the group,
1473          *   we need to check other locations.
1474          */
1475         if (env.best_cpu == -1 || (p->numa_group &&
1476                         nodes_weight(p->numa_group->active_nodes) > 1)) {
1477                 for_each_online_node(nid) {
1478                         if (nid == env.src_nid || nid == p->numa_preferred_nid)
1479                                 continue;
1480 
1481                         dist = node_distance(env.src_nid, env.dst_nid);
1482                         if (sched_numa_topology_type == NUMA_BACKPLANE &&
1483                                                 dist != env.dist) {
1484                                 taskweight = task_weight(p, env.src_nid, dist);
1485                                 groupweight = group_weight(p, env.src_nid, dist);
1486                         }
1487 
1488                         /* Only consider nodes where both task and groups benefit */
1489                         taskimp = task_weight(p, nid, dist) - taskweight;
1490                         groupimp = group_weight(p, nid, dist) - groupweight;
1491                         if (taskimp < 0 && groupimp < 0)
1492                                 continue;
1493 
1494                         env.dist = dist;
1495                         env.dst_nid = nid;
1496                         update_numa_stats(&env.dst_stats, env.dst_nid);
1497                         task_numa_find_cpu(&env, taskimp, groupimp);
1498                 }
1499         }
1500 
1501         /*
1502          * If the task is part of a workload that spans multiple NUMA nodes,
1503          * and is migrating into one of the workload's active nodes, remember
1504          * this node as the task's preferred numa node, so the workload can
1505          * settle down.
1506          * A task that migrated to a second choice node will be better off
1507          * trying for a better one later. Do not set the preferred node here.
1508          */
1509         if (p->numa_group) {
1510                 if (env.best_cpu == -1)
1511                         nid = env.src_nid;
1512                 else
1513                         nid = env.dst_nid;
1514 
1515                 if (node_isset(nid, p->numa_group->active_nodes))
1516                         sched_setnuma(p, env.dst_nid);
1517         }
1518 
1519         /* No better CPU than the current one was found. */
1520         if (env.best_cpu == -1)
1521                 return -EAGAIN;
1522 
1523         /*
1524          * Reset the scan period if the task is being rescheduled on an
1525          * alternative node to recheck if the tasks is now properly placed.
1526          */
1527         p->numa_scan_period = task_scan_min(p);
1528 
1529         if (env.best_task == NULL) {
1530                 ret = migrate_task_to(p, env.best_cpu);
1531                 if (ret != 0)
1532                         trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1533                 return ret;
1534         }
1535 
1536         ret = migrate_swap(p, env.best_task);
1537         if (ret != 0)
1538                 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1539         put_task_struct(env.best_task);
1540         return ret;
1541 }
1542 
1543 /* Attempt to migrate a task to a CPU on the preferred node. */
1544 static void numa_migrate_preferred(struct task_struct *p)
1545 {
1546         unsigned long interval = HZ;
1547 
1548         /* This task has no NUMA fault statistics yet */
1549         if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1550                 return;
1551 
1552         /* Periodically retry migrating the task to the preferred node */
1553         interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1554         p->numa_migrate_retry = jiffies + interval;
1555 
1556         /* Success if task is already running on preferred CPU */
1557         if (task_node(p) == p->numa_preferred_nid)
1558                 return;
1559 
1560         /* Otherwise, try migrate to a CPU on the preferred node */
1561         task_numa_migrate(p);
1562 }
1563 
1564 /*
1565  * Find the nodes on which the workload is actively running. We do this by
1566  * tracking the nodes from which NUMA hinting faults are triggered. This can
1567  * be different from the set of nodes where the workload's memory is currently
1568  * located.
1569  *
1570  * The bitmask is used to make smarter decisions on when to do NUMA page
1571  * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1572  * are added when they cause over 6/16 of the maximum number of faults, but
1573  * only removed when they drop below 3/16.
1574  */
1575 static void update_numa_active_node_mask(struct numa_group *numa_group)
1576 {
1577         unsigned long faults, max_faults = 0;
1578         int nid;
1579 
1580         for_each_online_node(nid) {
1581                 faults = group_faults_cpu(numa_group, nid);
1582                 if (faults > max_faults)
1583                         max_faults = faults;
1584         }
1585 
1586         for_each_online_node(nid) {
1587                 faults = group_faults_cpu(numa_group, nid);
1588                 if (!node_isset(nid, numa_group->active_nodes)) {
1589                         if (faults > max_faults * 6 / 16)
1590                                 node_set(nid, numa_group->active_nodes);
1591                 } else if (faults < max_faults * 3 / 16)
1592                         node_clear(nid, numa_group->active_nodes);
1593         }
1594 }
1595 
1596 /*
1597  * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1598  * increments. The more local the fault statistics are, the higher the scan
1599  * period will be for the next scan window. If local/(local+remote) ratio is
1600  * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1601  * the scan period will decrease. Aim for 70% local accesses.
1602  */
1603 #define NUMA_PERIOD_SLOTS 10
1604 #define NUMA_PERIOD_THRESHOLD 7
1605 
1606 /*
1607  * Increase the scan period (slow down scanning) if the majority of
1608  * our memory is already on our local node, or if the majority of
1609  * the page accesses are shared with other processes.
1610  * Otherwise, decrease the scan period.
1611  */
1612 static void update_task_scan_period(struct task_struct *p,
1613                         unsigned long shared, unsigned long private)
1614 {
1615         unsigned int period_slot;
1616         int ratio;
1617         int diff;
1618 
1619         unsigned long remote = p->numa_faults_locality[0];
1620         unsigned long local = p->numa_faults_locality[1];
1621 
1622         /*
1623          * If there were no record hinting faults then either the task is
1624          * completely idle or all activity is areas that are not of interest
1625          * to automatic numa balancing. Related to that, if there were failed
1626          * migration then it implies we are migrating too quickly or the local
1627          * node is overloaded. In either case, scan slower
1628          */
1629         if (local + shared == 0 || p->numa_faults_locality[2]) {
1630                 p->numa_scan_period = min(p->numa_scan_period_max,
1631                         p->numa_scan_period << 1);
1632 
1633                 p->mm->numa_next_scan = jiffies +
1634                         msecs_to_jiffies(p->numa_scan_period);
1635 
1636                 return;
1637         }
1638 
1639         /*
1640          * Prepare to scale scan period relative to the current period.
1641          *       == NUMA_PERIOD_THRESHOLD scan period stays the same
1642          *       <  NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1643          *       >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1644          */
1645         period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1646         ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1647         if (ratio >= NUMA_PERIOD_THRESHOLD) {
1648                 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1649                 if (!slot)
1650                         slot = 1;
1651                 diff = slot * period_slot;
1652         } else {
1653                 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1654 
1655                 /*
1656                  * Scale scan rate increases based on sharing. There is an
1657                  * inverse relationship between the degree of sharing and
1658                  * the adjustment made to the scanning period. Broadly
1659                  * speaking the intent is that there is little point
1660                  * scanning faster if shared accesses dominate as it may
1661                  * simply bounce migrations uselessly
1662                  */
1663                 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1664                 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1665         }
1666 
1667         p->numa_scan_period = clamp(p->numa_scan_period + diff,
1668                         task_scan_min(p), task_scan_max(p));
1669         memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1670 }
1671 
1672 /*
1673  * Get the fraction of time the task has been running since the last
1674  * NUMA placement cycle. The scheduler keeps similar statistics, but
1675  * decays those on a 32ms period, which is orders of magnitude off
1676  * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1677  * stats only if the task is so new there are no NUMA statistics yet.
1678  */
1679 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1680 {
1681         u64 runtime, delta, now;
1682         /* Use the start of this time slice to avoid calculations. */
1683         now = p->se.exec_start;
1684         runtime = p->se.sum_exec_runtime;
1685 
1686         if (p->last_task_numa_placement) {
1687                 delta = runtime - p->last_sum_exec_runtime;
1688                 *period = now - p->last_task_numa_placement;
1689         } else {
1690                 delta = p->se.avg.runnable_avg_sum;
1691                 *period = p->se.avg.avg_period;
1692         }
1693 
1694         p->last_sum_exec_runtime = runtime;
1695         p->last_task_numa_placement = now;
1696 
1697         return delta;
1698 }
1699 
1700 /*
1701  * Determine the preferred nid for a task in a numa_group. This needs to
1702  * be done in a way that produces consistent results with group_weight,
1703  * otherwise workloads might not converge.
1704  */
1705 static int preferred_group_nid(struct task_struct *p, int nid)
1706 {
1707         nodemask_t nodes;
1708         int dist;
1709 
1710         /* Direct connections between all NUMA nodes. */
1711         if (sched_numa_topology_type == NUMA_DIRECT)
1712                 return nid;
1713 
1714         /*
1715          * On a system with glueless mesh NUMA topology, group_weight
1716          * scores nodes according to the number of NUMA hinting faults on
1717          * both the node itself, and on nearby nodes.
1718          */
1719         if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1720                 unsigned long score, max_score = 0;
1721                 int node, max_node = nid;
1722 
1723                 dist = sched_max_numa_distance;
1724 
1725                 for_each_online_node(node) {
1726                         score = group_weight(p, node, dist);
1727                         if (score > max_score) {
1728                                 max_score = score;
1729                                 max_node = node;
1730                         }
1731                 }
1732                 return max_node;
1733         }
1734 
1735         /*
1736          * Finding the preferred nid in a system with NUMA backplane
1737          * interconnect topology is more involved. The goal is to locate
1738          * tasks from numa_groups near each other in the system, and
1739          * untangle workloads from different sides of the system. This requires
1740          * searching down the hierarchy of node groups, recursively searching
1741          * inside the highest scoring group of nodes. The nodemask tricks
1742          * keep the complexity of the search down.
1743          */
1744         nodes = node_online_map;
1745         for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1746                 unsigned long max_faults = 0;
1747                 nodemask_t max_group = NODE_MASK_NONE;
1748                 int a, b;
1749 
1750                 /* Are there nodes at this distance from each other? */
1751                 if (!find_numa_distance(dist))
1752                         continue;
1753 
1754                 for_each_node_mask(a, nodes) {
1755                         unsigned long faults = 0;
1756                         nodemask_t this_group;
1757                         nodes_clear(this_group);
1758 
1759                         /* Sum group's NUMA faults; includes a==b case. */
1760                         for_each_node_mask(b, nodes) {
1761                                 if (node_distance(a, b) < dist) {
1762                                         faults += group_faults(p, b);
1763                                         node_set(b, this_group);
1764                                         node_clear(b, nodes);
1765                                 }
1766                         }
1767 
1768                         /* Remember the top group. */
1769                         if (faults > max_faults) {
1770                                 max_faults = faults;
1771                                 max_group = this_group;
1772                                 /*
1773                                  * subtle: at the smallest distance there is
1774                                  * just one node left in each "group", the
1775                                  * winner is the preferred nid.
1776                                  */
1777                                 nid = a;
1778                         }
1779                 }
1780                 /* Next round, evaluate the nodes within max_group. */
1781                 if (!max_faults)
1782                         break;
1783                 nodes = max_group;
1784         }
1785         return nid;
1786 }
1787 
1788 static void task_numa_placement(struct task_struct *p)
1789 {
1790         int seq, nid, max_nid = -1, max_group_nid = -1;
1791         unsigned long max_faults = 0, max_group_faults = 0;
1792         unsigned long fault_types[2] = { 0, 0 };
1793         unsigned long total_faults;
1794         u64 runtime, period;
1795         spinlock_t *group_lock = NULL;
1796 
1797         seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1798         if (p->numa_scan_seq == seq)
1799                 return;
1800         p->numa_scan_seq = seq;
1801         p->numa_scan_period_max = task_scan_max(p);
1802 
1803         total_faults = p->numa_faults_locality[0] +
1804                        p->numa_faults_locality[1];
1805         runtime = numa_get_avg_runtime(p, &period);
1806 
1807         /* If the task is part of a group prevent parallel updates to group stats */
1808         if (p->numa_group) {
1809                 group_lock = &p->numa_group->lock;
1810                 spin_lock_irq(group_lock);
1811         }
1812 
1813         /* Find the node with the highest number of faults */
1814         for_each_online_node(nid) {
1815                 /* Keep track of the offsets in numa_faults array */
1816                 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1817                 unsigned long faults = 0, group_faults = 0;
1818                 int priv;
1819 
1820                 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1821                         long diff, f_diff, f_weight;
1822 
1823                         mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1824                         membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1825                         cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1826                         cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1827 
1828                         /* Decay existing window, copy faults since last scan */
1829                         diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1830                         fault_types[priv] += p->numa_faults[membuf_idx];
1831                         p->numa_faults[membuf_idx] = 0;
1832 
1833                         /*
1834                          * Normalize the faults_from, so all tasks in a group
1835                          * count according to CPU use, instead of by the raw
1836                          * number of faults. Tasks with little runtime have
1837                          * little over-all impact on throughput, and thus their
1838                          * faults are less important.
1839                          */
1840                         f_weight = div64_u64(runtime << 16, period + 1);
1841                         f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1842                                    (total_faults + 1);
1843                         f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1844                         p->numa_faults[cpubuf_idx] = 0;
1845 
1846                         p->numa_faults[mem_idx] += diff;
1847                         p->numa_faults[cpu_idx] += f_diff;
1848                         faults += p->numa_faults[mem_idx];
1849                         p->total_numa_faults += diff;
1850                         if (p->numa_group) {
1851                                 /*
1852                                  * safe because we can only change our own group
1853                                  *
1854                                  * mem_idx represents the offset for a given
1855                                  * nid and priv in a specific region because it
1856                                  * is at the beginning of the numa_faults array.
1857                                  */
1858                                 p->numa_group->faults[mem_idx] += diff;
1859                                 p->numa_group->faults_cpu[mem_idx] += f_diff;
1860                                 p->numa_group->total_faults += diff;
1861                                 group_faults += p->numa_group->faults[mem_idx];
1862                         }
1863                 }
1864 
1865                 if (faults > max_faults) {
1866                         max_faults = faults;
1867                         max_nid = nid;
1868                 }
1869 
1870                 if (group_faults > max_group_faults) {
1871                         max_group_faults = group_faults;
1872                         max_group_nid = nid;
1873                 }
1874         }
1875 
1876         update_task_scan_period(p, fault_types[0], fault_types[1]);
1877 
1878         if (p->numa_group) {
1879                 update_numa_active_node_mask(p->numa_group);
1880                 spin_unlock_irq(group_lock);
1881                 max_nid = preferred_group_nid(p, max_group_nid);
1882         }
1883 
1884         if (max_faults) {
1885                 /* Set the new preferred node */
1886                 if (max_nid != p->numa_preferred_nid)
1887                         sched_setnuma(p, max_nid);
1888 
1889                 if (task_node(p) != p->numa_preferred_nid)
1890                         numa_migrate_preferred(p);
1891         }
1892 }
1893 
1894 static inline int get_numa_group(struct numa_group *grp)
1895 {
1896         return atomic_inc_not_zero(&grp->refcount);
1897 }
1898 
1899 static inline void put_numa_group(struct numa_group *grp)
1900 {
1901         if (atomic_dec_and_test(&grp->refcount))
1902                 kfree_rcu(grp, rcu);
1903 }
1904 
1905 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1906                         int *priv)
1907 {
1908         struct numa_group *grp, *my_grp;
1909         struct task_struct *tsk;
1910         bool join = false;
1911         int cpu = cpupid_to_cpu(cpupid);
1912         int i;
1913 
1914         if (unlikely(!p->numa_group)) {
1915                 unsigned int size = sizeof(struct numa_group) +
1916                                     4*nr_node_ids*sizeof(unsigned long);
1917 
1918                 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1919                 if (!grp)
1920                         return;
1921 
1922                 atomic_set(&grp->refcount, 1);
1923                 spin_lock_init(&grp->lock);
1924                 grp->gid = p->pid;
1925                 /* Second half of the array tracks nids where faults happen */
1926                 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1927                                                 nr_node_ids;
1928 
1929                 node_set(task_node(current), grp->active_nodes);
1930 
1931                 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1932                         grp->faults[i] = p->numa_faults[i];
1933 
1934                 grp->total_faults = p->total_numa_faults;
1935 
1936                 grp->nr_tasks++;
1937                 rcu_assign_pointer(p->numa_group, grp);
1938         }
1939 
1940         rcu_read_lock();
1941         tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1942 
1943         if (!cpupid_match_pid(tsk, cpupid))
1944                 goto no_join;
1945 
1946         grp = rcu_dereference(tsk->numa_group);
1947         if (!grp)
1948                 goto no_join;
1949 
1950         my_grp = p->numa_group;
1951         if (grp == my_grp)
1952                 goto no_join;
1953 
1954         /*
1955          * Only join the other group if its bigger; if we're the bigger group,
1956          * the other task will join us.
1957          */
1958         if (my_grp->nr_tasks > grp->nr_tasks)
1959                 goto no_join;
1960 
1961         /*
1962          * Tie-break on the grp address.
1963          */
1964         if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1965                 goto no_join;
1966 
1967         /* Always join threads in the same process. */
1968         if (tsk->mm == current->mm)
1969                 join = true;
1970 
1971         /* Simple filter to avoid false positives due to PID collisions */
1972         if (flags & TNF_SHARED)
1973                 join = true;
1974 
1975         /* Update priv based on whether false sharing was detected */
1976         *priv = !join;
1977 
1978         if (join && !get_numa_group(grp))
1979                 goto no_join;
1980 
1981         rcu_read_unlock();
1982 
1983         if (!join)
1984                 return;
1985 
1986         BUG_ON(irqs_disabled());
1987         double_lock_irq(&my_grp->lock, &grp->lock);
1988 
1989         for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
1990                 my_grp->faults[i] -= p->numa_faults[i];
1991                 grp->faults[i] += p->numa_faults[i];
1992         }
1993         my_grp->total_faults -= p->total_numa_faults;
1994         grp->total_faults += p->total_numa_faults;
1995 
1996         my_grp->nr_tasks--;
1997         grp->nr_tasks++;
1998 
1999         spin_unlock(&my_grp->lock);
2000         spin_unlock_irq(&grp->lock);
2001 
2002         rcu_assign_pointer(p->numa_group, grp);
2003 
2004         put_numa_group(my_grp);
2005         return;
2006 
2007 no_join:
2008         rcu_read_unlock();
2009         return;
2010 }
2011 
2012 void task_numa_free(struct task_struct *p)
2013 {
2014         struct numa_group *grp = p->numa_group;
2015         void *numa_faults = p->numa_faults;
2016         unsigned long flags;
2017         int i;
2018 
2019         if (grp) {
2020                 spin_lock_irqsave(&grp->lock, flags);
2021                 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2022                         grp->faults[i] -= p->numa_faults[i];
2023                 grp->total_faults -= p->total_numa_faults;
2024 
2025                 grp->nr_tasks--;
2026                 spin_unlock_irqrestore(&grp->lock, flags);
2027                 RCU_INIT_POINTER(p->numa_group, NULL);
2028                 put_numa_group(grp);
2029         }
2030 
2031         p->numa_faults = NULL;
2032         kfree(numa_faults);
2033 }
2034 
2035 /*
2036  * Got a PROT_NONE fault for a page on @node.
2037  */
2038 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2039 {
2040         struct task_struct *p = current;
2041         bool migrated = flags & TNF_MIGRATED;
2042         int cpu_node = task_node(current);
2043         int local = !!(flags & TNF_FAULT_LOCAL);
2044         int priv;
2045 
2046         if (!numabalancing_enabled)
2047                 return;
2048 
2049         /* for example, ksmd faulting in a user's mm */
2050         if (!p->mm)
2051                 return;
2052 
2053         /* Allocate buffer to track faults on a per-node basis */
2054         if (unlikely(!p->numa_faults)) {
2055                 int size = sizeof(*p->numa_faults) *
2056                            NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2057 
2058                 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2059                 if (!p->numa_faults)
2060                         return;
2061 
2062                 p->total_numa_faults = 0;
2063                 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2064         }
2065 
2066         /*
2067          * First accesses are treated as private, otherwise consider accesses
2068          * to be private if the accessing pid has not changed
2069          */
2070         if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2071                 priv = 1;
2072         } else {
2073                 priv = cpupid_match_pid(p, last_cpupid);
2074                 if (!priv && !(flags & TNF_NO_GROUP))
2075                         task_numa_group(p, last_cpupid, flags, &priv);
2076         }
2077 
2078         /*
2079          * If a workload spans multiple NUMA nodes, a shared fault that
2080          * occurs wholly within the set of nodes that the workload is
2081          * actively using should be counted as local. This allows the
2082          * scan rate to slow down when a workload has settled down.
2083          */
2084         if (!priv && !local && p->numa_group &&
2085                         node_isset(cpu_node, p->numa_group->active_nodes) &&
2086                         node_isset(mem_node, p->numa_group->active_nodes))
2087                 local = 1;
2088 
2089         task_numa_placement(p);
2090 
2091         /*
2092          * Retry task to preferred node migration periodically, in case it
2093          * case it previously failed, or the scheduler moved us.
2094          */
2095         if (time_after(jiffies, p->numa_migrate_retry))
2096                 numa_migrate_preferred(p);
2097 
2098         if (migrated)
2099                 p->numa_pages_migrated += pages;
2100         if (flags & TNF_MIGRATE_FAIL)
2101                 p->numa_faults_locality[2] += pages;
2102 
2103         p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2104         p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2105         p->numa_faults_locality[local] += pages;
2106 }
2107 
2108 static void reset_ptenuma_scan(struct task_struct *p)
2109 {
2110         ACCESS_ONCE(p->mm->numa_scan_seq)++;
2111         p->mm->numa_scan_offset = 0;
2112 }
2113 
2114 /*
2115  * The expensive part of numa migration is done from task_work context.
2116  * Triggered from task_tick_numa().
2117  */
2118 void task_numa_work(struct callback_head *work)
2119 {
2120         unsigned long migrate, next_scan, now = jiffies;
2121         struct task_struct *p = current;
2122         struct mm_struct *mm = p->mm;
2123         struct vm_area_struct *vma;
2124         unsigned long start, end;
2125         unsigned long nr_pte_updates = 0;
2126         long pages;
2127 
2128         WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2129 
2130         work->next = work; /* protect against double add */
2131         /*
2132          * Who cares about NUMA placement when they're dying.
2133          *
2134          * NOTE: make sure not to dereference p->mm before this check,
2135          * exit_task_work() happens _after_ exit_mm() so we could be called
2136          * without p->mm even though we still had it when we enqueued this
2137          * work.
2138          */
2139         if (p->flags & PF_EXITING)
2140                 return;
2141 
2142         if (!mm->numa_next_scan) {
2143                 mm->numa_next_scan = now +
2144                         msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2145         }
2146 
2147         /*
2148          * Enforce maximal scan/migration frequency..
2149          */
2150         migrate = mm->numa_next_scan;
2151         if (time_before(now, migrate))
2152                 return;
2153 
2154         if (p->numa_scan_period == 0) {
2155                 p->numa_scan_period_max = task_scan_max(p);
2156                 p->numa_scan_period = task_scan_min(p);
2157         }
2158 
2159         next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2160         if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2161                 return;
2162 
2163         /*
2164          * Delay this task enough that another task of this mm will likely win
2165          * the next time around.
2166          */
2167         p->node_stamp += 2 * TICK_NSEC;
2168 
2169         start = mm->numa_scan_offset;
2170         pages = sysctl_numa_balancing_scan_size;
2171         pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2172         if (!pages)
2173                 return;
2174 
2175         down_read(&mm->mmap_sem);
2176         vma = find_vma(mm, start);
2177         if (!vma) {
2178                 reset_ptenuma_scan(p);
2179                 start = 0;
2180                 vma = mm->mmap;
2181         }
2182         for (; vma; vma = vma->vm_next) {
2183                 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2184                         is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2185                         continue;
2186                 }
2187 
2188                 /*
2189                  * Shared library pages mapped by multiple processes are not
2190                  * migrated as it is expected they are cache replicated. Avoid
2191                  * hinting faults in read-only file-backed mappings or the vdso
2192                  * as migrating the pages will be of marginal benefit.
2193                  */
2194                 if (!vma->vm_mm ||
2195                     (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2196                         continue;
2197 
2198                 /*
2199                  * Skip inaccessible VMAs to avoid any confusion between
2200                  * PROT_NONE and NUMA hinting ptes
2201                  */
2202                 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2203                         continue;
2204 
2205                 do {
2206                         start = max(start, vma->vm_start);
2207                         end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2208                         end = min(end, vma->vm_end);
2209                         nr_pte_updates += change_prot_numa(vma, start, end);
2210 
2211                         /*
2212                          * Scan sysctl_numa_balancing_scan_size but ensure that
2213                          * at least one PTE is updated so that unused virtual
2214                          * address space is quickly skipped.
2215                          */
2216                         if (nr_pte_updates)
2217                                 pages -= (end - start) >> PAGE_SHIFT;
2218 
2219                         start = end;
2220                         if (pages <= 0)
2221                                 goto out;
2222 
2223                         cond_resched();
2224                 } while (end != vma->vm_end);
2225         }
2226 
2227 out:
2228         /*
2229          * It is possible to reach the end of the VMA list but the last few
2230          * VMAs are not guaranteed to the vma_migratable. If they are not, we
2231          * would find the !migratable VMA on the next scan but not reset the
2232          * scanner to the start so check it now.
2233          */
2234         if (vma)
2235                 mm->numa_scan_offset = start;
2236         else
2237                 reset_ptenuma_scan(p);
2238         up_read(&mm->mmap_sem);
2239 }
2240 
2241 /*
2242  * Drive the periodic memory faults..
2243  */
2244 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2245 {
2246         struct callback_head *work = &curr->numa_work;
2247         u64 period, now;
2248 
2249         /*
2250          * We don't care about NUMA placement if we don't have memory.
2251          */
2252         if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2253                 return;
2254 
2255         /*
2256          * Using runtime rather than walltime has the dual advantage that
2257          * we (mostly) drive the selection from busy threads and that the
2258          * task needs to have done some actual work before we bother with
2259          * NUMA placement.
2260          */
2261         now = curr->se.sum_exec_runtime;
2262         period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2263 
2264         if (now - curr->node_stamp > period) {
2265                 if (!curr->node_stamp)
2266                         curr->numa_scan_period = task_scan_min(curr);
2267                 curr->node_stamp += period;
2268 
2269                 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2270                         init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2271                         task_work_add(curr, work, true);
2272                 }
2273         }
2274 }
2275 #else
2276 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2277 {
2278 }
2279 
2280 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2281 {
2282 }
2283 
2284 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2285 {
2286 }
2287 #endif /* CONFIG_NUMA_BALANCING */
2288 
2289 static void
2290 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2291 {
2292         update_load_add(&cfs_rq->load, se->load.weight);
2293         if (!parent_entity(se))
2294                 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2295 #ifdef CONFIG_SMP
2296         if (entity_is_task(se)) {
2297                 struct rq *rq = rq_of(cfs_rq);
2298 
2299                 account_numa_enqueue(rq, task_of(se));
2300                 list_add(&se->group_node, &rq->cfs_tasks);
2301         }
2302 #endif
2303         cfs_rq->nr_running++;
2304 }
2305 
2306 static void
2307 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2308 {
2309         update_load_sub(&cfs_rq->load, se->load.weight);
2310         if (!parent_entity(se))
2311                 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2312         if (entity_is_task(se)) {
2313                 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2314                 list_del_init(&se->group_node);
2315         }
2316         cfs_rq->nr_running--;
2317 }
2318 
2319 #ifdef CONFIG_FAIR_GROUP_SCHED
2320 # ifdef CONFIG_SMP
2321 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2322 {
2323         long tg_weight;
2324 
2325         /*
2326          * Use this CPU's actual weight instead of the last load_contribution
2327          * to gain a more accurate current total weight. See
2328          * update_cfs_rq_load_contribution().
