~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/kernel/sched/fair.c

Version: ~ [ linux-5.13-rc5 ] ~ [ linux-5.12.9 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.42 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.124 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.193 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.235 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.271 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.271 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.18.140 ] ~ [ linux-3.16.85 ] ~ [ linux-3.14.79 ] ~ [ linux-3.12.74 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  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 unsigned int get_update_sysctl_factor(void)
145 {
146         unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
147         unsigned int factor;
148 
149         switch (sysctl_sched_tunable_scaling) {
150         case SCHED_TUNABLESCALING_NONE:
151                 factor = 1;
152                 break;
153         case SCHED_TUNABLESCALING_LINEAR:
154                 factor = cpus;
155                 break;
156         case SCHED_TUNABLESCALING_LOG:
157         default:
158                 factor = 1 + ilog2(cpus);
159                 break;
160         }
161 
162         return factor;
163 }
164 
165 static void update_sysctl(void)
166 {
167         unsigned int factor = get_update_sysctl_factor();
168 
169 #define SET_SYSCTL(name) \
170         (sysctl_##name = (factor) * normalized_sysctl_##name)
171         SET_SYSCTL(sched_min_granularity);
172         SET_SYSCTL(sched_latency);
173         SET_SYSCTL(sched_wakeup_granularity);
174 #undef SET_SYSCTL
175 }
176 
177 void sched_init_granularity(void)
178 {
179         update_sysctl();
180 }
181 
182 #define WMULT_CONST     (~0U)
183 #define WMULT_SHIFT     32
184 
185 static void __update_inv_weight(struct load_weight *lw)
186 {
187         unsigned long w;
188 
189         if (likely(lw->inv_weight))
190                 return;
191 
192         w = scale_load_down(lw->weight);
193 
194         if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
195                 lw->inv_weight = 1;
196         else if (unlikely(!w))
197                 lw->inv_weight = WMULT_CONST;
198         else
199                 lw->inv_weight = WMULT_CONST / w;
200 }
201 
202 /*
203  * delta_exec * weight / lw.weight
204  *   OR
205  * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
206  *
207  * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
208  * we're guaranteed shift stays positive because inv_weight is guaranteed to
209  * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
210  *
211  * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
212  * weight/lw.weight <= 1, and therefore our shift will also be positive.
213  */
214 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
215 {
216         u64 fact = scale_load_down(weight);
217         int shift = WMULT_SHIFT;
218 
219         __update_inv_weight(lw);
220 
221         if (unlikely(fact >> 32)) {
222                 while (fact >> 32) {
223                         fact >>= 1;
224                         shift--;
225                 }
226         }
227 
228         /* hint to use a 32x32->64 mul */
229         fact = (u64)(u32)fact * lw->inv_weight;
230 
231         while (fact >> 32) {
232                 fact >>= 1;
233                 shift--;
234         }
235 
236         return mul_u64_u32_shr(delta_exec, fact, shift);
237 }
238 
239 
240 const struct sched_class fair_sched_class;
241 
242 /**************************************************************
243  * CFS operations on generic schedulable entities:
244  */
245 
246 #ifdef CONFIG_FAIR_GROUP_SCHED
247 
248 /* cpu runqueue to which this cfs_rq is attached */
249 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
250 {
251         return cfs_rq->rq;
252 }
253 
254 /* An entity is a task if it doesn't "own" a runqueue */
255 #define entity_is_task(se)      (!se->my_q)
256 
257 static inline struct task_struct *task_of(struct sched_entity *se)
258 {
259 #ifdef CONFIG_SCHED_DEBUG
260         WARN_ON_ONCE(!entity_is_task(se));
261 #endif
262         return container_of(se, struct task_struct, se);
263 }
264 
265 /* Walk up scheduling entities hierarchy */
266 #define for_each_sched_entity(se) \
267                 for (; se; se = se->parent)
268 
269 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
270 {
271         return p->se.cfs_rq;
272 }
273 
274 /* runqueue on which this entity is (to be) queued */
275 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
276 {
277         return se->cfs_rq;
278 }
279 
280 /* runqueue "owned" by this group */
281 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
282 {
283         return grp->my_q;
284 }
285 
286 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
287 {
288         if (!cfs_rq->on_list) {
289                 /*
290                  * Ensure we either appear before our parent (if already
291                  * enqueued) or force our parent to appear after us when it is
292                  * enqueued.  The fact that we always enqueue bottom-up
293                  * reduces this to two cases.
294                  */
295                 if (cfs_rq->tg->parent &&
296                     cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
297                         list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
298                                 &rq_of(cfs_rq)->leaf_cfs_rq_list);
299                 } else {
300                         list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
301                                 &rq_of(cfs_rq)->leaf_cfs_rq_list);
302                 }
303 
304                 cfs_rq->on_list = 1;
305         }
306 }
307 
308 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
309 {
310         if (cfs_rq->on_list) {
311                 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
312                 cfs_rq->on_list = 0;
313         }
314 }
315 
316 /* Iterate thr' all leaf cfs_rq's on a runqueue */
317 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
318         list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
319 
320 /* Do the two (enqueued) entities belong to the same group ? */
321 static inline struct cfs_rq *
322 is_same_group(struct sched_entity *se, struct sched_entity *pse)
323 {
324         if (se->cfs_rq == pse->cfs_rq)
325                 return se->cfs_rq;
326 
327         return NULL;
328 }
329 
330 static inline struct sched_entity *parent_entity(struct sched_entity *se)
331 {
332         return se->parent;
333 }
334 
335 static void
336 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
337 {
338         int se_depth, pse_depth;
339 
340         /*
341          * preemption test can be made between sibling entities who are in the
342          * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
343          * both tasks until we find their ancestors who are siblings of common
344          * parent.
345          */
346 
347         /* First walk up until both entities are at same depth */
348         se_depth = (*se)->depth;
349         pse_depth = (*pse)->depth;
350 
351         while (se_depth > pse_depth) {
352                 se_depth--;
353                 *se = parent_entity(*se);
354         }
355 
356         while (pse_depth > se_depth) {
357                 pse_depth--;
358                 *pse = parent_entity(*pse);
359         }
360 
361         while (!is_same_group(*se, *pse)) {
362                 *se = parent_entity(*se);
363                 *pse = parent_entity(*pse);
364         }
365 }
366 
367 #else   /* !CONFIG_FAIR_GROUP_SCHED */
368 
369 static inline struct task_struct *task_of(struct sched_entity *se)
370 {
371         return container_of(se, struct task_struct, se);
372 }
373 
374 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
375 {
376         return container_of(cfs_rq, struct rq, cfs);
377 }
378 
379 #define entity_is_task(se)      1
380 
381 #define for_each_sched_entity(se) \
382                 for (; se; se = NULL)
383 
384 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
385 {
386         return &task_rq(p)->cfs;
387 }
388 
389 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
390 {
391         struct task_struct *p = task_of(se);
392         struct rq *rq = task_rq(p);
393 
394         return &rq->cfs;
395 }
396 
397 /* runqueue "owned" by this group */
398 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
399 {
400         return NULL;
401 }
402 
403 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
404 {
405 }
406 
407 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
408 {
409 }
410 
411 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
412                 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
413 
414 static inline struct sched_entity *parent_entity(struct sched_entity *se)
415 {
416         return NULL;
417 }
418 
419 static inline void
420 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
421 {
422 }
423 
424 #endif  /* CONFIG_FAIR_GROUP_SCHED */
425 
426 static __always_inline
427 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
428 
429 /**************************************************************
430  * Scheduling class tree data structure manipulation methods:
431  */
432 
433 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
434 {
435         s64 delta = (s64)(vruntime - max_vruntime);
436         if (delta > 0)
437                 max_vruntime = vruntime;
438 
439         return max_vruntime;
440 }
441 
442 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
443 {
444         s64 delta = (s64)(vruntime - min_vruntime);
445         if (delta < 0)
446                 min_vruntime = vruntime;
447 
448         return min_vruntime;
449 }
450 
451 static inline int entity_before(struct sched_entity *a,
452                                 struct sched_entity *b)
453 {
454         return (s64)(a->vruntime - b->vruntime) < 0;
455 }
456 
457 static void update_min_vruntime(struct cfs_rq *cfs_rq)
458 {
459         u64 vruntime = cfs_rq->min_vruntime;
460 
461         if (cfs_rq->curr)
462                 vruntime = cfs_rq->curr->vruntime;
463 
464         if (cfs_rq->rb_leftmost) {
465                 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
466                                                    struct sched_entity,
467                                                    run_node);
468 
469                 if (!cfs_rq->curr)
470                         vruntime = se->vruntime;
471                 else
472                         vruntime = min_vruntime(vruntime, se->vruntime);
473         }
474 
475         /* ensure we never gain time by being placed backwards. */
476         cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
477 #ifndef CONFIG_64BIT
478         smp_wmb();
479         cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
480 #endif
481 }
482 
483 /*
484  * Enqueue an entity into the rb-tree:
485  */
486 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
487 {
488         struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
489         struct rb_node *parent = NULL;
490         struct sched_entity *entry;
491         int leftmost = 1;
492 
493         /*
494          * Find the right place in the rbtree:
495          */
496         while (*link) {
497                 parent = *link;
498                 entry = rb_entry(parent, struct sched_entity, run_node);
499                 /*
500                  * We dont care about collisions. Nodes with
501                  * the same key stay together.
502                  */
503                 if (entity_before(se, entry)) {
504                         link = &parent->rb_left;
505                 } else {
506                         link = &parent->rb_right;
507                         leftmost = 0;
508                 }
509         }
510 
511         /*
512          * Maintain a cache of leftmost tree entries (it is frequently
513          * used):
514          */
515         if (leftmost)
516                 cfs_rq->rb_leftmost = &se->run_node;
517 
518         rb_link_node(&se->run_node, parent, link);
519         rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
520 }
521 
522 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
523 {
524         if (cfs_rq->rb_leftmost == &se->run_node) {
525                 struct rb_node *next_node;
526 
527                 next_node = rb_next(&se->run_node);
528                 cfs_rq->rb_leftmost = next_node;
529         }
530 
531         rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
532 }
533 
534 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
535 {
536         struct rb_node *left = cfs_rq->rb_leftmost;
537 
538         if (!left)
539                 return NULL;
540 
541         return rb_entry(left, struct sched_entity, run_node);
542 }
543 
544 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
545 {
546         struct rb_node *next = rb_next(&se->run_node);
547 
548         if (!next)
549                 return NULL;
550 
551         return rb_entry(next, struct sched_entity, run_node);
552 }
553 
554 #ifdef CONFIG_SCHED_DEBUG
555 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
556 {
557         struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
558 
559         if (!last)
560                 return NULL;
561 
562         return rb_entry(last, struct sched_entity, run_node);
563 }
564 
565 /**************************************************************
566  * Scheduling class statistics methods:
567  */
568 
569 int sched_proc_update_handler(struct ctl_table *table, int write,
570                 void __user *buffer, size_t *lenp,
571                 loff_t *ppos)
572 {
573         int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
574         unsigned int factor = get_update_sysctl_factor();
575 
576         if (ret || !write)
577                 return ret;
578 
579         sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
580                                         sysctl_sched_min_granularity);
581 
582 #define WRT_SYSCTL(name) \
583         (normalized_sysctl_##name = sysctl_##name / (factor))
584         WRT_SYSCTL(sched_min_granularity);
585         WRT_SYSCTL(sched_latency);
586         WRT_SYSCTL(sched_wakeup_granularity);
587 #undef WRT_SYSCTL
588 
589         return 0;
590 }
591 #endif
592 
593 /*
594  * delta /= w
595  */
596 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
597 {
598         if (unlikely(se->load.weight != NICE_0_LOAD))
599                 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
600 
601         return delta;
602 }
603 
604 /*
605  * The idea is to set a period in which each task runs once.
606  *
607  * When there are too many tasks (sched_nr_latency) we have to stretch
608  * this period because otherwise the slices get too small.
609  *
610  * p = (nr <= nl) ? l : l*nr/nl
611  */
612 static u64 __sched_period(unsigned long nr_running)
613 {
614         if (unlikely(nr_running > sched_nr_latency))
615                 return nr_running * sysctl_sched_min_granularity;
616         else
617                 return sysctl_sched_latency;
618 }
619 
620 /*
621  * We calculate the wall-time slice from the period by taking a part
622  * proportional to the weight.
623  *
624  * s = p*P[w/rw]
625  */
626 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
627 {
628         u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
629 
630         for_each_sched_entity(se) {
631                 struct load_weight *load;
632                 struct load_weight lw;
633 
634                 cfs_rq = cfs_rq_of(se);
635                 load = &cfs_rq->load;
636 
637                 if (unlikely(!se->on_rq)) {
638                         lw = cfs_rq->load;
639 
640                         update_load_add(&lw, se->load.weight);
641                         load = &lw;
642                 }
643                 slice = __calc_delta(slice, se->load.weight, load);
644         }
645         return slice;
646 }
647 
648 /*
649  * We calculate the vruntime slice of a to-be-inserted task.
650  *
651  * vs = s/w
652  */
653 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
654 {
655         return calc_delta_fair(sched_slice(cfs_rq, se), se);
656 }
657 
658 #ifdef CONFIG_SMP
659 static int select_idle_sibling(struct task_struct *p, int cpu);
660 static unsigned long task_h_load(struct task_struct *p);
661 
662 /*
663  * We choose a half-life close to 1 scheduling period.
664  * Note: The tables below are dependent on this value.
665  */
666 #define LOAD_AVG_PERIOD 32
667 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
668 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
669 
670 /* Give new sched_entity start runnable values to heavy its load in infant time */
671 void init_entity_runnable_average(struct sched_entity *se)
672 {
673         struct sched_avg *sa = &se->avg;
674 
675         sa->last_update_time = 0;
676         /*
677          * sched_avg's period_contrib should be strictly less then 1024, so
678          * we give it 1023 to make sure it is almost a period (1024us), and
679          * will definitely be update (after enqueue).
680          */
681         sa->period_contrib = 1023;
682         sa->load_avg = scale_load_down(se->load.weight);
683         sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
684         sa->util_avg = scale_load_down(SCHED_LOAD_SCALE);
685         sa->util_sum = LOAD_AVG_MAX;
686         /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
687 }
688 
689 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq);
690 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq);
691 #else
692 void init_entity_runnable_average(struct sched_entity *se)
693 {
694 }
695 #endif
696 
697 /*
698  * Update the current task's runtime statistics.
699  */
700 static void update_curr(struct cfs_rq *cfs_rq)
701 {
702         struct sched_entity *curr = cfs_rq->curr;
703         u64 now = rq_clock_task(rq_of(cfs_rq));
704         u64 delta_exec;
705 
706         if (unlikely(!curr))
707                 return;
708 
709         delta_exec = now - curr->exec_start;
710         if (unlikely((s64)delta_exec <= 0))
711                 return;
712 
713         curr->exec_start = now;
714 
715         schedstat_set(curr->statistics.exec_max,
716                       max(delta_exec, curr->statistics.exec_max));
717 
718         curr->sum_exec_runtime += delta_exec;
719         schedstat_add(cfs_rq, exec_clock, delta_exec);
720 
721         curr->vruntime += calc_delta_fair(delta_exec, curr);
722         update_min_vruntime(cfs_rq);
723 
724         if (entity_is_task(curr)) {
725                 struct task_struct *curtask = task_of(curr);
726 
727                 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
728                 cpuacct_charge(curtask, delta_exec);
729                 account_group_exec_runtime(curtask, delta_exec);
730         }
731 
732         account_cfs_rq_runtime(cfs_rq, delta_exec);
733 }
734 
735 static void update_curr_fair(struct rq *rq)
736 {
737         update_curr(cfs_rq_of(&rq->curr->se));
738 }
739 
740 static inline void
741 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
742 {
743         schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
744 }
745 
746 /*
747  * Task is being enqueued - update stats:
748  */
749 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
750 {
751         /*
752          * Are we enqueueing a waiting task? (for current tasks
753          * a dequeue/enqueue event is a NOP)
754          */
755         if (se != cfs_rq->curr)
756                 update_stats_wait_start(cfs_rq, se);
757 }
758 
759 static void
760 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
761 {
762         schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
763                         rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
764         schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
765         schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
766                         rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
767 #ifdef CONFIG_SCHEDSTATS
768         if (entity_is_task(se)) {
769                 trace_sched_stat_wait(task_of(se),
770                         rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
771         }
772 #endif
773         schedstat_set(se->statistics.wait_start, 0);
774 }
775 
776 static inline void
777 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
778 {
779         /*
780          * Mark the end of the wait period if dequeueing a
781          * waiting task:
782          */
783         if (se != cfs_rq->curr)
784                 update_stats_wait_end(cfs_rq, se);
785 }
786 
787 /*
788  * We are picking a new current task - update its stats:
789  */
790 static inline void
791 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
792 {
793         /*
794          * We are starting a new run period:
795          */
796         se->exec_start = rq_clock_task(rq_of(cfs_rq));
797 }
798 
799 /**************************************************
800  * Scheduling class queueing methods:
801  */
802 
803 #ifdef CONFIG_NUMA_BALANCING
804 /*
805  * Approximate time to scan a full NUMA task in ms. The task scan period is
806  * calculated based on the tasks virtual memory size and
807  * numa_balancing_scan_size.
808  */
809 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
810 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
811 
812 /* Portion of address space to scan in MB */
813 unsigned int sysctl_numa_balancing_scan_size = 256;
814 
815 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
816 unsigned int sysctl_numa_balancing_scan_delay = 1000;
817 
818 static unsigned int task_nr_scan_windows(struct task_struct *p)
819 {
820         unsigned long rss = 0;
821         unsigned long nr_scan_pages;
822 
823         /*
824          * Calculations based on RSS as non-present and empty pages are skipped
825          * by the PTE scanner and NUMA hinting faults should be trapped based
826          * on resident pages
827          */
828         nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
829         rss = get_mm_rss(p->mm);
830         if (!rss)
831                 rss = nr_scan_pages;
832 
833         rss = round_up(rss, nr_scan_pages);
834         return rss / nr_scan_pages;
835 }
836 
837 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
838 #define MAX_SCAN_WINDOW 2560
839 
840 static unsigned int task_scan_min(struct task_struct *p)
841 {
842         unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
843         unsigned int scan, floor;
844         unsigned int windows = 1;
845 
846         if (scan_size < MAX_SCAN_WINDOW)
847                 windows = MAX_SCAN_WINDOW / scan_size;
848         floor = 1000 / windows;
849 
850         scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
851         return max_t(unsigned int, floor, scan);
852 }
853 
854 static unsigned int task_scan_max(struct task_struct *p)
855 {
856         unsigned int smin = task_scan_min(p);
857         unsigned int smax;
858 
859         /* Watch for min being lower than max due to floor calculations */
860         smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
861         return max(smin, smax);
862 }
863 
864 static void account_numa_enqueue(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 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
871 {
872         rq->nr_numa_running -= (p->numa_preferred_nid != -1);
873         rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
874 }
875 
876 struct numa_group {
877         atomic_t refcount;
878 
879         spinlock_t lock; /* nr_tasks, tasks */
880         int nr_tasks;
881         pid_t gid;
882 
883         struct rcu_head rcu;
884         nodemask_t active_nodes;
885         unsigned long total_faults;
886         /*
887          * Faults_cpu is used to decide whether memory should move
888          * towards the CPU. As a consequence, these stats are weighted
889          * more by CPU use than by memory faults.
890          */
891         unsigned long *faults_cpu;
892         unsigned long faults[0];
893 };
894 
895 /* Shared or private faults. */
896 #define NR_NUMA_HINT_FAULT_TYPES 2
897 
898 /* Memory and CPU locality */
899 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
900 
901 /* Averaged statistics, and temporary buffers. */
902 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
903 
904 pid_t task_numa_group_id(struct task_struct *p)
905 {
906         return p->numa_group ? p->numa_group->gid : 0;
907 }
908 
909 /*
910  * The averaged statistics, shared & private, memory & cpu,
911  * occupy the first half of the array. The second half of the
912  * array is for current counters, which are averaged into the
913  * first set by task_numa_placement.
