Version:  2.0.40 2.2.26 2.4.37 3.3 3.4 3.5 3.6 3.7 3.8 3.9 3.10 3.11 3.12 3.13 3.14 3.15 3.16 3.17 3.18 3.19

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

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