2329          */
2330         tg_weight = atomic_long_read(&tg->load_avg);
2331         tg_weight -= cfs_rq->tg_load_contrib;
2332         tg_weight += cfs_rq->load.weight;
2333 
2334         return tg_weight;
2335 }
2336 
2337 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2338 {
2339         long tg_weight, load, shares;
2340 
2341         tg_weight = calc_tg_weight(tg, cfs_rq);
2342         load = cfs_rq->load.weight;
2343 
2344         shares = (tg->shares * load);
2345         if (tg_weight)
2346                 shares /= tg_weight;
2347 
2348         if (shares < MIN_SHARES)
2349                 shares = MIN_SHARES;
2350         if (shares > tg->shares)
2351                 shares = tg->shares;
2352 
2353         return shares;
2354 }
2355 # else /* CONFIG_SMP */
2356 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2357 {
2358         return tg->shares;
2359 }
2360 # endif /* CONFIG_SMP */
2361 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2362                             unsigned long weight)
2363 {
2364         if (se->on_rq) {
2365                 /* commit outstanding execution time */
2366                 if (cfs_rq->curr == se)
2367                         update_curr(cfs_rq);
2368                 account_entity_dequeue(cfs_rq, se);
2369         }
2370 
2371         update_load_set(&se->load, weight);
2372 
2373         if (se->on_rq)
2374                 account_entity_enqueue(cfs_rq, se);
2375 }
2376 
2377 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2378 
2379 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2380 {
2381         struct task_group *tg;
2382         struct sched_entity *se;
2383         long shares;
2384 
2385         tg = cfs_rq->tg;
2386         se = tg->se[cpu_of(rq_of(cfs_rq))];
2387         if (!se || throttled_hierarchy(cfs_rq))
2388                 return;
2389 #ifndef CONFIG_SMP
2390         if (likely(se->load.weight == tg->shares))
2391                 return;
2392 #endif
2393         shares = calc_cfs_shares(cfs_rq, tg);
2394 
2395         reweight_entity(cfs_rq_of(se), se, shares);
2396 }
2397 #else /* CONFIG_FAIR_GROUP_SCHED */
2398 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2399 {
2400 }
2401 #endif /* CONFIG_FAIR_GROUP_SCHED */
2402 
2403 #ifdef CONFIG_SMP
2404 /*
2405  * We choose a half-life close to 1 scheduling period.
2406  * Note: The tables below are dependent on this value.
2407  */
2408 #define LOAD_AVG_PERIOD 32
2409 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2410 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2411 
2412 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2413 static const u32 runnable_avg_yN_inv[] = {
2414         0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2415         0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2416         0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2417         0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2418         0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2419         0x85aac367, 0x82cd8698,
2420 };
2421 
2422 /*
2423  * Precomputed \Sum y^k { 1<=k<=n }.  These are floor(true_value) to prevent
2424  * over-estimates when re-combining.
2425  */
2426 static const u32 runnable_avg_yN_sum[] = {
2427             0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2428          9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2429         17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2430 };
2431 
2432 /*
2433  * Approximate:
2434  *   val * y^n,    where y^32 ~= 0.5 (~1 scheduling period)
2435  */
2436 static __always_inline u64 decay_load(u64 val, u64 n)
2437 {
2438         unsigned int local_n;
2439 
2440         if (!n)
2441                 return val;
2442         else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2443                 return 0;
2444 
2445         /* after bounds checking we can collapse to 32-bit */
2446         local_n = n;
2447 
2448         /*
2449          * As y^PERIOD = 1/2, we can combine
2450          *    y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2451          * With a look-up table which covers y^n (n<PERIOD)
2452          *
2453          * To achieve constant time decay_load.
2454          */
2455         if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2456                 val >>= local_n / LOAD_AVG_PERIOD;
2457                 local_n %= LOAD_AVG_PERIOD;
2458         }
2459 
2460         val *= runnable_avg_yN_inv[local_n];
2461         /* We don't use SRR here since we always want to round down. */
2462         return val >> 32;
2463 }
2464 
2465 /*
2466  * For updates fully spanning n periods, the contribution to runnable
2467  * average will be: \Sum 1024*y^n
2468  *
2469  * We can compute this reasonably efficiently by combining:
2470  *   y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for  n <PERIOD}
2471  */
2472 static u32 __compute_runnable_contrib(u64 n)
2473 {
2474         u32 contrib = 0;
2475 
2476         if (likely(n <= LOAD_AVG_PERIOD))
2477                 return runnable_avg_yN_sum[n];
2478         else if (unlikely(n >= LOAD_AVG_MAX_N))
2479                 return LOAD_AVG_MAX;
2480 
2481         /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2482         do {
2483                 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2484                 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2485 
2486                 n -= LOAD_AVG_PERIOD;
2487         } while (n > LOAD_AVG_PERIOD);
2488 
2489         contrib = decay_load(contrib, n);
2490         return contrib + runnable_avg_yN_sum[n];
2491 }
2492 
2493 /*
2494  * We can represent the historical contribution to runnable average as the
2495  * coefficients of a geometric series.  To do this we sub-divide our runnable
2496  * history into segments of approximately 1ms (1024us); label the segment that
2497  * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2498  *
2499  * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2500  *      p0            p1           p2
2501  *     (now)       (~1ms ago)  (~2ms ago)
2502  *
2503  * Let u_i denote the fraction of p_i that the entity was runnable.
2504  *
2505  * We then designate the fractions u_i as our co-efficients, yielding the
2506  * following representation of historical load:
2507  *   u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2508  *
2509  * We choose y based on the with of a reasonably scheduling period, fixing:
2510  *   y^32 = 0.5
2511  *
2512  * This means that the contribution to load ~32ms ago (u_32) will be weighted
2513  * approximately half as much as the contribution to load within the last ms
2514  * (u_0).
2515  *
2516  * When a period "rolls over" and we have new u_0`, multiplying the previous
2517  * sum again by y is sufficient to update:
2518  *   load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2519  *            = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2520  */
2521 static __always_inline int __update_entity_runnable_avg(u64 now, int cpu,
2522                                                         struct sched_avg *sa,
2523                                                         int runnable,
2524                                                         int running)
2525 {
2526         u64 delta, periods;
2527         u32 runnable_contrib;
2528         int delta_w, decayed = 0;
2529         unsigned long scale_freq = arch_scale_freq_capacity(NULL, cpu);
2530 
2531         delta = now - sa->last_runnable_update;
2532         /*
2533          * This should only happen when time goes backwards, which it
2534          * unfortunately does during sched clock init when we swap over to TSC.
2535          */
2536         if ((s64)delta < 0) {
2537                 sa->last_runnable_update = now;
2538                 return 0;
2539         }
2540 
2541         /*
2542          * Use 1024ns as the unit of measurement since it's a reasonable
2543          * approximation of 1us and fast to compute.
2544          */
2545         delta >>= 10;
2546         if (!delta)
2547                 return 0;
2548         sa->last_runnable_update = now;
2549 
2550         /* delta_w is the amount already accumulated against our next period */
2551         delta_w = sa->avg_period % 1024;
2552         if (delta + delta_w >= 1024) {
2553                 /* period roll-over */
2554                 decayed = 1;
2555 
2556                 /*
2557                  * Now that we know we're crossing a period boundary, figure
2558                  * out how much from delta we need to complete the current
2559                  * period and accrue it.
2560                  */
2561                 delta_w = 1024 - delta_w;
2562                 if (runnable)
2563                         sa->runnable_avg_sum += delta_w;
2564                 if (running)
2565                         sa->running_avg_sum += delta_w * scale_freq
2566                                 >> SCHED_CAPACITY_SHIFT;
2567                 sa->avg_period += delta_w;
2568 
2569                 delta -= delta_w;
2570 
2571                 /* Figure out how many additional periods this update spans */
2572                 periods = delta / 1024;
2573                 delta %= 1024;
2574 
2575                 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2576                                                   periods + 1);
2577                 sa->running_avg_sum = decay_load(sa->running_avg_sum,
2578                                                   periods + 1);
2579                 sa->avg_period = decay_load(sa->avg_period,
2580                                                      periods + 1);
2581 
2582                 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2583                 runnable_contrib = __compute_runnable_contrib(periods);
2584                 if (runnable)
2585                         sa->runnable_avg_sum += runnable_contrib;
2586                 if (running)
2587                         sa->running_avg_sum += runnable_contrib * scale_freq
2588                                 >> SCHED_CAPACITY_SHIFT;
2589                 sa->avg_period += runnable_contrib;
2590         }
2591 
2592         /* Remainder of delta accrued against u_0` */
2593         if (runnable)
2594                 sa->runnable_avg_sum += delta;
2595         if (running)
2596                 sa->running_avg_sum += delta * scale_freq
2597                         >> SCHED_CAPACITY_SHIFT;
2598         sa->avg_period += delta;
2599 
2600         return decayed;
2601 }
2602 
2603 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2604 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2605 {
2606         struct cfs_rq *cfs_rq = cfs_rq_of(se);
2607         u64 decays = atomic64_read(&cfs_rq->decay_counter);
2608 
2609         decays -= se->avg.decay_count;
2610         se->avg.decay_count = 0;
2611         if (!decays)
2612                 return 0;
2613 
2614         se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2615         se->avg.utilization_avg_contrib =
2616                 decay_load(se->avg.utilization_avg_contrib, decays);
2617 
2618         return decays;
2619 }
2620 
2621 #ifdef CONFIG_FAIR_GROUP_SCHED
2622 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2623                                                  int force_update)
2624 {
2625         struct task_group *tg = cfs_rq->tg;
2626         long tg_contrib;
2627 
2628         tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2629         tg_contrib -= cfs_rq->tg_load_contrib;
2630 
2631         if (!tg_contrib)
2632                 return;
2633 
2634         if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2635                 atomic_long_add(tg_contrib, &tg->load_avg);
2636                 cfs_rq->tg_load_contrib += tg_contrib;
2637         }
2638 }
2639 
2640 /*
2641  * Aggregate cfs_rq runnable averages into an equivalent task_group
2642  * representation for computing load contributions.
2643  */
2644 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2645                                                   struct cfs_rq *cfs_rq)
2646 {
2647         struct task_group *tg = cfs_rq->tg;
2648         long contrib;
2649 
2650         /* The fraction of a cpu used by this cfs_rq */
2651         contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT,
2652                           sa->avg_period + 1);
2653         contrib -= cfs_rq->tg_runnable_contrib;
2654 
2655         if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2656                 atomic_add(contrib, &tg->runnable_avg);
2657                 cfs_rq->tg_runnable_contrib += contrib;
2658         }
2659 }
2660 
2661 static inline void __update_group_entity_contrib(struct sched_entity *se)
2662 {
2663         struct cfs_rq *cfs_rq = group_cfs_rq(se);
2664         struct task_group *tg = cfs_rq->tg;
2665         int runnable_avg;
2666 
2667         u64 contrib;
2668 
2669         contrib = cfs_rq->tg_load_contrib * tg->shares;
2670         se->avg.load_avg_contrib = div_u64(contrib,
2671                                      atomic_long_read(&tg->load_avg) + 1);
2672 
2673         /*
2674          * For group entities we need to compute a correction term in the case
2675          * that they are consuming <1 cpu so that we would contribute the same
2676          * load as a task of equal weight.
2677          *
2678          * Explicitly co-ordinating this measurement would be expensive, but
2679          * fortunately the sum of each cpus contribution forms a usable
2680          * lower-bound on the true value.
2681          *
2682          * Consider the aggregate of 2 contributions.  Either they are disjoint
2683          * (and the sum represents true value) or they are disjoint and we are
2684          * understating by the aggregate of their overlap.
2685          *
2686          * Extending this to N cpus, for a given overlap, the maximum amount we
2687          * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2688          * cpus that overlap for this interval and w_i is the interval width.
2689          *
2690          * On a small machine; the first term is well-bounded which bounds the
2691          * total error since w_i is a subset of the period.  Whereas on a
2692          * larger machine, while this first term can be larger, if w_i is the
2693          * of consequential size guaranteed to see n_i*w_i quickly converge to
2694          * our upper bound of 1-cpu.
2695          */
2696         runnable_avg = atomic_read(&tg->runnable_avg);
2697         if (runnable_avg < NICE_0_LOAD) {
2698                 se->avg.load_avg_contrib *= runnable_avg;
2699                 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2700         }
2701 }
2702 
2703 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2704 {
2705         __update_entity_runnable_avg(rq_clock_task(rq), cpu_of(rq), &rq->avg,
2706                         runnable, runnable);
2707         __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2708 }
2709 #else /* CONFIG_FAIR_GROUP_SCHED */
2710 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2711                                                  int force_update) {}
2712 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2713                                                   struct cfs_rq *cfs_rq) {}
2714 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2715 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2716 #endif /* CONFIG_FAIR_GROUP_SCHED */
2717 
2718 static inline void __update_task_entity_contrib(struct sched_entity *se)
2719 {
2720         u32 contrib;
2721 
2722         /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2723         contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2724         contrib /= (se->avg.avg_period + 1);
2725         se->avg.load_avg_contrib = scale_load(contrib);
2726 }
2727 
2728 /* Compute the current contribution to load_avg by se, return any delta */
2729 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2730 {
2731         long old_contrib = se->avg.load_avg_contrib;
2732 
2733         if (entity_is_task(se)) {
2734                 __update_task_entity_contrib(se);
2735         } else {
2736                 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2737                 __update_group_entity_contrib(se);
2738         }
2739 
2740         return se->avg.load_avg_contrib - old_contrib;
2741 }
2742 
2743 
2744 static inline void __update_task_entity_utilization(struct sched_entity *se)
2745 {
2746         u32 contrib;
2747 
2748         /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2749         contrib = se->avg.running_avg_sum * scale_load_down(SCHED_LOAD_SCALE);
2750         contrib /= (se->avg.avg_period + 1);
2751         se->avg.utilization_avg_contrib = scale_load(contrib);
2752 }
2753 
2754 static long __update_entity_utilization_avg_contrib(struct sched_entity *se)
2755 {
2756         long old_contrib = se->avg.utilization_avg_contrib;
2757 
2758         if (entity_is_task(se))
2759                 __update_task_entity_utilization(se);
2760         else
2761                 se->avg.utilization_avg_contrib =
2762                                         group_cfs_rq(se)->utilization_load_avg;
2763 
2764         return se->avg.utilization_avg_contrib - old_contrib;
2765 }
2766 
2767 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2768                                                  long load_contrib)
2769 {
2770         if (likely(load_contrib < cfs_rq->blocked_load_avg))
2771                 cfs_rq->blocked_load_avg -= load_contrib;
2772         else
2773                 cfs_rq->blocked_load_avg = 0;
2774 }
2775 
2776 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2777 
2778 /* Update a sched_entity's runnable average */
2779 static inline void update_entity_load_avg(struct sched_entity *se,
2780                                           int update_cfs_rq)
2781 {
2782         struct cfs_rq *cfs_rq = cfs_rq_of(se);
2783         long contrib_delta, utilization_delta;
2784         int cpu = cpu_of(rq_of(cfs_rq));
2785         u64 now;
2786 
2787         /*
2788          * For a group entity we need to use their owned cfs_rq_clock_task() in
2789          * case they are the parent of a throttled hierarchy.
2790          */
2791         if (entity_is_task(se))
2792                 now = cfs_rq_clock_task(cfs_rq);
2793         else
2794                 now = cfs_rq_clock_task(group_cfs_rq(se));
2795 
2796         if (!__update_entity_runnable_avg(now, cpu, &se->avg, se->on_rq,
2797                                         cfs_rq->curr == se))
2798                 return;
2799 
2800         contrib_delta = __update_entity_load_avg_contrib(se);
2801         utilization_delta = __update_entity_utilization_avg_contrib(se);
2802 
2803         if (!update_cfs_rq)
2804                 return;
2805 
2806         if (se->on_rq) {
2807                 cfs_rq->runnable_load_avg += contrib_delta;
2808                 cfs_rq->utilization_load_avg += utilization_delta;
2809         } else {
2810                 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2811         }
2812 }
2813 
2814 /*
2815  * Decay the load contributed by all blocked children and account this so that
2816  * their contribution may appropriately discounted when they wake up.
2817  */
2818 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2819 {
2820         u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2821         u64 decays;
2822 
2823         decays = now - cfs_rq->last_decay;
2824         if (!decays && !force_update)
2825                 return;
2826 
2827         if (atomic_long_read(&cfs_rq->removed_load)) {
2828                 unsigned long removed_load;
2829                 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2830                 subtract_blocked_load_contrib(cfs_rq, removed_load);
2831         }
2832 
2833         if (decays) {
2834                 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2835                                                       decays);
2836                 atomic64_add(decays, &cfs_rq->decay_counter);
2837                 cfs_rq->last_decay = now;
2838         }
2839 
2840         __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2841 }
2842 
2843 /* Add the load generated by se into cfs_rq's child load-average */
2844 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2845                                                   struct sched_entity *se,
2846                                                   int wakeup)
2847 {
2848         /*
2849          * We track migrations using entity decay_count <= 0, on a wake-up
2850          * migration we use a negative decay count to track the remote decays
2851          * accumulated while sleeping.
2852          *
2853          * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2854          * are seen by enqueue_entity_load_avg() as a migration with an already
2855          * constructed load_avg_contrib.
2856          */
2857         if (unlikely(se->avg.decay_count <= 0)) {
2858                 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2859                 if (se->avg.decay_count) {
2860                         /*
2861                          * In a wake-up migration we have to approximate the
2862                          * time sleeping.  This is because we can't synchronize
2863                          * clock_task between the two cpus, and it is not
2864                          * guaranteed to be read-safe.  Instead, we can
2865                          * approximate this using our carried decays, which are
2866                          * explicitly atomically readable.
2867                          */
2868                         se->avg.last_runnable_update -= (-se->avg.decay_count)
2869                                                         << 20;
2870                         update_entity_load_avg(se, 0);
2871                         /* Indicate that we're now synchronized and on-rq */
2872                         se->avg.decay_count = 0;
2873                 }
2874                 wakeup = 0;
2875         } else {
2876                 __synchronize_entity_decay(se);
2877         }
2878 
2879         /* migrated tasks did not contribute to our blocked load */
2880         if (wakeup) {
2881                 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2882                 update_entity_load_avg(se, 0);
2883         }
2884 
2885         cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2886         cfs_rq->utilization_load_avg += se->avg.utilization_avg_contrib;
2887         /* we force update consideration on load-balancer moves */
2888         update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2889 }
2890 
2891 /*
2892  * Remove se's load from this cfs_rq child load-average, if the entity is
2893  * transitioning to a blocked state we track its projected decay using
2894  * blocked_load_avg.
2895  */
2896 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2897                                                   struct sched_entity *se,
2898                                                   int sleep)
2899 {
2900         update_entity_load_avg(se, 1);
2901         /* we force update consideration on load-balancer moves */
2902         update_cfs_rq_blocked_load(cfs_rq, !sleep);
2903 
2904         cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2905         cfs_rq->utilization_load_avg -= se->avg.utilization_avg_contrib;
2906         if (sleep) {
2907                 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2908                 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2909         } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2910 }
2911 
2912 /*
2913  * Update the rq's load with the elapsed running time before entering
2914  * idle. if the last scheduled task is not a CFS task, idle_enter will
2915  * be the only way to update the runnable statistic.
2916  */
2917 void idle_enter_fair(struct rq *this_rq)
2918 {
2919         update_rq_runnable_avg(this_rq, 1);
2920 }
2921 
2922 /*
2923  * Update the rq's load with the elapsed idle time before a task is
2924  * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2925  * be the only way to update the runnable statistic.
2926  */
2927 void idle_exit_fair(struct rq *this_rq)
2928 {
2929         update_rq_runnable_avg(this_rq, 0);
2930 }
2931 
2932 static int idle_balance(struct rq *this_rq);
2933 
2934 #else /* CONFIG_SMP */
2935 
2936 static inline void update_entity_load_avg(struct sched_entity *se,
2937                                           int update_cfs_rq) {}
2938 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2939 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2940                                            struct sched_entity *se,
2941                                            int wakeup) {}
2942 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2943                                            struct sched_entity *se,
2944                                            int sleep) {}
2945 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2946                                               int force_update) {}
2947 
2948 static inline int idle_balance(struct rq *rq)
2949 {
2950         return 0;
2951 }
2952 
2953 #endif /* CONFIG_SMP */
2954 
2955 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2956 {
2957 #ifdef CONFIG_SCHEDSTATS
2958         struct task_struct *tsk = NULL;
2959 
2960         if (entity_is_task(se))
2961                 tsk = task_of(se);
2962 
2963         if (se->statistics.sleep_start) {
2964                 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2965 
2966                 if ((s64)delta < 0)
2967                         delta = 0;
2968 
2969                 if (unlikely(delta > se->statistics.sleep_max))
2970                         se->statistics.sleep_max = delta;
2971 
2972                 se->statistics.sleep_start = 0;
2973                 se->statistics.sum_sleep_runtime += delta;
2974 
2975                 if (tsk) {
2976                         account_scheduler_latency(tsk, delta >> 10, 1);
2977                         trace_sched_stat_sleep(tsk, delta);
2978                 }
2979         }
2980         if (se->statistics.block_start) {
2981                 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2982 
2983                 if ((s64)delta < 0)
2984                         delta = 0;
2985 
2986                 if (unlikely(delta > se->statistics.block_max))
2987                         se->statistics.block_max = delta;
2988 
2989                 se->statistics.block_start = 0;
2990                 se->statistics.sum_sleep_runtime += delta;
2991 
2992                 if (tsk) {
2993                         if (tsk->in_iowait) {
2994                                 se->statistics.iowait_sum += delta;
2995                                 se->statistics.iowait_count++;
2996                                 trace_sched_stat_iowait(tsk, delta);
2997                         }
2998 
2999                         trace_sched_stat_blocked(tsk, delta);
3000 
3001                         /*
3002                          * Blocking time is in units of nanosecs, so shift by
3003                          * 20 to get a milliseconds-range estimation of the
3004                          * amount of time that the task spent sleeping:
3005                          */
3006                         if (unlikely(prof_on == SLEEP_PROFILING)) {
3007                                 profile_hits(SLEEP_PROFILING,
3008                                                 (void *)get_wchan(tsk),
3009                                                 delta >> 20);
3010                         }
3011                         account_scheduler_latency(tsk, delta >> 10, 0);
3012                 }
3013         }
3014 #endif
3015 }
3016 
3017 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3018 {
3019 #ifdef CONFIG_SCHED_DEBUG
3020         s64 d = se->vruntime - cfs_rq->min_vruntime;
3021 
3022         if (d < 0)
3023                 d = -d;
3024 
3025         if (d > 3*sysctl_sched_latency)
3026                 schedstat_inc(cfs_rq, nr_spread_over);
3027 #endif
3028 }
3029 
3030 static void
3031 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3032 {
3033         u64 vruntime = cfs_rq->min_vruntime;
3034 
3035         /*
3036          * The 'current' period is already promised to the current tasks,
3037          * however the extra weight of the new task will slow them down a
3038          * little, place the new task so that it fits in the slot that
3039          * stays open at the end.
3040          */
3041         if (initial && sched_feat(START_DEBIT))
3042                 vruntime += sched_vslice(cfs_rq, se);
3043 
3044         /* sleeps up to a single latency don't count. */
3045         if (!initial) {
3046                 unsigned long thresh = sysctl_sched_latency;
3047 
3048                 /*
3049                  * Halve their sleep time's effect, to allow
3050                  * for a gentler effect of sleepers:
3051                  */
3052                 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3053                         thresh >>= 1;
3054 
3055                 vruntime -= thresh;
3056         }
3057 
3058         /* ensure we never gain time by being placed backwards. */
3059         se->vruntime = max_vruntime(se->vruntime, vruntime);
3060 }
3061 
3062 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3063 
3064 static void
3065 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3066 {
3067         /*
3068          * Update the normalized vruntime before updating min_vruntime
3069          * through calling update_curr().
3070          */
3071         if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3072                 se->vruntime += cfs_rq->min_vruntime;
3073 
3074         /*
3075          * Update run-time statistics of the 'current'.
3076          */
3077         update_curr(cfs_rq);
3078         enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
3079         account_entity_enqueue(cfs_rq, se);
3080         update_cfs_shares(cfs_rq);
3081 
3082         if (flags & ENQUEUE_WAKEUP) {
3083                 place_entity(cfs_rq, se, 0);
3084                 enqueue_sleeper(cfs_rq, se);
3085         }
3086 
3087         update_stats_enqueue(cfs_rq, se);
3088         check_spread(cfs_rq, se);
3089         if (se != cfs_rq->curr)
3090                 __enqueue_entity(cfs_rq, se);
3091         se->on_rq = 1;
3092 
3093         if (cfs_rq->nr_running == 1) {
3094                 list_add_leaf_cfs_rq(cfs_rq);
3095                 check_enqueue_throttle(cfs_rq);
3096         }
3097 }
3098 
3099 static void __clear_buddies_last(struct sched_entity *se)
3100 {
3101         for_each_sched_entity(se) {
3102                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3103                 if (cfs_rq->last != se)
3104                         break;
3105 
3106                 cfs_rq->last = NULL;
3107         }
3108 }
3109 
3110 static void __clear_buddies_next(struct sched_entity *se)
3111 {
3112         for_each_sched_entity(se) {
3113                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3114                 if (cfs_rq->next != se)
3115                         break;
3116 
3117                 cfs_rq->next = NULL;
3118         }
3119 }
3120 
3121 static void __clear_buddies_skip(struct sched_entity *se)
3122 {
3123         for_each_sched_entity(se) {
3124                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3125                 if (cfs_rq->skip != se)
3126                         break;
3127 
3128                 cfs_rq->skip = NULL;
3129         }
3130 }
3131 
3132 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3133 {
3134         if (cfs_rq->last == se)
3135                 __clear_buddies_last(se);
3136 
3137         if (cfs_rq->next == se)
3138                 __clear_buddies_next(se);
3139 
3140         if (cfs_rq->skip == se)
3141                 __clear_buddies_skip(se);
3142 }
3143 
3144 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3145 
3146 static void
3147 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3148 {
3149         /*
3150          * Update run-time statistics of the 'current'.
3151          */
3152         update_curr(cfs_rq);
3153         dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
3154 
3155         update_stats_dequeue(cfs_rq, se);
3156         if (flags & DEQUEUE_SLEEP) {
3157 #ifdef CONFIG_SCHEDSTATS
3158                 if (entity_is_task(se)) {
3159                         struct task_struct *tsk = task_of(se);
3160 
3161                         if (tsk->state & TASK_INTERRUPTIBLE)
3162                                 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3163                         if (tsk->state & TASK_UNINTERRUPTIBLE)
3164                                 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3165                 }
3166 #endif
3167         }
3168 
3169         clear_buddies(cfs_rq, se);
3170 
3171         if (se != cfs_rq->curr)
3172                 __dequeue_entity(cfs_rq, se);
3173         se->on_rq = 0;
3174         account_entity_dequeue(cfs_rq, se);
3175 
3176         /*
3177          * Normalize the entity after updating the min_vruntime because the
3178          * update can refer to the ->curr item and we need to reflect this
3179          * movement in our normalized position.