914  */
915 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
916 {
917         return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
918 }
919 
920 static inline unsigned long task_faults(struct task_struct *p, int nid)
921 {
922         if (!p->numa_faults)
923                 return 0;
924 
925         return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
926                 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
927 }
928 
929 static inline unsigned long group_faults(struct task_struct *p, int nid)
930 {
931         if (!p->numa_group)
932                 return 0;
933 
934         return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
935                 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
936 }
937 
938 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
939 {
940         return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
941                 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
942 }
943 
944 /* Handle placement on systems where not all nodes are directly connected. */
945 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
946                                         int maxdist, bool task)
947 {
948         unsigned long score = 0;
949         int node;
950 
951         /*
952          * All nodes are directly connected, and the same distance
953          * from each other. No need for fancy placement algorithms.
954          */
955         if (sched_numa_topology_type == NUMA_DIRECT)
956                 return 0;
957 
958         /*
959          * This code is called for each node, introducing N^2 complexity,
960          * which should be ok given the number of nodes rarely exceeds 8.
961          */
962         for_each_online_node(node) {
963                 unsigned long faults;
964                 int dist = node_distance(nid, node);
965 
966                 /*
967                  * The furthest away nodes in the system are not interesting
968                  * for placement; nid was already counted.
969                  */
970                 if (dist == sched_max_numa_distance || node == nid)
971                         continue;
972 
973                 /*
974                  * On systems with a backplane NUMA topology, compare groups
975                  * of nodes, and move tasks towards the group with the most
976                  * memory accesses. When comparing two nodes at distance
977                  * "hoplimit", only nodes closer by than "hoplimit" are part
978                  * of each group. Skip other nodes.
979                  */
980                 if (sched_numa_topology_type == NUMA_BACKPLANE &&
981                                         dist > maxdist)
982                         continue;
983 
984                 /* Add up the faults from nearby nodes. */
985                 if (task)
986                         faults = task_faults(p, node);
987                 else
988                         faults = group_faults(p, node);
989 
990                 /*
991                  * On systems with a glueless mesh NUMA topology, there are
992                  * no fixed "groups of nodes". Instead, nodes that are not
993                  * directly connected bounce traffic through intermediate
994                  * nodes; a numa_group can occupy any set of nodes.
995                  * The further away a node is, the less the faults count.
996                  * This seems to result in good task placement.
997                  */
998                 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
999                         faults *= (sched_max_numa_distance - dist);
1000                         faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1001                 }
1002 
1003                 score += faults;
1004         }
1005 
1006         return score;
1007 }
1008 
1009 /*
1010  * These return the fraction of accesses done by a particular task, or
1011  * task group, on a particular numa node.  The group weight is given a
1012  * larger multiplier, in order to group tasks together that are almost
1013  * evenly spread out between numa nodes.
1014  */
1015 static inline unsigned long task_weight(struct task_struct *p, int nid,
1016                                         int dist)
1017 {
1018         unsigned long faults, total_faults;
1019 
1020         if (!p->numa_faults)
1021                 return 0;
1022 
1023         total_faults = p->total_numa_faults;
1024 
1025         if (!total_faults)
1026                 return 0;
1027 
1028         faults = task_faults(p, nid);
1029         faults += score_nearby_nodes(p, nid, dist, true);
1030 
1031         return 1000 * faults / total_faults;
1032 }
1033 
1034 static inline unsigned long group_weight(struct task_struct *p, int nid,
1035                                          int dist)
1036 {
1037         unsigned long faults, total_faults;
1038 
1039         if (!p->numa_group)
1040                 return 0;
1041 
1042         total_faults = p->numa_group->total_faults;
1043 
1044         if (!total_faults)
1045                 return 0;
1046 
1047         faults = group_faults(p, nid);
1048         faults += score_nearby_nodes(p, nid, dist, false);
1049 
1050         return 1000 * faults / total_faults;
1051 }
1052 
1053 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1054                                 int src_nid, int dst_cpu)
1055 {
1056         struct numa_group *ng = p->numa_group;
1057         int dst_nid = cpu_to_node(dst_cpu);
1058         int last_cpupid, this_cpupid;
1059 
1060         this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1061 
1062         /*
1063          * Multi-stage node selection is used in conjunction with a periodic
1064          * migration fault to build a temporal task<->page relation. By using
1065          * a two-stage filter we remove short/unlikely relations.
1066          *
1067          * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1068          * a task's usage of a particular page (n_p) per total usage of this
1069          * page (n_t) (in a given time-span) to a probability.
1070          *
1071          * Our periodic faults will sample this probability and getting the
1072          * same result twice in a row, given these samples are fully
1073          * independent, is then given by P(n)^2, provided our sample period
1074          * is sufficiently short compared to the usage pattern.
1075          *
1076          * This quadric squishes small probabilities, making it less likely we
1077          * act on an unlikely task<->page relation.
1078          */
1079         last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1080         if (!cpupid_pid_unset(last_cpupid) &&
1081                                 cpupid_to_nid(last_cpupid) != dst_nid)
1082                 return false;
1083 
1084         /* Always allow migrate on private faults */
1085         if (cpupid_match_pid(p, last_cpupid))
1086                 return true;
1087 
1088         /* A shared fault, but p->numa_group has not been set up yet. */
1089         if (!ng)
1090                 return true;
1091 
1092         /*
1093          * Do not migrate if the destination is not a node that
1094          * is actively used by this numa group.
1095          */
1096         if (!node_isset(dst_nid, ng->active_nodes))
1097                 return false;
1098 
1099         /*
1100          * Source is a node that is not actively used by this
1101          * numa group, while the destination is. Migrate.
1102          */
1103         if (!node_isset(src_nid, ng->active_nodes))
1104                 return true;
1105 
1106         /*
1107          * Both source and destination are nodes in active
1108          * use by this numa group. Maximize memory bandwidth
1109          * by migrating from more heavily used groups, to less
1110          * heavily used ones, spreading the load around.
1111          * Use a 1/4 hysteresis to avoid spurious page movement.
1112          */
1113         return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1114 }
1115 
1116 static unsigned long weighted_cpuload(const int cpu);
1117 static unsigned long source_load(int cpu, int type);
1118 static unsigned long target_load(int cpu, int type);
1119 static unsigned long capacity_of(int cpu);
1120 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1121 
1122 /* Cached statistics for all CPUs within a node */
1123 struct numa_stats {
1124         unsigned long nr_running;
1125         unsigned long load;
1126 
1127         /* Total compute capacity of CPUs on a node */
1128         unsigned long compute_capacity;
1129 
1130         /* Approximate capacity in terms of runnable tasks on a node */
1131         unsigned long task_capacity;
1132         int has_free_capacity;
1133 };
1134 
1135 /*
1136  * XXX borrowed from update_sg_lb_stats
1137  */
1138 static void update_numa_stats(struct numa_stats *ns, int nid)
1139 {
1140         int smt, cpu, cpus = 0;
1141         unsigned long capacity;
1142 
1143         memset(ns, 0, sizeof(*ns));
1144         for_each_cpu(cpu, cpumask_of_node(nid)) {
1145                 struct rq *rq = cpu_rq(cpu);
1146 
1147                 ns->nr_running += rq->nr_running;
1148                 ns->load += weighted_cpuload(cpu);
1149                 ns->compute_capacity += capacity_of(cpu);
1150 
1151                 cpus++;
1152         }
1153 
1154         /*
1155          * If we raced with hotplug and there are no CPUs left in our mask
1156          * the @ns structure is NULL'ed and task_numa_compare() will
1157          * not find this node attractive.
1158          *
1159          * We'll either bail at !has_free_capacity, or we'll detect a huge
1160          * imbalance and bail there.
1161          */
1162         if (!cpus)
1163                 return;
1164 
1165         /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1166         smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1167         capacity = cpus / smt; /* cores */
1168 
1169         ns->task_capacity = min_t(unsigned, capacity,
1170                 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1171         ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1172 }
1173 
1174 struct task_numa_env {
1175         struct task_struct *p;
1176 
1177         int src_cpu, src_nid;
1178         int dst_cpu, dst_nid;
1179 
1180         struct numa_stats src_stats, dst_stats;
1181 
1182         int imbalance_pct;
1183         int dist;
1184 
1185         struct task_struct *best_task;
1186         long best_imp;
1187         int best_cpu;
1188 };
1189 
1190 static void task_numa_assign(struct task_numa_env *env,
1191                              struct task_struct *p, long imp)
1192 {
1193         if (env->best_task)
1194                 put_task_struct(env->best_task);
1195         if (p)
1196                 get_task_struct(p);
1197 
1198         env->best_task = p;
1199         env->best_imp = imp;
1200         env->best_cpu = env->dst_cpu;
1201 }
1202 
1203 static bool load_too_imbalanced(long src_load, long dst_load,
1204                                 struct task_numa_env *env)
1205 {
1206         long imb, old_imb;
1207         long orig_src_load, orig_dst_load;
1208         long src_capacity, dst_capacity;
1209 
1210         /*
1211          * The load is corrected for the CPU capacity available on each node.
1212          *
1213          * src_load        dst_load
1214          * ------------ vs ---------
1215          * src_capacity    dst_capacity
1216          */
1217         src_capacity = env->src_stats.compute_capacity;
1218         dst_capacity = env->dst_stats.compute_capacity;
1219 
1220         /* We care about the slope of the imbalance, not the direction. */
1221         if (dst_load < src_load)
1222                 swap(dst_load, src_load);
1223 
1224         /* Is the difference below the threshold? */
1225         imb = dst_load * src_capacity * 100 -
1226               src_load * dst_capacity * env->imbalance_pct;
1227         if (imb <= 0)
1228                 return false;
1229 
1230         /*
1231          * The imbalance is above the allowed threshold.
1232          * Compare it with the old imbalance.
1233          */
1234         orig_src_load = env->src_stats.load;
1235         orig_dst_load = env->dst_stats.load;
1236 
1237         if (orig_dst_load < orig_src_load)
1238                 swap(orig_dst_load, orig_src_load);
1239 
1240         old_imb = orig_dst_load * src_capacity * 100 -
1241                   orig_src_load * dst_capacity * env->imbalance_pct;
1242 
1243         /* Would this change make things worse? */
1244         return (imb > old_imb);
1245 }
1246 
1247 /*
1248  * This checks if the overall compute and NUMA accesses of the system would
1249  * be improved if the source tasks was migrated to the target dst_cpu taking
1250  * into account that it might be best if task running on the dst_cpu should
1251  * be exchanged with the source task
1252  */
1253 static void task_numa_compare(struct task_numa_env *env,
1254                               long taskimp, long groupimp)
1255 {
1256         struct rq *src_rq = cpu_rq(env->src_cpu);
1257         struct rq *dst_rq = cpu_rq(env->dst_cpu);
1258         struct task_struct *cur;
1259         long src_load, dst_load;
1260         long load;
1261         long imp = env->p->numa_group ? groupimp : taskimp;
1262         long moveimp = imp;
1263         int dist = env->dist;
1264 
1265         rcu_read_lock();
1266 
1267         raw_spin_lock_irq(&dst_rq->lock);
1268         cur = dst_rq->curr;
1269         /*
1270          * No need to move the exiting task, and this ensures that ->curr
1271          * wasn't reaped and thus get_task_struct() in task_numa_assign()
1272          * is safe under RCU read lock.
1273          * Note that rcu_read_lock() itself can't protect from the final
1274          * put_task_struct() after the last schedule().
1275          */
1276         if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1277                 cur = NULL;
1278         raw_spin_unlock_irq(&dst_rq->lock);
1279 
1280         /*
1281          * Because we have preemption enabled we can get migrated around and
1282          * end try selecting ourselves (current == env->p) as a swap candidate.
1283          */
1284         if (cur == env->p)
1285                 goto unlock;
1286 
1287         /*
1288          * "imp" is the fault differential for the source task between the
1289          * source and destination node. Calculate the total differential for
1290          * the source task and potential destination task. The more negative
1291          * the value is, the more rmeote accesses that would be expected to
1292          * be incurred if the tasks were swapped.
1293          */
1294         if (cur) {
1295                 /* Skip this swap candidate if cannot move to the source cpu */
1296                 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1297                         goto unlock;
1298 
1299                 /*
1300                  * If dst and source tasks are in the same NUMA group, or not
1301                  * in any group then look only at task weights.
1302                  */
1303                 if (cur->numa_group == env->p->numa_group) {
1304                         imp = taskimp + task_weight(cur, env->src_nid, dist) -
1305                               task_weight(cur, env->dst_nid, dist);
1306                         /*
1307                          * Add some hysteresis to prevent swapping the
1308                          * tasks within a group over tiny differences.
1309                          */
1310                         if (cur->numa_group)
1311                                 imp -= imp/16;
1312                 } else {
1313                         /*
1314                          * Compare the group weights. If a task is all by
1315                          * itself (not part of a group), use the task weight
1316                          * instead.
1317                          */
1318                         if (cur->numa_group)
1319                                 imp += group_weight(cur, env->src_nid, dist) -
1320                                        group_weight(cur, env->dst_nid, dist);
1321                         else
1322                                 imp += task_weight(cur, env->src_nid, dist) -
1323                                        task_weight(cur, env->dst_nid, dist);
1324                 }
1325         }
1326 
1327         if (imp <= env->best_imp && moveimp <= env->best_imp)
1328                 goto unlock;
1329 
1330         if (!cur) {
1331                 /* Is there capacity at our destination? */
1332                 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1333                     !env->dst_stats.has_free_capacity)
1334                         goto unlock;
1335 
1336                 goto balance;
1337         }
1338 
1339         /* Balance doesn't matter much if we're running a task per cpu */
1340         if (imp > env->best_imp && src_rq->nr_running == 1 &&
1341                         dst_rq->nr_running == 1)
1342                 goto assign;
1343 
1344         /*
1345          * In the overloaded case, try and keep the load balanced.
1346          */
1347 balance:
1348         load = task_h_load(env->p);
1349         dst_load = env->dst_stats.load + load;
1350         src_load = env->src_stats.load - load;
1351 
1352         if (moveimp > imp && moveimp > env->best_imp) {
1353                 /*
1354                  * If the improvement from just moving env->p direction is
1355                  * better than swapping tasks around, check if a move is
1356                  * possible. Store a slightly smaller score than moveimp,
1357                  * so an actually idle CPU will win.
1358                  */
1359                 if (!load_too_imbalanced(src_load, dst_load, env)) {
1360                         imp = moveimp - 1;
1361                         cur = NULL;
1362                         goto assign;
1363                 }
1364         }
1365 
1366         if (imp <= env->best_imp)
1367                 goto unlock;
1368 
1369         if (cur) {
1370                 load = task_h_load(cur);
1371                 dst_load -= load;
1372                 src_load += load;
1373         }
1374 
1375         if (load_too_imbalanced(src_load, dst_load, env))
1376                 goto unlock;
1377 
1378         /*
1379          * One idle CPU per node is evaluated for a task numa move.
1380          * Call select_idle_sibling to maybe find a better one.
1381          */
1382         if (!cur)
1383                 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1384 
1385 assign:
1386         task_numa_assign(env, cur, imp);
1387 unlock:
1388         rcu_read_unlock();
1389 }
1390 
1391 static void task_numa_find_cpu(struct task_numa_env *env,
1392                                 long taskimp, long groupimp)
1393 {
1394         int cpu;
1395 
1396         for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1397                 /* Skip this CPU if the source task cannot migrate */
1398                 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1399                         continue;
1400 
1401                 env->dst_cpu = cpu;
1402                 task_numa_compare(env, taskimp, groupimp);
1403         }
1404 }
1405 
1406 /* Only move tasks to a NUMA node less busy than the current node. */
1407 static bool numa_has_capacity(struct task_numa_env *env)
1408 {
1409         struct numa_stats *src = &env->src_stats;
1410         struct numa_stats *dst = &env->dst_stats;
1411 
1412         if (src->has_free_capacity && !dst->has_free_capacity)
1413                 return false;
1414 
1415         /*
1416          * Only consider a task move if the source has a higher load
1417          * than the destination, corrected for CPU capacity on each node.
1418          *
1419          *      src->load                dst->load
1420          * --------------------- vs ---------------------
1421          * src->compute_capacity    dst->compute_capacity
1422          */
1423         if (src->load * dst->compute_capacity * env->imbalance_pct >
1424 
1425             dst->load * src->compute_capacity * 100)
1426                 return true;
1427 
1428         return false;
1429 }
1430 
1431 static int task_numa_migrate(struct task_struct *p)
1432 {
1433         struct task_numa_env env = {
1434                 .p = p,
1435 
1436                 .src_cpu = task_cpu(p),
1437                 .src_nid = task_node(p),
1438 
1439                 .imbalance_pct = 112,
1440 
1441                 .best_task = NULL,
1442                 .best_imp = 0,
1443                 .best_cpu = -1
1444         };
1445         struct sched_domain *sd;
1446         unsigned long taskweight, groupweight;
1447         int nid, ret, dist;
1448         long taskimp, groupimp;
1449 
1450         /*
1451          * Pick the lowest SD_NUMA domain, as that would have the smallest
1452          * imbalance and would be the first to start moving tasks about.
1453          *
1454          * And we want to avoid any moving of tasks about, as that would create
1455          * random movement of tasks -- counter the numa conditions we're trying
1456          * to satisfy here.
1457          */
1458         rcu_read_lock();
1459         sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1460         if (sd)
1461                 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1462         rcu_read_unlock();
1463 
1464         /*
1465          * Cpusets can break the scheduler domain tree into smaller
1466          * balance domains, some of which do not cross NUMA boundaries.
1467          * Tasks that are "trapped" in such domains cannot be migrated
1468          * elsewhere, so there is no point in (re)trying.
1469          */
1470         if (unlikely(!sd)) {
1471                 p->numa_preferred_nid = task_node(p);
1472                 return -EINVAL;
1473         }
1474 
1475         env.dst_nid = p->numa_preferred_nid;
1476         dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1477         taskweight = task_weight(p, env.src_nid, dist);
1478         groupweight = group_weight(p, env.src_nid, dist);
1479         update_numa_stats(&env.src_stats, env.src_nid);
1480         taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1481         groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1482         update_numa_stats(&env.dst_stats, env.dst_nid);
1483 
1484         /* Try to find a spot on the preferred nid. */
1485         if (numa_has_capacity(&env))
1486                 task_numa_find_cpu(&env, taskimp, groupimp);
1487 
1488         /*
1489          * Look at other nodes in these cases:
1490          * - there is no space available on the preferred_nid
1491          * - the task is part of a numa_group that is interleaved across
1492          *   multiple NUMA nodes; in order to better consolidate the group,
1493          *   we need to check other locations.
1494          */
1495         if (env.best_cpu == -1 || (p->numa_group &&
1496                         nodes_weight(p->numa_group->active_nodes) > 1)) {
1497                 for_each_online_node(nid) {
1498                         if (nid == env.src_nid || nid == p->numa_preferred_nid)
1499                                 continue;
1500 
1501                         dist = node_distance(env.src_nid, env.dst_nid);
1502                         if (sched_numa_topology_type == NUMA_BACKPLANE &&
1503                                                 dist != env.dist) {
1504                                 taskweight = task_weight(p, env.src_nid, dist);
1505                                 groupweight = group_weight(p, env.src_nid, dist);
1506                         }
1507 
1508                         /* Only consider nodes where both task and groups benefit */
1509                         taskimp = task_weight(p, nid, dist) - taskweight;
1510                         groupimp = group_weight(p, nid, dist) - groupweight;
1511                         if (taskimp < 0 && groupimp < 0)
1512                                 continue;
1513 
1514                         env.dist = dist;
1515                         env.dst_nid = nid;
1516                         update_numa_stats(&env.dst_stats, env.dst_nid);
1517                         if (numa_has_capacity(&env))
1518                                 task_numa_find_cpu(&env, taskimp, groupimp);
1519                 }
1520         }
1521 
1522         /*
1523          * If the task is part of a workload that spans multiple NUMA nodes,
1524          * and is migrating into one of the workload's active nodes, remember
1525          * this node as the task's preferred numa node, so the workload can
1526          * settle down.
1527          * A task that migrated to a second choice node will be better off
1528          * trying for a better one later. Do not set the preferred node here.
1529          */
1530         if (p->numa_group) {
1531                 if (env.best_cpu == -1)
1532                         nid = env.src_nid;
1533                 else
1534                         nid = env.dst_nid;
1535 
1536                 if (node_isset(nid, p->numa_group->active_nodes))
1537                         sched_setnuma(p, env.dst_nid);
1538         }
1539 
1540         /* No better CPU than the current one was found. */
1541         if (env.best_cpu == -1)
1542                 return -EAGAIN;
1543 
1544         /*
1545          * Reset the scan period if the task is being rescheduled on an
1546          * alternative node to recheck if the tasks is now properly placed.
