Version:  2.0.40 2.2.26 2.4.37 3.4 3.5 3.6 3.7 3.8 3.9 3.10 3.11 3.12 3.13 3.14 3.15 3.16 3.17 3.18 3.19 4.0

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

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