3180          */
3181         if (!(flags & DEQUEUE_SLEEP))
3182                 se->vruntime -= cfs_rq->min_vruntime;
3183 
3184         /* return excess runtime on last dequeue */
3185         return_cfs_rq_runtime(cfs_rq);
3186 
3187         update_min_vruntime(cfs_rq);
3188         update_cfs_shares(cfs_rq);
3189 }
3190 
3191 /*
3192  * Preempt the current task with a newly woken task if needed:
3193  */
3194 static void
3195 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3196 {
3197         unsigned long ideal_runtime, delta_exec;
3198         struct sched_entity *se;
3199         s64 delta;
3200 
3201         ideal_runtime = sched_slice(cfs_rq, curr);
3202         delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3203         if (delta_exec > ideal_runtime) {
3204                 resched_curr(rq_of(cfs_rq));
3205                 /*
3206                  * The current task ran long enough, ensure it doesn't get
3207                  * re-elected due to buddy favours.
3208                  */
3209                 clear_buddies(cfs_rq, curr);
3210                 return;
3211         }
3212 
3213         /*
3214          * Ensure that a task that missed wakeup preemption by a
3215          * narrow margin doesn't have to wait for a full slice.
3216          * This also mitigates buddy induced latencies under load.
3217          */
3218         if (delta_exec < sysctl_sched_min_granularity)
3219                 return;
3220 
3221         se = __pick_first_entity(cfs_rq);
3222         delta = curr->vruntime - se->vruntime;
3223 
3224         if (delta < 0)
3225                 return;
3226 
3227         if (delta > ideal_runtime)
3228                 resched_curr(rq_of(cfs_rq));
3229 }
3230 
3231 static void
3232 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3233 {
3234         /* 'current' is not kept within the tree. */
3235         if (se->on_rq) {
3236                 /*
3237                  * Any task has to be enqueued before it get to execute on
3238                  * a CPU. So account for the time it spent waiting on the
3239                  * runqueue.
3240                  */
3241                 update_stats_wait_end(cfs_rq, se);
3242                 __dequeue_entity(cfs_rq, se);
3243                 update_entity_load_avg(se, 1);
3244         }
3245 
3246         update_stats_curr_start(cfs_rq, se);
3247         cfs_rq->curr = se;
3248 #ifdef CONFIG_SCHEDSTATS
3249         /*
3250          * Track our maximum slice length, if the CPU's load is at
3251          * least twice that of our own weight (i.e. dont track it
3252          * when there are only lesser-weight tasks around):
3253          */
3254         if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3255                 se->statistics.slice_max = max(se->statistics.slice_max,
3256                         se->sum_exec_runtime - se->prev_sum_exec_runtime);
3257         }
3258 #endif
3259         se->prev_sum_exec_runtime = se->sum_exec_runtime;
3260 }
3261 
3262 static int
3263 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3264 
3265 /*
3266  * Pick the next process, keeping these things in mind, in this order:
3267  * 1) keep things fair between processes/task groups
3268  * 2) pick the "next" process, since someone really wants that to run
3269  * 3) pick the "last" process, for cache locality
3270  * 4) do not run the "skip" process, if something else is available
3271  */
3272 static struct sched_entity *
3273 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3274 {
3275         struct sched_entity *left = __pick_first_entity(cfs_rq);
3276         struct sched_entity *se;
3277 
3278         /*
3279          * If curr is set we have to see if its left of the leftmost entity
3280          * still in the tree, provided there was anything in the tree at all.
3281          */
3282         if (!left || (curr && entity_before(curr, left)))
3283                 left = curr;
3284 
3285         se = left; /* ideally we run the leftmost entity */
3286 
3287         /*
3288          * Avoid running the skip buddy, if running something else can
3289          * be done without getting too unfair.
3290          */
3291         if (cfs_rq->skip == se) {
3292                 struct sched_entity *second;
3293 
3294                 if (se == curr) {
3295                         second = __pick_first_entity(cfs_rq);
3296                 } else {
3297                         second = __pick_next_entity(se);
3298                         if (!second || (curr && entity_before(curr, second)))
3299                                 second = curr;
3300                 }
3301 
3302                 if (second && wakeup_preempt_entity(second, left) < 1)
3303                         se = second;
3304         }
3305 
3306         /*
3307          * Prefer last buddy, try to return the CPU to a preempted task.
3308          */
3309         if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3310                 se = cfs_rq->last;
3311 
3312         /*
3313          * Someone really wants this to run. If it's not unfair, run it.
3314          */
3315         if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3316                 se = cfs_rq->next;
3317 
3318         clear_buddies(cfs_rq, se);
3319 
3320         return se;
3321 }
3322 
3323 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3324 
3325 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3326 {
3327         /*
3328          * If still on the runqueue then deactivate_task()
3329          * was not called and update_curr() has to be done:
3330          */
3331         if (prev->on_rq)
3332                 update_curr(cfs_rq);
3333 
3334         /* throttle cfs_rqs exceeding runtime */
3335         check_cfs_rq_runtime(cfs_rq);
3336 
3337         check_spread(cfs_rq, prev);
3338         if (prev->on_rq) {
3339                 update_stats_wait_start(cfs_rq, prev);
3340                 /* Put 'current' back into the tree. */
3341                 __enqueue_entity(cfs_rq, prev);
3342                 /* in !on_rq case, update occurred at dequeue */
3343                 update_entity_load_avg(prev, 1);
3344         }
3345         cfs_rq->curr = NULL;
3346 }
3347 
3348 static void
3349 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3350 {
3351         /*
3352          * Update run-time statistics of the 'current'.
3353          */
3354         update_curr(cfs_rq);
3355 
3356         /*
3357          * Ensure that runnable average is periodically updated.
3358          */
3359         update_entity_load_avg(curr, 1);
3360         update_cfs_rq_blocked_load(cfs_rq, 1);
3361         update_cfs_shares(cfs_rq);
3362 
3363 #ifdef CONFIG_SCHED_HRTICK
3364         /*
3365          * queued ticks are scheduled to match the slice, so don't bother
3366          * validating it and just reschedule.
3367          */
3368         if (queued) {
3369                 resched_curr(rq_of(cfs_rq));
3370                 return;
3371         }
3372         /*
3373          * don't let the period tick interfere with the hrtick preemption
3374          */
3375         if (!sched_feat(DOUBLE_TICK) &&
3376                         hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3377                 return;
3378 #endif
3379 
3380         if (cfs_rq->nr_running > 1)
3381                 check_preempt_tick(cfs_rq, curr);
3382 }
3383 
3384 
3385 /**************************************************
3386  * CFS bandwidth control machinery
3387  */
3388 
3389 #ifdef CONFIG_CFS_BANDWIDTH
3390 
3391 #ifdef HAVE_JUMP_LABEL
3392 static struct static_key __cfs_bandwidth_used;
3393 
3394 static inline bool cfs_bandwidth_used(void)
3395 {
3396         return static_key_false(&__cfs_bandwidth_used);
3397 }
3398 
3399 void cfs_bandwidth_usage_inc(void)
3400 {
3401         static_key_slow_inc(&__cfs_bandwidth_used);
3402 }
3403 
3404 void cfs_bandwidth_usage_dec(void)
3405 {
3406         static_key_slow_dec(&__cfs_bandwidth_used);
3407 }
3408 #else /* HAVE_JUMP_LABEL */
3409 static bool cfs_bandwidth_used(void)
3410 {
3411         return true;
3412 }
3413 
3414 void cfs_bandwidth_usage_inc(void) {}
3415 void cfs_bandwidth_usage_dec(void) {}
3416 #endif /* HAVE_JUMP_LABEL */
3417 
3418 /*
3419  * default period for cfs group bandwidth.
3420  * default: 0.1s, units: nanoseconds
3421  */
3422 static inline u64 default_cfs_period(void)
3423 {
3424         return 100000000ULL;
3425 }
3426 
3427 static inline u64 sched_cfs_bandwidth_slice(void)
3428 {
3429         return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3430 }
3431 
3432 /*
3433  * Replenish runtime according to assigned quota and update expiration time.
3434  * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3435  * additional synchronization around rq->lock.
3436  *
3437  * requires cfs_b->lock
3438  */
3439 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3440 {
3441         u64 now;
3442 
3443         if (cfs_b->quota == RUNTIME_INF)
3444                 return;
3445 
3446         now = sched_clock_cpu(smp_processor_id());
3447         cfs_b->runtime = cfs_b->quota;
3448         cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3449 }
3450 
3451 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3452 {
3453         return &tg->cfs_bandwidth;
3454 }
3455 
3456 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3457 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3458 {
3459         if (unlikely(cfs_rq->throttle_count))
3460                 return cfs_rq->throttled_clock_task;
3461 
3462         return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3463 }
3464 
3465 /* returns 0 on failure to allocate runtime */
3466 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3467 {
3468         struct task_group *tg = cfs_rq->tg;
3469         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3470         u64 amount = 0, min_amount, expires;
3471 
3472         /* note: this is a positive sum as runtime_remaining <= 0 */
3473         min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3474 
3475         raw_spin_lock(&cfs_b->lock);
3476         if (cfs_b->quota == RUNTIME_INF)
3477                 amount = min_amount;
3478         else {
3479                 /*
3480                  * If the bandwidth pool has become inactive, then at least one
3481                  * period must have elapsed since the last consumption.
3482                  * Refresh the global state and ensure bandwidth timer becomes
3483                  * active.
3484                  */
3485                 if (!cfs_b->timer_active) {
3486                         __refill_cfs_bandwidth_runtime(cfs_b);
3487                         __start_cfs_bandwidth(cfs_b, false);
3488                 }
3489 
3490                 if (cfs_b->runtime > 0) {
3491                         amount = min(cfs_b->runtime, min_amount);
3492                         cfs_b->runtime -= amount;
3493                         cfs_b->idle = 0;
3494                 }
3495         }
3496         expires = cfs_b->runtime_expires;
3497         raw_spin_unlock(&cfs_b->lock);
3498 
3499         cfs_rq->runtime_remaining += amount;
3500         /*
3501          * we may have advanced our local expiration to account for allowed
3502          * spread between our sched_clock and the one on which runtime was
3503          * issued.
3504          */
3505         if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3506                 cfs_rq->runtime_expires = expires;
3507 
3508         return cfs_rq->runtime_remaining > 0;
3509 }
3510 
3511 /*
3512  * Note: This depends on the synchronization provided by sched_clock and the
3513  * fact that rq->clock snapshots this value.
3514  */
3515 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3516 {
3517         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3518 
3519         /* if the deadline is ahead of our clock, nothing to do */
3520         if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3521                 return;
3522 
3523         if (cfs_rq->runtime_remaining < 0)
3524                 return;
3525 
3526         /*
3527          * If the local deadline has passed we have to consider the
3528          * possibility that our sched_clock is 'fast' and the global deadline
3529          * has not truly expired.
3530          *
3531          * Fortunately we can check determine whether this the case by checking
3532          * whether the global deadline has advanced. It is valid to compare
3533          * cfs_b->runtime_expires without any locks since we only care about
3534          * exact equality, so a partial write will still work.
3535          */
3536 
3537         if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3538                 /* extend local deadline, drift is bounded above by 2 ticks */
3539                 cfs_rq->runtime_expires += TICK_NSEC;
3540         } else {
3541                 /* global deadline is ahead, expiration has passed */
3542                 cfs_rq->runtime_remaining = 0;
3543         }
3544 }
3545 
3546 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3547 {
3548         /* dock delta_exec before expiring quota (as it could span periods) */
3549         cfs_rq->runtime_remaining -= delta_exec;
3550         expire_cfs_rq_runtime(cfs_rq);
3551 
3552         if (likely(cfs_rq->runtime_remaining > 0))
3553                 return;
3554 
3555         /*
3556          * if we're unable to extend our runtime we resched so that the active
3557          * hierarchy can be throttled
3558          */
3559         if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3560                 resched_curr(rq_of(cfs_rq));
3561 }
3562 
3563 static __always_inline
3564 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3565 {
3566         if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3567                 return;
3568 
3569         __account_cfs_rq_runtime(cfs_rq, delta_exec);
3570 }
3571 
3572 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3573 {
3574         return cfs_bandwidth_used() && cfs_rq->throttled;
3575 }
3576 
3577 /* check whether cfs_rq, or any parent, is throttled */
3578 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3579 {
3580         return cfs_bandwidth_used() && cfs_rq->throttle_count;
3581 }
3582 
3583 /*
3584  * Ensure that neither of the group entities corresponding to src_cpu or
3585  * dest_cpu are members of a throttled hierarchy when performing group
3586  * load-balance operations.
3587  */
3588 static inline int throttled_lb_pair(struct task_group *tg,
3589                                     int src_cpu, int dest_cpu)
3590 {
3591         struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3592 
3593         src_cfs_rq = tg->cfs_rq[src_cpu];
3594         dest_cfs_rq = tg->cfs_rq[dest_cpu];
3595 
3596         return throttled_hierarchy(src_cfs_rq) ||
3597                throttled_hierarchy(dest_cfs_rq);
3598 }
3599 
3600 /* updated child weight may affect parent so we have to do this bottom up */
3601 static int tg_unthrottle_up(struct task_group *tg, void *data)
3602 {
3603         struct rq *rq = data;
3604         struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3605 
3606         cfs_rq->throttle_count--;
3607 #ifdef CONFIG_SMP
3608         if (!cfs_rq->throttle_count) {
3609                 /* adjust cfs_rq_clock_task() */
3610                 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3611                                              cfs_rq->throttled_clock_task;
3612         }
3613 #endif
3614 
3615         return 0;
3616 }
3617 
3618 static int tg_throttle_down(struct task_group *tg, void *data)
3619 {
3620         struct rq *rq = data;
3621         struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3622 
3623         /* group is entering throttled state, stop time */
3624         if (!cfs_rq->throttle_count)
3625                 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3626         cfs_rq->throttle_count++;
3627 
3628         return 0;
3629 }
3630 
3631 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3632 {
3633         struct rq *rq = rq_of(cfs_rq);
3634         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3635         struct sched_entity *se;
3636         long task_delta, dequeue = 1;
3637 
3638         se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3639 
3640         /* freeze hierarchy runnable averages while throttled */
3641         rcu_read_lock();
3642         walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3643         rcu_read_unlock();
3644 
3645         task_delta = cfs_rq->h_nr_running;
3646         for_each_sched_entity(se) {
3647                 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3648                 /* throttled entity or throttle-on-deactivate */
3649                 if (!se->on_rq)
3650                         break;
3651 
3652                 if (dequeue)
3653                         dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3654                 qcfs_rq->h_nr_running -= task_delta;
3655 
3656                 if (qcfs_rq->load.weight)
3657                         dequeue = 0;
3658         }
3659 
3660         if (!se)
3661                 sub_nr_running(rq, task_delta);
3662 
3663         cfs_rq->throttled = 1;
3664         cfs_rq->throttled_clock = rq_clock(rq);
3665         raw_spin_lock(&cfs_b->lock);
3666         /*
3667          * Add to the _head_ of the list, so that an already-started
3668          * distribute_cfs_runtime will not see us
3669          */
3670         list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3671         if (!cfs_b->timer_active)
3672                 __start_cfs_bandwidth(cfs_b, false);
3673         raw_spin_unlock(&cfs_b->lock);
3674 }
3675 
3676 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3677 {
3678         struct rq *rq = rq_of(cfs_rq);
3679         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3680         struct sched_entity *se;
3681         int enqueue = 1;
3682         long task_delta;
3683 
3684         se = cfs_rq->tg->se[cpu_of(rq)];
3685 
3686         cfs_rq->throttled = 0;
3687 
3688         update_rq_clock(rq);
3689 
3690         raw_spin_lock(&cfs_b->lock);
3691         cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3692         list_del_rcu(&cfs_rq->throttled_list);
3693         raw_spin_unlock(&cfs_b->lock);
3694 
3695         /* update hierarchical throttle state */
3696         walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3697 
3698         if (!cfs_rq->load.weight)
3699                 return;
3700 
3701         task_delta = cfs_rq->h_nr_running;
3702         for_each_sched_entity(se) {
3703                 if (se->on_rq)
3704                         enqueue = 0;
3705 
3706                 cfs_rq = cfs_rq_of(se);
3707                 if (enqueue)
3708                         enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3709                 cfs_rq->h_nr_running += task_delta;
3710 
3711                 if (cfs_rq_throttled(cfs_rq))
3712                         break;
3713         }
3714 
3715         if (!se)
3716                 add_nr_running(rq, task_delta);
3717 
3718         /* determine whether we need to wake up potentially idle cpu */
3719         if (rq->curr == rq->idle && rq->cfs.nr_running)
3720                 resched_curr(rq);
3721 }
3722 
3723 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3724                 u64 remaining, u64 expires)
3725 {
3726         struct cfs_rq *cfs_rq;
3727         u64 runtime;
3728         u64 starting_runtime = remaining;
3729 
3730         rcu_read_lock();
3731         list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3732                                 throttled_list) {
3733                 struct rq *rq = rq_of(cfs_rq);
3734 
3735                 raw_spin_lock(&rq->lock);
3736                 if (!cfs_rq_throttled(cfs_rq))
3737                         goto next;
3738 
3739                 runtime = -cfs_rq->runtime_remaining + 1;
3740                 if (runtime > remaining)
3741                         runtime = remaining;
3742                 remaining -= runtime;
3743 
3744                 cfs_rq->runtime_remaining += runtime;
3745                 cfs_rq->runtime_expires = expires;
3746 
3747                 /* we check whether we're throttled above */
3748                 if (cfs_rq->runtime_remaining > 0)
3749                         unthrottle_cfs_rq(cfs_rq);
3750 
3751 next:
3752                 raw_spin_unlock(&rq->lock);
3753 
3754                 if (!remaining)
3755                         break;
3756         }
3757         rcu_read_unlock();
3758 
3759         return starting_runtime - remaining;
3760 }
3761 
3762 /*
3763  * Responsible for refilling a task_group's bandwidth and unthrottling its
3764  * cfs_rqs as appropriate. If there has been no activity within the last
3765  * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3766  * used to track this state.
3767  */
3768 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3769 {
3770         u64 runtime, runtime_expires;
3771         int throttled;
3772 
3773         /* no need to continue the timer with no bandwidth constraint */
3774         if (cfs_b->quota == RUNTIME_INF)
3775                 goto out_deactivate;
3776 
3777         throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3778         cfs_b->nr_periods += overrun;
3779 
3780         /*
3781          * idle depends on !throttled (for the case of a large deficit), and if
3782          * we're going inactive then everything else can be deferred
3783          */
3784         if (cfs_b->idle && !throttled)
3785                 goto out_deactivate;
3786 
3787         /*
3788          * if we have relooped after returning idle once, we need to update our
3789          * status as actually running, so that other cpus doing
3790          * __start_cfs_bandwidth will stop trying to cancel us.
3791          */
3792         cfs_b->timer_active = 1;
3793 
3794         __refill_cfs_bandwidth_runtime(cfs_b);
3795 
3796         if (!throttled) {
3797                 /* mark as potentially idle for the upcoming period */
3798                 cfs_b->idle = 1;
3799                 return 0;
3800         }
3801 
3802         /* account preceding periods in which throttling occurred */
3803         cfs_b->nr_throttled += overrun;
3804 
3805         runtime_expires = cfs_b->runtime_expires;
3806 
3807         /*
3808          * This check is repeated as we are holding onto the new bandwidth while
3809          * we unthrottle. This can potentially race with an unthrottled group
3810          * trying to acquire new bandwidth from the global pool. This can result
3811          * in us over-using our runtime if it is all used during this loop, but
3812          * only by limited amounts in that extreme case.
3813          */
3814         while (throttled && cfs_b->runtime > 0) {
3815                 runtime = cfs_b->runtime;
3816                 raw_spin_unlock(&cfs_b->lock);
3817                 /* we can't nest cfs_b->lock while distributing bandwidth */
3818                 runtime = distribute_cfs_runtime(cfs_b, runtime,
3819                                                  runtime_expires);
3820                 raw_spin_lock(&cfs_b->lock);
3821 
3822                 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3823 
3824                 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3825         }
3826 
3827         /*
3828          * While we are ensured activity in the period following an
3829          * unthrottle, this also covers the case in which the new bandwidth is
3830          * insufficient to cover the existing bandwidth deficit.  (Forcing the
3831          * timer to remain active while there are any throttled entities.)
3832          */
3833         cfs_b->idle = 0;
3834 
3835         return 0;
3836 
3837 out_deactivate:
3838         cfs_b->timer_active = 0;
3839         return 1;
3840 }
3841 
3842 /* a cfs_rq won't donate quota below this amount */
3843 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3844 /* minimum remaining period time to redistribute slack quota */
3845 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3846 /* how long we wait to gather additional slack before distributing */
3847 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3848 
3849 /*
3850  * Are we near the end of the current quota period?
3851  *
3852  * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3853  * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3854  * migrate_hrtimers, base is never cleared, so we are fine.
3855  */
3856 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3857 {
3858         struct hrtimer *refresh_timer = &cfs_b->period_timer;
3859         u64 remaining;
3860 
3861         /* if the call-back is running a quota refresh is already occurring */
3862         if (hrtimer_callback_running(refresh_timer))
3863                 return 1;
3864 
3865         /* is a quota refresh about to occur? */
3866         remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3867         if (remaining < min_expire)
3868                 return 1;
3869 
3870         return 0;
3871 }
3872 
3873 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3874 {
3875         u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3876 
3877         /* if there's a quota refresh soon don't bother with slack */
3878         if (runtime_refresh_within(cfs_b, min_left))
3879                 return;
3880 
3881         start_bandwidth_timer(&cfs_b->slack_timer,
3882                                 ns_to_ktime(cfs_bandwidth_slack_period));
3883 }
3884 
3885 /* we know any runtime found here is valid as update_curr() precedes return */
3886 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3887 {
3888         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3889         s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3890 
3891         if (slack_runtime <= 0)
3892                 return;
3893 
3894         raw_spin_lock(&cfs_b->lock);
3895         if (cfs_b->quota != RUNTIME_INF &&
3896             cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3897                 cfs_b->runtime += slack_runtime;
3898 
3899                 /* we are under rq->lock, defer unthrottling using a timer */
3900                 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3901                     !list_empty(&cfs_b->throttled_cfs_rq))
3902                         start_cfs_slack_bandwidth(cfs_b);
3903         }
3904         raw_spin_unlock(&cfs_b->lock);
3905 
3906         /* even if it's not valid for return we don't want to try again */
3907         cfs_rq->runtime_remaining -= slack_runtime;
3908 }
3909 
3910 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3911 {
3912         if (!cfs_bandwidth_used())
3913                 return;
3914 
3915         if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3916                 return;
3917 
3918         __return_cfs_rq_runtime(cfs_rq);
3919 }
3920 
3921 /*
3922  * This is done with a timer (instead of inline with bandwidth return) since
3923  * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3924  */
3925 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3926 {
3927         u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3928         u64 expires;
3929 
3930         /* confirm we're still not at a refresh boundary */
3931         raw_spin_lock(&cfs_b->lock);
3932         if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3933                 raw_spin_unlock(&cfs_b->lock);
3934                 return;
3935         }
3936 
3937         if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3938                 runtime = cfs_b->runtime;
3939 
3940         expires = cfs_b->runtime_expires;
3941         raw_spin_unlock(&cfs_b->lock);
3942 
3943         if (!runtime)
3944                 return;
3945 
3946         runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3947 
3948         raw_spin_lock(&cfs_b->lock);
3949         if (expires == cfs_b->runtime_expires)
3950                 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3951         raw_spin_unlock(&cfs_b->lock);
3952 }
3953 
3954 /*
3955  * When a group wakes up we want to make sure that its quota is not already
3956  * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3957  * runtime as update_curr() throttling can not not trigger until it's on-rq.
3958  */
3959 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3960 {
3961         if (!cfs_bandwidth_used())
3962                 return;
3963 
3964         /* an active group must be handled by the update_curr()->put() path */
3965         if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3966                 return;
3967 
3968         /* ensure the group is not already throttled */
3969         if (cfs_rq_throttled(cfs_rq))
3970                 return;
3971 
3972         /* update runtime allocation */
3973         account_cfs_rq_runtime(cfs_rq, 0);
3974         if (cfs_rq->runtime_remaining <= 0)
3975                 throttle_cfs_rq(cfs_rq);
3976 }
3977 
3978 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3979 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3980 {
3981         if (!cfs_bandwidth_used())
3982                 return false;
3983 
3984         if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3985                 return false;
3986 
3987         /*
3988          * it's possible for a throttled entity to be forced into a running
3989          * state (e.g. set_curr_task), in this case we're finished.
3990          */
3991         if (cfs_rq_throttled(cfs_rq))
3992                 return true;
3993 
3994         throttle_cfs_rq(cfs_rq);
3995         return true;
3996 }
3997 
3998 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3999 {
4000         struct cfs_bandwidth *cfs_b =
4001                 container_of(timer, struct cfs_bandwidth, slack_timer);
4002         do_sched_cfs_slack_timer(cfs_b);
4003 
4004         return HRTIMER_NORESTART;
4005 }
4006 
4007 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4008 {
4009         struct cfs_bandwidth *cfs_b =
4010                 container_of(timer, struct cfs_bandwidth, period_timer);
4011         ktime_t now;
4012         int overrun;
4013         int idle = 0;
4014 
4015         raw_spin_lock(&cfs_b->lock);
4016         for (;;) {
4017                 now = hrtimer_cb_get_time(timer);
4018                 overrun = hrtimer_forward(timer, now, cfs_b->period);
4019 
4020                 if (!overrun)
4021                         break;
4022 
4023                 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4024         }
4025         raw_spin_unlock(&cfs_b->lock);
4026 
4027         return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4028 }
4029 
4030 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4031 {
4032         raw_spin_lock_init(&cfs_b->lock);
4033         cfs_b->runtime = 0;
4034         cfs_b->quota = RUNTIME_INF;
4035         cfs_b->period = ns_to_ktime(default_cfs_period());
4036 
4037         INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4038         hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4039         cfs_b->period_timer.function = sched_cfs_period_timer;
4040         hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4041         cfs_b->slack_timer.function = sched_cfs_slack_timer;
4042 }
4043 
4044 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4045 {
4046         cfs_rq->runtime_enabled = 0;
4047         INIT_LIST_HEAD(&cfs_rq->throttled_list);
4048 }
4049 
4050 /* requires cfs_b->lock, may release to reprogram timer */
4051 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b, bool force)
4052 {
4053         /*
4054          * The timer may be active because we're trying to set a new bandwidth
4055          * period or because we're racing with the tear-down path
4056          * (timer_active==0 becomes visible before the hrtimer call-back
4057          * terminates).  In either case we ensure that it's re-programmed
4058          */
4059         while (unlikely(hrtimer_active(&cfs_b->period_timer)) &&
4060                hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) {
4061                 /* bounce the lock to allow do_sched_cfs_period_timer to run */
4062                 raw_spin_unlock(&cfs_b->lock);
4063                 cpu_relax();
4064                 raw_spin_lock(&cfs_b->lock);
4065                 /* if someone else restarted the timer then we're done */
4066                 if (!force && cfs_b->timer_active)
4067                         return;
4068         }
4069 
4070         cfs_b->timer_active = 1;
4071         start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
4072 }
4073 
4074 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4075 {
4076         /* init_cfs_bandwidth() was not called */
4077         if (!cfs_b->throttled_cfs_rq.next)
4078                 return;
4079 
4080         hrtimer_cancel(&cfs_b->period_timer);
4081         hrtimer_cancel(&cfs_b->slack_timer);
4082 }
4083 
4084 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4085 {
4086         struct cfs_rq *cfs_rq;
4087 
4088         for_each_leaf_cfs_rq(rq, cfs_rq) {
4089                 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4090 
4091                 raw_spin_lock(&cfs_b->lock);
4092                 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4093                 raw_spin_unlock(&cfs_b->lock);
4094         }
4095 }
4096 
4097 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4098 {
4099         struct cfs_rq *cfs_rq;
4100 
4101         for_each_leaf_cfs_rq(rq, cfs_rq) {
4102                 if (!cfs_rq->runtime_enabled)
4103                         continue;
4104 
4105                 /*
4106                  * clock_task is not advancing so we just need to make sure
4107                  * there's some valid quota amount
4108                  */
4109                 cfs_rq->runtime_remaining = 1;
4110                 /*
4111                  * Offline rq is schedulable till cpu is completely disabled
4112                  * in take_cpu_down(), so we prevent new cfs throttling here.