1547          */
1548         p->numa_scan_period = task_scan_min(p);
1549 
1550         if (env.best_task == NULL) {
1551                 ret = migrate_task_to(p, env.best_cpu);
1552                 if (ret != 0)
1553                         trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1554                 return ret;
1555         }
1556 
1557         ret = migrate_swap(p, env.best_task);
1558         if (ret != 0)
1559                 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1560         put_task_struct(env.best_task);
1561         return ret;
1562 }
1563 
1564 /* Attempt to migrate a task to a CPU on the preferred node. */
1565 static void numa_migrate_preferred(struct task_struct *p)
1566 {
1567         unsigned long interval = HZ;
1568 
1569         /* This task has no NUMA fault statistics yet */
1570         if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1571                 return;
1572 
1573         /* Periodically retry migrating the task to the preferred node */
1574         interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1575         p->numa_migrate_retry = jiffies + interval;
1576 
1577         /* Success if task is already running on preferred CPU */
1578         if (task_node(p) == p->numa_preferred_nid)
1579                 return;
1580 
1581         /* Otherwise, try migrate to a CPU on the preferred node */
1582         task_numa_migrate(p);
1583 }
1584 
1585 /*
1586  * Find the nodes on which the workload is actively running. We do this by
1587  * tracking the nodes from which NUMA hinting faults are triggered. This can
1588  * be different from the set of nodes where the workload's memory is currently
1589  * located.
1590  *
1591  * The bitmask is used to make smarter decisions on when to do NUMA page
1592  * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1593  * are added when they cause over 6/16 of the maximum number of faults, but
1594  * only removed when they drop below 3/16.
1595  */
1596 static void update_numa_active_node_mask(struct numa_group *numa_group)
1597 {
1598         unsigned long faults, max_faults = 0;
1599         int nid;
1600 
1601         for_each_online_node(nid) {
1602                 faults = group_faults_cpu(numa_group, nid);
1603                 if (faults > max_faults)
1604                         max_faults = faults;
1605         }
1606 
1607         for_each_online_node(nid) {
1608                 faults = group_faults_cpu(numa_group, nid);
1609                 if (!node_isset(nid, numa_group->active_nodes)) {
1610                         if (faults > max_faults * 6 / 16)
1611                                 node_set(nid, numa_group->active_nodes);
1612                 } else if (faults < max_faults * 3 / 16)
1613                         node_clear(nid, numa_group->active_nodes);
1614         }
1615 }
1616 
1617 /*
1618  * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1619  * increments. The more local the fault statistics are, the higher the scan
1620  * period will be for the next scan window. If local/(local+remote) ratio is
1621  * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1622  * the scan period will decrease. Aim for 70% local accesses.
1623  */
1624 #define NUMA_PERIOD_SLOTS 10
1625 #define NUMA_PERIOD_THRESHOLD 7
1626 
1627 /*
1628  * Increase the scan period (slow down scanning) if the majority of
1629  * our memory is already on our local node, or if the majority of
1630  * the page accesses are shared with other processes.
1631  * Otherwise, decrease the scan period.
1632  */
1633 static void update_task_scan_period(struct task_struct *p,
1634                         unsigned long shared, unsigned long private)
1635 {
1636         unsigned int period_slot;
1637         int ratio;
1638         int diff;
1639 
1640         unsigned long remote = p->numa_faults_locality[0];
1641         unsigned long local = p->numa_faults_locality[1];
1642 
1643         /*
1644          * If there were no record hinting faults then either the task is
1645          * completely idle or all activity is areas that are not of interest
1646          * to automatic numa balancing. Related to that, if there were failed
1647          * migration then it implies we are migrating too quickly or the local
1648          * node is overloaded. In either case, scan slower
1649          */
1650         if (local + shared == 0 || p->numa_faults_locality[2]) {
1651                 p->numa_scan_period = min(p->numa_scan_period_max,
1652                         p->numa_scan_period << 1);
1653 
1654                 p->mm->numa_next_scan = jiffies +
1655                         msecs_to_jiffies(p->numa_scan_period);
1656 
1657                 return;
1658         }
1659 
1660         /*
1661          * Prepare to scale scan period relative to the current period.
1662          *       == NUMA_PERIOD_THRESHOLD scan period stays the same
1663          *       <  NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1664          *       >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1665          */
1666         period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1667         ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1668         if (ratio >= NUMA_PERIOD_THRESHOLD) {
1669                 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1670                 if (!slot)
1671                         slot = 1;
1672                 diff = slot * period_slot;
1673         } else {
1674                 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1675 
1676                 /*
1677                  * Scale scan rate increases based on sharing. There is an
1678                  * inverse relationship between the degree of sharing and
1679                  * the adjustment made to the scanning period. Broadly
1680                  * speaking the intent is that there is little point
1681                  * scanning faster if shared accesses dominate as it may
1682                  * simply bounce migrations uselessly
1683                  */
1684                 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1685                 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1686         }
1687 
1688         p->numa_scan_period = clamp(p->numa_scan_period + diff,
1689                         task_scan_min(p), task_scan_max(p));
1690         memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1691 }
1692 
1693 /*
1694  * Get the fraction of time the task has been running since the last
1695  * NUMA placement cycle. The scheduler keeps similar statistics, but
1696  * decays those on a 32ms period, which is orders of magnitude off
1697  * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1698  * stats only if the task is so new there are no NUMA statistics yet.
1699  */
1700 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1701 {
1702         u64 runtime, delta, now;
1703         /* Use the start of this time slice to avoid calculations. */
1704         now = p->se.exec_start;
1705         runtime = p->se.sum_exec_runtime;
1706 
1707         if (p->last_task_numa_placement) {
1708                 delta = runtime - p->last_sum_exec_runtime;
1709                 *period = now - p->last_task_numa_placement;
1710         } else {
1711                 delta = p->se.avg.load_sum / p->se.load.weight;
1712                 *period = LOAD_AVG_MAX;
1713         }
1714 
1715         p->last_sum_exec_runtime = runtime;
1716         p->last_task_numa_placement = now;
1717 
1718         return delta;
1719 }
1720 
1721 /*
1722  * Determine the preferred nid for a task in a numa_group. This needs to
1723  * be done in a way that produces consistent results with group_weight,
1724  * otherwise workloads might not converge.
1725  */
1726 static int preferred_group_nid(struct task_struct *p, int nid)
1727 {
1728         nodemask_t nodes;
1729         int dist;
1730 
1731         /* Direct connections between all NUMA nodes. */
1732         if (sched_numa_topology_type == NUMA_DIRECT)
1733                 return nid;
1734 
1735         /*
1736          * On a system with glueless mesh NUMA topology, group_weight
1737          * scores nodes according to the number of NUMA hinting faults on
1738          * both the node itself, and on nearby nodes.
1739          */
1740         if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1741                 unsigned long score, max_score = 0;
1742                 int node, max_node = nid;
1743 
1744                 dist = sched_max_numa_distance;
1745 
1746                 for_each_online_node(node) {
1747                         score = group_weight(p, node, dist);
1748                         if (score > max_score) {
1749                                 max_score = score;
1750                                 max_node = node;
1751                         }
1752                 }
1753                 return max_node;
1754         }
1755 
1756         /*
1757          * Finding the preferred nid in a system with NUMA backplane
1758          * interconnect topology is more involved. The goal is to locate
1759          * tasks from numa_groups near each other in the system, and
1760          * untangle workloads from different sides of the system. This requires
1761          * searching down the hierarchy of node groups, recursively searching
1762          * inside the highest scoring group of nodes. The nodemask tricks
1763          * keep the complexity of the search down.
1764          */
1765         nodes = node_online_map;
1766         for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1767                 unsigned long max_faults = 0;
1768                 nodemask_t max_group = NODE_MASK_NONE;
1769                 int a, b;
1770 
1771                 /* Are there nodes at this distance from each other? */
1772                 if (!find_numa_distance(dist))
1773                         continue;
1774 
1775                 for_each_node_mask(a, nodes) {
1776                         unsigned long faults = 0;
1777                         nodemask_t this_group;
1778                         nodes_clear(this_group);
1779 
1780                         /* Sum group's NUMA faults; includes a==b case. */
1781                         for_each_node_mask(b, nodes) {
1782                                 if (node_distance(a, b) < dist) {
1783                                         faults += group_faults(p, b);
1784                                         node_set(b, this_group);
1785                                         node_clear(b, nodes);
1786                                 }
1787                         }
1788 
1789                         /* Remember the top group. */
1790                         if (faults > max_faults) {
1791                                 max_faults = faults;
1792                                 max_group = this_group;
1793                                 /*
1794                                  * subtle: at the smallest distance there is
1795                                  * just one node left in each "group", the
1796                                  * winner is the preferred nid.
1797                                  */
1798                                 nid = a;
1799                         }
1800                 }
1801                 /* Next round, evaluate the nodes within max_group. */
1802                 if (!max_faults)
1803                         break;
1804                 nodes = max_group;
1805         }
1806         return nid;
1807 }
1808 
1809 static void task_numa_placement(struct task_struct *p)
1810 {
1811         int seq, nid, max_nid = -1, max_group_nid = -1;
1812         unsigned long max_faults = 0, max_group_faults = 0;
1813         unsigned long fault_types[2] = { 0, 0 };
1814         unsigned long total_faults;
1815         u64 runtime, period;
1816         spinlock_t *group_lock = NULL;
1817 
1818         /*
1819          * The p->mm->numa_scan_seq field gets updated without
1820          * exclusive access. Use READ_ONCE() here to ensure
1821          * that the field is read in a single access:
1822          */
1823         seq = READ_ONCE(p->mm->numa_scan_seq);
1824         if (p->numa_scan_seq == seq)
1825                 return;
1826         p->numa_scan_seq = seq;
1827         p->numa_scan_period_max = task_scan_max(p);
1828 
1829         total_faults = p->numa_faults_locality[0] +
1830                        p->numa_faults_locality[1];
1831         runtime = numa_get_avg_runtime(p, &period);
1832 
1833         /* If the task is part of a group prevent parallel updates to group stats */
1834         if (p->numa_group) {
1835                 group_lock = &p->numa_group->lock;
1836                 spin_lock_irq(group_lock);
1837         }
1838 
1839         /* Find the node with the highest number of faults */
1840         for_each_online_node(nid) {
1841                 /* Keep track of the offsets in numa_faults array */
1842                 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1843                 unsigned long faults = 0, group_faults = 0;
1844                 int priv;
1845 
1846                 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1847                         long diff, f_diff, f_weight;
1848 
1849                         mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1850                         membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1851                         cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1852                         cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1853 
1854                         /* Decay existing window, copy faults since last scan */
1855                         diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1856                         fault_types[priv] += p->numa_faults[membuf_idx];
1857                         p->numa_faults[membuf_idx] = 0;
1858 
1859                         /*
1860                          * Normalize the faults_from, so all tasks in a group
1861                          * count according to CPU use, instead of by the raw
1862                          * number of faults. Tasks with little runtime have
1863                          * little over-all impact on throughput, and thus their
1864                          * faults are less important.
1865                          */
1866                         f_weight = div64_u64(runtime << 16, period + 1);
1867                         f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1868                                    (total_faults + 1);
1869                         f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1870                         p->numa_faults[cpubuf_idx] = 0;
1871 
1872                         p->numa_faults[mem_idx] += diff;
1873                         p->numa_faults[cpu_idx] += f_diff;
1874                         faults += p->numa_faults[mem_idx];
1875                         p->total_numa_faults += diff;
1876                         if (p->numa_group) {
1877                                 /*
1878                                  * safe because we can only change our own group
1879                                  *
1880                                  * mem_idx represents the offset for a given
1881                                  * nid and priv in a specific region because it
1882                                  * is at the beginning of the numa_faults array.
1883                                  */
1884                                 p->numa_group->faults[mem_idx] += diff;
1885                                 p->numa_group->faults_cpu[mem_idx] += f_diff;
1886                                 p->numa_group->total_faults += diff;
1887                                 group_faults += p->numa_group->faults[mem_idx];
1888                         }
1889                 }
1890 
1891                 if (faults > max_faults) {
1892                         max_faults = faults;
1893                         max_nid = nid;
1894                 }
1895 
1896                 if (group_faults > max_group_faults) {
1897                         max_group_faults = group_faults;
1898                         max_group_nid = nid;
1899                 }
1900         }
1901 
1902         update_task_scan_period(p, fault_types[0], fault_types[1]);
1903 
1904         if (p->numa_group) {
1905                 update_numa_active_node_mask(p->numa_group);
1906                 spin_unlock_irq(group_lock);
1907                 max_nid = preferred_group_nid(p, max_group_nid);
1908         }
1909 
1910         if (max_faults) {
1911                 /* Set the new preferred node */
1912                 if (max_nid != p->numa_preferred_nid)
1913                         sched_setnuma(p, max_nid);
1914 
1915                 if (task_node(p) != p->numa_preferred_nid)
1916                         numa_migrate_preferred(p);
1917         }
1918 }
1919 
1920 static inline int get_numa_group(struct numa_group *grp)
1921 {
1922         return atomic_inc_not_zero(&grp->refcount);
1923 }
1924 
1925 static inline void put_numa_group(struct numa_group *grp)
1926 {
1927         if (atomic_dec_and_test(&grp->refcount))
1928                 kfree_rcu(grp, rcu);
1929 }
1930 
1931 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1932                         int *priv)
1933 {
1934         struct numa_group *grp, *my_grp;
1935         struct task_struct *tsk;
1936         bool join = false;
1937         int cpu = cpupid_to_cpu(cpupid);
1938         int i;
1939 
1940         if (unlikely(!p->numa_group)) {
1941                 unsigned int size = sizeof(struct numa_group) +
1942                                     4*nr_node_ids*sizeof(unsigned long);
1943 
1944                 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1945                 if (!grp)
1946                         return;
1947 
1948                 atomic_set(&grp->refcount, 1);
1949                 spin_lock_init(&grp->lock);
1950                 grp->gid = p->pid;
1951                 /* Second half of the array tracks nids where faults happen */
1952                 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1953                                                 nr_node_ids;
1954 
1955                 node_set(task_node(current), grp->active_nodes);
1956 
1957                 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1958                         grp->faults[i] = p->numa_faults[i];
1959 
1960                 grp->total_faults = p->total_numa_faults;
1961 
1962                 grp->nr_tasks++;
1963                 rcu_assign_pointer(p->numa_group, grp);
1964         }
1965 
1966         rcu_read_lock();
1967         tsk = READ_ONCE(cpu_rq(cpu)->curr);
1968 
1969         if (!cpupid_match_pid(tsk, cpupid))
1970                 goto no_join;
1971 
1972         grp = rcu_dereference(tsk->numa_group);
1973         if (!grp)
1974                 goto no_join;
1975 
1976         my_grp = p->numa_group;
1977         if (grp == my_grp)
1978                 goto no_join;
1979 
1980         /*
1981          * Only join the other group if its bigger; if we're the bigger group,
1982          * the other task will join us.
1983          */
1984         if (my_grp->nr_tasks > grp->nr_tasks)
1985                 goto no_join;
1986 
1987         /*
1988          * Tie-break on the grp address.
1989          */
1990         if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1991                 goto no_join;
1992 
1993         /* Always join threads in the same process. */
1994         if (tsk->mm == current->mm)
1995                 join = true;
1996 
1997         /* Simple filter to avoid false positives due to PID collisions */
1998         if (flags & TNF_SHARED)
1999                 join = true;
2000 
2001         /* Update priv based on whether false sharing was detected */
2002         *priv = !join;
2003 
2004         if (join && !get_numa_group(grp))
2005                 goto no_join;
2006 
2007         rcu_read_unlock();
2008 
2009         if (!join)
2010                 return;
2011 
2012         BUG_ON(irqs_disabled());
2013         double_lock_irq(&my_grp->lock, &grp->lock);
2014 
2015         for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2016                 my_grp->faults[i] -= p->numa_faults[i];
2017                 grp->faults[i] += p->numa_faults[i];
2018         }
2019         my_grp->total_faults -= p->total_numa_faults;
2020         grp->total_faults += p->total_numa_faults;
2021 
2022         my_grp->nr_tasks--;
2023         grp->nr_tasks++;
2024 
2025         spin_unlock(&my_grp->lock);
2026         spin_unlock_irq(&grp->lock);
2027 
2028         rcu_assign_pointer(p->numa_group, grp);
2029 
2030         put_numa_group(my_grp);
2031         return;
2032 
2033 no_join:
2034         rcu_read_unlock();
2035         return;
2036 }
2037 
2038 void task_numa_free(struct task_struct *p)
2039 {
2040         struct numa_group *grp = p->numa_group;
2041         void *numa_faults = p->numa_faults;
2042         unsigned long flags;
2043         int i;
2044 
2045         if (grp) {
2046                 spin_lock_irqsave(&grp->lock, flags);
2047                 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2048                         grp->faults[i] -= p->numa_faults[i];
2049                 grp->total_faults -= p->total_numa_faults;
2050 
2051                 grp->nr_tasks--;
2052                 spin_unlock_irqrestore(&grp->lock, flags);
2053                 RCU_INIT_POINTER(p->numa_group, NULL);
2054                 put_numa_group(grp);
2055         }
2056 
2057         p->numa_faults = NULL;
2058         kfree(numa_faults);
2059 }
2060 
2061 /*
2062  * Got a PROT_NONE fault for a page on @node.
2063  */
2064 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2065 {
2066         struct task_struct *p = current;
2067         bool migrated = flags & TNF_MIGRATED;
2068         int cpu_node = task_node(current);
2069         int local = !!(flags & TNF_FAULT_LOCAL);
2070         int priv;
2071 
2072         if (!numabalancing_enabled)
2073                 return;
2074 
2075         /* for example, ksmd faulting in a user's mm */
2076         if (!p->mm)
2077                 return;
2078 
2079         /* Allocate buffer to track faults on a per-node basis */
2080         if (unlikely(!p->numa_faults)) {
2081                 int size = sizeof(*p->numa_faults) *
2082                            NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2083 
2084                 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2085                 if (!p->numa_faults)
2086                         return;
2087 
2088                 p->total_numa_faults = 0;
2089                 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2090         }
2091 
2092         /*
2093          * First accesses are treated as private, otherwise consider accesses
2094          * to be private if the accessing pid has not changed
2095          */
2096         if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2097                 priv = 1;
2098         } else {
2099                 priv = cpupid_match_pid(p, last_cpupid);
2100                 if (!priv && !(flags & TNF_NO_GROUP))
2101                         task_numa_group(p, last_cpupid, flags, &priv);
2102         }
2103 
2104         /*
2105          * If a workload spans multiple NUMA nodes, a shared fault that
2106          * occurs wholly within the set of nodes that the workload is
2107          * actively using should be counted as local. This allows the
2108          * scan rate to slow down when a workload has settled down.
2109          */
2110         if (!priv && !local && p->numa_group &&
2111                         node_isset(cpu_node, p->numa_group->active_nodes) &&
2112                         node_isset(mem_node, p->numa_group->active_nodes))
2113                 local = 1;
2114 
2115         task_numa_placement(p);
2116 
2117         /*
2118          * Retry task to preferred node migration periodically, in case it
2119          * case it previously failed, or the scheduler moved us.
2120          */
2121         if (time_after(jiffies, p->numa_migrate_retry))
2122                 numa_migrate_preferred(p);
2123 
2124         if (migrated)
2125                 p->numa_pages_migrated += pages;
2126         if (flags & TNF_MIGRATE_FAIL)
2127                 p->numa_faults_locality[2] += pages;
2128 
2129         p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2130         p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2131         p->numa_faults_locality[local] += pages;
2132 }
2133 
2134 static void reset_ptenuma_scan(struct task_struct *p)
2135 {
2136         /*
2137          * We only did a read acquisition of the mmap sem, so
2138          * p->mm->numa_scan_seq is written to without exclusive access
2139          * and the update is not guaranteed to be atomic. That's not
2140          * much of an issue though, since this is just used for
2141          * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2142          * expensive, to avoid any form of compiler optimizations:
2143          */
2144         WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2145         p->mm->numa_scan_offset = 0;
2146 }
2147 
2148 /*
2149  * The expensive part of numa migration is done from task_work context.