4113                  */
4114                 cfs_rq->runtime_enabled = 0;
4115 
4116                 if (cfs_rq_throttled(cfs_rq))
4117                         unthrottle_cfs_rq(cfs_rq);
4118         }
4119 }
4120 
4121 #else /* CONFIG_CFS_BANDWIDTH */
4122 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4123 {
4124         return rq_clock_task(rq_of(cfs_rq));
4125 }
4126 
4127 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4128 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4129 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4130 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4131 
4132 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4133 {
4134         return 0;
4135 }
4136 
4137 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4138 {
4139         return 0;
4140 }
4141 
4142 static inline int throttled_lb_pair(struct task_group *tg,
4143                                     int src_cpu, int dest_cpu)
4144 {
4145         return 0;
4146 }
4147 
4148 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4149 
4150 #ifdef CONFIG_FAIR_GROUP_SCHED
4151 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4152 #endif
4153 
4154 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4155 {
4156         return NULL;
4157 }
4158 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4159 static inline void update_runtime_enabled(struct rq *rq) {}
4160 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4161 
4162 #endif /* CONFIG_CFS_BANDWIDTH */
4163 
4164 /**************************************************
4165  * CFS operations on tasks:
4166  */
4167 
4168 #ifdef CONFIG_SCHED_HRTICK
4169 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4170 {
4171         struct sched_entity *se = &p->se;
4172         struct cfs_rq *cfs_rq = cfs_rq_of(se);
4173 
4174         WARN_ON(task_rq(p) != rq);
4175 
4176         if (cfs_rq->nr_running > 1) {
4177                 u64 slice = sched_slice(cfs_rq, se);
4178                 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4179                 s64 delta = slice - ran;
4180 
4181                 if (delta < 0) {
4182                         if (rq->curr == p)
4183                                 resched_curr(rq);
4184                         return;
4185                 }
4186                 hrtick_start(rq, delta);
4187         }
4188 }
4189 
4190 /*
4191  * called from enqueue/dequeue and updates the hrtick when the
4192  * current task is from our class and nr_running is low enough
4193  * to matter.
4194  */
4195 static void hrtick_update(struct rq *rq)
4196 {
4197         struct task_struct *curr = rq->curr;
4198 
4199         if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4200                 return;
4201 
4202         if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4203                 hrtick_start_fair(rq, curr);
4204 }
4205 #else /* !CONFIG_SCHED_HRTICK */
4206 static inline void
4207 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4208 {
4209 }
4210 
4211 static inline void hrtick_update(struct rq *rq)
4212 {
4213 }
4214 #endif
4215 
4216 /*
4217  * The enqueue_task method is called before nr_running is
4218  * increased. Here we update the fair scheduling stats and
4219  * then put the task into the rbtree:
4220  */
4221 static void
4222 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4223 {
4224         struct cfs_rq *cfs_rq;
4225         struct sched_entity *se = &p->se;
4226 
4227         for_each_sched_entity(se) {
4228                 if (se->on_rq)
4229                         break;
4230                 cfs_rq = cfs_rq_of(se);
4231                 enqueue_entity(cfs_rq, se, flags);
4232 
4233                 /*
4234                  * end evaluation on encountering a throttled cfs_rq
4235                  *
4236                  * note: in the case of encountering a throttled cfs_rq we will
4237                  * post the final h_nr_running increment below.
4238                 */
4239                 if (cfs_rq_throttled(cfs_rq))
4240                         break;
4241                 cfs_rq->h_nr_running++;
4242 
4243                 flags = ENQUEUE_WAKEUP;
4244         }
4245 
4246         for_each_sched_entity(se) {
4247                 cfs_rq = cfs_rq_of(se);
4248                 cfs_rq->h_nr_running++;
4249 
4250                 if (cfs_rq_throttled(cfs_rq))
4251                         break;
4252 
4253                 update_cfs_shares(cfs_rq);
4254                 update_entity_load_avg(se, 1);
4255         }
4256 
4257         if (!se) {
4258                 update_rq_runnable_avg(rq, rq->nr_running);
4259                 add_nr_running(rq, 1);
4260         }
4261         hrtick_update(rq);
4262 }
4263 
4264 static void set_next_buddy(struct sched_entity *se);
4265 
4266 /*
4267  * The dequeue_task method is called before nr_running is
4268  * decreased. We remove the task from the rbtree and
4269  * update the fair scheduling stats:
4270  */
4271 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4272 {
4273         struct cfs_rq *cfs_rq;
4274         struct sched_entity *se = &p->se;
4275         int task_sleep = flags & DEQUEUE_SLEEP;
4276 
4277         for_each_sched_entity(se) {
4278                 cfs_rq = cfs_rq_of(se);
4279                 dequeue_entity(cfs_rq, se, flags);
4280 
4281                 /*
4282                  * end evaluation on encountering a throttled cfs_rq
4283                  *
4284                  * note: in the case of encountering a throttled cfs_rq we will
4285                  * post the final h_nr_running decrement below.
4286                 */
4287                 if (cfs_rq_throttled(cfs_rq))
4288                         break;
4289                 cfs_rq->h_nr_running--;
4290 
4291                 /* Don't dequeue parent if it has other entities besides us */
4292                 if (cfs_rq->load.weight) {
4293                         /*
4294                          * Bias pick_next to pick a task from this cfs_rq, as
4295                          * p is sleeping when it is within its sched_slice.
4296                          */
4297                         if (task_sleep && parent_entity(se))
4298                                 set_next_buddy(parent_entity(se));
4299 
4300                         /* avoid re-evaluating load for this entity */
4301                         se = parent_entity(se);
4302                         break;
4303                 }
4304                 flags |= DEQUEUE_SLEEP;
4305         }
4306 
4307         for_each_sched_entity(se) {
4308                 cfs_rq = cfs_rq_of(se);
4309                 cfs_rq->h_nr_running--;
4310 
4311                 if (cfs_rq_throttled(cfs_rq))
4312                         break;
4313 
4314                 update_cfs_shares(cfs_rq);
4315                 update_entity_load_avg(se, 1);
4316         }
4317 
4318         if (!se) {
4319                 sub_nr_running(rq, 1);
4320                 update_rq_runnable_avg(rq, 1);
4321         }
4322         hrtick_update(rq);
4323 }
4324 
4325 #ifdef CONFIG_SMP
4326 /* Used instead of source_load when we know the type == 0 */
4327 static unsigned long weighted_cpuload(const int cpu)
4328 {
4329         return cpu_rq(cpu)->cfs.runnable_load_avg;
4330 }
4331 
4332 /*
4333  * Return a low guess at the load of a migration-source cpu weighted
4334  * according to the scheduling class and "nice" value.
4335  *
4336  * We want to under-estimate the load of migration sources, to
4337  * balance conservatively.
4338  */
4339 static unsigned long source_load(int cpu, int type)
4340 {
4341         struct rq *rq = cpu_rq(cpu);
4342         unsigned long total = weighted_cpuload(cpu);
4343 
4344         if (type == 0 || !sched_feat(LB_BIAS))
4345                 return total;
4346 
4347         return min(rq->cpu_load[type-1], total);
4348 }
4349 
4350 /*
4351  * Return a high guess at the load of a migration-target cpu weighted
4352  * according to the scheduling class and "nice" value.
4353  */
4354 static unsigned long target_load(int cpu, int type)
4355 {
4356         struct rq *rq = cpu_rq(cpu);
4357         unsigned long total = weighted_cpuload(cpu);
4358 
4359         if (type == 0 || !sched_feat(LB_BIAS))
4360                 return total;
4361 
4362         return max(rq->cpu_load[type-1], total);
4363 }
4364 
4365 static unsigned long capacity_of(int cpu)
4366 {
4367         return cpu_rq(cpu)->cpu_capacity;
4368 }
4369 
4370 static unsigned long capacity_orig_of(int cpu)
4371 {
4372         return cpu_rq(cpu)->cpu_capacity_orig;
4373 }
4374 
4375 static unsigned long cpu_avg_load_per_task(int cpu)
4376 {
4377         struct rq *rq = cpu_rq(cpu);
4378         unsigned long nr_running = ACCESS_ONCE(rq->cfs.h_nr_running);
4379         unsigned long load_avg = rq->cfs.runnable_load_avg;
4380 
4381         if (nr_running)
4382                 return load_avg / nr_running;
4383 
4384         return 0;
4385 }
4386 
4387 static void record_wakee(struct task_struct *p)
4388 {
4389         /*
4390          * Rough decay (wiping) for cost saving, don't worry
4391          * about the boundary, really active task won't care
4392          * about the loss.
4393          */
4394         if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4395                 current->wakee_flips >>= 1;
4396                 current->wakee_flip_decay_ts = jiffies;
4397         }
4398 
4399         if (current->last_wakee != p) {
4400                 current->last_wakee = p;
4401                 current->wakee_flips++;
4402         }
4403 }
4404 
4405 static void task_waking_fair(struct task_struct *p)
4406 {
4407         struct sched_entity *se = &p->se;
4408         struct cfs_rq *cfs_rq = cfs_rq_of(se);
4409         u64 min_vruntime;
4410 
4411 #ifndef CONFIG_64BIT
4412         u64 min_vruntime_copy;
4413 
4414         do {
4415                 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4416                 smp_rmb();
4417                 min_vruntime = cfs_rq->min_vruntime;
4418         } while (min_vruntime != min_vruntime_copy);
4419 #else
4420         min_vruntime = cfs_rq->min_vruntime;
4421 #endif
4422 
4423         se->vruntime -= min_vruntime;
4424         record_wakee(p);
4425 }
4426 
4427 #ifdef CONFIG_FAIR_GROUP_SCHED
4428 /*
4429  * effective_load() calculates the load change as seen from the root_task_group
4430  *
4431  * Adding load to a group doesn't make a group heavier, but can cause movement
4432  * of group shares between cpus. Assuming the shares were perfectly aligned one
4433  * can calculate the shift in shares.
4434  *
4435  * Calculate the effective load difference if @wl is added (subtracted) to @tg
4436  * on this @cpu and results in a total addition (subtraction) of @wg to the
4437  * total group weight.
4438  *
4439  * Given a runqueue weight distribution (rw_i) we can compute a shares
4440  * distribution (s_i) using:
4441  *
4442  *   s_i = rw_i / \Sum rw_j                                             (1)
4443  *
4444  * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4445  * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4446  * shares distribution (s_i):
4447  *
4448  *   rw_i = {   2,   4,   1,   0 }
4449  *   s_i  = { 2/7, 4/7, 1/7,   0 }
4450  *
4451  * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4452  * task used to run on and the CPU the waker is running on), we need to
4453  * compute the effect of waking a task on either CPU and, in case of a sync
4454  * wakeup, compute the effect of the current task going to sleep.
4455  *
4456  * So for a change of @wl to the local @cpu with an overall group weight change
4457  * of @wl we can compute the new shares distribution (s'_i) using:
4458  *
4459  *   s'_i = (rw_i + @wl) / (@wg + \Sum rw_j)                            (2)
4460  *
4461  * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4462  * differences in waking a task to CPU 0. The additional task changes the
4463  * weight and shares distributions like:
4464  *
4465  *   rw'_i = {   3,   4,   1,   0 }
4466  *   s'_i  = { 3/8, 4/8, 1/8,   0 }
4467  *
4468  * We can then compute the difference in effective weight by using:
4469  *
4470  *   dw_i = S * (s'_i - s_i)                                            (3)
4471  *
4472  * Where 'S' is the group weight as seen by its parent.
4473  *
4474  * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4475  * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4476  * 4/7) times the weight of the group.
4477  */
4478 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4479 {
4480         struct sched_entity *se = tg->se[cpu];
4481 
4482         if (!tg->parent)        /* the trivial, non-cgroup case */
4483                 return wl;
4484 
4485         for_each_sched_entity(se) {
4486                 long w, W;
4487 
4488                 tg = se->my_q->tg;
4489 
4490                 /*
4491                  * W = @wg + \Sum rw_j
4492                  */
4493                 W = wg + calc_tg_weight(tg, se->my_q);
4494 
4495                 /*
4496                  * w = rw_i + @wl
4497                  */
4498                 w = se->my_q->load.weight + wl;
4499 
4500                 /*
4501                  * wl = S * s'_i; see (2)
4502                  */
4503                 if (W > 0 && w < W)
4504                         wl = (w * (long)tg->shares) / W;
4505                 else
4506                         wl = tg->shares;
4507 
4508                 /*
4509                  * Per the above, wl is the new se->load.weight value; since
4510                  * those are clipped to [MIN_SHARES, ...) do so now. See
4511                  * calc_cfs_shares().
4512                  */
4513                 if (wl < MIN_SHARES)
4514                         wl = MIN_SHARES;
4515 
4516                 /*
4517                  * wl = dw_i = S * (s'_i - s_i); see (3)
4518                  */
4519                 wl -= se->load.weight;
4520 
4521                 /*
4522                  * Recursively apply this logic to all parent groups to compute
4523                  * the final effective load change on the root group. Since
4524                  * only the @tg group gets extra weight, all parent groups can
4525                  * only redistribute existing shares. @wl is the shift in shares
4526                  * resulting from this level per the above.
4527                  */
4528                 wg = 0;
4529         }
4530 
4531         return wl;
4532 }
4533 #else
4534 
4535 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4536 {
4537         return wl;
4538 }
4539 
4540 #endif
4541 
4542 static int wake_wide(struct task_struct *p)
4543 {
4544         int factor = this_cpu_read(sd_llc_size);
4545 
4546         /*
4547          * Yeah, it's the switching-frequency, could means many wakee or
4548          * rapidly switch, use factor here will just help to automatically
4549          * adjust the loose-degree, so bigger node will lead to more pull.
4550          */
4551         if (p->wakee_flips > factor) {
4552                 /*
4553                  * wakee is somewhat hot, it needs certain amount of cpu
4554                  * resource, so if waker is far more hot, prefer to leave
4555                  * it alone.
4556                  */
4557                 if (current->wakee_flips > (factor * p->wakee_flips))
4558                         return 1;
4559         }
4560 
4561         return 0;
4562 }
4563 
4564 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4565 {
4566         s64 this_load, load;
4567         s64 this_eff_load, prev_eff_load;
4568         int idx, this_cpu, prev_cpu;
4569         struct task_group *tg;
4570         unsigned long weight;
4571         int balanced;
4572 
4573         /*
4574          * If we wake multiple tasks be careful to not bounce
4575          * ourselves around too much.
4576          */
4577         if (wake_wide(p))
4578                 return 0;
4579 
4580         idx       = sd->wake_idx;
4581         this_cpu  = smp_processor_id();
4582         prev_cpu  = task_cpu(p);
4583         load      = source_load(prev_cpu, idx);
4584         this_load = target_load(this_cpu, idx);
4585 
4586         /*
4587          * If sync wakeup then subtract the (maximum possible)
4588          * effect of the currently running task from the load
4589          * of the current CPU:
4590          */
4591         if (sync) {
4592                 tg = task_group(current);
4593                 weight = current->se.load.weight;
4594 
4595                 this_load += effective_load(tg, this_cpu, -weight, -weight);
4596                 load += effective_load(tg, prev_cpu, 0, -weight);
4597         }
4598 
4599         tg = task_group(p);
4600         weight = p->se.load.weight;
4601 
4602         /*
4603          * In low-load situations, where prev_cpu is idle and this_cpu is idle
4604          * due to the sync cause above having dropped this_load to 0, we'll
4605          * always have an imbalance, but there's really nothing you can do
4606          * about that, so that's good too.
4607          *
4608          * Otherwise check if either cpus are near enough in load to allow this
4609          * task to be woken on this_cpu.
4610          */
4611         this_eff_load = 100;
4612         this_eff_load *= capacity_of(prev_cpu);
4613 
4614         prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4615         prev_eff_load *= capacity_of(this_cpu);
4616 
4617         if (this_load > 0) {
4618                 this_eff_load *= this_load +
4619                         effective_load(tg, this_cpu, weight, weight);
4620 
4621                 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4622         }
4623 
4624         balanced = this_eff_load <= prev_eff_load;
4625 
4626         schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4627 
4628         if (!balanced)
4629                 return 0;
4630 
4631         schedstat_inc(sd, ttwu_move_affine);
4632         schedstat_inc(p, se.statistics.nr_wakeups_affine);
4633 
4634         return 1;
4635 }
4636 
4637 /*
4638  * find_idlest_group finds and returns the least busy CPU group within the
4639  * domain.
4640  */
4641 static struct sched_group *
4642 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4643                   int this_cpu, int sd_flag)
4644 {
4645         struct sched_group *idlest = NULL, *group = sd->groups;
4646         unsigned long min_load = ULONG_MAX, this_load = 0;
4647         int load_idx = sd->forkexec_idx;
4648         int imbalance = 100 + (sd->imbalance_pct-100)/2;
4649 
4650         if (sd_flag & SD_BALANCE_WAKE)
4651                 load_idx = sd->wake_idx;
4652 
4653         do {
4654                 unsigned long load, avg_load;
4655                 int local_group;
4656                 int i;
4657 
4658                 /* Skip over this group if it has no CPUs allowed */
4659                 if (!cpumask_intersects(sched_group_cpus(group),
4660                                         tsk_cpus_allowed(p)))
4661                         continue;
4662 
4663                 local_group = cpumask_test_cpu(this_cpu,
4664                                                sched_group_cpus(group));
4665 
4666                 /* Tally up the load of all CPUs in the group */
4667                 avg_load = 0;
4668 
4669                 for_each_cpu(i, sched_group_cpus(group)) {
4670                         /* Bias balancing toward cpus of our domain */
4671                         if (local_group)
4672                                 load = source_load(i, load_idx);
4673                         else
4674                                 load = target_load(i, load_idx);
4675 
4676                         avg_load += load;
4677                 }
4678 
4679                 /* Adjust by relative CPU capacity of the group */
4680                 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
4681 
4682                 if (local_group) {
4683                         this_load = avg_load;
4684                 } else if (avg_load < min_load) {
4685                         min_load = avg_load;
4686                         idlest = group;
4687                 }
4688         } while (group = group->next, group != sd->groups);
4689 
4690         if (!idlest || 100*this_load < imbalance*min_load)
4691                 return NULL;
4692         return idlest;
4693 }
4694 
4695 /*
4696  * find_idlest_cpu - find the idlest cpu among the cpus in group.
4697  */
4698 static int
4699 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4700 {
4701         unsigned long load, min_load = ULONG_MAX;
4702         unsigned int min_exit_latency = UINT_MAX;
4703         u64 latest_idle_timestamp = 0;
4704         int least_loaded_cpu = this_cpu;
4705         int shallowest_idle_cpu = -1;
4706         int i;
4707 
4708         /* Traverse only the allowed CPUs */
4709         for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4710                 if (idle_cpu(i)) {
4711                         struct rq *rq = cpu_rq(i);
4712                         struct cpuidle_state *idle = idle_get_state(rq);
4713                         if (idle && idle->exit_latency < min_exit_latency) {
4714                                 /*
4715                                  * We give priority to a CPU whose idle state
4716                                  * has the smallest exit latency irrespective
4717                                  * of any idle timestamp.
4718                                  */
4719                                 min_exit_latency = idle->exit_latency;
4720                                 latest_idle_timestamp = rq->idle_stamp;
4721                                 shallowest_idle_cpu = i;
4722                         } else if ((!idle || idle->exit_latency == min_exit_latency) &&
4723                                    rq->idle_stamp > latest_idle_timestamp) {
4724                                 /*
4725                                  * If equal or no active idle state, then
4726                                  * the most recently idled CPU might have
4727                                  * a warmer cache.
4728                                  */
4729                                 latest_idle_timestamp = rq->idle_stamp;
4730                                 shallowest_idle_cpu = i;
4731                         }
4732                 } else if (shallowest_idle_cpu == -1) {
4733                         load = weighted_cpuload(i);
4734                         if (load < min_load || (load == min_load && i == this_cpu)) {
4735                                 min_load = load;
4736                                 least_loaded_cpu = i;
4737                         }
4738                 }
4739         }
4740 
4741         return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
4742 }
4743 
4744 /*
4745  * Try and locate an idle CPU in the sched_domain.
4746  */
4747 static int select_idle_sibling(struct task_struct *p, int target)
4748 {
4749         struct sched_domain *sd;
4750         struct sched_group *sg;
4751         int i = task_cpu(p);
4752 
4753         if (idle_cpu(target))
4754                 return target;
4755 
4756         /*
4757          * If the prevous cpu is cache affine and idle, don't be stupid.
4758          */
4759         if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4760                 return i;
4761 
4762         /*
4763          * Otherwise, iterate the domains and find an elegible idle cpu.
4764          */
4765         sd = rcu_dereference(per_cpu(sd_llc, target));
4766         for_each_lower_domain(sd) {
4767                 sg = sd->groups;
4768                 do {
4769                         if (!cpumask_intersects(sched_group_cpus(sg),
4770                                                 tsk_cpus_allowed(p)))
4771                                 goto next;
4772 
4773                         for_each_cpu(i, sched_group_cpus(sg)) {
4774                                 if (i == target || !idle_cpu(i))
4775                                         goto next;
4776                         }
4777 
4778                         target = cpumask_first_and(sched_group_cpus(sg),
4779                                         tsk_cpus_allowed(p));
4780                         goto done;
4781 next:
4782                         sg = sg->next;
4783                 } while (sg != sd->groups);
4784         }
4785 done:
4786         return target;
4787 }
4788 /*
4789  * get_cpu_usage returns the amount of capacity of a CPU that is used by CFS
4790  * tasks. The unit of the return value must be the one of capacity so we can
4791  * compare the usage with the capacity of the CPU that is available for CFS
4792  * task (ie cpu_capacity).
4793  * cfs.utilization_load_avg is the sum of running time of runnable tasks on a
4794  * CPU. It represents the amount of utilization of a CPU in the range
4795  * [0..SCHED_LOAD_SCALE].  The usage of a CPU can't be higher than the full
4796  * capacity of the CPU because it's about the running time on this CPU.
4797  * Nevertheless, cfs.utilization_load_avg can be higher than SCHED_LOAD_SCALE
4798  * because of unfortunate rounding in avg_period and running_load_avg or just
4799  * after migrating tasks until the average stabilizes with the new running
4800  * time. So we need to check that the usage stays into the range
4801  * [0..cpu_capacity_orig] and cap if necessary.
4802  * Without capping the usage, a group could be seen as overloaded (CPU0 usage
4803  * at 121% + CPU1 usage at 80%) whereas CPU1 has 20% of available capacity
4804  */
4805 static int get_cpu_usage(int cpu)
4806 {
4807         unsigned long usage = cpu_rq(cpu)->cfs.utilization_load_avg;
4808         unsigned long capacity = capacity_orig_of(cpu);
4809 
4810         if (usage >= SCHED_LOAD_SCALE)
4811                 return capacity;
4812 
4813         return (usage * capacity) >> SCHED_LOAD_SHIFT;
4814 }
4815 
4816 /*
4817  * select_task_rq_fair: Select target runqueue for the waking task in domains
4818  * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4819  * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4820  *
4821  * Balances load by selecting the idlest cpu in the idlest group, or under
4822  * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4823  *
4824  * Returns the target cpu number.
4825  *
4826  * preempt must be disabled.
4827  */
4828 static int
4829 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4830 {
4831         struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4832         int cpu = smp_processor_id();
4833         int new_cpu = cpu;
4834         int want_affine = 0;
4835         int sync = wake_flags & WF_SYNC;
4836 
4837         if (sd_flag & SD_BALANCE_WAKE)
4838                 want_affine = cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
4839 
4840         rcu_read_lock();
4841         for_each_domain(cpu, tmp) {
4842                 if (!(tmp->flags & SD_LOAD_BALANCE))
4843                         continue;
4844 
4845                 /*
4846                  * If both cpu and prev_cpu are part of this domain,
4847                  * cpu is a valid SD_WAKE_AFFINE target.
4848                  */
4849                 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4850                     cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4851                         affine_sd = tmp;
4852                         break;
4853                 }
4854 
4855                 if (tmp->flags & sd_flag)
4856                         sd = tmp;
4857         }
4858 
4859         if (affine_sd && cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4860                 prev_cpu = cpu;
4861 
4862         if (sd_flag & SD_BALANCE_WAKE) {
4863                 new_cpu = select_idle_sibling(p, prev_cpu);
4864                 goto unlock;
4865         }
4866 
4867         while (sd) {
4868                 struct sched_group *group;
4869                 int weight;
4870 
4871                 if (!(sd->flags & sd_flag)) {
4872                         sd = sd->child;
4873                         continue;
4874                 }
4875 
4876                 group = find_idlest_group(sd, p, cpu, sd_flag);
4877                 if (!group) {
4878                         sd = sd->child;
4879                         continue;
4880                 }
4881 
4882                 new_cpu = find_idlest_cpu(group, p, cpu);
4883                 if (new_cpu == -1 || new_cpu == cpu) {
4884                         /* Now try balancing at a lower domain level of cpu */
4885                         sd = sd->child;
4886                         continue;
4887                 }
4888 
4889                 /* Now try balancing at a lower domain level of new_cpu */
4890                 cpu = new_cpu;
4891                 weight = sd->span_weight;
4892                 sd = NULL;
4893                 for_each_domain(cpu, tmp) {
4894                         if (weight <= tmp->span_weight)
4895                                 break;
4896                         if (tmp->flags & sd_flag)
4897                                 sd = tmp;
4898                 }
4899                 /* while loop will break here if sd == NULL */
4900         }
4901 unlock:
4902         rcu_read_unlock();
4903 
4904         return new_cpu;
4905 }
4906 
4907 /*
4908  * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4909  * cfs_rq_of(p) references at time of call are still valid and identify the
4910  * previous cpu.  However, the caller only guarantees p->pi_lock is held; no
4911  * other assumptions, including the state of rq->lock, should be made.
4912  */
4913 static void
4914 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4915 {
4916         struct sched_entity *se = &p->se;
4917         struct cfs_rq *cfs_rq = cfs_rq_of(se);
4918 
4919         /*
4920          * Load tracking: accumulate removed load so that it can be processed
4921          * when we next update owning cfs_rq under rq->lock.  Tasks contribute
4922          * to blocked load iff they have a positive decay-count.  It can never
4923          * be negative here since on-rq tasks have decay-count == 0.
4924          */
4925         if (se->avg.decay_count) {
4926                 se->avg.decay_count = -__synchronize_entity_decay(se);
4927                 atomic_long_add(se->avg.load_avg_contrib,
4928                                                 &cfs_rq->removed_load);
4929         }
4930 
4931         /* We have migrated, no longer consider this task hot */
4932         se->exec_start = 0;
4933 }
4934 #endif /* CONFIG_SMP */
4935 
4936 static unsigned long
4937 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4938 {
4939         unsigned long gran = sysctl_sched_wakeup_granularity;
4940 
4941         /*
4942          * Since its curr running now, convert the gran from real-time
4943          * to virtual-time in his units.
4944          *
4945          * By using 'se' instead of 'curr' we penalize light tasks, so
4946          * they get preempted easier. That is, if 'se' < 'curr' then
4947          * the resulting gran will be larger, therefore penalizing the
4948          * lighter, if otoh 'se' > 'curr' then the resulting gran will
4949          * be smaller, again penalizing the lighter task.
4950          *
4951          * This is especially important for buddies when the leftmost
4952          * task is higher priority than the buddy.
4953          */
4954         return calc_delta_fair(gran, se);
4955 }
4956 
4957 /*
4958  * Should 'se' preempt 'curr'.
4959  *
4960  *             |s1
4961  *        |s2
4962  *   |s3
4963  *         g
4964  *      |<--->|c
4965  *
4966  *  w(c, s1) = -1
4967  *  w(c, s2) =  0
4968  *  w(c, s3) =  1
4969  *
4970  */
4971 static int
4972 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4973 {
4974         s64 gran, vdiff = curr->vruntime - se->vruntime;
4975 
4976         if (vdiff <= 0)
4977                 return -1;
4978 
4979         gran = wakeup_gran(curr, se);
4980         if (vdiff > gran)
4981                 return 1;
4982 
4983         return 0;
4984 }
4985 
4986 static void set_last_buddy(struct sched_entity *se)
4987 {
4988         if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4989                 return;
4990 
4991         for_each_sched_entity(se)
4992                 cfs_rq_of(se)->last = se;
4993 }
4994 
4995 static void set_next_buddy(struct sched_entity *se)
4996 {
4997         if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4998                 return;
4999 
5000         for_each_sched_entity(se)
5001                 cfs_rq_of(se)->next = se;
5002 }
5003 
5004 static void set_skip_buddy(struct sched_entity *se)
5005 {
5006         for_each_sched_entity(se)
5007                 cfs_rq_of(se)->skip = se;
5008 }
5009 
5010 /*
5011  * Preempt the current task with a newly woken task if needed:
5012  */
5013 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5014 {
5015         struct task_struct *curr = rq->curr;
5016         struct sched_entity *se = &curr->se, *pse = &p->se;
5017         struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5018         int scale = cfs_rq->nr_running >= sched_nr_latency;
5019         int next_buddy_marked = 0;
5020 
5021         if (unlikely(se == pse))
5022                 return;
5023 
5024         /*
5025          * This is possible from callers such as attach_tasks(), in which we
5026          * unconditionally check_prempt_curr() after an enqueue (which may have
5027          * lead to a throttle).  This both saves work and prevents false
5028          * next-buddy nomination below.
5029          */
5030         if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5031                 return;
5032 
5033         if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5034                 set_next_buddy(pse);
5035                 next_buddy_marked = 1;
5036         }
5037 
5038         /*
5039          * We can come here with TIF_NEED_RESCHED already set from new task
5040          * wake up path.
5041          *
5042          * Note: this also catches the edge-case of curr being in a throttled
5043          * group (e.g. via set_curr_task), since update_curr() (in the
5044          * enqueue of curr) will have resulted in resched being set.  This
5045          * prevents us from potentially nominating it as a false LAST_BUDDY
5046          * below.
5047          */
5048         if (test_tsk_need_resched(curr))
5049                 return;
5050 
5051         /* Idle tasks are by definition preempted by non-idle tasks. */
5052         if (unlikely(curr->policy == SCHED_IDLE) &&
5053             likely(p->policy != SCHED_IDLE))
5054                 goto preempt;
5055 
5056         /*
5057          * Batch and idle tasks do not preempt non-idle tasks (their preemption
5058          * is driven by the tick):
5059          */
5060         if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5061                 return;
5062 
5063         find_matching_se(&se, &pse);
5064         update_curr(cfs_rq_of(se));
5065         BUG_ON(!pse);
5066         if (wakeup_preempt_entity(se, pse) == 1) {
5067                 /*
5068                  * Bias pick_next to pick the sched entity that is
5069                  * triggering this preemption.
5070                  */
5071                 if (!next_buddy_marked)
5072                         set_next_buddy(pse);
5073                 goto preempt;
5074         }
5075 
5076         return;
5077 
5078 preempt:
5079         resched_curr(rq);
5080         /*
5081          * Only set the backward buddy when the current task is still
5082          * on the rq. This can happen when a wakeup gets interleaved
5083          * with schedule on the ->pre_schedule() or idle_balance()
5084          * point, either of which can * drop the rq lock.
5085          *
5086          * Also, during early boot the idle thread is in the fair class,
5087          * for obvious reasons its a bad idea to schedule back to it.
5088          */
5089         if (unlikely(!se->on_rq || curr == rq->idle))
5090                 return;
5091 
5092         if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5093                 set_last_buddy(se);
5094 }
5095 
5096 static struct task_struct *
5097 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
5098 {
5099         struct cfs_rq *cfs_rq = &rq->cfs;
5100         struct sched_entity *se;
5101         struct task_struct *p;
5102         int new_tasks;
5103 
5104 again:
5105 #ifdef CONFIG_FAIR_GROUP_SCHED
5106         if (!cfs_rq->nr_running)
5107                 goto idle;
5108 
5109         if (prev->sched_class != &fair_sched_class)
5110                 goto simple;
5111 
5112         /*
5113          * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5114          * likely that a next task is from the same cgroup as the current.
5115          *
5116          * Therefore attempt to avoid putting and setting the entire cgroup
5117          * hierarchy, only change the part that actually changes.
5118          */
5119 
5120         do {
5121                 struct sched_entity *curr = cfs_rq->curr;
5122 
5123                 /*
5124                  * Since we got here without doing put_prev_entity() we also
5125                  * have to consider cfs_rq->curr. If it is still a runnable
5126                  * entity, update_curr() will update its vruntime, otherwise
5127                  * forget we've ever seen it.
5128                  */
5129                 if (curr && curr->on_rq)
5130                         update_curr(cfs_rq);
5131                 else
5132                         curr = NULL;
5133 
5134                 /*
5135                  * This call to check_cfs_rq_runtime() will do the throttle and
5136                  * dequeue its entity in the parent(s). Therefore the 'simple'
5137                  * nr_running test will indeed be correct.
5138                  */
5139                 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5140                         goto simple;
5141 
5142                 se = pick_next_entity(cfs_rq, curr);
5143                 cfs_rq = group_cfs_rq(se);
5144         } while (cfs_rq);
5145 
5146         p = task_of(se);
5147 
5148         /*
5149          * Since we haven't yet done put_prev_entity and if the selected task
5150          * is a different task than we started out with, try and touch the
5151          * least amount of cfs_rqs.
5152          */
5153         if (prev != p) {
5154                 struct sched_entity *pse = &prev->se;
5155 
5156                 while (!(cfs_rq = is_same_group(se, pse))) {
5157                         int se_depth = se->depth;
5158                         int pse_depth = pse->depth;
5159 
5160                         if (se_depth <= pse_depth) {
5161                                 put_prev_entity(cfs_rq_of(pse), pse);
5162                                 pse = parent_entity(pse);
5163                         }
5164                         if (se_depth >= pse_depth) {
5165                                 set_next_entity(cfs_rq_of(se), se);
5166                                 se = parent_entity(se);
5167                         }
5168                 }
5169 
5170                 put_prev_entity(cfs_rq, pse);
5171                 set_next_entity(cfs_rq, se);
5172         }
5173 
5174         if (hrtick_enabled(rq))
5175                 hrtick_start_fair(rq, p);
5176 
5177         return p;
5178 simple:
5179         cfs_rq = &rq->cfs;
5180 #endif
5181 
5182         if (!cfs_rq->nr_running)
5183                 goto idle;
5184 
5185         put_prev_task(rq, prev);
5186 
5187         do {
5188                 se = pick_next_entity(cfs_rq, NULL);
5189                 set_next_entity(cfs_rq, se);
5190                 cfs_rq = group_cfs_rq(se);
5191         } while (cfs_rq);
5192 
5193         p = task_of(se);
5194 
5195         if (hrtick_enabled(rq))
5196                 hrtick_start_fair(rq, p);
5197 
5198         return p;
5199 
5200 idle:
5201         new_tasks = idle_balance(rq);
5202         /*
5203          * Because idle_balance() releases (and re-acquires) rq->lock, it is
5204          * possible for any higher priority task to appear. In that case we
5205          * must re-start the pick_next_entity() loop.
5206          */
5207         if (new_tasks < 0)
5208                 return RETRY_TASK;
5209 
5210         if (new_tasks > 0)
5211                 goto again;
5212 
5213         return NULL;
5214 }
5215 
5216 /*
5217  * Account for a descheduled task:
5218  */
5219 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5220 {
5221         struct sched_entity *se = &prev->se;
5222         struct cfs_rq *cfs_rq;
5223 
5224         for_each_sched_entity(se) {
5225                 cfs_rq = cfs_rq_of(se);
5226                 put_prev_entity(cfs_rq, se);
5227         }
5228 }
5229 
5230 /*
5231  * sched_yield() is very simple
5232  *
5233  * The magic of dealing with the ->skip buddy is in pick_next_entity.
5234  */
5235 static void yield_task_fair(struct rq *rq)
5236 {
5237         struct task_struct *curr = rq->curr;
5238         struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5239         struct sched_entity *se = &curr->se;
5240 
5241         /*
5242          * Are we the only task in the tree?
5243          */
5244         if (unlikely(rq->nr_running == 1))
5245                 return;
5246 
5247         clear_buddies(cfs_rq, se);
5248 
5249         if (curr->policy != SCHED_BATCH) {
5250                 update_rq_clock(rq);
5251                 /*
5252                  * Update run-time statistics of the 'current'.
5253                  */
5254                 update_curr(cfs_rq);
5255                 /*
5256                  * Tell update_rq_clock() that we've just updated,
5257                  * so we don't do microscopic update in schedule()
5258                  * and double the fastpath cost.
5259                  */
5260                 rq_clock_skip_update(rq, true);
5261         }
5262 
5263         set_skip_buddy(se);
5264 }
5265 
5266 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5267 {
5268         struct sched_entity *se = &p->se;
5269 
5270         /* throttled hierarchies are not runnable */
5271         if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5272                 return false;
5273 
5274         /* Tell the scheduler that we'd really like pse to run next. */
5275         set_next_buddy(se);
5276 
5277         yield_task_fair(rq);
5278 
5279         return true;
5280 }
5281 
5282 #ifdef CONFIG_SMP
5283 /**************************************************
5284  * Fair scheduling class load-balancing methods.
5285  *
5286  * BASICS
5287  *
5288  * The purpose of load-balancing is to achieve the same basic fairness the
5289  * per-cpu scheduler provides, namely provide a proportional amount of compute
5290  * time to each task. This is expressed in the following equation:
5291  *
5292  *   W_i,n/P_i == W_j,n/P_j for all i,j                               (1)
5293  *
5294  * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5295  * W_i,0 is defined as:
5296  *
5297  *   W_i,0 = \Sum_j w_i,j                                             (2)
5298  *
5299  * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5300  * is derived from the nice value as per prio_to_weight[].
5301  *
5302  * The weight average is an exponential decay average of the instantaneous
5303  * weight:
5304  *
5305  *   W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0               (3)
5306  *
5307  * C_i is the compute capacity of cpu i, typically it is the
5308  * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5309  * can also include other factors [XXX].
5310  *
5311  * To achieve this balance we define a measure of imbalance which follows
5312  * directly from (1):
5313  *
5314  *   imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j }    (4)
5315  *
5316  * We them move tasks around to minimize the imbalance. In the continuous
5317  * function space it is obvious this converges, in the discrete case we get
5318  * a few fun cases generally called infeasible weight scenarios.
5319  *
5320  * [XXX expand on:
5321  *     - infeasible weights;
5322  *     - local vs global optima in the discrete case. ]
5323  *
5324  *
5325  * SCHED DOMAINS
5326  *
5327  * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5328  * for all i,j solution, we create a tree of cpus that follows the hardware
5329  * topology where each level pairs two lower groups (or better). This results
5330  * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5331  * tree to only the first of the previous level and we decrease the frequency
5332  * of load-balance at each level inv. proportional to the number of cpus in
5333  * the groups.
5334  *
5335  * This yields:
5336  *
5337  *     log_2 n     1     n
5338  *   \Sum       { --- * --- * 2^i } = O(n)                            (5)
5339  *     i = 0      2^i   2^i
5340  *                               `- size of each group
5341  *         |         |     `- number of cpus doing load-balance
5342  *         |         `- freq
5343  *         `- sum over all levels
5344  *
5345  * Coupled with a limit on how many tasks we can migrate every balance pass,
5346  * this makes (5) the runtime complexity of the balancer.
5347  *
5348  * An important property here is that each CPU is still (indirectly) connected
5349  * to every other cpu in at most O(log n) steps:
5350  *
5351  * The adjacency matrix of the resulting graph is given by:
5352  *
5353  *             log_2 n     
5354  *   A_i,j = \Union     (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1)  (6)
5355  *             k = 0
5356  *
5357  * And you'll find that:
5358  *
5359  *   A^(log_2 n)_i,j != 0  for all i,j                                (7)
5360  *
5361  * Showing there's indeed a path between every cpu in at most O(log n) steps.
5362  * The task movement gives a factor of O(m), giving a convergence complexity
5363  * of:
5364  *
5365  *   O(nm log n),  n := nr_cpus, m := nr_tasks                        (8)
5366  *
5367  *
5368  * WORK CONSERVING
5369  *
5370  * In order to avoid CPUs going idle while there's still work to do, new idle
5371  * balancing is more aggressive and has the newly idle cpu iterate up the domain
5372  * tree itself instead of relying on other CPUs to bring it work.
5373  *
5374  * This adds some complexity to both (5) and (8) but it reduces the total idle
5375  * time.
5376  *
5377  * [XXX more?]
5378  *
5379  *
5380  * CGROUPS
5381  *
5382  * Cgroups make a horror show out of (2), instead of a simple sum we get:
5383  *
5384  *                                s_k,i
5385  *   W_i,0 = \Sum_j \Prod_k w_k * -----                               (9)
5386  *                                 S_k
5387  *
5388  * Where
5389  *
5390  *   s_k,i = \Sum_j w_i,j,k  and  S_k = \Sum_i s_k,i                 (10)
5391  *
5392  * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5393  *
5394  * The big problem is S_k, its a global sum needed to compute a local (W_i)
5395  * property.
5396  *
5397  * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5398  *      rewrite all of this once again.]
5399  */ 
5400 
5401 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5402 
5403 enum fbq_type { regular, remote, all };
5404 
5405 #define LBF_ALL_PINNED  0x01
5406 #define LBF_NEED_BREAK  0x02
5407 #define LBF_DST_PINNED  0x04
5408 #define LBF_SOME_PINNED 0x08
5409 
5410 struct lb_env {
5411         struct sched_domain     *sd;
5412 
5413         struct rq               *src_rq;
5414         int                     src_cpu;
5415 
5416         int                     dst_cpu;
5417         struct rq               *dst_rq;
5418 
5419         struct cpumask          *dst_grpmask;
5420         int                     new_dst_cpu;
5421         enum cpu_idle_type      idle;
5422         long                    imbalance;
5423         /* The set of CPUs under consideration for load-balancing */
5424         struct cpumask          *cpus;
5425 
5426         unsigned int            flags;
5427 
5428         unsigned int            loop;
5429         unsigned int            loop_break;
5430         unsigned int            loop_max;
5431 
5432         enum fbq_type           fbq_type;
5433         struct list_head        tasks;
5434 };
5435 
5436 /*
5437  * Is this task likely cache-hot:
5438  */
5439 static int task_hot(struct task_struct *p, struct lb_env *env)
5440 {
5441         s64 delta;
5442 
5443         lockdep_assert_held(&env->src_rq->lock);
5444 
5445         if (p->sched_class != &fair_sched_class)
5446                 return 0;
5447 
5448         if (unlikely(p->policy == SCHED_IDLE))
5449                 return 0;
5450 
5451         /*
5452          * Buddy candidates are cache hot:
5453          */
5454         if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
5455                         (&p->se == cfs_rq_of(&p->se)->next ||
5456                          &p->se == cfs_rq_of(&p->se)->last))
5457                 return 1;
5458 
5459         if (sysctl_sched_migration_cost == -1)
5460                 return 1;
5461         if (sysctl_sched_migration_cost == 0)
5462                 return 0;
5463 
5464         delta = rq_clock_task(env->src_rq) - p->se.exec_start;
5465 
5466         return delta < (s64)sysctl_sched_migration_cost;
5467 }
5468 
5469 #ifdef CONFIG_NUMA_BALANCING
5470 /* Returns true if the destination node has incurred more faults */
5471 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
5472 {
5473         struct numa_group *numa_group = rcu_dereference(p->numa_group);
5474         int src_nid, dst_nid;
5475 
5476         if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults ||
5477             !(env->sd->flags & SD_NUMA)) {
5478                 return false;
5479         }
5480 
5481         src_nid = cpu_to_node(env->src_cpu);
5482         dst_nid = cpu_to_node(env->dst_cpu);
5483 
5484         if (src_nid == dst_nid)
5485                 return false;
5486 
5487         if (numa_group) {
5488                 /* Task is already in the group's interleave set. */
5489                 if (node_isset(src_nid, numa_group->active_nodes))
5490                         return false;
5491 
5492                 /* Task is moving into the group's interleave set. */
5493                 if (node_isset(dst_nid, numa_group->active_nodes))
5494                         return true;
5495 
5496                 return group_faults(p, dst_nid) > group_faults(p, src_nid);
5497         }
5498 
5499         /* Encourage migration to the preferred node. */
5500         if (dst_nid == p->numa_preferred_nid)
5501                 return true;
5502 
5503         return task_faults(p, dst_nid) > task_faults(p, src_nid);
5504 }
5505 
5506 
5507 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5508 {
5509         struct numa_group *numa_group = rcu_dereference(p->numa_group);
5510         int src_nid, dst_nid;
5511 
5512         if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
5513                 return false;
5514 
5515         if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
5516                 return false;
5517 
5518         src_nid = cpu_to_node(env->src_cpu);
5519         dst_nid = cpu_to_node(env->dst_cpu);
5520 
5521         if (src_nid == dst_nid)
5522                 return false;
5523 
5524         if (numa_group) {
5525                 /* Task is moving within/into the group's interleave set. */
5526                 if (node_isset(dst_nid, numa_group->active_nodes))
5527                         return false;
5528 
5529                 /* Task is moving out of the group's interleave set. */
5530                 if (node_isset(src_nid, numa_group->active_nodes))
5531                         return true;
5532 
5533                 return group_faults(p, dst_nid) < group_faults(p, src_nid);
5534         }
5535 
5536         /* Migrating away from the preferred node is always bad. */
5537         if (src_nid == p->numa_preferred_nid)
5538                 return true;
5539 
5540         return task_faults(p, dst_nid) < task_faults(p, src_nid);
5541 }
5542 
5543 #else
5544 static inline bool migrate_improves_locality(struct task_struct *p,
5545                                              struct lb_env *env)
5546 {
5547         return false;
5548 }
5549 
5550 static inline bool migrate_degrades_locality(struct task_struct *p,
5551                                              struct lb_env *env)
5552 {
5553         return false;
5554 }
5555 #endif
5556 
5557 /*
5558  * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5559  */
5560 static
5561 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5562 {
5563         int tsk_cache_hot = 0;
5564 
5565         lockdep_assert_held(&env->src_rq->lock);
5566 
5567         /*
5568          * We do not migrate tasks that are:
5569          * 1) throttled_lb_pair, or
5570          * 2) cannot be migrated to this CPU due to cpus_allowed, or
5571          * 3) running (obviously), or
5572          * 4) are cache-hot on their current CPU.
5573          */
5574         if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5575                 return 0;
5576 
5577         if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5578                 int cpu;
5579 
5580                 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5581 
5582                 env->flags |= LBF_SOME_PINNED;
5583 
5584                 /*
5585                  * Remember if this task can be migrated to any other cpu in
5586                  * our sched_group. We may want to revisit it if we couldn't
5587                  * meet load balance goals by pulling other tasks on src_cpu.
5588                  *
5589                  * Also avoid computing new_dst_cpu if we have already computed
5590                  * one in current iteration.
5591                  */
5592                 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5593                         return 0;
5594 
5595                 /* Prevent to re-select dst_cpu via env's cpus */
5596                 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5597                         if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5598                                 env->flags |= LBF_DST_PINNED;
5599                                 env->new_dst_cpu = cpu;
5600                                 break;
5601                         }
5602                 }
5603 
5604                 return 0;
5605         }
5606 
5607         /* Record that we found atleast one task that could run on dst_cpu */
5608         env->flags &= ~LBF_ALL_PINNED;
5609 
5610         if (task_running(env->src_rq, p)) {
5611                 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5612                 return 0;
5613         }
5614 
5615         /*
5616          * Aggressive migration if:
5617          * 1) destination numa is preferred
5618          * 2) task is cache cold, or
5619          * 3) too many balance attempts have failed.
5620          */
5621         tsk_cache_hot = task_hot(p, env);
5622         if (!tsk_cache_hot)
5623                 tsk_cache_hot = migrate_degrades_locality(p, env);
5624 
5625         if (migrate_improves_locality(p, env) || !tsk_cache_hot ||
5626             env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5627                 if (tsk_cache_hot) {
5628                         schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5629                         schedstat_inc(p, se.statistics.nr_forced_migrations);
5630                 }
5631                 return 1;
5632         }
5633 
5634         schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5635         return 0;
5636 }
5637 
5638 /*
5639  * detach_task() -- detach the task for the migration specified in env
5640  */
5641 static void detach_task(struct task_struct *p, struct lb_env *env)
5642 {
5643         lockdep_assert_held(&env->src_rq->lock);
5644 
5645         deactivate_task(env->src_rq, p, 0);
5646         p->on_rq = TASK_ON_RQ_MIGRATING;
5647         set_task_cpu(p, env->dst_cpu);
5648 }
5649 
5650 /*
5651  * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5652  * part of active balancing operations within "domain".
5653  *
5654  * Returns a task if successful and NULL otherwise.
5655  */
5656 static struct task_struct *detach_one_task(struct lb_env *env)
5657 {
5658         struct task_struct *p, *n;
5659 
5660         lockdep_assert_held(&env->src_rq->lock);
5661 
5662         list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5663                 if (!can_migrate_task(p, env))
5664                         continue;
5665 
5666                 detach_task(p, env);
5667 
5668                 /*
5669                  * Right now, this is only the second place where
5670                  * lb_gained[env->idle] is updated (other is detach_tasks)
5671                  * so we can safely collect stats here rather than
5672                  * inside detach_tasks().
5673                  */
5674                 schedstat_inc(env->sd, lb_gained[env->idle]);
5675                 return p;
5676         }
5677         return NULL;
5678 }
5679 
5680 static const unsigned int sched_nr_migrate_break = 32;
5681 
5682 /*
5683  * detach_tasks() -- tries to detach up to imbalance weighted load from
5684  * busiest_rq, as part of a balancing operation within domain "sd".
5685  *
5686  * Returns number of detached tasks if successful and 0 otherwise.
5687  */
5688 static int detach_tasks(struct lb_env *env)
5689 {
5690         struct list_head *tasks = &env->src_rq->cfs_tasks;
5691         struct task_struct *p;
5692         unsigned long load;
5693         int detached = 0;
5694 
5695         lockdep_assert_held(&env->src_rq->lock);
5696 
5697         if (env->imbalance <= 0)
5698                 return 0;
5699 
5700         while (!list_empty(tasks)) {
5701                 p = list_first_entry(tasks, struct task_struct, se.group_node);
5702 
5703                 env->loop++;
5704                 /* We've more or less seen every task there is, call it quits */
5705                 if (env->loop > env->loop_max)
5706                         break;
5707 
5708                 /* take a breather every nr_migrate tasks */
5709                 if (env->loop > env->loop_break) {
5710                         env->loop_break += sched_nr_migrate_break;
5711                         env->flags |= LBF_NEED_BREAK;
5712                         break;
5713                 }
5714 
5715                 if (!can_migrate_task(p, env))
5716                         goto next;
5717 
5718                 load = task_h_load(p);
5719 
5720                 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5721                         goto next;
5722 
5723                 if ((load / 2) > env->imbalance)
5724                         goto next;
5725 
5726                 detach_task(p, env);
5727                 list_add(&p->se.group_node, &env->tasks);
5728 
5729                 detached++;
5730                 env->imbalance -= load;
5731 
5732 #ifdef CONFIG_PREEMPT
5733                 /*
5734                  * NEWIDLE balancing is a source of latency, so preemptible
5735                  * kernels will stop after the first task is detached to minimize
5736                  * the critical section.
5737                  */
5738                 if (env->idle == CPU_NEWLY_IDLE)
5739                         break;
5740 #endif
5741 
5742                 /*
5743                  * We only want to steal up to the prescribed amount of
5744                  * weighted load.
5745                  */
5746                 if (env->imbalance <= 0)
5747                         break;
5748 
5749                 continue;
5750 next:
5751                 list_move_tail(&p->se.group_node, tasks);
5752         }
5753 
5754         /*
5755          * Right now, this is one of only two places we collect this stat
5756          * so we can safely collect detach_one_task() stats here rather
5757          * than inside detach_one_task().
5758          */
5759         schedstat_add(env->sd, lb_gained[env->idle], detached);
5760 
5761         return detached;
5762 }
5763 
5764 /*
5765  * attach_task() -- attach the task detached by detach_task() to its new rq.
5766  */
5767 static void attach_task(struct rq *rq, struct task_struct *p)
5768 {
5769         lockdep_assert_held(&rq->lock);
5770 
5771         BUG_ON(task_rq(p) != rq);
5772         p->on_rq = TASK_ON_RQ_QUEUED;
5773         activate_task(rq, p, 0);
5774         check_preempt_curr(rq, p, 0);
5775 }
5776 
5777 /*
5778  * attach_one_task() -- attaches the task returned from detach_one_task() to
5779  * its new rq.
5780  */
5781 static void attach_one_task(struct rq *rq, struct task_struct *p)
5782 {
5783         raw_spin_lock(&rq->lock);
5784         attach_task(rq, p);
5785         raw_spin_unlock(&rq->lock);
5786 }
5787 
5788 /*
5789  * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
5790  * new rq.