2150  * Triggered from task_tick_numa().
2151  */
2152 void task_numa_work(struct callback_head *work)
2153 {
2154         unsigned long migrate, next_scan, now = jiffies;
2155         struct task_struct *p = current;
2156         struct mm_struct *mm = p->mm;
2157         struct vm_area_struct *vma;
2158         unsigned long start, end;
2159         unsigned long nr_pte_updates = 0;
2160         long pages;
2161 
2162         WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2163 
2164         work->next = work; /* protect against double add */
2165         /*
2166          * Who cares about NUMA placement when they're dying.
2167          *
2168          * NOTE: make sure not to dereference p->mm before this check,
2169          * exit_task_work() happens _after_ exit_mm() so we could be called
2170          * without p->mm even though we still had it when we enqueued this
2171          * work.
2172          */
2173         if (p->flags & PF_EXITING)
2174                 return;
2175 
2176         if (!mm->numa_next_scan) {
2177                 mm->numa_next_scan = now +
2178                         msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2179         }
2180 
2181         /*
2182          * Enforce maximal scan/migration frequency..
2183          */
2184         migrate = mm->numa_next_scan;
2185         if (time_before(now, migrate))
2186                 return;
2187 
2188         if (p->numa_scan_period == 0) {
2189                 p->numa_scan_period_max = task_scan_max(p);
2190                 p->numa_scan_period = task_scan_min(p);
2191         }
2192 
2193         next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2194         if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2195                 return;
2196 
2197         /*
2198          * Delay this task enough that another task of this mm will likely win
2199          * the next time around.
2200          */
2201         p->node_stamp += 2 * TICK_NSEC;
2202 
2203         start = mm->numa_scan_offset;
2204         pages = sysctl_numa_balancing_scan_size;
2205         pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2206         if (!pages)
2207                 return;
2208 
2209         down_read(&mm->mmap_sem);
2210         vma = find_vma(mm, start);
2211         if (!vma) {
2212                 reset_ptenuma_scan(p);
2213                 start = 0;
2214                 vma = mm->mmap;
2215         }
2216         for (; vma; vma = vma->vm_next) {
2217                 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2218                         is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2219                         continue;
2220                 }
2221 
2222                 /*
2223                  * Shared library pages mapped by multiple processes are not
2224                  * migrated as it is expected they are cache replicated. Avoid
2225                  * hinting faults in read-only file-backed mappings or the vdso
2226                  * as migrating the pages will be of marginal benefit.
2227                  */
2228                 if (!vma->vm_mm ||
2229                     (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2230                         continue;
2231 
2232                 /*
2233                  * Skip inaccessible VMAs to avoid any confusion between
2234                  * PROT_NONE and NUMA hinting ptes
2235                  */
2236                 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2237                         continue;
2238 
2239                 do {
2240                         start = max(start, vma->vm_start);
2241                         end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2242                         end = min(end, vma->vm_end);
2243                         nr_pte_updates += change_prot_numa(vma, start, end);
2244 
2245                         /*
2246                          * Scan sysctl_numa_balancing_scan_size but ensure that
2247                          * at least one PTE is updated so that unused virtual
2248                          * address space is quickly skipped.
2249                          */
2250                         if (nr_pte_updates)
2251                                 pages -= (end - start) >> PAGE_SHIFT;
2252 
2253                         start = end;
2254                         if (pages <= 0)
2255                                 goto out;
2256 
2257                         cond_resched();
2258                 } while (end != vma->vm_end);
2259         }
2260 
2261 out:
2262         /*
2263          * It is possible to reach the end of the VMA list but the last few
2264          * VMAs are not guaranteed to the vma_migratable. If they are not, we
2265          * would find the !migratable VMA on the next scan but not reset the
2266          * scanner to the start so check it now.
2267          */
2268         if (vma)
2269                 mm->numa_scan_offset = start;
2270         else
2271                 reset_ptenuma_scan(p);
2272         up_read(&mm->mmap_sem);
2273 }
2274 
2275 /*
2276  * Drive the periodic memory faults..
2277  */
2278 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2279 {
2280         struct callback_head *work = &curr->numa_work;
2281         u64 period, now;
2282 
2283         /*
2284          * We don't care about NUMA placement if we don't have memory.
2285          */
2286         if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2287                 return;
2288 
2289         /*
2290          * Using runtime rather than walltime has the dual advantage that
2291          * we (mostly) drive the selection from busy threads and that the
2292          * task needs to have done some actual work before we bother with
2293          * NUMA placement.
2294          */
2295         now = curr->se.sum_exec_runtime;
2296         period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2297 
2298         if (now - curr->node_stamp > period) {
2299                 if (!curr->node_stamp)
2300                         curr->numa_scan_period = task_scan_min(curr);
2301                 curr->node_stamp += period;
2302 
2303                 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2304                         init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2305                         task_work_add(curr, work, true);
2306                 }
2307         }
2308 }
2309 #else
2310 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2311 {
2312 }
2313 
2314 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2315 {
2316 }
2317 
2318 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2319 {
2320 }
2321 #endif /* CONFIG_NUMA_BALANCING */
2322 
2323 static void
2324 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2325 {
2326         update_load_add(&cfs_rq->load, se->load.weight);
2327         if (!parent_entity(se))
2328                 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2329 #ifdef CONFIG_SMP
2330         if (entity_is_task(se)) {
2331                 struct rq *rq = rq_of(cfs_rq);
2332 
2333                 account_numa_enqueue(rq, task_of(se));
2334                 list_add(&se->group_node, &rq->cfs_tasks);
2335         }
2336 #endif
2337         cfs_rq->nr_running++;
2338 }
2339 
2340 static void
2341 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2342 {
2343         update_load_sub(&cfs_rq->load, se->load.weight);
2344         if (!parent_entity(se))
2345                 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2346         if (entity_is_task(se)) {
2347                 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2348                 list_del_init(&se->group_node);
2349         }
2350         cfs_rq->nr_running--;
2351 }
2352 
2353 #ifdef CONFIG_FAIR_GROUP_SCHED
2354 # ifdef CONFIG_SMP
2355 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2356 {
2357         long tg_weight;
2358 
2359         /*
2360          * Use this CPU's real-time load instead of the last load contribution
2361          * as the updating of the contribution is delayed, and we will use the
2362          * the real-time load to calc the share. See update_tg_load_avg().
2363          */
2364         tg_weight = atomic_long_read(&tg->load_avg);
2365         tg_weight -= cfs_rq->tg_load_avg_contrib;
2366         tg_weight += cfs_rq->load.weight;
2367 
2368         return tg_weight;
2369 }
2370 
2371 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2372 {
2373         long tg_weight, load, shares;
2374 
2375         tg_weight = calc_tg_weight(tg, cfs_rq);
2376         load = cfs_rq->load.weight;
2377 
2378         shares = (tg->shares * load);
2379         if (tg_weight)
2380                 shares /= tg_weight;
2381 
2382         if (shares < MIN_SHARES)
2383                 shares = MIN_SHARES;
2384         if (shares > tg->shares)
2385                 shares = tg->shares;
2386 
2387         return shares;
2388 }
2389 # else /* CONFIG_SMP */
2390 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2391 {
2392         return tg->shares;
2393 }
2394 # endif /* CONFIG_SMP */
2395 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2396                             unsigned long weight)
2397 {
2398         if (se->on_rq) {
2399                 /* commit outstanding execution time */
2400                 if (cfs_rq->curr == se)
2401                         update_curr(cfs_rq);
2402                 account_entity_dequeue(cfs_rq, se);
2403         }
2404 
2405         update_load_set(&se->load, weight);
2406 
2407         if (se->on_rq)
2408                 account_entity_enqueue(cfs_rq, se);
2409 }
2410 
2411 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2412 
2413 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2414 {
2415         struct task_group *tg;
2416         struct sched_entity *se;
2417         long shares;
2418 
2419         tg = cfs_rq->tg;
2420         se = tg->se[cpu_of(rq_of(cfs_rq))];
2421         if (!se || throttled_hierarchy(cfs_rq))
2422                 return;
2423 #ifndef CONFIG_SMP
2424         if (likely(se->load.weight == tg->shares))
2425                 return;
2426 #endif
2427         shares = calc_cfs_shares(cfs_rq, tg);
2428 
2429         reweight_entity(cfs_rq_of(se), se, shares);
2430 }
2431 #else /* CONFIG_FAIR_GROUP_SCHED */
2432 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2433 {
2434 }
2435 #endif /* CONFIG_FAIR_GROUP_SCHED */
2436 
2437 #ifdef CONFIG_SMP
2438 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2439 static const u32 runnable_avg_yN_inv[] = {
2440         0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2441         0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2442         0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2443         0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2444         0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2445         0x85aac367, 0x82cd8698,
2446 };
2447 
2448 /*
2449  * Precomputed \Sum y^k { 1<=k<=n }.  These are floor(true_value) to prevent
2450  * over-estimates when re-combining.
2451  */
2452 static const u32 runnable_avg_yN_sum[] = {
2453             0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2454          9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2455         17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2456 };
2457 
2458 /*
2459  * Approximate:
2460  *   val * y^n,    where y^32 ~= 0.5 (~1 scheduling period)
2461  */
2462 static __always_inline u64 decay_load(u64 val, u64 n)
2463 {
2464         unsigned int local_n;
2465 
2466         if (!n)
2467                 return val;
2468         else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2469                 return 0;
2470 
2471         /* after bounds checking we can collapse to 32-bit */
2472         local_n = n;
2473 
2474         /*
2475          * As y^PERIOD = 1/2, we can combine
2476          *    y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2477          * With a look-up table which covers y^n (n<PERIOD)
2478          *
2479          * To achieve constant time decay_load.
2480          */
2481         if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2482                 val >>= local_n / LOAD_AVG_PERIOD;
2483                 local_n %= LOAD_AVG_PERIOD;
2484         }
2485 
2486         val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2487         return val;
2488 }
2489 
2490 /*
2491  * For updates fully spanning n periods, the contribution to runnable
2492  * average will be: \Sum 1024*y^n
2493  *
2494  * We can compute this reasonably efficiently by combining:
2495  *   y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for  n <PERIOD}
2496  */
2497 static u32 __compute_runnable_contrib(u64 n)
2498 {
2499         u32 contrib = 0;
2500 
2501         if (likely(n <= LOAD_AVG_PERIOD))
2502                 return runnable_avg_yN_sum[n];
2503         else if (unlikely(n >= LOAD_AVG_MAX_N))
2504                 return LOAD_AVG_MAX;
2505 
2506         /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2507         do {
2508                 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2509                 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2510 
2511                 n -= LOAD_AVG_PERIOD;
2512         } while (n > LOAD_AVG_PERIOD);
2513 
2514         contrib = decay_load(contrib, n);
2515         return contrib + runnable_avg_yN_sum[n];
2516 }
2517 
2518 /*
2519  * We can represent the historical contribution to runnable average as the
2520  * coefficients of a geometric series.  To do this we sub-divide our runnable
2521  * history into segments of approximately 1ms (1024us); label the segment that
2522  * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2523  *
2524  * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2525  *      p0            p1           p2
2526  *     (now)       (~1ms ago)  (~2ms ago)
2527  *
2528  * Let u_i denote the fraction of p_i that the entity was runnable.
2529  *
2530  * We then designate the fractions u_i as our co-efficients, yielding the
2531  * following representation of historical load:
2532  *   u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2533  *
2534  * We choose y based on the with of a reasonably scheduling period, fixing:
2535  *   y^32 = 0.5
2536  *
2537  * This means that the contribution to load ~32ms ago (u_32) will be weighted
2538  * approximately half as much as the contribution to load within the last ms
2539  * (u_0).
2540  *
2541  * When a period "rolls over" and we have new u_0`, multiplying the previous
2542  * sum again by y is sufficient to update:
2543  *   load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2544  *            = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2545  */
2546 static __always_inline int
2547 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2548                   unsigned long weight, int running, struct cfs_rq *cfs_rq)
2549 {
2550         u64 delta, periods;
2551         u32 contrib;
2552         int delta_w, decayed = 0;
2553         unsigned long scale_freq = arch_scale_freq_capacity(NULL, cpu);
2554 
2555         delta = now - sa->last_update_time;
2556         /*
2557          * This should only happen when time goes backwards, which it
2558          * unfortunately does during sched clock init when we swap over to TSC.
2559          */
2560         if ((s64)delta < 0) {
2561                 sa->last_update_time = now;
2562                 return 0;
2563         }
2564 
2565         /*
2566          * Use 1024ns as the unit of measurement since it's a reasonable
2567          * approximation of 1us and fast to compute.
2568          */
2569         delta >>= 10;
2570         if (!delta)
2571                 return 0;
2572         sa->last_update_time = now;
2573 
2574         /* delta_w is the amount already accumulated against our next period */
2575         delta_w = sa->period_contrib;
2576         if (delta + delta_w >= 1024) {
2577                 decayed = 1;
2578 
2579                 /* how much left for next period will start over, we don't know yet */
2580                 sa->period_contrib = 0;
2581 
2582                 /*
2583                  * Now that we know we're crossing a period boundary, figure
2584                  * out how much from delta we need to complete the current
2585                  * period and accrue it.
2586                  */
2587                 delta_w = 1024 - delta_w;
2588                 if (weight) {
2589                         sa->load_sum += weight * delta_w;
2590                         if (cfs_rq)
2591                                 cfs_rq->runnable_load_sum += weight * delta_w;
2592                 }
2593                 if (running)
2594                         sa->util_sum += delta_w * scale_freq >> SCHED_CAPACITY_SHIFT;
2595 
2596                 delta -= delta_w;
2597 
2598                 /* Figure out how many additional periods this update spans */
2599                 periods = delta / 1024;
2600                 delta %= 1024;
2601 
2602                 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2603                 if (cfs_rq) {
2604                         cfs_rq->runnable_load_sum =
2605                                 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2606                 }
2607                 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2608 
2609                 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2610                 contrib = __compute_runnable_contrib(periods);
2611                 if (weight) {
2612                         sa->load_sum += weight * contrib;
2613                         if (cfs_rq)
2614                                 cfs_rq->runnable_load_sum += weight * contrib;
2615                 }
2616                 if (running)
2617                         sa->util_sum += contrib * scale_freq >> SCHED_CAPACITY_SHIFT;
2618         }
2619 
2620         /* Remainder of delta accrued against u_0` */
2621         if (weight) {
2622                 sa->load_sum += weight * delta;
2623                 if (cfs_rq)
2624                         cfs_rq->runnable_load_sum += weight * delta;
2625         }
2626         if (running)
2627                 sa->util_sum += delta * scale_freq >> SCHED_CAPACITY_SHIFT;
2628 
2629         sa->period_contrib += delta;
2630 
2631         if (decayed) {
2632                 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2633                 if (cfs_rq) {
2634                         cfs_rq->runnable_load_avg =
2635                                 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2636                 }
2637                 sa->util_avg = (sa->util_sum << SCHED_LOAD_SHIFT) / LOAD_AVG_MAX;
2638         }
2639 
2640         return decayed;
2641 }
2642 
2643 #ifdef CONFIG_FAIR_GROUP_SCHED
2644 /*
2645  * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2646  * and effective_load (which is not done because it is too costly).
2647  */
2648 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2649 {
2650         long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2651 
2652         if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2653                 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2654                 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2655         }
2656 }
2657 
2658 #else /* CONFIG_FAIR_GROUP_SCHED */
2659 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2660 #endif /* CONFIG_FAIR_GROUP_SCHED */
2661 
2662 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2663 
2664 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2665 static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
2666 {
2667         struct sched_avg *sa = &cfs_rq->avg;
2668         int decayed, removed = 0;
2669 
2670         if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2671                 long r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2672                 sa->load_avg = max_t(long, sa->load_avg - r, 0);
2673                 sa->load_sum = max_t(s64, sa->load_sum - r * LOAD_AVG_MAX, 0);
2674                 removed = 1;
2675         }
2676 
2677         if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2678                 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2679                 sa->util_avg = max_t(long, sa->util_avg - r, 0);
2680                 sa->util_sum = max_t(s32, sa->util_sum -
2681                         ((r * LOAD_AVG_MAX) >> SCHED_LOAD_SHIFT), 0);
2682         }
2683 
2684         decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2685                 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2686 
2687 #ifndef CONFIG_64BIT
2688         smp_wmb();
2689         cfs_rq->load_last_update_time_copy = sa->last_update_time;
2690 #endif
2691 
2692         return decayed || removed;
2693 }
2694 
2695 /* Update task and its cfs_rq load average */
2696 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2697 {
2698         struct cfs_rq *cfs_rq = cfs_rq_of(se);
2699         int cpu = cpu_of(rq_of(cfs_rq));
2700         u64 now = cfs_rq_clock_task(cfs_rq);
2701 
2702         /*
2703          * Track task load average for carrying it to new CPU after migrated, and
2704          * track group sched_entity load average for task_h_load calc in migration
2705          */
2706         __update_load_avg(now, cpu, &se->avg,
2707                 se->on_rq * scale_load_down(se->load.weight), cfs_rq->curr == se, NULL);
2708 
2709         if (update_cfs_rq_load_avg(now, cfs_rq) && update_tg)
2710                 update_tg_load_avg(cfs_rq, 0);
2711 }
2712 
2713 /* Add the load generated by se into cfs_rq's load average */
2714 static inline void
2715 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2716 {
2717         struct sched_avg *sa = &se->avg;
2718         u64 now = cfs_rq_clock_task(cfs_rq);
2719         int migrated = 0, decayed;
2720 
2721         if (sa->last_update_time == 0) {
2722                 sa->last_update_time = now;
2723                 migrated = 1;
2724         }
2725         else {
2726                 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2727                         se->on_rq * scale_load_down(se->load.weight),
2728                         cfs_rq->curr == se, NULL);
2729         }
2730 
2731         decayed = update_cfs_rq_load_avg(now, cfs_rq);
2732 
2733         cfs_rq->runnable_load_avg += sa->load_avg;
2734         cfs_rq->runnable_load_sum += sa->load_sum;
2735 
2736         if (migrated) {
2737                 cfs_rq->avg.load_avg += sa->load_avg;
2738                 cfs_rq->avg.load_sum += sa->load_sum;
2739                 cfs_rq->avg.util_avg += sa->util_avg;
2740                 cfs_rq->avg.util_sum += sa->util_sum;
2741         }
2742 
2743         if (decayed || migrated)
2744                 update_tg_load_avg(cfs_rq, 0);
2745 }
2746 
2747 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2748 static inline void
2749 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2750 {
2751         update_load_avg(se, 1);
2752 
2753         cfs_rq->runnable_load_avg =
2754                 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2755         cfs_rq->runnable_load_sum =
2756                 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2757 }
2758 
2759 /*
2760  * Task first catches up with cfs_rq, and then subtract
2761  * itself from the cfs_rq (task must be off the queue now).
2762  */
2763 void remove_entity_load_avg(struct sched_entity *se)
2764 {
2765         struct cfs_rq *cfs_rq = cfs_rq_of(se);
2766         u64 last_update_time;
2767 
2768 #ifndef CONFIG_64BIT
2769         u64 last_update_time_copy;
2770 
2771         do {
2772                 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2773                 smp_rmb();
2774                 last_update_time = cfs_rq->avg.last_update_time;
2775         } while (last_update_time != last_update_time_copy);
2776 #else
2777         last_update_time = cfs_rq->avg.last_update_time;
2778 #endif
2779 
2780         __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
2781         atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2782         atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2783 }
2784 
2785 /*
2786  * Update the rq's load with the elapsed running time before entering
2787  * idle. if the last scheduled task is not a CFS task, idle_enter will
2788  * be the only way to update the runnable statistic.
2789  */
2790 void idle_enter_fair(struct rq *this_rq)
2791 {
2792 }
2793 
2794 /*
2795  * Update the rq's load with the elapsed idle time before a task is
2796  * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2797  * be the only way to update the runnable statistic.
2798  */
2799 void idle_exit_fair(struct rq *this_rq)
2800 {
2801 }
2802 
2803 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
2804 {
2805         return cfs_rq->runnable_load_avg;
2806 }
2807 
2808 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
2809 {
2810         return cfs_rq->avg.load_avg;
2811 }
2812 
2813 static int idle_balance(struct rq *this_rq);
2814 
2815 #else /* CONFIG_SMP */
2816 
2817 static inline void update_load_avg(struct sched_entity *se, int update_tg) {}
2818 static inline void
2819 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2820 static inline void
2821 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2822 static inline void remove_entity_load_avg(struct sched_entity *se) {}
2823 
2824 static inline int idle_balance(struct rq *rq)
2825 {
2826         return 0;
2827 }
2828 
2829 #endif /* CONFIG_SMP */
2830 
2831 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2832 {
2833 #ifdef CONFIG_SCHEDSTATS
2834         struct task_struct *tsk = NULL;
2835 
2836         if (entity_is_task(se))
2837                 tsk = task_of(se);
2838 
2839         if (se->statistics.sleep_start) {
2840                 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2841 
2842                 if ((s64)delta < 0)
2843                         delta = 0;
2844 
2845                 if (unlikely(delta > se->statistics.sleep_max))
2846                         se->statistics.sleep_max = delta;
2847 
2848                 se->statistics.sleep_start = 0;
2849                 se->statistics.sum_sleep_runtime += delta;
2850 
2851                 if (tsk) {
2852                         account_scheduler_latency(tsk, delta >> 10, 1);
2853                         trace_sched_stat_sleep(tsk, delta);
2854                 }
2855         }
2856         if (se->statistics.block_start) {
2857                 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2858 
2859                 if ((s64)delta < 0)
2860                         delta = 0;
2861 
2862                 if (unlikely(delta > se->statistics.block_max))
2863                         se->statistics.block_max = delta;
2864 
2865                 se->statistics.block_start = 0;
2866                 se->statistics.sum_sleep_runtime += delta;
2867 
2868                 if (tsk) {
2869                         if (tsk->in_iowait) {
2870                                 se->statistics.iowait_sum += delta;
2871                                 se->statistics.iowait_count++;
2872                                 trace_sched_stat_iowait(tsk, delta);
2873                         }
2874 
2875                         trace_sched_stat_blocked(tsk, delta);
2876 
2877                         /*
2878                          * Blocking time is in units of nanosecs, so shift by
2879                          * 20 to get a milliseconds-range estimation of the
2880                          * amount of time that the task spent sleeping:
2881                          */
2882                         if (unlikely(prof_on == SLEEP_PROFILING)) {
2883                                 profile_hits(SLEEP_PROFILING,
2884                                                 (void *)get_wchan(tsk),
2885                                                 delta >> 20);
2886                         }
2887                         account_scheduler_latency(tsk, delta >> 10, 0);
2888                 }
2889         }
2890 #endif
2891 }
2892 
2893 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2894 {
2895 #ifdef CONFIG_SCHED_DEBUG
2896         s64 d = se->vruntime - cfs_rq->min_vruntime;
2897 
2898         if (d < 0)
2899                 d = -d;
2900 
2901         if (d > 3*sysctl_sched_latency)
2902                 schedstat_inc(cfs_rq, nr_spread_over);
2903 #endif
2904 }
2905 
2906 static void
2907 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2908 {
2909         u64 vruntime = cfs_rq->min_vruntime;
2910 
2911         /*
2912          * The 'current' period is already promised to the current tasks,
2913          * however the extra weight of the new task will slow them down a
2914          * little, place the new task so that it fits in the slot that
2915          * stays open at the end.
2916          */
2917         if (initial && sched_feat(START_DEBIT))
2918                 vruntime += sched_vslice(cfs_rq, se);
2919 
2920         /* sleeps up to a single latency don't count. */
2921         if (!initial) {
2922                 unsigned long thresh = sysctl_sched_latency;
2923 
2924                 /*
2925                  * Halve their sleep time's effect, to allow
2926                  * for a gentler effect of sleepers:
2927                  */
2928                 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2929                         thresh >>= 1;
2930 
2931                 vruntime -= thresh;
2932         }
2933 
2934         /* ensure we never gain time by being placed backwards. */
2935         se->vruntime = max_vruntime(se->vruntime, vruntime);
2936 }
2937 
2938 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2939 
2940 static void
2941 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2942 {
2943         /*
2944          * Update the normalized vruntime before updating min_vruntime
2945          * through calling update_curr().
2946          */
2947         if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2948                 se->vruntime += cfs_rq->min_vruntime;
2949 
2950         /*
2951          * Update run-time statistics of the 'current'.
2952          */
2953         update_curr(cfs_rq);
2954         enqueue_entity_load_avg(cfs_rq, se);
2955         account_entity_enqueue(cfs_rq, se);
2956         update_cfs_shares(cfs_rq);
2957 
2958         if (flags & ENQUEUE_WAKEUP) {
2959                 place_entity(cfs_rq, se, 0);
2960                 enqueue_sleeper(cfs_rq, se);
2961         }
2962 
2963         update_stats_enqueue(cfs_rq, se);
2964         check_spread(cfs_rq, se);
2965         if (se != cfs_rq->curr)
2966                 __enqueue_entity(cfs_rq, se);
2967         se->on_rq = 1;
2968 
2969         if (cfs_rq->nr_running == 1) {
2970                 list_add_leaf_cfs_rq(cfs_rq);
2971                 check_enqueue_throttle(cfs_rq);
2972         }
2973 }
2974 
2975 static void __clear_buddies_last(struct sched_entity *se)
2976 {
2977         for_each_sched_entity(se) {
2978                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2979                 if (cfs_rq->last != se)
2980                         break;
2981 
2982                 cfs_rq->last = NULL;
2983         }
2984 }
2985 
2986 static void __clear_buddies_next(struct sched_entity *se)
2987 {
2988         for_each_sched_entity(se) {
2989                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2990                 if (cfs_rq->next != se)
2991                         break;
2992 
2993                 cfs_rq->next = NULL;
2994         }
2995 }
2996 
2997 static void __clear_buddies_skip(struct sched_entity *se)
2998 {
2999         for_each_sched_entity(se) {
3000                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3001                 if (cfs_rq->skip != se)
3002                         break;
3003 
3004                 cfs_rq->skip = NULL;
3005         }
3006 }
3007 
3008 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3009 {
3010         if (cfs_rq->last == se)
3011                 __clear_buddies_last(se);
3012 
3013         if (cfs_rq->next == se)
3014                 __clear_buddies_next(se);
3015 
3016         if (cfs_rq->skip == se)
3017                 __clear_buddies_skip(se);
3018 }
3019 
3020 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3021 
3022 static void
3023 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3024 {
3025         /*
3026          * Update run-time statistics of the 'current'.
3027          */
3028         update_curr(cfs_rq);
3029         dequeue_entity_load_avg(cfs_rq, se);
3030 
3031         update_stats_dequeue(cfs_rq, se);
3032         if (flags & DEQUEUE_SLEEP) {
3033 #ifdef CONFIG_SCHEDSTATS
3034                 if (entity_is_task(se)) {
3035                         struct task_struct *tsk = task_of(se);
3036 
3037                         if (tsk->state & TASK_INTERRUPTIBLE)
3038                                 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3039                         if (tsk->state & TASK_UNINTERRUPTIBLE)
3040                                 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3041                 }
3042 #endif
3043         }
3044 
3045         clear_buddies(cfs_rq, se);
3046 
3047         if (se != cfs_rq->curr)
3048                 __dequeue_entity(cfs_rq, se);
3049         se->on_rq = 0;
3050         account_entity_dequeue(cfs_rq, se);
3051 
3052         /*
3053          * Normalize the entity after updating the min_vruntime because the
3054          * update can refer to the ->curr item and we need to reflect this
3055          * movement in our normalized position.
3056          */
3057         if (!(flags & DEQUEUE_SLEEP))
3058                 se->vruntime -= cfs_rq->min_vruntime;
3059 
3060         /* return excess runtime on last dequeue */
3061         return_cfs_rq_runtime(cfs_rq);
3062 
3063         update_min_vruntime(cfs_rq);
3064         update_cfs_shares(cfs_rq);
3065 }
3066 
3067 /*
3068  * Preempt the current task with a newly woken task if needed:
3069  */
3070 static void
3071 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3072 {
3073         unsigned long ideal_runtime, delta_exec;
3074         struct sched_entity *se;
3075         s64 delta;
3076 
3077         ideal_runtime = sched_slice(cfs_rq, curr);
3078         delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3079         if (delta_exec > ideal_runtime) {
3080                 resched_curr(rq_of(cfs_rq));
3081                 /*
3082                  * The current task ran long enough, ensure it doesn't get
3083                  * re-elected due to buddy favours.
3084                  */
3085                 clear_buddies(cfs_rq, curr);
3086                 return;
3087         }
3088 
3089         /*
3090          * Ensure that a task that missed wakeup preemption by a
3091          * narrow margin doesn't have to wait for a full slice.
3092          * This also mitigates buddy induced latencies under load.
3093          */
3094         if (delta_exec < sysctl_sched_min_granularity)
3095                 return;
3096 
3097         se = __pick_first_entity(cfs_rq);
3098         delta = curr->vruntime - se->vruntime;
3099 
3100         if (delta < 0)
3101                 return;
3102 
3103         if (delta > ideal_runtime)
3104                 resched_curr(rq_of(cfs_rq));
3105 }
3106 
3107 static void
3108 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3109 {
3110         /* 'current' is not kept within the tree. */
3111         if (se->on_rq) {
3112                 /*
3113                  * Any task has to be enqueued before it get to execute on
3114                  * a CPU. So account for the time it spent waiting on the
3115                  * runqueue.
3116                  */
3117                 update_stats_wait_end(cfs_rq, se);
3118                 __dequeue_entity(cfs_rq, se);
3119                 update_load_avg(se, 1);
3120         }
3121 
3122         update_stats_curr_start(cfs_rq, se);
3123         cfs_rq->curr = se;
3124 #ifdef CONFIG_SCHEDSTATS
3125         /*
3126          * Track our maximum slice length, if the CPU's load is at
3127          * least twice that of our own weight (i.e. dont track it
3128          * when there are only lesser-weight tasks around):
3129          */
3130         if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3131                 se->statistics.slice_max = max(se->statistics.slice_max,
3132                         se->sum_exec_runtime - se->prev_sum_exec_runtime);
3133         }
3134 #endif
3135         se->prev_sum_exec_runtime = se->sum_exec_runtime;
3136 }
3137 
3138 static int
3139 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3140 
3141 /*
3142  * Pick the next process, keeping these things in mind, in this order:
3143  * 1) keep things fair between processes/task groups
3144  * 2) pick the "next" process, since someone really wants that to run
3145  * 3) pick the "last" process, for cache locality
3146  * 4) do not run the "skip" process, if something else is available
3147  */
3148 static struct sched_entity *
3149 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3150 {
3151         struct sched_entity *left = __pick_first_entity(cfs_rq);
3152         struct sched_entity *se;
3153 
3154         /*
3155          * If curr is set we have to see if its left of the leftmost entity
3156          * still in the tree, provided there was anything in the tree at all.
3157          */
3158         if (!left || (curr && entity_before(curr, left)))
3159                 left = curr;
3160 
3161         se = left; /* ideally we run the leftmost entity */
3162 
3163         /*
3164          * Avoid running the skip buddy, if running something else can
3165          * be done without getting too unfair.
3166          */
3167         if (cfs_rq->skip == se) {
3168                 struct sched_entity *second;
3169 
3170                 if (se == curr) {
3171                         second = __pick_first_entity(cfs_rq);
3172                 } else {
3173                         second = __pick_next_entity(se);
3174                         if (!second || (curr && entity_before(curr, second)))
3175                                 second = curr;
3176                 }
3177 
3178                 if (second && wakeup_preempt_entity(second, left) < 1)
3179                         se = second;
3180         }
3181 
3182         /*
3183          * Prefer last buddy, try to return the CPU to a preempted task.
3184          */
3185         if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3186                 se = cfs_rq->last;
3187 
3188         /*
3189          * Someone really wants this to run. If it's not unfair, run it.
3190          */
3191         if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3192                 se = cfs_rq->next;
3193 
3194         clear_buddies(cfs_rq, se);
3195 
3196         return se;
3197 }
3198 
3199 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3200 
3201 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3202 {
3203         /*
3204          * If still on the runqueue then deactivate_task()
3205          * was not called and update_curr() has to be done:
3206          */
3207         if (prev->on_rq)
3208                 update_curr(cfs_rq);
3209 
3210         /* throttle cfs_rqs exceeding runtime */
3211         check_cfs_rq_runtime(cfs_rq);
3212 
3213         check_spread(cfs_rq, prev);
3214         if (prev->on_rq) {
3215                 update_stats_wait_start(cfs_rq, prev);
3216                 /* Put 'current' back into the tree. */
3217                 __enqueue_entity(cfs_rq, prev);
3218                 /* in !on_rq case, update occurred at dequeue */
3219                 update_load_avg(prev, 0);
3220         }
3221         cfs_rq->curr = NULL;
3222 }
3223 
3224 static void
3225 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3226 {
3227         /*
3228          * Update run-time statistics of the 'current'.
3229          */
3230         update_curr(cfs_rq);
3231 
3232         /*
3233          * Ensure that runnable average is periodically updated.
3234          */
3235         update_load_avg(curr, 1);
3236         update_cfs_shares(cfs_rq);
3237 
3238 #ifdef CONFIG_SCHED_HRTICK
3239         /*
3240          * queued ticks are scheduled to match the slice, so don't bother
3241          * validating it and just reschedule.
3242          */
3243         if (queued) {
3244                 resched_curr(rq_of(cfs_rq));
3245                 return;
3246         }
3247         /*
3248          * don't let the period tick interfere with the hrtick preemption
3249          */
3250         if (!sched_feat(DOUBLE_TICK) &&
3251                         hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3252                 return;
3253 #endif
3254 
3255         if (cfs_rq->nr_running > 1)
3256                 check_preempt_tick(cfs_rq, curr);
3257 }
3258 
3259 
3260 /**************************************************
3261  * CFS bandwidth control machinery
3262  */
3263 
3264 #ifdef CONFIG_CFS_BANDWIDTH
3265 
3266 #ifdef HAVE_JUMP_LABEL
3267 static struct static_key __cfs_bandwidth_used;
3268 
3269 static inline bool cfs_bandwidth_used(void)
3270 {
3271         return static_key_false(&__cfs_bandwidth_used);
3272 }
3273 
3274 void cfs_bandwidth_usage_inc(void)
3275 {
3276         static_key_slow_inc(&__cfs_bandwidth_used);
3277 }
3278 
3279 void cfs_bandwidth_usage_dec(void)
3280 {
3281         static_key_slow_dec(&__cfs_bandwidth_used);
3282 }
3283 #else /* HAVE_JUMP_LABEL */
3284 static bool cfs_bandwidth_used(void)
3285 {
3286         return true;
3287 }
3288 
3289 void cfs_bandwidth_usage_inc(void) {}
3290 void cfs_bandwidth_usage_dec(void) {}
3291 #endif /* HAVE_JUMP_LABEL */
3292 
3293 /*
3294  * default period for cfs group bandwidth.
3295  * default: 0.1s, units: nanoseconds
3296  */
3297 static inline u64 default_cfs_period(void)
3298 {
3299         return 100000000ULL;
3300 }
3301 
3302 static inline u64 sched_cfs_bandwidth_slice(void)
3303 {
3304         return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3305 }
3306 
3307 /*
3308  * Replenish runtime according to assigned quota and update expiration time.
3309  * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3310  * additional synchronization around rq->lock.
3311  *
3312  * requires cfs_b->lock
3313  */
3314 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3315 {
3316         u64 now;
3317 
3318         if (cfs_b->quota == RUNTIME_INF)
3319                 return;
3320 
3321         now = sched_clock_cpu(smp_processor_id());
3322         cfs_b->runtime = cfs_b->quota;
3323         cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3324 }
3325 
3326 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3327 {
3328         return &tg->cfs_bandwidth;
3329 }
3330 
3331 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3332 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3333 {
3334         if (unlikely(cfs_rq->throttle_count))
3335                 return cfs_rq->throttled_clock_task;
3336 
3337         return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3338 }
3339 
3340 /* returns 0 on failure to allocate runtime */
3341 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3342 {
3343         struct task_group *tg = cfs_rq->tg;
3344         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3345         u64 amount = 0, min_amount, expires;
3346 
3347         /* note: this is a positive sum as runtime_remaining <= 0 */
3348         min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3349 
3350         raw_spin_lock(&cfs_b->lock);
3351         if (cfs_b->quota == RUNTIME_INF)
3352                 amount = min_amount;
3353         else {
3354                 start_cfs_bandwidth(cfs_b);
3355 
3356                 if (cfs_b->runtime > 0) {
3357                         amount = min(cfs_b->runtime, min_amount);
3358                         cfs_b->runtime -= amount;
3359                         cfs_b->idle = 0;
3360                 }
3361         }
3362         expires = cfs_b->runtime_expires;
3363         raw_spin_unlock(&cfs_b->lock);
3364 
3365         cfs_rq->runtime_remaining += amount;
3366         /*
3367          * we may have advanced our local expiration to account for allowed
3368          * spread between our sched_clock and the one on which runtime was
3369          * issued.
3370          */
3371         if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3372                 cfs_rq->runtime_expires = expires;
3373 
3374         return cfs_rq->runtime_remaining > 0;
3375 }
3376 
3377 /*
3378  * Note: This depends on the synchronization provided by sched_clock and the
3379  * fact that rq->clock snapshots this value.
3380  */
3381 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3382 {
3383         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3384 
3385         /* if the deadline is ahead of our clock, nothing to do */
3386         if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3387                 return;
3388 
3389         if (cfs_rq->runtime_remaining < 0)
3390                 return;
3391 
3392         /*
3393          * If the local deadline has passed we have to consider the
3394          * possibility that our sched_clock is 'fast' and the global deadline
3395          * has not truly expired.
3396          *
3397          * Fortunately we can check determine whether this the case by checking
3398          * whether the global deadline has advanced. It is valid to compare
3399          * cfs_b->runtime_expires without any locks since we only care about
3400          * exact equality, so a partial write will still work.
3401          */
3402 
3403         if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3404                 /* extend local deadline, drift is bounded above by 2 ticks */
3405                 cfs_rq->runtime_expires += TICK_NSEC;
3406         } else {
3407                 /* global deadline is ahead, expiration has passed */
3408                 cfs_rq->runtime_remaining = 0;
3409         }
3410 }
3411 
3412 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3413 {
3414         /* dock delta_exec before expiring quota (as it could span periods) */
3415         cfs_rq->runtime_remaining -= delta_exec;
3416         expire_cfs_rq_runtime(cfs_rq);
3417 
3418         if (likely(cfs_rq->runtime_remaining > 0))
3419                 return;
3420 
3421         /*
3422          * if we're unable to extend our runtime we resched so that the active
3423          * hierarchy can be throttled
3424          */
3425         if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3426                 resched_curr(rq_of(cfs_rq));
3427 }
3428 
3429 static __always_inline
3430 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3431 {
3432         if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3433                 return;
3434 
3435         __account_cfs_rq_runtime(cfs_rq, delta_exec);
3436 }
3437 
3438 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3439 {
3440         return cfs_bandwidth_used() && cfs_rq->throttled;
3441 }
3442 
3443 /* check whether cfs_rq, or any parent, is throttled */
3444 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3445 {
3446         return cfs_bandwidth_used() && cfs_rq->throttle_count;
3447 }
3448 
3449 /*
3450  * Ensure that neither of the group entities corresponding to src_cpu or
3451  * dest_cpu are members of a throttled hierarchy when performing group
3452  * load-balance operations.