5791  */
5792 static void attach_tasks(struct lb_env *env)
5793 {
5794         struct list_head *tasks = &env->tasks;
5795         struct task_struct *p;
5796 
5797         raw_spin_lock(&env->dst_rq->lock);
5798 
5799         while (!list_empty(tasks)) {
5800                 p = list_first_entry(tasks, struct task_struct, se.group_node);
5801                 list_del_init(&p->se.group_node);
5802 
5803                 attach_task(env->dst_rq, p);
5804         }
5805 
5806         raw_spin_unlock(&env->dst_rq->lock);
5807 }
5808 
5809 #ifdef CONFIG_FAIR_GROUP_SCHED
5810 /*
5811  * update tg->load_weight by folding this cpu's load_avg
5812  */
5813 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5814 {
5815         struct sched_entity *se = tg->se[cpu];
5816         struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5817 
5818         /* throttled entities do not contribute to load */
5819         if (throttled_hierarchy(cfs_rq))
5820                 return;
5821 
5822         update_cfs_rq_blocked_load(cfs_rq, 1);
5823 
5824         if (se) {
5825                 update_entity_load_avg(se, 1);
5826                 /*
5827                  * We pivot on our runnable average having decayed to zero for
5828                  * list removal.  This generally implies that all our children
5829                  * have also been removed (modulo rounding error or bandwidth
5830                  * control); however, such cases are rare and we can fix these
5831                  * at enqueue.
5832                  *
5833                  * TODO: fix up out-of-order children on enqueue.
5834                  */
5835                 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5836                         list_del_leaf_cfs_rq(cfs_rq);
5837         } else {
5838                 struct rq *rq = rq_of(cfs_rq);
5839                 update_rq_runnable_avg(rq, rq->nr_running);
5840         }
5841 }
5842 
5843 static void update_blocked_averages(int cpu)
5844 {
5845         struct rq *rq = cpu_rq(cpu);
5846         struct cfs_rq *cfs_rq;
5847         unsigned long flags;
5848 
5849         raw_spin_lock_irqsave(&rq->lock, flags);
5850         update_rq_clock(rq);
5851         /*
5852          * Iterates the task_group tree in a bottom up fashion, see
5853          * list_add_leaf_cfs_rq() for details.
5854          */
5855         for_each_leaf_cfs_rq(rq, cfs_rq) {
5856                 /*
5857                  * Note: We may want to consider periodically releasing
5858                  * rq->lock about these updates so that creating many task
5859                  * groups does not result in continually extending hold time.
5860                  */
5861                 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5862         }
5863 
5864         raw_spin_unlock_irqrestore(&rq->lock, flags);
5865 }
5866 
5867 /*
5868  * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5869  * This needs to be done in a top-down fashion because the load of a child
5870  * group is a fraction of its parents load.
5871  */
5872 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5873 {
5874         struct rq *rq = rq_of(cfs_rq);
5875         struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5876         unsigned long now = jiffies;
5877         unsigned long load;
5878 
5879         if (cfs_rq->last_h_load_update == now)
5880                 return;
5881 
5882         cfs_rq->h_load_next = NULL;
5883         for_each_sched_entity(se) {
5884                 cfs_rq = cfs_rq_of(se);
5885                 cfs_rq->h_load_next = se;
5886                 if (cfs_rq->last_h_load_update == now)
5887                         break;
5888         }
5889 
5890         if (!se) {
5891                 cfs_rq->h_load = cfs_rq->runnable_load_avg;
5892                 cfs_rq->last_h_load_update = now;
5893         }
5894 
5895         while ((se = cfs_rq->h_load_next) != NULL) {
5896                 load = cfs_rq->h_load;
5897                 load = div64_ul(load * se->avg.load_avg_contrib,
5898                                 cfs_rq->runnable_load_avg + 1);
5899                 cfs_rq = group_cfs_rq(se);
5900                 cfs_rq->h_load = load;
5901                 cfs_rq->last_h_load_update = now;
5902         }
5903 }
5904 
5905 static unsigned long task_h_load(struct task_struct *p)
5906 {
5907         struct cfs_rq *cfs_rq = task_cfs_rq(p);
5908 
5909         update_cfs_rq_h_load(cfs_rq);
5910         return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
5911                         cfs_rq->runnable_load_avg + 1);
5912 }
5913 #else
5914 static inline void update_blocked_averages(int cpu)
5915 {
5916 }
5917 
5918 static unsigned long task_h_load(struct task_struct *p)
5919 {
5920         return p->se.avg.load_avg_contrib;
5921 }
5922 #endif
5923 
5924 /********** Helpers for find_busiest_group ************************/
5925 
5926 enum group_type {
5927         group_other = 0,
5928         group_imbalanced,
5929         group_overloaded,
5930 };
5931 
5932 /*
5933  * sg_lb_stats - stats of a sched_group required for load_balancing
5934  */
5935 struct sg_lb_stats {
5936         unsigned long avg_load; /*Avg load across the CPUs of the group */
5937         unsigned long group_load; /* Total load over the CPUs of the group */
5938         unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5939         unsigned long load_per_task;
5940         unsigned long group_capacity;
5941         unsigned long group_usage; /* Total usage of the group */
5942         unsigned int sum_nr_running; /* Nr tasks running in the group */
5943         unsigned int idle_cpus;
5944         unsigned int group_weight;
5945         enum group_type group_type;
5946         int group_no_capacity;
5947 #ifdef CONFIG_NUMA_BALANCING
5948         unsigned int nr_numa_running;
5949         unsigned int nr_preferred_running;
5950 #endif
5951 };
5952 
5953 /*
5954  * sd_lb_stats - Structure to store the statistics of a sched_domain
5955  *               during load balancing.
5956  */
5957 struct sd_lb_stats {
5958         struct sched_group *busiest;    /* Busiest group in this sd */
5959         struct sched_group *local;      /* Local group in this sd */
5960         unsigned long total_load;       /* Total load of all groups in sd */
5961         unsigned long total_capacity;   /* Total capacity of all groups in sd */
5962         unsigned long avg_load; /* Average load across all groups in sd */
5963 
5964         struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5965         struct sg_lb_stats local_stat;  /* Statistics of the local group */
5966 };
5967 
5968 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5969 {
5970         /*
5971          * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5972          * local_stat because update_sg_lb_stats() does a full clear/assignment.
5973          * We must however clear busiest_stat::avg_load because
5974          * update_sd_pick_busiest() reads this before assignment.
5975          */
5976         *sds = (struct sd_lb_stats){
5977                 .busiest = NULL,
5978                 .local = NULL,
5979                 .total_load = 0UL,
5980                 .total_capacity = 0UL,
5981                 .busiest_stat = {
5982                         .avg_load = 0UL,
5983                         .sum_nr_running = 0,
5984                         .group_type = group_other,
5985                 },
5986         };
5987 }
5988 
5989 /**
5990  * get_sd_load_idx - Obtain the load index for a given sched domain.
5991  * @sd: The sched_domain whose load_idx is to be obtained.
5992  * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5993  *
5994  * Return: The load index.
5995  */
5996 static inline int get_sd_load_idx(struct sched_domain *sd,
5997                                         enum cpu_idle_type idle)
5998 {
5999         int load_idx;
6000 
6001         switch (idle) {
6002         case CPU_NOT_IDLE:
6003                 load_idx = sd->busy_idx;
6004                 break;
6005 
6006         case CPU_NEWLY_IDLE:
6007                 load_idx = sd->newidle_idx;
6008                 break;
6009         default:
6010                 load_idx = sd->idle_idx;
6011                 break;
6012         }
6013 
6014         return load_idx;
6015 }
6016 
6017 static unsigned long default_scale_cpu_capacity(struct sched_domain *sd, int cpu)
6018 {
6019         if ((sd->flags & SD_SHARE_CPUCAPACITY) && (sd->span_weight > 1))
6020                 return sd->smt_gain / sd->span_weight;
6021 
6022         return SCHED_CAPACITY_SCALE;
6023 }
6024 
6025 unsigned long __weak arch_scale_cpu_capacity(struct sched_domain *sd, int cpu)
6026 {
6027         return default_scale_cpu_capacity(sd, cpu);
6028 }
6029 
6030 static unsigned long scale_rt_capacity(int cpu)
6031 {
6032         struct rq *rq = cpu_rq(cpu);
6033         u64 total, used, age_stamp, avg;
6034         s64 delta;
6035 
6036         /*
6037          * Since we're reading these variables without serialization make sure
6038          * we read them once before doing sanity checks on them.
6039          */
6040         age_stamp = ACCESS_ONCE(rq->age_stamp);
6041         avg = ACCESS_ONCE(rq->rt_avg);
6042         delta = __rq_clock_broken(rq) - age_stamp;
6043 
6044         if (unlikely(delta < 0))
6045                 delta = 0;
6046 
6047         total = sched_avg_period() + delta;
6048 
6049         used = div_u64(avg, total);
6050 
6051         if (likely(used < SCHED_CAPACITY_SCALE))
6052                 return SCHED_CAPACITY_SCALE - used;
6053 
6054         return 1;
6055 }
6056 
6057 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6058 {
6059         unsigned long capacity = SCHED_CAPACITY_SCALE;
6060         struct sched_group *sdg = sd->groups;
6061 
6062         if (sched_feat(ARCH_CAPACITY))
6063                 capacity *= arch_scale_cpu_capacity(sd, cpu);
6064         else
6065                 capacity *= default_scale_cpu_capacity(sd, cpu);
6066 
6067         capacity >>= SCHED_CAPACITY_SHIFT;
6068 
6069         cpu_rq(cpu)->cpu_capacity_orig = capacity;
6070 
6071         capacity *= scale_rt_capacity(cpu);
6072         capacity >>= SCHED_CAPACITY_SHIFT;
6073 
6074         if (!capacity)
6075                 capacity = 1;
6076 
6077         cpu_rq(cpu)->cpu_capacity = capacity;
6078         sdg->sgc->capacity = capacity;
6079 }
6080 
6081 void update_group_capacity(struct sched_domain *sd, int cpu)
6082 {
6083         struct sched_domain *child = sd->child;
6084         struct sched_group *group, *sdg = sd->groups;
6085         unsigned long capacity;
6086         unsigned long interval;
6087 
6088         interval = msecs_to_jiffies(sd->balance_interval);
6089         interval = clamp(interval, 1UL, max_load_balance_interval);
6090         sdg->sgc->next_update = jiffies + interval;
6091 
6092         if (!child) {
6093                 update_cpu_capacity(sd, cpu);
6094                 return;
6095         }
6096 
6097         capacity = 0;
6098 
6099         if (child->flags & SD_OVERLAP) {
6100                 /*
6101                  * SD_OVERLAP domains cannot assume that child groups
6102                  * span the current group.
6103                  */
6104 
6105                 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6106                         struct sched_group_capacity *sgc;
6107                         struct rq *rq = cpu_rq(cpu);
6108 
6109                         /*
6110                          * build_sched_domains() -> init_sched_groups_capacity()
6111                          * gets here before we've attached the domains to the
6112                          * runqueues.
6113                          *
6114                          * Use capacity_of(), which is set irrespective of domains
6115                          * in update_cpu_capacity().
6116                          *
6117                          * This avoids capacity from being 0 and
6118                          * causing divide-by-zero issues on boot.
6119                          */
6120                         if (unlikely(!rq->sd)) {
6121                                 capacity += capacity_of(cpu);
6122                                 continue;
6123                         }
6124 
6125                         sgc = rq->sd->groups->sgc;
6126                         capacity += sgc->capacity;
6127                 }
6128         } else  {
6129                 /*
6130                  * !SD_OVERLAP domains can assume that child groups
6131                  * span the current group.
6132                  */ 
6133 
6134                 group = child->groups;
6135                 do {
6136                         capacity += group->sgc->capacity;
6137                         group = group->next;
6138                 } while (group != child->groups);
6139         }
6140 
6141         sdg->sgc->capacity = capacity;
6142 }
6143 
6144 /*
6145  * Check whether the capacity of the rq has been noticeably reduced by side
6146  * activity. The imbalance_pct is used for the threshold.
6147  * Return true is the capacity is reduced
6148  */
6149 static inline int
6150 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6151 {
6152         return ((rq->cpu_capacity * sd->imbalance_pct) <
6153                                 (rq->cpu_capacity_orig * 100));
6154 }
6155 
6156 /*
6157  * Group imbalance indicates (and tries to solve) the problem where balancing
6158  * groups is inadequate due to tsk_cpus_allowed() constraints.
6159  *
6160  * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6161  * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6162  * Something like:
6163  *
6164  *      { 0 1 2 3 } { 4 5 6 7 }
6165  *              *     * * *
6166  *
6167  * If we were to balance group-wise we'd place two tasks in the first group and
6168  * two tasks in the second group. Clearly this is undesired as it will overload
6169  * cpu 3 and leave one of the cpus in the second group unused.
6170  *
6171  * The current solution to this issue is detecting the skew in the first group
6172  * by noticing the lower domain failed to reach balance and had difficulty
6173  * moving tasks due to affinity constraints.
6174  *
6175  * When this is so detected; this group becomes a candidate for busiest; see
6176  * update_sd_pick_busiest(). And calculate_imbalance() and
6177  * find_busiest_group() avoid some of the usual balance conditions to allow it
6178  * to create an effective group imbalance.
6179  *
6180  * This is a somewhat tricky proposition since the next run might not find the
6181  * group imbalance and decide the groups need to be balanced again. A most
6182  * subtle and fragile situation.
6183  */
6184 
6185 static inline int sg_imbalanced(struct sched_group *group)
6186 {
6187         return group->sgc->imbalance;
6188 }
6189 
6190 /*
6191  * group_has_capacity returns true if the group has spare capacity that could
6192  * be used by some tasks.
6193  * We consider that a group has spare capacity if the  * number of task is
6194  * smaller than the number of CPUs or if the usage is lower than the available
6195  * capacity for CFS tasks.
6196  * For the latter, we use a threshold to stabilize the state, to take into
6197  * account the variance of the tasks' load and to return true if the available
6198  * capacity in meaningful for the load balancer.
6199  * As an example, an available capacity of 1% can appear but it doesn't make
6200  * any benefit for the load balance.
6201  */
6202 static inline bool
6203 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6204 {
6205         if (sgs->sum_nr_running < sgs->group_weight)
6206                 return true;
6207 
6208         if ((sgs->group_capacity * 100) >
6209                         (sgs->group_usage * env->sd->imbalance_pct))
6210                 return true;
6211 
6212         return false;
6213 }
6214 
6215 /*
6216  *  group_is_overloaded returns true if the group has more tasks than it can
6217  *  handle.
6218  *  group_is_overloaded is not equals to !group_has_capacity because a group
6219  *  with the exact right number of tasks, has no more spare capacity but is not
6220  *  overloaded so both group_has_capacity and group_is_overloaded return
6221  *  false.
6222  */
6223 static inline bool
6224 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6225 {
6226         if (sgs->sum_nr_running <= sgs->group_weight)
6227                 return false;
6228 
6229         if ((sgs->group_capacity * 100) <
6230                         (sgs->group_usage * env->sd->imbalance_pct))
6231                 return true;
6232 
6233         return false;
6234 }
6235 
6236 static enum group_type group_classify(struct lb_env *env,
6237                 struct sched_group *group,
6238                 struct sg_lb_stats *sgs)
6239 {
6240         if (sgs->group_no_capacity)
6241                 return group_overloaded;
6242 
6243         if (sg_imbalanced(group))
6244                 return group_imbalanced;
6245 
6246         return group_other;
6247 }
6248 
6249 /**
6250  * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6251  * @env: The load balancing environment.
6252  * @group: sched_group whose statistics are to be updated.
6253  * @load_idx: Load index of sched_domain of this_cpu for load calc.
6254  * @local_group: Does group contain this_cpu.
6255  * @sgs: variable to hold the statistics for this group.
6256  * @overload: Indicate more than one runnable task for any CPU.
6257  */
6258 static inline void update_sg_lb_stats(struct lb_env *env,
6259                         struct sched_group *group, int load_idx,
6260                         int local_group, struct sg_lb_stats *sgs,
6261                         bool *overload)
6262 {
6263         unsigned long load;
6264         int i;
6265 
6266         memset(sgs, 0, sizeof(*sgs));
6267 
6268         for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6269                 struct rq *rq = cpu_rq(i);
6270 
6271                 /* Bias balancing toward cpus of our domain */
6272                 if (local_group)
6273                         load = target_load(i, load_idx);
6274                 else
6275                         load = source_load(i, load_idx);
6276 
6277                 sgs->group_load += load;
6278                 sgs->group_usage += get_cpu_usage(i);
6279                 sgs->sum_nr_running += rq->cfs.h_nr_running;
6280 
6281                 if (rq->nr_running > 1)
6282                         *overload = true;
6283 
6284 #ifdef CONFIG_NUMA_BALANCING
6285                 sgs->nr_numa_running += rq->nr_numa_running;
6286                 sgs->nr_preferred_running += rq->nr_preferred_running;
6287 #endif
6288                 sgs->sum_weighted_load += weighted_cpuload(i);
6289                 if (idle_cpu(i))
6290                         sgs->idle_cpus++;
6291         }
6292 
6293         /* Adjust by relative CPU capacity of the group */
6294         sgs->group_capacity = group->sgc->capacity;
6295         sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6296 
6297         if (sgs->sum_nr_running)
6298                 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6299 
6300         sgs->group_weight = group->group_weight;
6301 
6302         sgs->group_no_capacity = group_is_overloaded(env, sgs);
6303         sgs->group_type = group_classify(env, group, sgs);
6304 }
6305 
6306 /**
6307  * update_sd_pick_busiest - return 1 on busiest group
6308  * @env: The load balancing environment.
6309  * @sds: sched_domain statistics
6310  * @sg: sched_group candidate to be checked for being the busiest
6311  * @sgs: sched_group statistics
6312  *
6313  * Determine if @sg is a busier group than the previously selected
6314  * busiest group.
6315  *
6316  * Return: %true if @sg is a busier group than the previously selected
6317  * busiest group. %false otherwise.
6318  */
6319 static bool update_sd_pick_busiest(struct lb_env *env,
6320                                    struct sd_lb_stats *sds,
6321                                    struct sched_group *sg,
6322                                    struct sg_lb_stats *sgs)
6323 {
6324         struct sg_lb_stats *busiest = &sds->busiest_stat;
6325 
6326         if (sgs->group_type > busiest->group_type)
6327                 return true;
6328 
6329         if (sgs->group_type < busiest->group_type)
6330                 return false;
6331 
6332         if (sgs->avg_load <= busiest->avg_load)
6333                 return false;
6334 
6335         /* This is the busiest node in its class. */
6336         if (!(env->sd->flags & SD_ASYM_PACKING))
6337                 return true;
6338 
6339         /*
6340          * ASYM_PACKING needs to move all the work to the lowest
6341          * numbered CPUs in the group, therefore mark all groups
6342          * higher than ourself as busy.
6343          */
6344         if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6345                 if (!sds->busiest)
6346                         return true;
6347 
6348                 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
6349                         return true;
6350         }
6351 
6352         return false;
6353 }
6354 
6355 #ifdef CONFIG_NUMA_BALANCING
6356 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6357 {
6358         if (sgs->sum_nr_running > sgs->nr_numa_running)
6359                 return regular;
6360         if (sgs->sum_nr_running > sgs->nr_preferred_running)
6361                 return remote;
6362         return all;
6363 }
6364 
6365 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6366 {
6367         if (rq->nr_running > rq->nr_numa_running)
6368                 return regular;
6369         if (rq->nr_running > rq->nr_preferred_running)
6370                 return remote;
6371         return all;
6372 }
6373 #else
6374 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6375 {
6376         return all;
6377 }
6378 
6379 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6380 {
6381         return regular;
6382 }
6383 #endif /* CONFIG_NUMA_BALANCING */
6384 
6385 /**
6386  * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6387  * @env: The load balancing environment.
6388  * @sds: variable to hold the statistics for this sched_domain.
6389  */
6390 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6391 {
6392         struct sched_domain *child = env->sd->child;
6393         struct sched_group *sg = env->sd->groups;
6394         struct sg_lb_stats tmp_sgs;
6395         int load_idx, prefer_sibling = 0;
6396         bool overload = false;
6397 
6398         if (child && child->flags & SD_PREFER_SIBLING)
6399                 prefer_sibling = 1;
6400 
6401         load_idx = get_sd_load_idx(env->sd, env->idle);
6402 
6403         do {
6404                 struct sg_lb_stats *sgs = &tmp_sgs;
6405                 int local_group;
6406 
6407                 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6408                 if (local_group) {
6409                         sds->local = sg;
6410                         sgs = &sds->local_stat;
6411 
6412                         if (env->idle != CPU_NEWLY_IDLE ||
6413                             time_after_eq(jiffies, sg->sgc->next_update))
6414                                 update_group_capacity(env->sd, env->dst_cpu);
6415                 }
6416 
6417                 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6418                                                 &overload);
6419 
6420                 if (local_group)
6421                         goto next_group;
6422 
6423                 /*
6424                  * In case the child domain prefers tasks go to siblings
6425                  * first, lower the sg capacity so that we'll try
6426                  * and move all the excess tasks away. We lower the capacity
6427                  * of a group only if the local group has the capacity to fit
6428                  * these excess tasks. The extra check prevents the case where
6429                  * you always pull from the heaviest group when it is already
6430                  * under-utilized (possible with a large weight task outweighs
6431                  * the tasks on the system).
6432                  */
6433                 if (prefer_sibling && sds->local &&
6434                     group_has_capacity(env, &sds->local_stat) &&
6435                     (sgs->sum_nr_running > 1)) {
6436                         sgs->group_no_capacity = 1;
6437                         sgs->group_type = group_overloaded;
6438                 }
6439 
6440                 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6441                         sds->busiest = sg;
6442                         sds->busiest_stat = *sgs;
6443                 }
6444 
6445 next_group:
6446                 /* Now, start updating sd_lb_stats */
6447                 sds->total_load += sgs->group_load;
6448                 sds->total_capacity += sgs->group_capacity;
6449 
6450                 sg = sg->next;
6451         } while (sg != env->sd->groups);
6452 
6453         if (env->sd->flags & SD_NUMA)
6454                 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6455 
6456         if (!env->sd->parent) {
6457                 /* update overload indicator if we are at root domain */
6458                 if (env->dst_rq->rd->overload != overload)
6459                         env->dst_rq->rd->overload = overload;
6460         }
6461 
6462 }
6463 
6464 /**
6465  * check_asym_packing - Check to see if the group is packed into the
6466  *                      sched doman.
6467  *
6468  * This is primarily intended to used at the sibling level.  Some
6469  * cores like POWER7 prefer to use lower numbered SMT threads.  In the
6470  * case of POWER7, it can move to lower SMT modes only when higher
6471  * threads are idle.  When in lower SMT modes, the threads will
6472  * perform better since they share less core resources.  Hence when we
6473  * have idle threads, we want them to be the higher ones.
6474  *
6475  * This packing function is run on idle threads.  It checks to see if
6476  * the busiest CPU in this domain (core in the P7 case) has a higher
6477  * CPU number than the packing function is being run on.  Here we are
6478  * assuming lower CPU number will be equivalent to lower a SMT thread
6479  * number.
6480  *
6481  * Return: 1 when packing is required and a task should be moved to
6482  * this CPU.  The amount of the imbalance is returned in *imbalance.
6483  *
6484  * @env: The load balancing environment.
6485  * @sds: Statistics of the sched_domain which is to be packed
6486  */
6487 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6488 {
6489         int busiest_cpu;
6490 
6491         if (!(env->sd->flags & SD_ASYM_PACKING))
6492                 return 0;
6493 
6494         if (!sds->busiest)
6495                 return 0;
6496 
6497         busiest_cpu = group_first_cpu(sds->busiest);
6498         if (env->dst_cpu > busiest_cpu)
6499                 return 0;
6500 
6501         env->imbalance = DIV_ROUND_CLOSEST(
6502                 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6503                 SCHED_CAPACITY_SCALE);
6504 
6505         return 1;
6506 }
6507 
6508 /**
6509  * fix_small_imbalance - Calculate the minor imbalance that exists
6510  *                      amongst the groups of a sched_domain, during
6511  *                      load balancing.
6512  * @env: The load balancing environment.
6513  * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6514  */
6515 static inline
6516 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6517 {
6518         unsigned long tmp, capa_now = 0, capa_move = 0;
6519         unsigned int imbn = 2;
6520         unsigned long scaled_busy_load_per_task;
6521         struct sg_lb_stats *local, *busiest;
6522 
6523         local = &sds->local_stat;
6524         busiest = &sds->busiest_stat;
6525 
6526         if (!local->sum_nr_running)
6527                 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6528         else if (busiest->load_per_task > local->load_per_task)
6529                 imbn = 1;
6530 
6531         scaled_busy_load_per_task =
6532                 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6533                 busiest->group_capacity;
6534 
6535         if (busiest->avg_load + scaled_busy_load_per_task >=
6536             local->avg_load + (scaled_busy_load_per_task * imbn)) {
6537                 env->imbalance = busiest->load_per_task;
6538                 return;
6539         }
6540 
6541         /*
6542          * OK, we don't have enough imbalance to justify moving tasks,
6543          * however we may be able to increase total CPU capacity used by
6544          * moving them.
6545          */
6546 
6547         capa_now += busiest->group_capacity *
6548                         min(busiest->load_per_task, busiest->avg_load);
6549         capa_now += local->group_capacity *
6550                         min(local->load_per_task, local->avg_load);
6551         capa_now /= SCHED_CAPACITY_SCALE;
6552 
6553         /* Amount of load we'd subtract */
6554         if (busiest->avg_load > scaled_busy_load_per_task) {
6555                 capa_move += busiest->group_capacity *
6556                             min(busiest->load_per_task,
6557                                 busiest->avg_load - scaled_busy_load_per_task);
6558         }
6559 
6560         /* Amount of load we'd add */
6561         if (busiest->avg_load * busiest->group_capacity <
6562             busiest->load_per_task * SCHED_CAPACITY_SCALE) {
6563                 tmp = (busiest->avg_load * busiest->group_capacity) /
6564                       local->group_capacity;
6565         } else {
6566                 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6567                       local->group_capacity;
6568         }
6569         capa_move += local->group_capacity *
6570                     min(local->load_per_task, local->avg_load + tmp);
6571         capa_move /= SCHED_CAPACITY_SCALE;
6572 
6573         /* Move if we gain throughput */
6574         if (capa_move > capa_now)
6575                 env->imbalance = busiest->load_per_task;
6576 }
6577 
6578 /**
6579  * calculate_imbalance - Calculate the amount of imbalance present within the
6580  *                       groups of a given sched_domain during load balance.
6581  * @env: load balance environment
6582  * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6583  */
6584 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6585 {
6586         unsigned long max_pull, load_above_capacity = ~0UL;
6587         struct sg_lb_stats *local, *busiest;
6588 
6589         local = &sds->local_stat;
6590         busiest = &sds->busiest_stat;
6591 
6592         if (busiest->group_type == group_imbalanced) {
6593                 /*
6594                  * In the group_imb case we cannot rely on group-wide averages
6595                  * to ensure cpu-load equilibrium, look at wider averages. XXX
6596                  */
6597                 busiest->load_per_task =
6598                         min(busiest->load_per_task, sds->avg_load);
6599         }
6600 
6601         /*
6602          * In the presence of smp nice balancing, certain scenarios can have
6603          * max load less than avg load(as we skip the groups at or below
6604          * its cpu_capacity, while calculating max_load..)
6605          */
6606         if (busiest->avg_load <= sds->avg_load ||
6607             local->avg_load >= sds->avg_load) {
6608                 env->imbalance = 0;
6609                 return fix_small_imbalance(env, sds);
6610         }
6611 
6612         /*
6613          * If there aren't any idle cpus, avoid creating some.