3453  */
3454 static inline int throttled_lb_pair(struct task_group *tg,
3455                                     int src_cpu, int dest_cpu)
3456 {
3457         struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3458 
3459         src_cfs_rq = tg->cfs_rq[src_cpu];
3460         dest_cfs_rq = tg->cfs_rq[dest_cpu];
3461 
3462         return throttled_hierarchy(src_cfs_rq) ||
3463                throttled_hierarchy(dest_cfs_rq);
3464 }
3465 
3466 /* updated child weight may affect parent so we have to do this bottom up */
3467 static int tg_unthrottle_up(struct task_group *tg, void *data)
3468 {
3469         struct rq *rq = data;
3470         struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3471 
3472         cfs_rq->throttle_count--;
3473 #ifdef CONFIG_SMP
3474         if (!cfs_rq->throttle_count) {
3475                 /* adjust cfs_rq_clock_task() */
3476                 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3477                                              cfs_rq->throttled_clock_task;
3478         }
3479 #endif
3480 
3481         return 0;
3482 }
3483 
3484 static int tg_throttle_down(struct task_group *tg, void *data)
3485 {
3486         struct rq *rq = data;
3487         struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3488 
3489         /* group is entering throttled state, stop time */
3490         if (!cfs_rq->throttle_count)
3491                 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3492         cfs_rq->throttle_count++;
3493 
3494         return 0;
3495 }
3496 
3497 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3498 {
3499         struct rq *rq = rq_of(cfs_rq);
3500         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3501         struct sched_entity *se;
3502         long task_delta, dequeue = 1;
3503         bool empty;
3504 
3505         se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3506 
3507         /* freeze hierarchy runnable averages while throttled */
3508         rcu_read_lock();
3509         walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3510         rcu_read_unlock();
3511 
3512         task_delta = cfs_rq->h_nr_running;
3513         for_each_sched_entity(se) {
3514                 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3515                 /* throttled entity or throttle-on-deactivate */
3516                 if (!se->on_rq)
3517                         break;
3518 
3519                 if (dequeue)
3520                         dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3521                 qcfs_rq->h_nr_running -= task_delta;
3522 
3523                 if (qcfs_rq->load.weight)
3524                         dequeue = 0;
3525         }
3526 
3527         if (!se)
3528                 sub_nr_running(rq, task_delta);
3529 
3530         cfs_rq->throttled = 1;
3531         cfs_rq->throttled_clock = rq_clock(rq);
3532         raw_spin_lock(&cfs_b->lock);
3533         empty = list_empty(&cfs_b->throttled_cfs_rq);
3534 
3535         /*
3536          * Add to the _head_ of the list, so that an already-started
3537          * distribute_cfs_runtime will not see us
3538          */
3539         list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3540 
3541         /*
3542          * If we're the first throttled task, make sure the bandwidth
3543          * timer is running.
3544          */
3545         if (empty)
3546                 start_cfs_bandwidth(cfs_b);
3547 
3548         raw_spin_unlock(&cfs_b->lock);
3549 }
3550 
3551 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3552 {
3553         struct rq *rq = rq_of(cfs_rq);
3554         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3555         struct sched_entity *se;
3556         int enqueue = 1;
3557         long task_delta;
3558 
3559         se = cfs_rq->tg->se[cpu_of(rq)];
3560 
3561         cfs_rq->throttled = 0;
3562 
3563         update_rq_clock(rq);
3564 
3565         raw_spin_lock(&cfs_b->lock);
3566         cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3567         list_del_rcu(&cfs_rq->throttled_list);
3568         raw_spin_unlock(&cfs_b->lock);
3569 
3570         /* update hierarchical throttle state */
3571         walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3572 
3573         if (!cfs_rq->load.weight)
3574                 return;
3575 
3576         task_delta = cfs_rq->h_nr_running;
3577         for_each_sched_entity(se) {
3578                 if (se->on_rq)
3579                         enqueue = 0;
3580 
3581                 cfs_rq = cfs_rq_of(se);
3582                 if (enqueue)
3583                         enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3584                 cfs_rq->h_nr_running += task_delta;
3585 
3586                 if (cfs_rq_throttled(cfs_rq))
3587                         break;
3588         }
3589 
3590         if (!se)
3591                 add_nr_running(rq, task_delta);
3592 
3593         /* determine whether we need to wake up potentially idle cpu */
3594         if (rq->curr == rq->idle && rq->cfs.nr_running)
3595                 resched_curr(rq);
3596 }
3597 
3598 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3599                 u64 remaining, u64 expires)
3600 {
3601         struct cfs_rq *cfs_rq;
3602         u64 runtime;
3603         u64 starting_runtime = remaining;
3604 
3605         rcu_read_lock();
3606         list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3607                                 throttled_list) {
3608                 struct rq *rq = rq_of(cfs_rq);
3609 
3610                 raw_spin_lock(&rq->lock);
3611                 if (!cfs_rq_throttled(cfs_rq))
3612                         goto next;
3613 
3614                 runtime = -cfs_rq->runtime_remaining + 1;
3615                 if (runtime > remaining)
3616                         runtime = remaining;
3617                 remaining -= runtime;
3618 
3619                 cfs_rq->runtime_remaining += runtime;
3620                 cfs_rq->runtime_expires = expires;
3621 
3622                 /* we check whether we're throttled above */
3623                 if (cfs_rq->runtime_remaining > 0)
3624                         unthrottle_cfs_rq(cfs_rq);
3625 
3626 next:
3627                 raw_spin_unlock(&rq->lock);
3628 
3629                 if (!remaining)
3630                         break;
3631         }
3632         rcu_read_unlock();
3633 
3634         return starting_runtime - remaining;
3635 }
3636 
3637 /*
3638  * Responsible for refilling a task_group's bandwidth and unthrottling its
3639  * cfs_rqs as appropriate. If there has been no activity within the last
3640  * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3641  * used to track this state.
3642  */
3643 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3644 {
3645         u64 runtime, runtime_expires;
3646         int throttled;
3647 
3648         /* no need to continue the timer with no bandwidth constraint */
3649         if (cfs_b->quota == RUNTIME_INF)
3650                 goto out_deactivate;
3651 
3652         throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3653         cfs_b->nr_periods += overrun;
3654 
3655         /*
3656          * idle depends on !throttled (for the case of a large deficit), and if
3657          * we're going inactive then everything else can be deferred
3658          */
3659         if (cfs_b->idle && !throttled)
3660                 goto out_deactivate;
3661 
3662         __refill_cfs_bandwidth_runtime(cfs_b);
3663 
3664         if (!throttled) {
3665                 /* mark as potentially idle for the upcoming period */
3666                 cfs_b->idle = 1;
3667                 return 0;
3668         }
3669 
3670         /* account preceding periods in which throttling occurred */
3671         cfs_b->nr_throttled += overrun;
3672 
3673         runtime_expires = cfs_b->runtime_expires;
3674 
3675         /*
3676          * This check is repeated as we are holding onto the new bandwidth while
3677          * we unthrottle. This can potentially race with an unthrottled group
3678          * trying to acquire new bandwidth from the global pool. This can result
3679          * in us over-using our runtime if it is all used during this loop, but
3680          * only by limited amounts in that extreme case.
3681          */
3682         while (throttled && cfs_b->runtime > 0) {
3683                 runtime = cfs_b->runtime;
3684                 raw_spin_unlock(&cfs_b->lock);
3685                 /* we can't nest cfs_b->lock while distributing bandwidth */
3686                 runtime = distribute_cfs_runtime(cfs_b, runtime,
3687                                                  runtime_expires);
3688                 raw_spin_lock(&cfs_b->lock);
3689 
3690                 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3691 
3692                 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3693         }
3694 
3695         /*
3696          * While we are ensured activity in the period following an
3697          * unthrottle, this also covers the case in which the new bandwidth is
3698          * insufficient to cover the existing bandwidth deficit.  (Forcing the
3699          * timer to remain active while there are any throttled entities.)
3700          */
3701         cfs_b->idle = 0;
3702 
3703         return 0;
3704 
3705 out_deactivate:
3706         return 1;
3707 }
3708 
3709 /* a cfs_rq won't donate quota below this amount */
3710 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3711 /* minimum remaining period time to redistribute slack quota */
3712 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3713 /* how long we wait to gather additional slack before distributing */
3714 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3715 
3716 /*
3717  * Are we near the end of the current quota period?
3718  *
3719  * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3720  * hrtimer base being cleared by hrtimer_start. In the case of
3721  * migrate_hrtimers, base is never cleared, so we are fine.
3722  */
3723 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3724 {
3725         struct hrtimer *refresh_timer = &cfs_b->period_timer;
3726         u64 remaining;
3727 
3728         /* if the call-back is running a quota refresh is already occurring */
3729         if (hrtimer_callback_running(refresh_timer))
3730                 return 1;
3731 
3732         /* is a quota refresh about to occur? */
3733         remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3734         if (remaining < min_expire)
3735                 return 1;
3736 
3737         return 0;
3738 }
3739 
3740 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3741 {
3742         u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3743 
3744         /* if there's a quota refresh soon don't bother with slack */
3745         if (runtime_refresh_within(cfs_b, min_left))
3746                 return;
3747 
3748         hrtimer_start(&cfs_b->slack_timer,
3749                         ns_to_ktime(cfs_bandwidth_slack_period),
3750                         HRTIMER_MODE_REL);
3751 }
3752 
3753 /* we know any runtime found here is valid as update_curr() precedes return */
3754 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3755 {
3756         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3757         s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3758 
3759         if (slack_runtime <= 0)
3760                 return;
3761 
3762         raw_spin_lock(&cfs_b->lock);
3763         if (cfs_b->quota != RUNTIME_INF &&
3764             cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3765                 cfs_b->runtime += slack_runtime;
3766 
3767                 /* we are under rq->lock, defer unthrottling using a timer */
3768                 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3769                     !list_empty(&cfs_b->throttled_cfs_rq))
3770                         start_cfs_slack_bandwidth(cfs_b);
3771         }
3772         raw_spin_unlock(&cfs_b->lock);
3773 
3774         /* even if it's not valid for return we don't want to try again */
3775         cfs_rq->runtime_remaining -= slack_runtime;
3776 }
3777 
3778 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3779 {
3780         if (!cfs_bandwidth_used())
3781                 return;
3782 
3783         if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3784                 return;
3785 
3786         __return_cfs_rq_runtime(cfs_rq);
3787 }
3788 
3789 /*
3790  * This is done with a timer (instead of inline with bandwidth return) since
3791  * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3792  */
3793 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3794 {
3795         u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3796         u64 expires;
3797 
3798         /* confirm we're still not at a refresh boundary */
3799         raw_spin_lock(&cfs_b->lock);
3800         if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3801                 raw_spin_unlock(&cfs_b->lock);
3802                 return;
3803         }
3804 
3805         if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3806                 runtime = cfs_b->runtime;
3807 
3808         expires = cfs_b->runtime_expires;
3809         raw_spin_unlock(&cfs_b->lock);
3810 
3811         if (!runtime)
3812                 return;
3813 
3814         runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3815 
3816         raw_spin_lock(&cfs_b->lock);
3817         if (expires == cfs_b->runtime_expires)
3818                 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3819         raw_spin_unlock(&cfs_b->lock);
3820 }
3821 
3822 /*
3823  * When a group wakes up we want to make sure that its quota is not already
3824  * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3825  * runtime as update_curr() throttling can not not trigger until it's on-rq.
3826  */
3827 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3828 {
3829         if (!cfs_bandwidth_used())
3830                 return;
3831 
3832         /* an active group must be handled by the update_curr()->put() path */
3833         if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3834                 return;
3835 
3836         /* ensure the group is not already throttled */
3837         if (cfs_rq_throttled(cfs_rq))
3838                 return;
3839 
3840         /* update runtime allocation */
3841         account_cfs_rq_runtime(cfs_rq, 0);
3842         if (cfs_rq->runtime_remaining <= 0)
3843                 throttle_cfs_rq(cfs_rq);
3844 }
3845 
3846 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3847 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3848 {
3849         if (!cfs_bandwidth_used())
3850                 return false;
3851 
3852         if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3853                 return false;
3854 
3855         /*
3856          * it's possible for a throttled entity to be forced into a running
3857          * state (e.g. set_curr_task), in this case we're finished.
3858          */
3859         if (cfs_rq_throttled(cfs_rq))
3860                 return true;
3861 
3862         throttle_cfs_rq(cfs_rq);
3863         return true;
3864 }
3865 
3866 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3867 {
3868         struct cfs_bandwidth *cfs_b =
3869                 container_of(timer, struct cfs_bandwidth, slack_timer);
3870 
3871         do_sched_cfs_slack_timer(cfs_b);
3872 
3873         return HRTIMER_NORESTART;
3874 }
3875 
3876 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3877 {
3878         struct cfs_bandwidth *cfs_b =
3879                 container_of(timer, struct cfs_bandwidth, period_timer);
3880         int overrun;
3881         int idle = 0;
3882 
3883         raw_spin_lock(&cfs_b->lock);
3884         for (;;) {
3885                 overrun = hrtimer_forward_now(timer, cfs_b->period);
3886                 if (!overrun)
3887                         break;
3888 
3889                 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3890         }
3891         if (idle)
3892                 cfs_b->period_active = 0;
3893         raw_spin_unlock(&cfs_b->lock);
3894 
3895         return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3896 }
3897 
3898 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3899 {
3900         raw_spin_lock_init(&cfs_b->lock);
3901         cfs_b->runtime = 0;
3902         cfs_b->quota = RUNTIME_INF;
3903         cfs_b->period = ns_to_ktime(default_cfs_period());
3904 
3905         INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3906         hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
3907         cfs_b->period_timer.function = sched_cfs_period_timer;
3908         hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3909         cfs_b->slack_timer.function = sched_cfs_slack_timer;
3910 }
3911 
3912 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3913 {
3914         cfs_rq->runtime_enabled = 0;
3915         INIT_LIST_HEAD(&cfs_rq->throttled_list);
3916 }
3917 
3918 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3919 {
3920         lockdep_assert_held(&cfs_b->lock);
3921 
3922         if (!cfs_b->period_active) {
3923                 cfs_b->period_active = 1;
3924                 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
3925                 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
3926         }
3927 }
3928 
3929 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3930 {
3931         /* init_cfs_bandwidth() was not called */
3932         if (!cfs_b->throttled_cfs_rq.next)
3933                 return;
3934 
3935         hrtimer_cancel(&cfs_b->period_timer);
3936         hrtimer_cancel(&cfs_b->slack_timer);
3937 }
3938 
3939 static void __maybe_unused update_runtime_enabled(struct rq *rq)
3940 {
3941         struct cfs_rq *cfs_rq;
3942 
3943         for_each_leaf_cfs_rq(rq, cfs_rq) {
3944                 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
3945 
3946                 raw_spin_lock(&cfs_b->lock);
3947                 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
3948                 raw_spin_unlock(&cfs_b->lock);
3949         }
3950 }
3951 
3952 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3953 {
3954         struct cfs_rq *cfs_rq;
3955 
3956         for_each_leaf_cfs_rq(rq, cfs_rq) {
3957                 if (!cfs_rq->runtime_enabled)
3958                         continue;
3959 
3960                 /*
3961                  * clock_task is not advancing so we just need to make sure
3962                  * there's some valid quota amount
3963                  */
3964                 cfs_rq->runtime_remaining = 1;
3965                 /*
3966                  * Offline rq is schedulable till cpu is completely disabled
3967                  * in take_cpu_down(), so we prevent new cfs throttling here.
3968                  */
3969                 cfs_rq->runtime_enabled = 0;
3970 
3971                 if (cfs_rq_throttled(cfs_rq))
3972                         unthrottle_cfs_rq(cfs_rq);
3973         }
3974 }
3975 
3976 #else /* CONFIG_CFS_BANDWIDTH */
3977 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3978 {
3979         return rq_clock_task(rq_of(cfs_rq));
3980 }
3981 
3982 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
3983 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
3984 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3985 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3986 
3987 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3988 {
3989         return 0;
3990 }
3991 
3992 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3993 {
3994         return 0;
3995 }
3996 
3997 static inline int throttled_lb_pair(struct task_group *tg,
3998                                     int src_cpu, int dest_cpu)
3999 {
4000         return 0;
4001 }
4002 
4003 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4004 
4005 #ifdef CONFIG_FAIR_GROUP_SCHED
4006 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4007 #endif
4008 
4009 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4010 {
4011         return NULL;
4012 }
4013 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4014 static inline void update_runtime_enabled(struct rq *rq) {}
4015 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4016 
4017 #endif /* CONFIG_CFS_BANDWIDTH */
4018 
4019 /**************************************************
4020  * CFS operations on tasks:
4021  */
4022 
4023 #ifdef CONFIG_SCHED_HRTICK
4024 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4025 {
4026         struct sched_entity *se = &p->se;
4027         struct cfs_rq *cfs_rq = cfs_rq_of(se);
4028 
4029         WARN_ON(task_rq(p) != rq);
4030 
4031         if (cfs_rq->nr_running > 1) {
4032                 u64 slice = sched_slice(cfs_rq, se);
4033                 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4034                 s64 delta = slice - ran;
4035 
4036                 if (delta < 0) {
4037                         if (rq->curr == p)
4038                                 resched_curr(rq);
4039                         return;
4040                 }
4041                 hrtick_start(rq, delta);
4042         }
4043 }
4044 
4045 /*
4046  * called from enqueue/dequeue and updates the hrtick when the
4047  * current task is from our class and nr_running is low enough
4048  * to matter.
4049  */
4050 static void hrtick_update(struct rq *rq)
4051 {
4052         struct task_struct *curr = rq->curr;
4053 
4054         if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4055                 return;
4056 
4057         if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4058                 hrtick_start_fair(rq, curr);
4059 }
4060 #else /* !CONFIG_SCHED_HRTICK */
4061 static inline void
4062 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4063 {
4064 }
4065 
4066 static inline void hrtick_update(struct rq *rq)
4067 {
4068 }
4069 #endif
4070 
4071 /*
4072  * The enqueue_task method is called before nr_running is
4073  * increased. Here we update the fair scheduling stats and
4074  * then put the task into the rbtree:
4075  */
4076 static void
4077 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4078 {
4079         struct cfs_rq *cfs_rq;
4080         struct sched_entity *se = &p->se;
4081 
4082         for_each_sched_entity(se) {
4083                 if (se->on_rq)
4084                         break;
4085                 cfs_rq = cfs_rq_of(se);
4086                 enqueue_entity(cfs_rq, se, flags);
4087 
4088                 /*
4089                  * end evaluation on encountering a throttled cfs_rq
4090                  *
4091                  * note: in the case of encountering a throttled cfs_rq we will
4092                  * post the final h_nr_running increment below.
4093                 */
4094                 if (cfs_rq_throttled(cfs_rq))
4095                         break;
4096                 cfs_rq->h_nr_running++;
4097 
4098                 flags = ENQUEUE_WAKEUP;
4099         }
4100 
4101         for_each_sched_entity(se) {
4102                 cfs_rq = cfs_rq_of(se);
4103                 cfs_rq->h_nr_running++;
4104 
4105                 if (cfs_rq_throttled(cfs_rq))
4106                         break;
4107 
4108                 update_load_avg(se, 1);
4109                 update_cfs_shares(cfs_rq);
4110         }
4111 
4112         if (!se)
4113                 add_nr_running(rq, 1);
4114 
4115         hrtick_update(rq);
4116 }
4117 
4118 static void set_next_buddy(struct sched_entity *se);
4119 
4120 /*
4121  * The dequeue_task method is called before nr_running is
4122  * decreased. We remove the task from the rbtree and
4123  * update the fair scheduling stats:
4124  */
4125 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4126 {
4127         struct cfs_rq *cfs_rq;
4128         struct sched_entity *se = &p->se;
4129         int task_sleep = flags & DEQUEUE_SLEEP;
4130 
4131         for_each_sched_entity(se) {
4132                 cfs_rq = cfs_rq_of(se);
4133                 dequeue_entity(cfs_rq, se, flags);
4134 
4135                 /*
4136                  * end evaluation on encountering a throttled cfs_rq
4137                  *
4138                  * note: in the case of encountering a throttled cfs_rq we will
4139                  * post the final h_nr_running decrement below.
4140                 */
4141                 if (cfs_rq_throttled(cfs_rq))
4142                         break;
4143                 cfs_rq->h_nr_running--;
4144 
4145                 /* Don't dequeue parent if it has other entities besides us */
4146                 if (cfs_rq->load.weight) {
4147                         /*
4148                          * Bias pick_next to pick a task from this cfs_rq, as
4149                          * p is sleeping when it is within its sched_slice.
4150                          */
4151                         if (task_sleep && parent_entity(se))
4152                                 set_next_buddy(parent_entity(se));
4153 
4154                         /* avoid re-evaluating load for this entity */
4155                         se = parent_entity(se);
4156                         break;
4157                 }
4158                 flags |= DEQUEUE_SLEEP;
4159         }
4160 
4161         for_each_sched_entity(se) {
4162                 cfs_rq = cfs_rq_of(se);
4163                 cfs_rq->h_nr_running--;
4164 
4165                 if (cfs_rq_throttled(cfs_rq))
4166                         break;
4167 
4168                 update_load_avg(se, 1);
4169                 update_cfs_shares(cfs_rq);
4170         }
4171 
4172         if (!se)
4173                 sub_nr_running(rq, 1);
4174 
4175         hrtick_update(rq);
4176 }
4177 
4178 #ifdef CONFIG_SMP
4179 
4180 /*
4181  * per rq 'load' arrray crap; XXX kill this.
4182  */
4183 
4184 /*
4185  * The exact cpuload at various idx values, calculated at every tick would be
4186  * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4187  *
4188  * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4189  * on nth tick when cpu may be busy, then we have:
4190  * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4191  * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4192  *
4193  * decay_load_missed() below does efficient calculation of
4194  * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4195  * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4196  *
4197  * The calculation is approximated on a 128 point scale.
4198  * degrade_zero_ticks is the number of ticks after which load at any
4199  * particular idx is approximated to be zero.
4200  * degrade_factor is a precomputed table, a row for each load idx.
4201  * Each column corresponds to degradation factor for a power of two ticks,
4202  * based on 128 point scale.
4203  * Example:
4204  * row 2, col 3 (=12) says that the degradation at load idx 2 after
4205  * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4206  *
4207  * With this power of 2 load factors, we can degrade the load n times
4208  * by looking at 1 bits in n and doing as many mult/shift instead of
4209  * n mult/shifts needed by the exact degradation.
4210  */
4211 #define DEGRADE_SHIFT           7
4212 static const unsigned char
4213                 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4214 static const unsigned char
4215                 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4216                                         {0, 0, 0, 0, 0, 0, 0, 0},
4217                                         {64, 32, 8, 0, 0, 0, 0, 0},
4218                                         {96, 72, 40, 12, 1, 0, 0},
4219                                         {112, 98, 75, 43, 15, 1, 0},
4220                                         {120, 112, 98, 76, 45, 16, 2} };
4221 
4222 /*
4223  * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4224  * would be when CPU is idle and so we just decay the old load without
4225  * adding any new load.
4226  */
4227 static unsigned long
4228 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4229 {
4230         int j = 0;
4231 
4232         if (!missed_updates)
4233                 return load;
4234 
4235         if (missed_updates >= degrade_zero_ticks[idx])
4236                 return 0;
4237 
4238         if (idx == 1)
4239                 return load >> missed_updates;
4240 
4241         while (missed_updates) {
4242                 if (missed_updates % 2)
4243                         load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4244 
4245                 missed_updates >>= 1;
4246                 j++;
4247         }
4248         return load;
4249 }
4250 
4251 /*
4252  * Update rq->cpu_load[] statistics. This function is usually called every
4253  * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4254  * every tick. We fix it up based on jiffies.
4255  */
4256 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4257                               unsigned long pending_updates)
4258 {
4259         int i, scale;
4260 
4261         this_rq->nr_load_updates++;
4262 
4263         /* Update our load: */
4264         this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4265         for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4266                 unsigned long old_load, new_load;
4267 
4268                 /* scale is effectively 1 << i now, and >> i divides by scale */
4269 
4270                 old_load = this_rq->cpu_load[i];
4271                 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4272                 new_load = this_load;
4273                 /*
4274                  * Round up the averaging division if load is increasing. This
4275                  * prevents us from getting stuck on 9 if the load is 10, for
4276                  * example.
4277                  */
4278                 if (new_load > old_load)
4279                         new_load += scale - 1;
4280 
4281                 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4282         }
4283 
4284         sched_avg_update(this_rq);
4285 }
4286 
4287 /* Used instead of source_load when we know the type == 0 */
4288 static unsigned long weighted_cpuload(const int cpu)
4289 {
4290         return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4291 }
4292 
4293 #ifdef CONFIG_NO_HZ_COMMON
4294 /*
4295  * There is no sane way to deal with nohz on smp when using jiffies because the
4296  * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4297  * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4298  *
4299  * Therefore we cannot use the delta approach from the regular tick since that
4300  * would seriously skew the load calculation. However we'll make do for those
4301  * updates happening while idle (nohz_idle_balance) or coming out of idle
4302  * (tick_nohz_idle_exit).
4303  *
4304  * This means we might still be one tick off for nohz periods.
4305  */
4306 
4307 /*
4308  * Called from nohz_idle_balance() to update the load ratings before doing the
4309  * idle balance.
4310  */
4311 static void update_idle_cpu_load(struct rq *this_rq)
4312 {
4313         unsigned long curr_jiffies = READ_ONCE(jiffies);
4314         unsigned long load = weighted_cpuload(cpu_of(this_rq));
4315         unsigned long pending_updates;
4316 
4317         /*
4318          * bail if there's load or we're actually up-to-date.
4319          */
4320         if (load || curr_jiffies == this_rq->last_load_update_tick)
4321                 return;
4322 
4323         pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4324         this_rq->last_load_update_tick = curr_jiffies;
4325 
4326         __update_cpu_load(this_rq, load, pending_updates);
4327 }
4328 
4329 /*
4330  * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4331  */
4332 void update_cpu_load_nohz(void)
4333 {
4334         struct rq *this_rq = this_rq();
4335         unsigned long curr_jiffies = READ_ONCE(jiffies);
4336         unsigned long pending_updates;
4337 
4338         if (curr_jiffies == this_rq->last_load_update_tick)
4339                 return;
4340 
4341         raw_spin_lock(&this_rq->lock);
4342         pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4343         if (pending_updates) {
4344                 this_rq->last_load_update_tick = curr_jiffies;
4345                 /*
4346                  * We were idle, this means load 0, the current load might be
4347                  * !0 due to remote wakeups and the sort.
4348                  */
4349                 __update_cpu_load(this_rq, 0, pending_updates);
4350         }
4351         raw_spin_unlock(&this_rq->lock);
4352 }
4353 #endif /* CONFIG_NO_HZ */
4354 
4355 /*
4356  * Called from scheduler_tick()
4357  */
4358 void update_cpu_load_active(struct rq *this_rq)
4359 {
4360         unsigned long load = weighted_cpuload(cpu_of(this_rq));
4361         /*
4362          * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4363          */
4364         this_rq->last_load_update_tick = jiffies;
4365         __update_cpu_load(this_rq, load, 1);
4366 }
4367 
4368 /*
4369  * Return a low guess at the load of a migration-source cpu weighted
4370  * according to the scheduling class and "nice" value.
4371  *
4372  * We want to under-estimate the load of migration sources, to
4373  * balance conservatively.
4374  */
4375 static unsigned long source_load(int cpu, int type)
4376 {
4377         struct rq *rq = cpu_rq(cpu);
4378         unsigned long total = weighted_cpuload(cpu);
4379 
4380         if (type == 0 || !sched_feat(LB_BIAS))
4381                 return total;
4382 
4383         return min(rq->cpu_load[type-1], total);
4384 }
4385 
4386 /*
4387  * Return a high guess at the load of a migration-target cpu weighted
4388  * according to the scheduling class and "nice" value.
4389  */
4390 static unsigned long target_load(int cpu, int type)
4391 {
4392         struct rq *rq = cpu_rq(cpu);
4393         unsigned long total = weighted_cpuload(cpu);
4394 
4395         if (type == 0 || !sched_feat(LB_BIAS))
4396                 return total;
4397 
4398         return max(rq->cpu_load[type-1], total);
4399 }
4400 
4401 static unsigned long capacity_of(int cpu)
4402 {
4403         return cpu_rq(cpu)->cpu_capacity;
4404 }
4405 
4406 static unsigned long capacity_orig_of(int cpu)
4407 {
4408         return cpu_rq(cpu)->cpu_capacity_orig;
4409 }
4410 
4411 static unsigned long cpu_avg_load_per_task(int cpu)
4412 {
4413         struct rq *rq = cpu_rq(cpu);
4414         unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4415         unsigned long load_avg = weighted_cpuload(cpu);
4416 
4417         if (nr_running)
4418                 return load_avg / nr_running;
4419 
4420         return 0;
4421 }
4422 
4423 static void record_wakee(struct task_struct *p)
4424 {
4425         /*
4426          * Rough decay (wiping) for cost saving, don't worry
4427          * about the boundary, really active task won't care
4428          * about the loss.
4429          */
4430         if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4431                 current->wakee_flips >>= 1;
4432                 current->wakee_flip_decay_ts = jiffies;
4433         }
4434 
4435         if (current->last_wakee != p) {
4436                 current->last_wakee = p;
4437                 current->wakee_flips++;
4438         }
4439 }
4440 
4441 static void task_waking_fair(struct task_struct *p)
4442 {
4443         struct sched_entity *se = &p->se;
4444         struct cfs_rq *cfs_rq = cfs_rq_of(se);
4445         u64 min_vruntime;
4446 
4447 #ifndef CONFIG_64BIT
4448         u64 min_vruntime_copy;
4449 
4450         do {
4451                 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4452                 smp_rmb();
4453                 min_vruntime = cfs_rq->min_vruntime;
4454         } while (min_vruntime != min_vruntime_copy);
4455 #else
4456         min_vruntime = cfs_rq->min_vruntime;
4457 #endif
4458 
4459         se->vruntime -= min_vruntime;
4460         record_wakee(p);
4461 }
4462 
4463 #ifdef CONFIG_FAIR_GROUP_SCHED
4464 /*
4465  * effective_load() calculates the load change as seen from the root_task_group
4466  *
4467  * Adding load to a group doesn't make a group heavier, but can cause movement
4468  * of group shares between cpus. Assuming the shares were perfectly aligned one
4469  * can calculate the shift in shares.
4470  *
4471  * Calculate the effective load difference if @wl is added (subtracted) to @tg
4472  * on this @cpu and results in a total addition (subtraction) of @wg to the
4473  * total group weight.
4474  *
4475  * Given a runqueue weight distribution (rw_i) we can compute a shares
4476  * distribution (s_i) using:
4477  *
4478  *   s_i = rw_i / \Sum rw_j                                             (1)
4479  *
4480  * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4481  * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4482  * shares distribution (s_i):
4483  *
4484  *   rw_i = {   2,   4,   1,   0 }
4485  *   s_i  = { 2/7, 4/7, 1/7,   0 }
4486  *
4487  * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4488  * task used to run on and the CPU the waker is running on), we need to
4489  * compute the effect of waking a task on either CPU and, in case of a sync
4490  * wakeup, compute the effect of the current task going to sleep.
4491  *
4492  * So for a change of @wl to the local @cpu with an overall group weight change
4493  * of @wl we can compute the new shares distribution (s'_i) using:
4494  *
4495  *   s'_i = (rw_i + @wl) / (@wg + \Sum rw_j)                            (2)
4496  *
4497  * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4498  * differences in waking a task to CPU 0. The additional task changes the
4499  * weight and shares distributions like:
4500  *
4501  *   rw'_i = {   3,   4,   1,   0 }
4502  *   s'_i  = { 3/8, 4/8, 1/8,   0 }
4503  *
4504  * We can then compute the difference in effective weight by using:
4505  *
4506  *   dw_i = S * (s'_i - s_i)                                            (3)
4507  *
4508  * Where 'S' is the group weight as seen by its parent.
4509  *
4510  * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4511  * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4512  * 4/7) times the weight of the group.
4513  */
4514 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4515 {
4516         struct sched_entity *se = tg->se[cpu];
4517 
4518         if (!tg->parent)        /* the trivial, non-cgroup case */
4519                 return wl;
4520 
4521         for_each_sched_entity(se) {
4522                 long w, W;
4523 
4524                 tg = se->my_q->tg;
4525 
4526                 /*
4527                  * W = @wg + \Sum rw_j
4528                  */
4529                 W = wg + calc_tg_weight(tg, se->my_q);
4530 
4531                 /*
4532                  * w = rw_i + @wl
4533                  */
4534                 w = cfs_rq_load_avg(se->my_q) + wl;
4535 
4536                 /*
4537                  * wl = S * s'_i; see (2)
4538                  */
4539                 if (W > 0 && w < W)
4540                         wl = (w * (long)tg->shares) / W;
4541                 else
4542                         wl = tg->shares;
4543 
4544                 /*
4545                  * Per the above, wl is the new se->load.weight value; since
4546                  * those are clipped to [MIN_SHARES, ...) do so now. See
4547                  * calc_cfs_shares().
4548                  */
4549                 if (wl < MIN_SHARES)
4550                         wl = MIN_SHARES;
4551 
4552                 /*
4553                  * wl = dw_i = S * (s'_i - s_i); see (3)
4554                  */
4555                 wl -= se->avg.load_avg;
4556 
4557                 /*
4558                  * Recursively apply this logic to all parent groups to compute
4559                  * the final effective load change on the root group. Since
4560                  * only the @tg group gets extra weight, all parent groups can
4561                  * only redistribute existing shares. @wl is the shift in shares
4562                  * resulting from this level per the above.
4563                  */
4564                 wg = 0;
4565         }
4566 
4567         return wl;
4568 }
4569 #else
4570 
4571 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4572 {
4573         return wl;
4574 }
4575 
4576 #endif
4577 
4578 /*
4579  * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
4580  * A waker of many should wake a different task than the one last awakened
4581  * at a frequency roughly N times higher than one of its wakees.  In order
4582  * to determine whether we should let the load spread vs consolodating to
4583  * shared cache, we look for a minimum 'flip' frequency of llc_size in one
4584  * partner, and a factor of lls_size higher frequency in the other.  With
4585  * both conditions met, we can be relatively sure that the relationship is
4586  * non-monogamous, with partner count exceeding socket size.  Waker/wakee
4587  * being client/server, worker/dispatcher, interrupt source or whatever is
4588  * irrelevant, spread criteria is apparent partner count exceeds socket size.
4589  */
4590 static int wake_wide(struct task_struct *p)
4591 {
4592         unsigned int master = current->wakee_flips;
4593         unsigned int slave = p->wakee_flips;
4594         int factor = this_cpu_read(sd_llc_size);
4595 
4596         if (master < slave)
4597                 swap(master, slave);
4598         if (slave < factor || master < slave * factor)
4599                 return 0;
4600         return 1;
4601 }
4602 
4603 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4604 {
4605         s64 this_load, load;
4606         s64 this_eff_load, prev_eff_load;
4607         int idx, this_cpu, prev_cpu;
4608         struct task_group *tg;
4609         unsigned long weight;
4610         int balanced;
4611 
4612         idx       = sd->wake_idx;
4613         this_cpu  = smp_processor_id();
4614         prev_cpu  = task_cpu(p);
4615         load      = source_load(prev_cpu, idx);
4616         this_load = target_load(this_cpu, idx);
4617 
4618         /*
4619          * If sync wakeup then subtract the (maximum possible)
4620          * effect of the currently running task from the load
4621          * of the current CPU:
4622          */
4623         if (sync) {
4624                 tg = task_group(current);
4625                 weight = current->se.avg.load_avg;
4626 
4627                 this_load += effective_load(tg, this_cpu, -weight, -weight);
4628                 load += effective_load(tg, prev_cpu, 0, -weight);
4629         }
4630 
4631         tg = task_group(p);
4632         weight = p->se.avg.load_avg;
4633 
4634         /*
4635          * In low-load situations, where prev_cpu is idle and this_cpu is idle
4636          * due to the sync cause above having dropped this_load to 0, we'll
4637          * always have an imbalance, but there's really nothing you can do
4638          * about that, so that's good too.
4639          *
4640          * Otherwise check if either cpus are near enough in load to allow this
4641          * task to be woken on this_cpu.
4642          */
4643         this_eff_load = 100;
4644         this_eff_load *= capacity_of(prev_cpu);
4645 
4646         prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4647         prev_eff_load *= capacity_of(this_cpu);
4648 
4649         if (this_load > 0) {
4650                 this_eff_load *= this_load +
4651                         effective_load(tg, this_cpu, weight, weight);
4652 
4653                 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4654         }
4655 
4656         balanced = this_eff_load <= prev_eff_load;
4657 
4658         schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4659 
4660         if (!balanced)
4661                 return 0;
4662 
4663         schedstat_inc(sd, ttwu_move_affine);
4664         schedstat_inc(p, se.statistics.nr_wakeups_affine);
4665 
4666         return 1;
4667 }
4668 
4669 /*
4670  * find_idlest_group finds and returns the least busy CPU group within the
4671  * domain.
4672  */
4673 static struct sched_group *
4674 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4675                   int this_cpu, int sd_flag)
4676 {
4677         struct sched_group *idlest = NULL, *group = sd->groups;
4678         unsigned long min_load = ULONG_MAX, this_load = 0;
4679         int load_idx = sd->forkexec_idx;
4680         int imbalance = 100 + (sd->imbalance_pct-100)/2;
4681 
4682         if (sd_flag & SD_BALANCE_WAKE)
4683                 load_idx = sd->wake_idx;
4684 
4685         do {
4686                 unsigned long load, avg_load;
4687                 int local_group;
4688                 int i;
4689 
4690                 /* Skip over this group if it has no CPUs allowed */
4691                 if (!cpumask_intersects(sched_group_cpus(group),
4692                                         tsk_cpus_allowed(p)))
4693                         continue;
4694 
4695                 local_group = cpumask_test_cpu(this_cpu,
4696                                                sched_group_cpus(group));
4697 
4698                 /* Tally up the load of all CPUs in the group */
4699                 avg_load = 0;
4700 
4701                 for_each_cpu(i, sched_group_cpus(group)) {
4702                         /* Bias balancing toward cpus of our domain */
4703                         if (local_group)
4704                                 load = source_load(i, load_idx);
4705                         else
4706                                 load = target_load(i, load_idx);
4707 
4708                         avg_load += load;
4709                 }
4710 
4711                 /* Adjust by relative CPU capacity of the group */
4712                 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
4713 
4714                 if (local_group) {
4715                         this_load = avg_load;
4716                 } else if (avg_load < min_load) {
4717                         min_load = avg_load;
4718                         idlest = group;
4719                 }
4720         } while (group = group->next, group != sd->groups);
4721 
4722         if (!idlest || 100*this_load < imbalance*min_load)
4723                 return NULL;
4724         return idlest;
4725 }
4726 
4727 /*
4728  * find_idlest_cpu - find the idlest cpu among the cpus in group.
4729  */
4730 static int
4731 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4732 {
4733         unsigned long load, min_load = ULONG_MAX;
4734         unsigned int min_exit_latency = UINT_MAX;
4735         u64 latest_idle_timestamp = 0;
4736         int least_loaded_cpu = this_cpu;
4737         int shallowest_idle_cpu = -1;
4738         int i;
4739 
4740         /* Traverse only the allowed CPUs */
4741         for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4742                 if (idle_cpu(i)) {
4743                         struct rq *rq = cpu_rq(i);
4744                         struct cpuidle_state *idle = idle_get_state(rq);
4745                         if (idle && idle->exit_latency < min_exit_latency) {
4746                                 /*
4747                                  * We give priority to a CPU whose idle state
4748                                  * has the smallest exit latency irrespective
4749                                  * of any idle timestamp.
4750                                  */
4751                                 min_exit_latency = idle->exit_latency;
4752                                 latest_idle_timestamp = rq->idle_stamp;
4753                                 shallowest_idle_cpu = i;
4754                         } else if ((!idle || idle->exit_latency == min_exit_latency) &&
4755                                    rq->idle_stamp > latest_idle_timestamp) {
4756                                 /*
4757                                  * If equal or no active idle state, then
4758                                  * the most recently idled CPU might have
4759                                  * a warmer cache.
4760                                  */
4761                                 latest_idle_timestamp = rq->idle_stamp;
4762                                 shallowest_idle_cpu = i;
4763                         }
4764                 } else if (shallowest_idle_cpu == -1) {
4765                         load = weighted_cpuload(i);
4766                         if (load < min_load || (load == min_load && i == this_cpu)) {
4767                                 min_load = load;
4768                                 least_loaded_cpu = i;
4769                         }
4770                 }
4771         }
4772 
4773         return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
4774 }
4775 
4776 /*
4777  * Try and locate an idle CPU in the sched_domain.
4778  */
4779 static int select_idle_sibling(struct task_struct *p, int target)
4780 {
4781         struct sched_domain *sd;
4782         struct sched_group *sg;
4783         int i = task_cpu(p);
4784 
4785         if (idle_cpu(target))
4786                 return target;
4787 
4788         /*
4789          * If the prevous cpu is cache affine and idle, don't be stupid.