6614          */
6615         if (busiest->group_type == group_overloaded &&
6616             local->group_type   == group_overloaded) {
6617                 load_above_capacity = busiest->sum_nr_running *
6618                                         SCHED_LOAD_SCALE;
6619                 if (load_above_capacity > busiest->group_capacity)
6620                         load_above_capacity -= busiest->group_capacity;
6621                 else
6622                         load_above_capacity = ~0UL;
6623         }
6624 
6625         /*
6626          * We're trying to get all the cpus to the average_load, so we don't
6627          * want to push ourselves above the average load, nor do we wish to
6628          * reduce the max loaded cpu below the average load. At the same time,
6629          * we also don't want to reduce the group load below the group capacity
6630          * (so that we can implement power-savings policies etc). Thus we look
6631          * for the minimum possible imbalance.
6632          */
6633         max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6634 
6635         /* How much load to actually move to equalise the imbalance */
6636         env->imbalance = min(
6637                 max_pull * busiest->group_capacity,
6638                 (sds->avg_load - local->avg_load) * local->group_capacity
6639         ) / SCHED_CAPACITY_SCALE;
6640 
6641         /*
6642          * if *imbalance is less than the average load per runnable task
6643          * there is no guarantee that any tasks will be moved so we'll have
6644          * a think about bumping its value to force at least one task to be
6645          * moved
6646          */
6647         if (env->imbalance < busiest->load_per_task)
6648                 return fix_small_imbalance(env, sds);
6649 }
6650 
6651 /******* find_busiest_group() helpers end here *********************/
6652 
6653 /**
6654  * find_busiest_group - Returns the busiest group within the sched_domain
6655  * if there is an imbalance. If there isn't an imbalance, and
6656  * the user has opted for power-savings, it returns a group whose
6657  * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6658  * such a group exists.
6659  *
6660  * Also calculates the amount of weighted load which should be moved
6661  * to restore balance.
6662  *
6663  * @env: The load balancing environment.
6664  *
6665  * Return:      - The busiest group if imbalance exists.
6666  *              - If no imbalance and user has opted for power-savings balance,
6667  *                 return the least loaded group whose CPUs can be
6668  *                 put to idle by rebalancing its tasks onto our group.
6669  */
6670 static struct sched_group *find_busiest_group(struct lb_env *env)
6671 {
6672         struct sg_lb_stats *local, *busiest;
6673         struct sd_lb_stats sds;
6674 
6675         init_sd_lb_stats(&sds);
6676 
6677         /*
6678          * Compute the various statistics relavent for load balancing at
6679          * this level.
6680          */
6681         update_sd_lb_stats(env, &sds);
6682         local = &sds.local_stat;
6683         busiest = &sds.busiest_stat;
6684 
6685         /* ASYM feature bypasses nice load balance check */
6686         if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6687             check_asym_packing(env, &sds))
6688                 return sds.busiest;
6689 
6690         /* There is no busy sibling group to pull tasks from */
6691         if (!sds.busiest || busiest->sum_nr_running == 0)
6692                 goto out_balanced;
6693 
6694         sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
6695                                                 / sds.total_capacity;
6696 
6697         /*
6698          * If the busiest group is imbalanced the below checks don't
6699          * work because they assume all things are equal, which typically
6700          * isn't true due to cpus_allowed constraints and the like.
6701          */
6702         if (busiest->group_type == group_imbalanced)
6703                 goto force_balance;
6704 
6705         /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6706         if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
6707             busiest->group_no_capacity)
6708                 goto force_balance;
6709 
6710         /*
6711          * If the local group is busier than the selected busiest group
6712          * don't try and pull any tasks.
6713          */
6714         if (local->avg_load >= busiest->avg_load)
6715                 goto out_balanced;
6716 
6717         /*
6718          * Don't pull any tasks if this group is already above the domain
6719          * average load.
6720          */
6721         if (local->avg_load >= sds.avg_load)
6722                 goto out_balanced;
6723 
6724         if (env->idle == CPU_IDLE) {
6725                 /*
6726                  * This cpu is idle. If the busiest group is not overloaded
6727                  * and there is no imbalance between this and busiest group
6728                  * wrt idle cpus, it is balanced. The imbalance becomes
6729                  * significant if the diff is greater than 1 otherwise we
6730                  * might end up to just move the imbalance on another group
6731                  */
6732                 if ((busiest->group_type != group_overloaded) &&
6733                                 (local->idle_cpus <= (busiest->idle_cpus + 1)))
6734                         goto out_balanced;
6735         } else {
6736                 /*
6737                  * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6738                  * imbalance_pct to be conservative.
6739                  */
6740                 if (100 * busiest->avg_load <=
6741                                 env->sd->imbalance_pct * local->avg_load)
6742                         goto out_balanced;
6743         }
6744 
6745 force_balance:
6746         /* Looks like there is an imbalance. Compute it */
6747         calculate_imbalance(env, &sds);
6748         return sds.busiest;
6749 
6750 out_balanced:
6751         env->imbalance = 0;
6752         return NULL;
6753 }
6754 
6755 /*
6756  * find_busiest_queue - find the busiest runqueue among the cpus in group.
6757  */
6758 static struct rq *find_busiest_queue(struct lb_env *env,
6759                                      struct sched_group *group)
6760 {
6761         struct rq *busiest = NULL, *rq;
6762         unsigned long busiest_load = 0, busiest_capacity = 1;
6763         int i;
6764 
6765         for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6766                 unsigned long capacity, wl;
6767                 enum fbq_type rt;
6768 
6769                 rq = cpu_rq(i);
6770                 rt = fbq_classify_rq(rq);
6771 
6772                 /*
6773                  * We classify groups/runqueues into three groups:
6774                  *  - regular: there are !numa tasks
6775                  *  - remote:  there are numa tasks that run on the 'wrong' node
6776                  *  - all:     there is no distinction
6777                  *
6778                  * In order to avoid migrating ideally placed numa tasks,
6779                  * ignore those when there's better options.
6780                  *
6781                  * If we ignore the actual busiest queue to migrate another
6782                  * task, the next balance pass can still reduce the busiest
6783                  * queue by moving tasks around inside the node.
6784                  *
6785                  * If we cannot move enough load due to this classification
6786                  * the next pass will adjust the group classification and
6787                  * allow migration of more tasks.
6788                  *
6789                  * Both cases only affect the total convergence complexity.
6790                  */
6791                 if (rt > env->fbq_type)
6792                         continue;
6793 
6794                 capacity = capacity_of(i);
6795 
6796                 wl = weighted_cpuload(i);
6797 
6798                 /*
6799                  * When comparing with imbalance, use weighted_cpuload()
6800                  * which is not scaled with the cpu capacity.
6801                  */
6802 
6803                 if (rq->nr_running == 1 && wl > env->imbalance &&
6804                     !check_cpu_capacity(rq, env->sd))
6805                         continue;
6806 
6807                 /*
6808                  * For the load comparisons with the other cpu's, consider
6809                  * the weighted_cpuload() scaled with the cpu capacity, so
6810                  * that the load can be moved away from the cpu that is
6811                  * potentially running at a lower capacity.
6812                  *
6813                  * Thus we're looking for max(wl_i / capacity_i), crosswise
6814                  * multiplication to rid ourselves of the division works out
6815                  * to: wl_i * capacity_j > wl_j * capacity_i;  where j is
6816                  * our previous maximum.
6817                  */
6818                 if (wl * busiest_capacity > busiest_load * capacity) {
6819                         busiest_load = wl;
6820                         busiest_capacity = capacity;
6821                         busiest = rq;
6822                 }
6823         }
6824 
6825         return busiest;
6826 }
6827 
6828 /*
6829  * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6830  * so long as it is large enough.
6831  */
6832 #define MAX_PINNED_INTERVAL     512
6833 
6834 /* Working cpumask for load_balance and load_balance_newidle. */
6835 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6836 
6837 static int need_active_balance(struct lb_env *env)
6838 {
6839         struct sched_domain *sd = env->sd;
6840 
6841         if (env->idle == CPU_NEWLY_IDLE) {
6842 
6843                 /*
6844                  * ASYM_PACKING needs to force migrate tasks from busy but
6845                  * higher numbered CPUs in order to pack all tasks in the
6846                  * lowest numbered CPUs.
6847                  */
6848                 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6849                         return 1;
6850         }
6851 
6852         /*
6853          * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
6854          * It's worth migrating the task if the src_cpu's capacity is reduced
6855          * because of other sched_class or IRQs if more capacity stays
6856          * available on dst_cpu.
6857          */
6858         if ((env->idle != CPU_NOT_IDLE) &&
6859             (env->src_rq->cfs.h_nr_running == 1)) {
6860                 if ((check_cpu_capacity(env->src_rq, sd)) &&
6861                     (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
6862                         return 1;
6863         }
6864 
6865         return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6866 }
6867 
6868 static int active_load_balance_cpu_stop(void *data);
6869 
6870 static int should_we_balance(struct lb_env *env)
6871 {
6872         struct sched_group *sg = env->sd->groups;
6873         struct cpumask *sg_cpus, *sg_mask;
6874         int cpu, balance_cpu = -1;
6875 
6876         /*
6877          * In the newly idle case, we will allow all the cpu's
6878          * to do the newly idle load balance.
6879          */
6880         if (env->idle == CPU_NEWLY_IDLE)
6881                 return 1;
6882 
6883         sg_cpus = sched_group_cpus(sg);
6884         sg_mask = sched_group_mask(sg);
6885         /* Try to find first idle cpu */
6886         for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6887                 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6888                         continue;
6889 
6890                 balance_cpu = cpu;
6891                 break;
6892         }
6893 
6894         if (balance_cpu == -1)
6895                 balance_cpu = group_balance_cpu(sg);
6896 
6897         /*
6898          * First idle cpu or the first cpu(busiest) in this sched group
6899          * is eligible for doing load balancing at this and above domains.
6900          */
6901         return balance_cpu == env->dst_cpu;
6902 }
6903 
6904 /*
6905  * Check this_cpu to ensure it is balanced within domain. Attempt to move
6906  * tasks if there is an imbalance.
6907  */
6908 static int load_balance(int this_cpu, struct rq *this_rq,
6909                         struct sched_domain *sd, enum cpu_idle_type idle,
6910                         int *continue_balancing)
6911 {
6912         int ld_moved, cur_ld_moved, active_balance = 0;
6913         struct sched_domain *sd_parent = sd->parent;
6914         struct sched_group *group;
6915         struct rq *busiest;
6916         unsigned long flags;
6917         struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
6918 
6919         struct lb_env env = {
6920                 .sd             = sd,
6921                 .dst_cpu        = this_cpu,
6922                 .dst_rq         = this_rq,
6923                 .dst_grpmask    = sched_group_cpus(sd->groups),
6924                 .idle           = idle,
6925                 .loop_break     = sched_nr_migrate_break,
6926                 .cpus           = cpus,
6927                 .fbq_type       = all,
6928                 .tasks          = LIST_HEAD_INIT(env.tasks),
6929         };
6930 
6931         /*
6932          * For NEWLY_IDLE load_balancing, we don't need to consider
6933          * other cpus in our group
6934          */
6935         if (idle == CPU_NEWLY_IDLE)
6936                 env.dst_grpmask = NULL;
6937 
6938         cpumask_copy(cpus, cpu_active_mask);
6939 
6940         schedstat_inc(sd, lb_count[idle]);
6941 
6942 redo:
6943         if (!should_we_balance(&env)) {
6944                 *continue_balancing = 0;
6945                 goto out_balanced;
6946         }
6947 
6948         group = find_busiest_group(&env);
6949         if (!group) {
6950                 schedstat_inc(sd, lb_nobusyg[idle]);
6951                 goto out_balanced;
6952         }
6953 
6954         busiest = find_busiest_queue(&env, group);
6955         if (!busiest) {
6956                 schedstat_inc(sd, lb_nobusyq[idle]);
6957                 goto out_balanced;
6958         }
6959 
6960         BUG_ON(busiest == env.dst_rq);
6961 
6962         schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6963 
6964         env.src_cpu = busiest->cpu;
6965         env.src_rq = busiest;
6966 
6967         ld_moved = 0;
6968         if (busiest->nr_running > 1) {
6969                 /*
6970                  * Attempt to move tasks. If find_busiest_group has found
6971                  * an imbalance but busiest->nr_running <= 1, the group is
6972                  * still unbalanced. ld_moved simply stays zero, so it is
6973                  * correctly treated as an imbalance.
6974                  */
6975                 env.flags |= LBF_ALL_PINNED;
6976                 env.loop_max  = min(sysctl_sched_nr_migrate, busiest->nr_running);
6977 
6978 more_balance:
6979                 raw_spin_lock_irqsave(&busiest->lock, flags);
6980 
6981                 /*
6982                  * cur_ld_moved - load moved in current iteration
6983                  * ld_moved     - cumulative load moved across iterations
6984                  */
6985                 cur_ld_moved = detach_tasks(&env);
6986 
6987                 /*
6988                  * We've detached some tasks from busiest_rq. Every
6989                  * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
6990                  * unlock busiest->lock, and we are able to be sure
6991                  * that nobody can manipulate the tasks in parallel.
6992                  * See task_rq_lock() family for the details.
6993                  */
6994 
6995                 raw_spin_unlock(&busiest->lock);
6996 
6997                 if (cur_ld_moved) {
6998                         attach_tasks(&env);
6999                         ld_moved += cur_ld_moved;
7000                 }
7001 
7002                 local_irq_restore(flags);
7003 
7004                 if (env.flags & LBF_NEED_BREAK) {
7005                         env.flags &= ~LBF_NEED_BREAK;
7006                         goto more_balance;
7007                 }
7008 
7009                 /*
7010                  * Revisit (affine) tasks on src_cpu that couldn't be moved to
7011                  * us and move them to an alternate dst_cpu in our sched_group
7012                  * where they can run. The upper limit on how many times we
7013                  * iterate on same src_cpu is dependent on number of cpus in our
7014                  * sched_group.
7015                  *
7016                  * This changes load balance semantics a bit on who can move
7017                  * load to a given_cpu. In addition to the given_cpu itself
7018                  * (or a ilb_cpu acting on its behalf where given_cpu is
7019                  * nohz-idle), we now have balance_cpu in a position to move
7020                  * load to given_cpu. In rare situations, this may cause
7021                  * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7022                  * _independently_ and at _same_ time to move some load to
7023                  * given_cpu) causing exceess load to be moved to given_cpu.
7024                  * This however should not happen so much in practice and
7025                  * moreover subsequent load balance cycles should correct the
7026                  * excess load moved.
7027                  */
7028                 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7029 
7030                         /* Prevent to re-select dst_cpu via env's cpus */
7031                         cpumask_clear_cpu(env.dst_cpu, env.cpus);
7032 
7033                         env.dst_rq       = cpu_rq(env.new_dst_cpu);
7034                         env.dst_cpu      = env.new_dst_cpu;
7035                         env.flags       &= ~LBF_DST_PINNED;
7036                         env.loop         = 0;
7037                         env.loop_break   = sched_nr_migrate_break;
7038 
7039                         /*
7040                          * Go back to "more_balance" rather than "redo" since we
7041                          * need to continue with same src_cpu.
7042                          */
7043                         goto more_balance;
7044                 }
7045 
7046                 /*
7047                  * We failed to reach balance because of affinity.
7048                  */
7049                 if (sd_parent) {
7050                         int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7051 
7052                         if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7053                                 *group_imbalance = 1;
7054                 }
7055 
7056                 /* All tasks on this runqueue were pinned by CPU affinity */
7057                 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7058                         cpumask_clear_cpu(cpu_of(busiest), cpus);
7059                         if (!cpumask_empty(cpus)) {
7060                                 env.loop = 0;
7061                                 env.loop_break = sched_nr_migrate_break;
7062                                 goto redo;
7063                         }
7064                         goto out_all_pinned;
7065                 }
7066         }
7067 
7068         if (!ld_moved) {
7069                 schedstat_inc(sd, lb_failed[idle]);
7070                 /*
7071                  * Increment the failure counter only on periodic balance.
7072                  * We do not want newidle balance, which can be very
7073                  * frequent, pollute the failure counter causing
7074                  * excessive cache_hot migrations and active balances.
7075                  */
7076                 if (idle != CPU_NEWLY_IDLE)
7077                         sd->nr_balance_failed++;
7078 
7079                 if (need_active_balance(&env)) {
7080                         raw_spin_lock_irqsave(&busiest->lock, flags);
7081 
7082                         /* don't kick the active_load_balance_cpu_stop,
7083                          * if the curr task on busiest cpu can't be
7084                          * moved to this_cpu
7085                          */
7086                         if (!cpumask_test_cpu(this_cpu,
7087                                         tsk_cpus_allowed(busiest->curr))) {
7088                                 raw_spin_unlock_irqrestore(&busiest->lock,
7089                                                             flags);
7090                                 env.flags |= LBF_ALL_PINNED;
7091                                 goto out_one_pinned;
7092                         }
7093 
7094                         /*
7095                          * ->active_balance synchronizes accesses to
7096                          * ->active_balance_work.  Once set, it's cleared
7097                          * only after active load balance is finished.
7098                          */
7099                         if (!busiest->active_balance) {
7100                                 busiest->active_balance = 1;
7101                                 busiest->push_cpu = this_cpu;
7102                                 active_balance = 1;
7103                         }
7104                         raw_spin_unlock_irqrestore(&busiest->lock, flags);
7105 
7106                         if (active_balance) {
7107                                 stop_one_cpu_nowait(cpu_of(busiest),
7108                                         active_load_balance_cpu_stop, busiest,
7109                                         &busiest->active_balance_work);
7110                         }
7111 
7112                         /*
7113                          * We've kicked active balancing, reset the failure
7114                          * counter.
7115                          */
7116                         sd->nr_balance_failed = sd->cache_nice_tries+1;
7117                 }
7118         } else
7119                 sd->nr_balance_failed = 0;
7120 
7121         if (likely(!active_balance)) {
7122                 /* We were unbalanced, so reset the balancing interval */
7123                 sd->balance_interval = sd->min_interval;
7124         } else {
7125                 /*
7126                  * If we've begun active balancing, start to back off. This
7127                  * case may not be covered by the all_pinned logic if there
7128                  * is only 1 task on the busy runqueue (because we don't call
7129                  * detach_tasks).
7130                  */
7131                 if (sd->balance_interval < sd->max_interval)
7132                         sd->balance_interval *= 2;
7133         }
7134 
7135         goto out;
7136 
7137 out_balanced:
7138         /*
7139          * We reach balance although we may have faced some affinity
7140          * constraints. Clear the imbalance flag if it was set.
7141          */
7142         if (sd_parent) {
7143                 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7144 
7145                 if (*group_imbalance)
7146                         *group_imbalance = 0;
7147         }
7148 
7149 out_all_pinned:
7150         /*
7151          * We reach balance because all tasks are pinned at this level so
7152          * we can't migrate them. Let the imbalance flag set so parent level
7153          * can try to migrate them.
7154          */
7155         schedstat_inc(sd, lb_balanced[idle]);
7156 
7157         sd->nr_balance_failed = 0;
7158 
7159 out_one_pinned:
7160         /* tune up the balancing interval */
7161         if (((env.flags & LBF_ALL_PINNED) &&
7162                         sd->balance_interval < MAX_PINNED_INTERVAL) ||
7163                         (sd->balance_interval < sd->max_interval))
7164                 sd->balance_interval *= 2;
7165 
7166         ld_moved = 0;
7167 out:
7168         return ld_moved;
7169 }
7170 
7171 static inline unsigned long
7172 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7173 {
7174         unsigned long interval = sd->balance_interval;
7175 
7176         if (cpu_busy)
7177                 interval *= sd->busy_factor;
7178 
7179         /* scale ms to jiffies */
7180         interval = msecs_to_jiffies(interval);
7181         interval = clamp(interval, 1UL, max_load_balance_interval);
7182 
7183         return interval;
7184 }
7185 
7186 static inline void
7187 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
7188 {
7189         unsigned long interval, next;
7190 
7191         interval = get_sd_balance_interval(sd, cpu_busy);
7192         next = sd->last_balance + interval;
7193 
7194         if (time_after(*next_balance, next))
7195                 *next_balance = next;
7196 }
7197 
7198 /*
7199  * idle_balance is called by schedule() if this_cpu is about to become
7200  * idle. Attempts to pull tasks from other CPUs.
7201  */
7202 static int idle_balance(struct rq *this_rq)
7203 {
7204         unsigned long next_balance = jiffies + HZ;
7205         int this_cpu = this_rq->cpu;
7206         struct sched_domain *sd;
7207         int pulled_task = 0;
7208         u64 curr_cost = 0;
7209 
7210         idle_enter_fair(this_rq);
7211 
7212         /*
7213          * We must set idle_stamp _before_ calling idle_balance(), such that we
7214          * measure the duration of idle_balance() as idle time.
7215          */
7216         this_rq->idle_stamp = rq_clock(this_rq);
7217 
7218         if (this_rq->avg_idle < sysctl_sched_migration_cost ||
7219             !this_rq->rd->overload) {
7220                 rcu_read_lock();
7221                 sd = rcu_dereference_check_sched_domain(this_rq->sd);
7222                 if (sd)
7223                         update_next_balance(sd, 0, &next_balance);
7224                 rcu_read_unlock();
7225 
7226                 goto out;
7227         }
7228 
7229         /*
7230          * Drop the rq->lock, but keep IRQ/preempt disabled.
7231          */
7232         raw_spin_unlock(&this_rq->lock);
7233 
7234         update_blocked_averages(this_cpu);
7235         rcu_read_lock();
7236         for_each_domain(this_cpu, sd) {
7237                 int continue_balancing = 1;
7238                 u64 t0, domain_cost;
7239 
7240                 if (!(sd->flags & SD_LOAD_BALANCE))
7241                         continue;
7242 
7243                 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7244                         update_next_balance(sd, 0, &next_balance);
7245                         break;
7246                 }
7247 
7248                 if (sd->flags & SD_BALANCE_NEWIDLE) {
7249                         t0 = sched_clock_cpu(this_cpu);
7250 
7251                         pulled_task = load_balance(this_cpu, this_rq,
7252                                                    sd, CPU_NEWLY_IDLE,
7253                                                    &continue_balancing);
7254 
7255                         domain_cost = sched_clock_cpu(this_cpu) - t0;
7256                         if (domain_cost > sd->max_newidle_lb_cost)
7257                                 sd->max_newidle_lb_cost = domain_cost;
7258 
7259                         curr_cost += domain_cost;
7260                 }
7261 
7262                 update_next_balance(sd, 0, &next_balance);
7263 
7264                 /*
7265                  * Stop searching for tasks to pull if there are
7266                  * now runnable tasks on this rq.
7267                  */
7268                 if (pulled_task || this_rq->nr_running > 0)
7269                         break;
7270         }
7271         rcu_read_unlock();
7272 
7273         raw_spin_lock(&this_rq->lock);
7274 
7275         if (curr_cost > this_rq->max_idle_balance_cost)
7276                 this_rq->max_idle_balance_cost = curr_cost;
7277 
7278         /*
7279          * While browsing the domains, we released the rq lock, a task could
7280          * have been enqueued in the meantime. Since we're not going idle,
7281          * pretend we pulled a task.
7282          */
7283         if (this_rq->cfs.h_nr_running && !pulled_task)
7284                 pulled_task = 1;
7285 
7286 out:
7287         /* Move the next balance forward */
7288         if (time_after(this_rq->next_balance, next_balance))
7289                 this_rq->next_balance = next_balance;
7290 
7291         /* Is there a task of a high priority class? */
7292         if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7293                 pulled_task = -1;
7294 
7295         if (pulled_task) {
7296                 idle_exit_fair(this_rq);
7297                 this_rq->idle_stamp = 0;
7298         }
7299 
7300         return pulled_task;
7301 }
7302 
7303 /*
7304  * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7305  * running tasks off the busiest CPU onto idle CPUs. It requires at
7306  * least 1 task to be running on each physical CPU where possible, and
7307  * avoids physical / logical imbalances.
7308  */
7309 static int active_load_balance_cpu_stop(void *data)
7310 {
7311         struct rq *busiest_rq = data;
7312         int busiest_cpu = cpu_of(busiest_rq);
7313         int target_cpu = busiest_rq->push_cpu;
7314         struct rq *target_rq = cpu_rq(target_cpu);
7315         struct sched_domain *sd;
7316         struct task_struct *p = NULL;
7317 
7318         raw_spin_lock_irq(&busiest_rq->lock);
7319 
7320         /* make sure the requested cpu hasn't gone down in the meantime */
7321         if (unlikely(busiest_cpu != smp_processor_id() ||
7322                      !busiest_rq->active_balance))
7323                 goto out_unlock;
7324 
7325         /* Is there any task to move? */
7326         if (busiest_rq->nr_running <= 1)
7327                 goto out_unlock;
7328 
7329         /*
7330          * This condition is "impossible", if it occurs
7331          * we need to fix it. Originally reported by
7332          * Bjorn Helgaas on a 128-cpu setup.
7333          */
7334         BUG_ON(busiest_rq == target_rq);
7335 
7336         /* Search for an sd spanning us and the target CPU. */
7337         rcu_read_lock();
7338         for_each_domain(target_cpu, sd) {
7339                 if ((sd->flags & SD_LOAD_BALANCE) &&
7340                     cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7341                                 break;
7342         }
7343 
7344         if (likely(sd)) {
7345                 struct lb_env env = {
7346                         .sd             = sd,
7347                         .dst_cpu        = target_cpu,
7348                         .dst_rq         = target_rq,
7349                         .src_cpu        = busiest_rq->cpu,
7350                         .src_rq         = busiest_rq,
7351                         .idle           = CPU_IDLE,
7352                 };
7353 
7354                 schedstat_inc(sd, alb_count);
7355 
7356                 p = detach_one_task(&env);
7357                 if (p)
7358                         schedstat_inc(sd, alb_pushed);
7359                 else
7360                         schedstat_inc(sd, alb_failed);
7361         }
7362         rcu_read_unlock();
7363 out_unlock:
7364         busiest_rq->active_balance = 0;
7365         raw_spin_unlock(&busiest_rq->lock);
7366 
7367         if (p)
7368                 attach_one_task(target_rq, p);
7369 
7370         local_irq_enable();
7371 
7372         return 0;
7373 }
7374 
7375 static inline int on_null_domain(struct rq *rq)
7376 {
7377         return unlikely(!rcu_dereference_sched(rq->sd));
7378 }
7379 
7380 #ifdef CONFIG_NO_HZ_COMMON
7381 /*
7382  * idle load balancing details
7383  * - When one of the busy CPUs notice that there may be an idle rebalancing
7384  *   needed, they will kick the idle load balancer, which then does idle
7385  *   load balancing for all the idle CPUs.
7386  */
7387 static struct {
7388         cpumask_var_t idle_cpus_mask;
7389         atomic_t nr_cpus;
7390         unsigned long next_balance;     /* in jiffy units */
7391 } nohz ____cacheline_aligned;
7392 
7393 static inline int find_new_ilb(void)
7394 {
7395         int ilb = cpumask_first(nohz.idle_cpus_mask);
7396 
7397         if (ilb < nr_cpu_ids && idle_cpu(ilb))
7398                 return ilb;
7399 
7400         return nr_cpu_ids;
7401 }
7402 
7403 /*
7404  * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7405  * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7406  * CPU (if there is one).
7407  */
7408 static void nohz_balancer_kick(void)
7409 {
7410         int ilb_cpu;
7411 
7412         nohz.next_balance++;
7413 
7414         ilb_cpu = find_new_ilb();
7415 
7416         if (ilb_cpu >= nr_cpu_ids)
7417                 return;
7418 
7419         if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7420                 return;
7421         /*
7422          * Use smp_send_reschedule() instead of resched_cpu().
7423          * This way we generate a sched IPI on the target cpu which
7424          * is idle. And the softirq performing nohz idle load balance
7425          * will be run before returning from the IPI.
7426          */
7427         smp_send_reschedule(ilb_cpu);
7428         return;
7429 }
7430 
7431 static inline void nohz_balance_exit_idle(int cpu)
7432 {
7433         if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7434                 /*
7435                  * Completely isolated CPUs don't ever set, so we must test.
7436                  */
7437                 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7438                         cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7439                         atomic_dec(&nohz.nr_cpus);
7440                 }
7441                 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7442         }
7443 }
7444 
7445 static inline void set_cpu_sd_state_busy(void)
7446 {
7447         struct sched_domain *sd;
7448         int cpu = smp_processor_id();
7449 
7450         rcu_read_lock();
7451         sd = rcu_dereference(per_cpu(sd_busy, cpu));
7452 
7453         if (!sd || !sd->nohz_idle)
7454                 goto unlock;
7455         sd->nohz_idle = 0;
7456 
7457         atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7458 unlock:
7459         rcu_read_unlock();
7460 }
7461 
7462 void set_cpu_sd_state_idle(void)
7463 {
7464         struct sched_domain *sd;
7465         int cpu = smp_processor_id();
7466 
7467         rcu_read_lock();
7468         sd = rcu_dereference(per_cpu(sd_busy, cpu));
7469 
7470         if (!sd || sd->nohz_idle)
7471                 goto unlock;
7472         sd->nohz_idle = 1;
7473 
7474         atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7475 unlock:
7476         rcu_read_unlock();
7477 }
7478 
7479 /*
7480  * This routine will record that the cpu is going idle with tick stopped.
7481  * This info will be used in performing idle load balancing in the future.
7482  */
7483 void nohz_balance_enter_idle(int cpu)
7484 {
7485         /*
7486          * If this cpu is going down, then nothing needs to be done.
7487          */
7488         if (!cpu_active(cpu))
7489                 return;
7490 
7491         if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7492                 return;
7493 
7494         /*
7495          * If we're a completely isolated CPU, we don't play.
7496          */
7497         if (on_null_domain(cpu_rq(cpu)))
7498                 return;
7499 
7500         cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7501         atomic_inc(&nohz.nr_cpus);
7502         set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7503 }
7504 
7505 static int sched_ilb_notifier(struct notifier_block *nfb,
7506                                         unsigned long action, void *hcpu)
7507 {
7508         switch (action & ~CPU_TASKS_FROZEN) {
7509         case CPU_DYING:
7510                 nohz_balance_exit_idle(smp_processor_id());
7511                 return NOTIFY_OK;
7512         default:
7513                 return NOTIFY_DONE;
7514         }
7515 }
7516 #endif
7517 
7518 static DEFINE_SPINLOCK(balancing);
7519 
7520 /*
7521  * Scale the max load_balance interval with the number of CPUs in the system.
7522  * This trades load-balance latency on larger machines for less cross talk.
7523  */
7524 void update_max_interval(void)
7525 {
7526         max_load_balance_interval = HZ*num_online_cpus()/10;
7527 }
7528 
7529 /*
7530  * It checks each scheduling domain to see if it is due to be balanced,
7531  * and initiates a balancing operation if so.
7532  *
7533  * Balancing parameters are set up in init_sched_domains.
7534  */
7535 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7536 {
7537         int continue_balancing = 1;
7538         int cpu = rq->cpu;
7539         unsigned long interval;
7540         struct sched_domain *sd;
7541         /* Earliest time when we have to do rebalance again */
7542         unsigned long next_balance = jiffies + 60*HZ;
7543         int update_next_balance = 0;
7544         int need_serialize, need_decay = 0;
7545         u64 max_cost = 0;
7546 
7547         update_blocked_averages(cpu);
7548 
7549         rcu_read_lock();
7550         for_each_domain(cpu, sd) {
7551                 /*
7552                  * Decay the newidle max times here because this is a regular
7553                  * visit to all the domains. Decay ~1% per second.
7554                  */
7555                 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7556                         sd->max_newidle_lb_cost =
7557                                 (sd->max_newidle_lb_cost * 253) / 256;
7558                         sd->next_decay_max_lb_cost = jiffies + HZ;
7559                         need_decay = 1;
7560                 }
7561                 max_cost += sd->max_newidle_lb_cost;
7562 
7563                 if (!(sd->flags & SD_LOAD_BALANCE))
7564                         continue;
7565 
7566                 /*
7567                  * Stop the load balance at this level. There is another
7568                  * CPU in our sched group which is doing load balancing more
7569                  * actively.
7570                  */
7571                 if (!continue_balancing) {
7572                         if (need_decay)
7573                                 continue;
7574                         break;
7575                 }
7576 
7577                 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7578 
7579                 need_serialize = sd->flags & SD_SERIALIZE;
7580                 if (need_serialize) {
7581                         if (!spin_trylock(&balancing))
7582                                 goto out;
7583                 }
7584 
7585                 if (time_after_eq(jiffies, sd->last_balance + interval)) {
7586                         if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7587                                 /*
7588                                  * The LBF_DST_PINNED logic could have changed
7589                                  * env->dst_cpu, so we can't know our idle
7590                                  * state even if we migrated tasks. Update it.
7591                                  */
7592                                 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7593                         }
7594                         sd->last_balance = jiffies;
7595                         interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7596                 }
7597                 if (need_serialize)
7598                         spin_unlock(&balancing);
7599 out:
7600                 if (time_after(next_balance, sd->last_balance + interval)) {
7601                         next_balance = sd->last_balance + interval;
7602                         update_next_balance = 1;
7603                 }
7604         }
7605         if (need_decay) {
7606                 /*
7607                  * Ensure the rq-wide value also decays but keep it at a
7608                  * reasonable floor to avoid funnies with rq->avg_idle.
7609                  */
7610                 rq->max_idle_balance_cost =
7611                         max((u64)sysctl_sched_migration_cost, max_cost);
7612         }
7613         rcu_read_unlock();
7614 
7615         /*
7616          * next_balance will be updated only when there is a need.
7617          * When the cpu is attached to null domain for ex, it will not be
7618          * updated.
7619          */
7620         if (likely(update_next_balance))
7621                 rq->next_balance = next_balance;
7622 }
7623 
7624 #ifdef CONFIG_NO_HZ_COMMON
7625 /*
7626  * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7627  * rebalancing for all the cpus for whom scheduler ticks are stopped.
7628  */
7629 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7630 {
7631         int this_cpu = this_rq->cpu;
7632         struct rq *rq;
7633         int balance_cpu;
7634 
7635         if (idle != CPU_IDLE ||
7636             !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7637                 goto end;
7638 
7639         for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7640                 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7641                         continue;
7642 
7643                 /*
7644                  * If this cpu gets work to do, stop the load balancing
7645                  * work being done for other cpus. Next load
7646                  * balancing owner will pick it up.
7647                  */
7648                 if (need_resched())
7649                         break;
7650 
7651                 rq = cpu_rq(balance_cpu);
7652 
7653                 /*
7654                  * If time for next balance is due,
7655                  * do the balance.
7656                  */
7657                 if (time_after_eq(jiffies, rq->next_balance)) {
7658                         raw_spin_lock_irq(&rq->lock);
7659                         update_rq_clock(rq);
7660                         update_idle_cpu_load(rq);
7661                         raw_spin_unlock_irq(&rq->lock);
7662                         rebalance_domains(rq, CPU_IDLE);
7663                 }
7664 
7665                 if (time_after(this_rq->next_balance, rq->next_balance))
7666                         this_rq->next_balance = rq->next_balance;
7667         }
7668         nohz.next_balance = this_rq->next_balance;
7669 end:
7670         clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7671 }
7672 
7673 /*
7674  * Current heuristic for kicking the idle load balancer in the presence
7675  * of an idle cpu in the system.
7676  *   - This rq has more than one task.
7677  *   - This rq has at least one CFS task and the capacity of the CPU is
7678  *     significantly reduced because of RT tasks or IRQs.
7679  *   - At parent of LLC scheduler domain level, this cpu's scheduler group has
7680  *     multiple busy cpu.
7681  *   - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7682  *     domain span are idle.
7683  */
7684 static inline bool nohz_kick_needed(struct rq *rq)
7685 {
7686         unsigned long now = jiffies;
7687         struct sched_domain *sd;
7688         struct sched_group_capacity *sgc;
7689         int nr_busy, cpu = rq->cpu;
7690         bool kick = false;
7691 
7692         if (unlikely(rq->idle_balance))
7693                 return false;
7694 
7695        /*
7696         * We may be recently in ticked or tickless idle mode. At the first
7697         * busy tick after returning from idle, we will update the busy stats.
7698         */
7699         set_cpu_sd_state_busy();
7700         nohz_balance_exit_idle(cpu);
7701 
7702         /*
7703          * None are in tickless mode and hence no need for NOHZ idle load
7704          * balancing.
7705          */
7706         if (likely(!atomic_read(&nohz.nr_cpus)))
7707                 return false;
7708 
7709         if (time_before(now, nohz.next_balance))
7710                 return false;
7711 
7712         if (rq->nr_running >= 2)
7713                 return true;
7714 
7715         rcu_read_lock();
7716         sd = rcu_dereference(per_cpu(sd_busy, cpu));
7717         if (sd) {
7718                 sgc = sd->groups->sgc;
7719                 nr_busy = atomic_read(&sgc->nr_busy_cpus);
7720 
7721                 if (nr_busy > 1) {
7722                         kick = true;
7723                         goto unlock;
7724                 }
7725 
7726         }
7727 
7728         sd = rcu_dereference(rq->sd);
7729         if (sd) {
7730                 if ((rq->cfs.h_nr_running >= 1) &&
7731                                 check_cpu_capacity(rq, sd)) {
7732                         kick = true;
7733                         goto unlock;
7734                 }
7735         }
7736 
7737         sd = rcu_dereference(per_cpu(sd_asym, cpu));
7738         if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7739                                   sched_domain_span(sd)) < cpu)) {
7740                 kick = true;
7741                 goto unlock;
7742         }
7743 
7744 unlock:
7745         rcu_read_unlock();
7746         return kick;
7747 }
7748 #else
7749 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7750 #endif
7751 
7752 /*
7753  * run_rebalance_domains is triggered when needed from the scheduler tick.
7754  * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7755  */
7756 static void run_rebalance_domains(struct softirq_action *h)
7757 {
7758         struct rq *this_rq = this_rq();
7759         enum cpu_idle_type idle = this_rq->idle_balance ?
7760                                                 CPU_IDLE : CPU_NOT_IDLE;
7761 
7762         /*
7763          * If this cpu has a pending nohz_balance_kick, then do the
7764          * balancing on behalf of the other idle cpus whose ticks are
7765          * stopped. Do nohz_idle_balance *before* rebalance_domains to
7766          * give the idle cpus a chance to load balance. Else we may
7767          * load balance only within the local sched_domain hierarchy
7768          * and abort nohz_idle_balance altogether if we pull some load.
7769          */
7770         nohz_idle_balance(this_rq, idle);
7771         rebalance_domains(this_rq, idle);
7772 }
7773 
7774 /*
7775  * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7776  */
7777 void trigger_load_balance(struct rq *rq)
7778 {
7779         /* Don't need to rebalance while attached to NULL domain */
7780         if (unlikely(on_null_domain(rq)))
7781                 return;
7782 
7783         if (time_after_eq(jiffies, rq->next_balance))
7784                 raise_softirq(SCHED_SOFTIRQ);
7785 #ifdef CONFIG_NO_HZ_COMMON
7786         if (nohz_kick_needed(rq))
7787                 nohz_balancer_kick();
7788 #endif
7789 }
7790 
7791 static void rq_online_fair(struct rq *rq)
7792 {
7793         update_sysctl();
7794 
7795         update_runtime_enabled(rq);
7796 }
7797 
7798 static void rq_offline_fair(struct rq *rq)
7799 {
7800         update_sysctl();
7801 
7802         /* Ensure any throttled groups are reachable by pick_next_task */
7803         unthrottle_offline_cfs_rqs(rq);
7804 }
7805 
7806 #endif /* CONFIG_SMP */
7807 
7808 /*
7809  * scheduler tick hitting a task of our scheduling class:
7810  */
7811 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7812 {
7813         struct cfs_rq *cfs_rq;
7814         struct sched_entity *se = &curr->se;
7815 
7816         for_each_sched_entity(se) {
7817                 cfs_rq = cfs_rq_of(se);
7818                 entity_tick(cfs_rq, se, queued);
7819         }
7820 
7821         if (numabalancing_enabled)
7822                 task_tick_numa(rq, curr);
7823 
7824         update_rq_runnable_avg(rq, 1);
7825 }
7826 
7827 /*
7828  * called on fork with the child task as argument from the parent's context
7829  *  - child not yet on the tasklist
7830  *  - preemption disabled
7831  */
7832 static void task_fork_fair(struct task_struct *p)
7833 {
7834         struct cfs_rq *cfs_rq;
7835         struct sched_entity *se = &p->se, *curr;
7836         int this_cpu = smp_processor_id();
7837         struct rq *rq = this_rq();
7838         unsigned long flags;
7839 
7840         raw_spin_lock_irqsave(&rq->lock, flags);
7841 
7842         update_rq_clock(rq);
7843 
7844         cfs_rq = task_cfs_rq(current);
7845         curr = cfs_rq->curr;
7846 
7847         /*
7848          * Not only the cpu but also the task_group of the parent might have
7849          * been changed after parent->se.parent,cfs_rq were copied to
7850          * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7851          * of child point to valid ones.
7852          */
7853         rcu_read_lock();
7854         __set_task_cpu(p, this_cpu);
7855         rcu_read_unlock();
7856 
7857         update_curr(cfs_rq);
7858 
7859         if (curr)
7860                 se->vruntime = curr->vruntime;
7861         place_entity(cfs_rq, se, 1);
7862 
7863         if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
7864                 /*
7865                  * Upon rescheduling, sched_class::put_prev_task() will place
7866                  * 'current' within the tree based on its new key value.
7867                  */
7868                 swap(curr->vruntime, se->vruntime);
7869                 resched_curr(rq);
7870         }
7871 
7872         se->vruntime -= cfs_rq->min_vruntime;
7873 
7874         raw_spin_unlock_irqrestore(&rq->lock, flags);
7875 }
7876 
7877 /*
7878  * Priority of the task has changed. Check to see if we preempt
7879  * the current task.
7880  */
7881 static void
7882 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
7883 {
7884         if (!task_on_rq_queued(p))
7885                 return;
7886 
7887         /*
7888          * Reschedule if we are currently running on this runqueue and
7889          * our priority decreased, or if we are not currently running on
7890          * this runqueue and our priority is higher than the current's
7891          */
7892         if (rq->curr == p) {
7893                 if (p->prio > oldprio)
7894                         resched_curr(rq);
7895         } else
7896                 check_preempt_curr(rq, p, 0);
7897 }
7898 
7899 static void switched_from_fair(struct rq *rq, struct task_struct *p)
7900 {
7901         struct sched_entity *se = &p->se;
7902         struct cfs_rq *cfs_rq = cfs_rq_of(se);
7903 
7904         /*
7905          * Ensure the task's vruntime is normalized, so that when it's
7906          * switched back to the fair class the enqueue_entity(.flags=0) will
7907          * do the right thing.
7908          *
7909          * If it's queued, then the dequeue_entity(.flags=0) will already
7910          * have normalized the vruntime, if it's !queued, then only when
7911          * the task is sleeping will it still have non-normalized vruntime.
7912          */
7913         if (!task_on_rq_queued(p) && p->state != TASK_RUNNING) {
7914                 /*
7915                  * Fix up our vruntime so that the current sleep doesn't
7916                  * cause 'unlimited' sleep bonus.
7917                  */
7918                 place_entity(cfs_rq, se, 0);
7919                 se->vruntime -= cfs_rq->min_vruntime;
7920         }
7921 
7922 #ifdef CONFIG_SMP
7923         /*
7924         * Remove our load from contribution when we leave sched_fair
7925         * and ensure we don't carry in an old decay_count if we
7926         * switch back.
7927         */
7928         if (se->avg.decay_count) {
7929                 __synchronize_entity_decay(se);
7930                 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
7931         }
7932 #endif
7933 }
7934 
7935 /*
7936  * We switched to the sched_fair class.
7937  */
7938 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7939 {
7940 #ifdef CONFIG_FAIR_GROUP_SCHED
7941         struct sched_entity *se = &p->se;
7942         /*
7943          * Since the real-depth could have been changed (only FAIR
7944          * class maintain depth value), reset depth properly.
7945          */
7946         se->depth = se->parent ? se->parent->depth + 1 : 0;
7947 #endif
7948         if (!task_on_rq_queued(p))
7949                 return;
7950 
7951         /*
7952          * We were most likely switched from sched_rt, so
7953          * kick off the schedule if running, otherwise just see
7954          * if we can still preempt the current task.
7955          */
7956         if (rq->curr == p)
7957                 resched_curr(rq);
7958         else
7959                 check_preempt_curr(rq, p, 0);
7960 }
7961 
7962 /* Account for a task changing its policy or group.
7963  *
7964  * This routine is mostly called to set cfs_rq->curr field when a task
7965  * migrates between groups/classes.
7966  */
7967 static void set_curr_task_fair(struct rq *rq)
7968 {
7969         struct sched_entity *se = &rq->curr->se;
7970 
7971         for_each_sched_entity(se) {
7972                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7973 
7974                 set_next_entity(cfs_rq, se);
7975                 /* ensure bandwidth has been allocated on our new cfs_rq */
7976                 account_cfs_rq_runtime(cfs_rq, 0);
7977         }
7978 }
7979 
7980 void init_cfs_rq(struct cfs_rq *cfs_rq)
7981 {
7982         cfs_rq->tasks_timeline = RB_ROOT;
7983         cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7984 #ifndef CONFIG_64BIT
7985         cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7986 #endif
7987 #ifdef CONFIG_SMP
7988         atomic64_set(&cfs_rq->decay_counter, 1);
7989         atomic_long_set(&cfs_rq->removed_load, 0);
7990 #endif
7991 }
7992 
7993 #ifdef CONFIG_FAIR_GROUP_SCHED
7994 static void task_move_group_fair(struct task_struct *p, int queued)
7995 {
7996         struct sched_entity *se = &p->se;
7997         struct cfs_rq *cfs_rq;
7998 
7999         /*
8000          * If the task was not on the rq at the time of this cgroup movement
8001          * it must have been asleep, sleeping tasks keep their ->vruntime
8002          * absolute on their old rq until wakeup (needed for the fair sleeper
8003          * bonus in place_entity()).
8004          *
8005          * If it was on the rq, we've just 'preempted' it, which does convert
8006          * ->vruntime to a relative base.
8007          *
8008          * Make sure both cases convert their relative position when migrating
8009          * to another cgroup's rq. This does somewhat interfere with the
8010          * fair sleeper stuff for the first placement, but who cares.
8011          */
8012         /*
8013          * When !queued, vruntime of the task has usually NOT been normalized.
8014          * But there are some cases where it has already been normalized:
8015          *
8016          * - Moving a forked child which is waiting for being woken up by
8017          *   wake_up_new_task().
8018          * - Moving a task which has been woken up by try_to_wake_up() and
8019          *   waiting for actually being woken up by sched_ttwu_pending().
8020          *
8021          * To prevent boost or penalty in the new cfs_rq caused by delta
8022          * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
8023          */
8024         if (!queued && (!se->sum_exec_runtime || p->state == TASK_WAKING))
8025                 queued = 1;
8026 
8027         if (!queued)
8028                 se->vruntime -= cfs_rq_of(se)->min_vruntime;
8029         set_task_rq(p, task_cpu(p));
8030         se->depth = se->parent ? se->parent->depth + 1 : 0;
8031         if (!queued) {
8032                 cfs_rq = cfs_rq_of(se);
8033                 se->vruntime += cfs_rq->min_vruntime;
8034 #ifdef CONFIG_SMP
8035                 /*
8036                  * migrate_task_rq_fair() will have removed our previous
8037                  * contribution, but we must synchronize for ongoing future
8038                  * decay.
8039                  */
8040                 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
8041                 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
8042 #endif
8043         }
8044 }
8045 
8046 void free_fair_sched_group(struct task_group *tg)
8047 {
8048         int i;
8049 
8050         destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8051 
8052         for_each_possible_cpu(i) {
8053                 if (tg->cfs_rq)
8054                         kfree(tg->cfs_rq[i]);
8055                 if (tg->se)
8056                         kfree(tg->se[i]);
8057         }
8058 
8059         kfree(tg->cfs_rq);
8060         kfree(tg->se);
8061 }
8062 
8063 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8064 {
8065         struct cfs_rq *cfs_rq;
8066         struct sched_entity *se;
8067         int i;
8068 
8069         tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8070         if (!tg->cfs_rq)
8071                 goto err;
8072         tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8073         if (!tg->se)
8074                 goto err;
8075 
8076         tg->shares = NICE_0_LOAD;
8077 
8078         init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8079 
8080         for_each_possible_cpu(i) {
8081                 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8082                                       GFP_KERNEL, cpu_to_node(i));
8083                 if (!cfs_rq)
8084                         goto err;
8085 
8086                 se = kzalloc_node(sizeof(struct sched_entity),
8087                                   GFP_KERNEL, cpu_to_node(i));
8088                 if (!se)
8089                         goto err_free_rq;
8090 
8091                 init_cfs_rq(cfs_rq);
8092                 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8093         }
8094 
8095         return 1;
8096 
8097 err_free_rq:
8098         kfree(cfs_rq);
8099 err:
8100         return 0;
8101 }
8102 
8103 void unregister_fair_sched_group(struct task_group *tg, int cpu)
8104 {
8105         struct rq *rq = cpu_rq(cpu);
8106         unsigned long flags;
8107 
8108         /*
8109         * Only empty task groups can be destroyed; so we can speculatively
8110         * check on_list without danger of it being re-added.
8111         */
8112         if (!tg->cfs_rq[cpu]->on_list)
8113                 return;
8114 
8115         raw_spin_lock_irqsave(&rq->lock, flags);
8116         list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8117         raw_spin_unlock_irqrestore(&rq->lock, flags);
8118 }
8119 
8120 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8121                         struct sched_entity *se, int cpu,
8122                         struct sched_entity *parent)
8123 {
8124         struct rq *rq = cpu_rq(cpu);
8125 
8126         cfs_rq->tg = tg;
8127         cfs_rq->rq = rq;
8128         init_cfs_rq_runtime(cfs_rq);
8129 
8130         tg->cfs_rq[cpu] = cfs_rq;
8131         tg->se[cpu] = se;
8132 
8133         /* se could be NULL for root_task_group */
8134         if (!se)
8135                 return;
8136 
8137         if (!parent) {
8138                 se->cfs_rq = &rq->cfs;
8139                 se->depth = 0;
8140         } else {
8141                 se->cfs_rq = parent->my_q;
8142                 se->depth = parent->depth + 1;
8143         }
8144 
8145         se->my_q = cfs_rq;
8146         /* guarantee group entities always have weight */
8147         update_load_set(&se->load, NICE_0_LOAD);
8148         se->parent = parent;
8149 }
8150 
8151 static DEFINE_MUTEX(shares_mutex);
8152 
8153 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8154 {
8155         int i;
8156         unsigned long flags;
8157 
8158         /*
8159          * We can't change the weight of the root cgroup.
8160          */
8161         if (!tg->se[0])
8162                 return -EINVAL;
8163 
8164         shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8165 
8166         mutex_lock(&shares_mutex);
8167         if (tg->shares == shares)
8168                 goto done;
8169 
8170         tg->shares = shares;
8171         for_each_possible_cpu(i) {
8172                 struct rq *rq = cpu_rq(i);
8173                 struct sched_entity *se;
8174 
8175                 se = tg->se[i];
8176                 /* Propagate contribution to hierarchy */
8177                 raw_spin_lock_irqsave(&rq->lock, flags);
8178 
8179                 /* Possible calls to update_curr() need rq clock */
8180                 update_rq_clock(rq);
8181                 for_each_sched_entity(se)
8182                         update_cfs_shares(group_cfs_rq(se));
8183                 raw_spin_unlock_irqrestore(&rq->lock, flags);
8184         }
8185 
8186 done:
8187         mutex_unlock(&shares_mutex);
8188         return 0;
8189 }
8190 #else /* CONFIG_FAIR_GROUP_SCHED */
8191 
8192 void free_fair_sched_group(struct task_group *tg) { }
8193 
8194 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8195 {
8196         return 1;
8197 }
8198 
8199 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
8200 
8201 #endif /* CONFIG_FAIR_GROUP_SCHED */
8202 
8203 
8204 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
8205 {
8206         struct sched_entity *se = &task->se;
8207         unsigned int rr_interval = 0;
8208 
8209         /*
8210          * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8211          * idle runqueue:
8212          */
8213         if (rq->cfs.load.weight)
8214                 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
8215 
8216         return rr_interval;
8217 }
8218 
8219 /*
8220  * All the scheduling class methods:
8221  */
8222 const struct sched_class fair_sched_class = {
8223         .next                   = &idle_sched_class,
8224         .enqueue_task           = enqueue_task_fair,
8225         .dequeue_task           = dequeue_task_fair,
8226         .yield_task             = yield_task_fair,
8227         .yield_to_task          = yield_to_task_fair,
8228 
8229         .check_preempt_curr     = check_preempt_wakeup,
8230 
8231         .pick_next_task         = pick_next_task_fair,
8232         .put_prev_task          = put_prev_task_fair,
8233 
8234 #ifdef CONFIG_SMP
8235         .select_task_rq         = select_task_rq_fair,
8236         .migrate_task_rq        = migrate_task_rq_fair,
8237 
8238         .rq_online              = rq_online_fair,
8239         .rq_offline             = rq_offline_fair,
8240 
8241         .task_waking            = task_waking_fair,
8242 #endif
8243 
8244         .set_curr_task          = set_curr_task_fair,
8245         .task_tick              = task_tick_fair,
8246         .task_fork              = task_fork_fair,
8247 
8248         .prio_changed           = prio_changed_fair,
8249         .switched_from          = switched_from_fair,
8250         .switched_to            = switched_to_fair,
8251 
8252         .get_rr_interval        = get_rr_interval_fair,
8253 
8254         .update_curr            = update_curr_fair,
8255 
8256 #ifdef CONFIG_FAIR_GROUP_SCHED
8257         .task_move_group        = task_move_group_fair,
8258 #endif
8259 };
8260 
8261 #ifdef CONFIG_SCHED_DEBUG
8262 void print_cfs_stats(struct seq_file *m, int cpu)
8263 {
8264         struct cfs_rq *cfs_rq;
8265 
8266         rcu_read_lock();
8267         for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
8268                 print_cfs_rq(m, cpu, cfs_rq);
8269         rcu_read_unlock();
8270 }
8271 #endif
8272 
8273 __init void init_sched_fair_class(void)
8274 {
8275 #ifdef CONFIG_SMP
8276         open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8277 
8278 #ifdef CONFIG_NO_HZ_COMMON
8279         nohz.next_balance = jiffies;
8280         zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8281         cpu_notifier(sched_ilb_notifier, 0);
8282 #endif
8283 #endif /* SMP */
8284 
8285 }
8286 

This page was automatically generated by LXR 0.3.1 (source).  &bu