4790          */
4791         if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4792                 return i;
4793 
4794         /*
4795          * Otherwise, iterate the domains and find an elegible idle cpu.
4796          */
4797         sd = rcu_dereference(per_cpu(sd_llc, target));
4798         for_each_lower_domain(sd) {
4799                 sg = sd->groups;
4800                 do {
4801                         if (!cpumask_intersects(sched_group_cpus(sg),
4802                                                 tsk_cpus_allowed(p)))
4803                                 goto next;
4804 
4805                         for_each_cpu(i, sched_group_cpus(sg)) {
4806                                 if (i == target || !idle_cpu(i))
4807                                         goto next;
4808                         }
4809 
4810                         target = cpumask_first_and(sched_group_cpus(sg),
4811                                         tsk_cpus_allowed(p));
4812                         goto done;
4813 next:
4814                         sg = sg->next;
4815                 } while (sg != sd->groups);
4816         }
4817 done:
4818         return target;
4819 }
4820 /*
4821  * get_cpu_usage returns the amount of capacity of a CPU that is used by CFS
4822  * tasks. The unit of the return value must be the one of capacity so we can
4823  * compare the usage with the capacity of the CPU that is available for CFS
4824  * task (ie cpu_capacity).
4825  * cfs.avg.util_avg is the sum of running time of runnable tasks on a
4826  * CPU. It represents the amount of utilization of a CPU in the range
4827  * [0..SCHED_LOAD_SCALE].  The usage of a CPU can't be higher than the full
4828  * capacity of the CPU because it's about the running time on this CPU.
4829  * Nevertheless, cfs.avg.util_avg can be higher than SCHED_LOAD_SCALE
4830  * because of unfortunate rounding in util_avg or just
4831  * after migrating tasks until the average stabilizes with the new running
4832  * time. So we need to check that the usage stays into the range
4833  * [0..cpu_capacity_orig] and cap if necessary.
4834  * Without capping the usage, a group could be seen as overloaded (CPU0 usage
4835  * at 121% + CPU1 usage at 80%) whereas CPU1 has 20% of available capacity
4836  */
4837 static int get_cpu_usage(int cpu)
4838 {
4839         unsigned long usage = cpu_rq(cpu)->cfs.avg.util_avg;
4840         unsigned long capacity = capacity_orig_of(cpu);
4841 
4842         if (usage >= SCHED_LOAD_SCALE)
4843                 return capacity;
4844 
4845         return (usage * capacity) >> SCHED_LOAD_SHIFT;
4846 }
4847 
4848 /*
4849  * select_task_rq_fair: Select target runqueue for the waking task in domains
4850  * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4851  * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4852  *
4853  * Balances load by selecting the idlest cpu in the idlest group, or under
4854  * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4855  *
4856  * Returns the target cpu number.
4857  *
4858  * preempt must be disabled.
4859  */
4860 static int
4861 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4862 {
4863         struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4864         int cpu = smp_processor_id();
4865         int new_cpu = prev_cpu;
4866         int want_affine = 0;
4867         int sync = wake_flags & WF_SYNC;
4868 
4869         if (sd_flag & SD_BALANCE_WAKE)
4870                 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
4871 
4872         rcu_read_lock();
4873         for_each_domain(cpu, tmp) {
4874                 if (!(tmp->flags & SD_LOAD_BALANCE))
4875                         break;
4876 
4877                 /*
4878                  * If both cpu and prev_cpu are part of this domain,
4879                  * cpu is a valid SD_WAKE_AFFINE target.
4880                  */
4881                 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4882                     cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4883                         affine_sd = tmp;
4884                         break;
4885                 }
4886 
4887                 if (tmp->flags & sd_flag)
4888                         sd = tmp;
4889                 else if (!want_affine)
4890                         break;
4891         }
4892 
4893         if (affine_sd) {
4894                 sd = NULL; /* Prefer wake_affine over balance flags */
4895                 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4896                         new_cpu = cpu;
4897         }
4898 
4899         if (!sd) {
4900                 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
4901                         new_cpu = select_idle_sibling(p, new_cpu);
4902 
4903         } else while (sd) {
4904                 struct sched_group *group;
4905                 int weight;
4906 
4907                 if (!(sd->flags & sd_flag)) {
4908                         sd = sd->child;
4909                         continue;
4910                 }
4911 
4912                 group = find_idlest_group(sd, p, cpu, sd_flag);
4913                 if (!group) {
4914                         sd = sd->child;
4915                         continue;
4916                 }
4917 
4918                 new_cpu = find_idlest_cpu(group, p, cpu);
4919                 if (new_cpu == -1 || new_cpu == cpu) {
4920                         /* Now try balancing at a lower domain level of cpu */
4921                         sd = sd->child;
4922                         continue;
4923                 }
4924 
4925                 /* Now try balancing at a lower domain level of new_cpu */
4926                 cpu = new_cpu;
4927                 weight = sd->span_weight;
4928                 sd = NULL;
4929                 for_each_domain(cpu, tmp) {
4930                         if (weight <= tmp->span_weight)
4931                                 break;
4932                         if (tmp->flags & sd_flag)
4933                                 sd = tmp;
4934                 }
4935                 /* while loop will break here if sd == NULL */
4936         }
4937         rcu_read_unlock();
4938 
4939         return new_cpu;
4940 }
4941 
4942 /*
4943  * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4944  * cfs_rq_of(p) references at time of call are still valid and identify the
4945  * previous cpu.  However, the caller only guarantees p->pi_lock is held; no
4946  * other assumptions, including the state of rq->lock, should be made.
4947  */
4948 static void migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4949 {
4950         /*
4951          * We are supposed to update the task to "current" time, then its up to date
4952          * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
4953          * what current time is, so simply throw away the out-of-date time. This
4954          * will result in the wakee task is less decayed, but giving the wakee more
4955          * load sounds not bad.
4956          */
4957         remove_entity_load_avg(&p->se);
4958 
4959         /* Tell new CPU we are migrated */
4960         p->se.avg.last_update_time = 0;
4961 
4962         /* We have migrated, no longer consider this task hot */
4963         p->se.exec_start = 0;
4964 }
4965 
4966 static void task_dead_fair(struct task_struct *p)
4967 {
4968         remove_entity_load_avg(&p->se);
4969 }
4970 #endif /* CONFIG_SMP */
4971 
4972 static unsigned long
4973 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4974 {
4975         unsigned long gran = sysctl_sched_wakeup_granularity;
4976 
4977         /*
4978          * Since its curr running now, convert the gran from real-time
4979          * to virtual-time in his units.
4980          *
4981          * By using 'se' instead of 'curr' we penalize light tasks, so
4982          * they get preempted easier. That is, if 'se' < 'curr' then
4983          * the resulting gran will be larger, therefore penalizing the
4984          * lighter, if otoh 'se' > 'curr' then the resulting gran will
4985          * be smaller, again penalizing the lighter task.
4986          *
4987          * This is especially important for buddies when the leftmost
4988          * task is higher priority than the buddy.
4989          */
4990         return calc_delta_fair(gran, se);
4991 }
4992 
4993 /*
4994  * Should 'se' preempt 'curr'.
4995  *
4996  *             |s1
4997  *        |s2
4998  *   |s3
4999  *         g
5000  *      |<--->|c
5001  *
5002  *  w(c, s1) = -1
5003  *  w(c, s2) =  0
5004  *  w(c, s3) =  1
5005  *
5006  */
5007 static int
5008 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5009 {
5010         s64 gran, vdiff = curr->vruntime - se->vruntime;
5011 
5012         if (vdiff <= 0)
5013                 return -1;
5014 
5015         gran = wakeup_gran(curr, se);
5016         if (vdiff > gran)
5017                 return 1;
5018 
5019         return 0;
5020 }
5021 
5022 static void set_last_buddy(struct sched_entity *se)
5023 {
5024         if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5025                 return;
5026 
5027         for_each_sched_entity(se)
5028                 cfs_rq_of(se)->last = se;
5029 }
5030 
5031 static void set_next_buddy(struct sched_entity *se)
5032 {
5033         if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5034                 return;
5035 
5036         for_each_sched_entity(se)
5037                 cfs_rq_of(se)->next = se;
5038 }
5039 
5040 static void set_skip_buddy(struct sched_entity *se)
5041 {
5042         for_each_sched_entity(se)
5043                 cfs_rq_of(se)->skip = se;
5044 }
5045 
5046 /*
5047  * Preempt the current task with a newly woken task if needed:
5048  */
5049 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5050 {
5051         struct task_struct *curr = rq->curr;
5052         struct sched_entity *se = &curr->se, *pse = &p->se;
5053         struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5054         int scale = cfs_rq->nr_running >= sched_nr_latency;
5055         int next_buddy_marked = 0;
5056 
5057         if (unlikely(se == pse))
5058                 return;
5059 
5060         /*
5061          * This is possible from callers such as attach_tasks(), in which we
5062          * unconditionally check_prempt_curr() after an enqueue (which may have
5063          * lead to a throttle).  This both saves work and prevents false
5064          * next-buddy nomination below.
5065          */
5066         if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5067                 return;
5068 
5069         if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5070                 set_next_buddy(pse);
5071                 next_buddy_marked = 1;
5072         }
5073 
5074         /*
5075          * We can come here with TIF_NEED_RESCHED already set from new task
5076          * wake up path.
5077          *
5078          * Note: this also catches the edge-case of curr being in a throttled
5079          * group (e.g. via set_curr_task), since update_curr() (in the
5080          * enqueue of curr) will have resulted in resched being set.  This
5081          * prevents us from potentially nominating it as a false LAST_BUDDY
5082          * below.
5083          */
5084         if (test_tsk_need_resched(curr))
5085                 return;
5086 
5087         /* Idle tasks are by definition preempted by non-idle tasks. */
5088         if (unlikely(curr->policy == SCHED_IDLE) &&
5089             likely(p->policy != SCHED_IDLE))
5090                 goto preempt;
5091 
5092         /*
5093          * Batch and idle tasks do not preempt non-idle tasks (their preemption
5094          * is driven by the tick):
5095          */
5096         if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5097                 return;
5098 
5099         find_matching_se(&se, &pse);
5100         update_curr(cfs_rq_of(se));
5101         BUG_ON(!pse);
5102         if (wakeup_preempt_entity(se, pse) == 1) {
5103                 /*
5104                  * Bias pick_next to pick the sched entity that is
5105                  * triggering this preemption.
5106                  */
5107                 if (!next_buddy_marked)
5108                         set_next_buddy(pse);
5109                 goto preempt;
5110         }
5111 
5112         return;
5113 
5114 preempt:
5115         resched_curr(rq);
5116         /*
5117          * Only set the backward buddy when the current task is still
5118          * on the rq. This can happen when a wakeup gets interleaved
5119          * with schedule on the ->pre_schedule() or idle_balance()
5120          * point, either of which can * drop the rq lock.
5121          *
5122          * Also, during early boot the idle thread is in the fair class,
5123          * for obvious reasons its a bad idea to schedule back to it.
5124          */
5125         if (unlikely(!se->on_rq || curr == rq->idle))
5126                 return;
5127 
5128         if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5129                 set_last_buddy(se);
5130 }
5131 
5132 static struct task_struct *
5133 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
5134 {
5135         struct cfs_rq *cfs_rq = &rq->cfs;
5136         struct sched_entity *se;
5137         struct task_struct *p;
5138         int new_tasks;
5139 
5140 again:
5141 #ifdef CONFIG_FAIR_GROUP_SCHED
5142         if (!cfs_rq->nr_running)
5143                 goto idle;
5144 
5145         if (prev->sched_class != &fair_sched_class)
5146                 goto simple;
5147 
5148         /*
5149          * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5150          * likely that a next task is from the same cgroup as the current.
5151          *
5152          * Therefore attempt to avoid putting and setting the entire cgroup
5153          * hierarchy, only change the part that actually changes.
5154          */
5155 
5156         do {
5157                 struct sched_entity *curr = cfs_rq->curr;
5158 
5159                 /*
5160                  * Since we got here without doing put_prev_entity() we also
5161                  * have to consider cfs_rq->curr. If it is still a runnable
5162                  * entity, update_curr() will update its vruntime, otherwise
5163                  * forget we've ever seen it.
5164                  */
5165                 if (curr) {
5166                         if (curr->on_rq)
5167                                 update_curr(cfs_rq);
5168                         else
5169                                 curr = NULL;
5170 
5171                         /*
5172                          * This call to check_cfs_rq_runtime() will do the
5173                          * throttle and dequeue its entity in the parent(s).
5174                          * Therefore the 'simple' nr_running test will indeed
5175                          * be correct.
5176                          */
5177                         if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5178                                 goto simple;
5179                 }
5180 
5181                 se = pick_next_entity(cfs_rq, curr);
5182                 cfs_rq = group_cfs_rq(se);
5183         } while (cfs_rq);
5184 
5185         p = task_of(se);
5186 
5187         /*
5188          * Since we haven't yet done put_prev_entity and if the selected task
5189          * is a different task than we started out with, try and touch the
5190          * least amount of cfs_rqs.
5191          */
5192         if (prev != p) {
5193                 struct sched_entity *pse = &prev->se;
5194 
5195                 while (!(cfs_rq = is_same_group(se, pse))) {
5196                         int se_depth = se->depth;
5197                         int pse_depth = pse->depth;
5198 
5199                         if (se_depth <= pse_depth) {
5200                                 put_prev_entity(cfs_rq_of(pse), pse);
5201                                 pse = parent_entity(pse);
5202                         }
5203                         if (se_depth >= pse_depth) {
5204                                 set_next_entity(cfs_rq_of(se), se);
5205                                 se = parent_entity(se);
5206                         }
5207                 }
5208 
5209                 put_prev_entity(cfs_rq, pse);
5210                 set_next_entity(cfs_rq, se);
5211         }
5212 
5213         if (hrtick_enabled(rq))
5214                 hrtick_start_fair(rq, p);
5215 
5216         return p;
5217 simple:
5218         cfs_rq = &rq->cfs;
5219 #endif
5220 
5221         if (!cfs_rq->nr_running)
5222                 goto idle;
5223 
5224         put_prev_task(rq, prev);
5225 
5226         do {
5227                 se = pick_next_entity(cfs_rq, NULL);
5228                 set_next_entity(cfs_rq, se);
5229                 cfs_rq = group_cfs_rq(se);
5230         } while (cfs_rq);
5231 
5232         p = task_of(se);
5233 
5234         if (hrtick_enabled(rq))
5235                 hrtick_start_fair(rq, p);
5236 
5237         return p;
5238 
5239 idle:
5240         /*
5241          * This is OK, because current is on_cpu, which avoids it being picked
5242          * for load-balance and preemption/IRQs are still disabled avoiding
5243          * further scheduler activity on it and we're being very careful to
5244          * re-start the picking loop.
5245          */
5246         lockdep_unpin_lock(&rq->lock);
5247         new_tasks = idle_balance(rq);
5248         lockdep_pin_lock(&rq->lock);
5249         /*
5250          * Because idle_balance() releases (and re-acquires) rq->lock, it is
5251          * possible for any higher priority task to appear. In that case we
5252          * must re-start the pick_next_entity() loop.
5253          */
5254         if (new_tasks < 0)
5255                 return RETRY_TASK;
5256 
5257         if (new_tasks > 0)
5258                 goto again;
5259 
5260         return NULL;
5261 }
5262 
5263 /*
5264  * Account for a descheduled task:
5265  */
5266 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5267 {
5268         struct sched_entity *se = &prev->se;
5269         struct cfs_rq *cfs_rq;
5270 
5271         for_each_sched_entity(se) {
5272                 cfs_rq = cfs_rq_of(se);
5273                 put_prev_entity(cfs_rq, se);
5274         }
5275 }
5276 
5277 /*
5278  * sched_yield() is very simple
5279  *
5280  * The magic of dealing with the ->skip buddy is in pick_next_entity.
5281  */
5282 static void yield_task_fair(struct rq *rq)
5283 {
5284         struct task_struct *curr = rq->curr;
5285         struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5286         struct sched_entity *se = &curr->se;
5287 
5288         /*
5289          * Are we the only task in the tree?
5290          */
5291         if (unlikely(rq->nr_running == 1))
5292                 return;
5293 
5294         clear_buddies(cfs_rq, se);
5295 
5296         if (curr->policy != SCHED_BATCH) {
5297                 update_rq_clock(rq);
5298                 /*
5299                  * Update run-time statistics of the 'current'.
5300                  */
5301                 update_curr(cfs_rq);
5302                 /*
5303                  * Tell update_rq_clock() that we've just updated,
5304                  * so we don't do microscopic update in schedule()
5305                  * and double the fastpath cost.
5306                  */
5307                 rq_clock_skip_update(rq, true);
5308         }
5309 
5310         set_skip_buddy(se);
5311 }
5312 
5313 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5314 {
5315         struct sched_entity *se = &p->se;
5316 
5317         /* throttled hierarchies are not runnable */
5318         if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5319                 return false;
5320 
5321         /* Tell the scheduler that we'd really like pse to run next. */
5322         set_next_buddy(se);
5323 
5324         yield_task_fair(rq);
5325 
5326         return true;
5327 }
5328 
5329 #ifdef CONFIG_SMP
5330 /**************************************************
5331  * Fair scheduling class load-balancing methods.
5332  *
5333  * BASICS
5334  *
5335  * The purpose of load-balancing is to achieve the same basic fairness the
5336  * per-cpu scheduler provides, namely provide a proportional amount of compute
5337  * time to each task. This is expressed in the following equation:
5338  *
5339  *   W_i,n/P_i == W_j,n/P_j for all i,j                               (1)
5340  *
5341  * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5342  * W_i,0 is defined as:
5343  *
5344  *   W_i,0 = \Sum_j w_i,j                                             (2)
5345  *
5346  * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5347  * is derived from the nice value as per prio_to_weight[].
5348  *
5349  * The weight average is an exponential decay average of the instantaneous
5350  * weight:
5351  *
5352  *   W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0               (3)
5353  *
5354  * C_i is the compute capacity of cpu i, typically it is the
5355  * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5356  * can also include other factors [XXX].
5357  *
5358  * To achieve this balance we define a measure of imbalance which follows
5359  * directly from (1):
5360  *
5361  *   imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j }    (4)
5362  *
5363  * We them move tasks around to minimize the imbalance. In the continuous
5364  * function space it is obvious this converges, in the discrete case we get
5365  * a few fun cases generally called infeasible weight scenarios.
5366  *
5367  * [XXX expand on:
5368  *     - infeasible weights;
5369  *     - local vs global optima in the discrete case. ]
5370  *
5371  *
5372  * SCHED DOMAINS
5373  *
5374  * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5375  * for all i,j solution, we create a tree of cpus that follows the hardware
5376  * topology where each level pairs two lower groups (or better). This results
5377  * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5378  * tree to only the first of the previous level and we decrease the frequency
5379  * of load-balance at each level inv. proportional to the number of cpus in
5380  * the groups.
5381  *
5382  * This yields:
5383  *
5384  *     log_2 n     1     n
5385  *   \Sum       { --- * --- * 2^i } = O(n)                            (5)
5386  *     i = 0      2^i   2^i
5387  *                               `- size of each group
5388  *         |         |     `- number of cpus doing load-balance
5389  *         |         `- freq
5390  *         `- sum over all levels
5391  *
5392  * Coupled with a limit on how many tasks we can migrate every balance pass,
5393  * this makes (5) the runtime complexity of the balancer.
5394  *
5395  * An important property here is that each CPU is still (indirectly) connected
5396  * to every other cpu in at most O(log n) steps:
5397  *
5398  * The adjacency matrix of the resulting graph is given by:
5399  *
5400  *             log_2 n     
5401  *   A_i,j = \Union     (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1)  (6)
5402  *             k = 0
5403  *
5404  * And you'll find that:
5405  *
5406  *   A^(log_2 n)_i,j != 0  for all i,j                                (7)
5407  *
5408  * Showing there's indeed a path between every cpu in at most O(log n) steps.
5409  * The task movement gives a factor of O(m), giving a convergence complexity
5410  * of:
5411  *
5412  *   O(nm log n),  n := nr_cpus, m := nr_tasks                        (8)
5413  *
5414  *
5415  * WORK CONSERVING
5416  *
5417  * In order to avoid CPUs going idle while there's still work to do, new idle
5418  * balancing is more aggressive and has the newly idle cpu iterate up the domain
5419  * tree itself instead of relying on other CPUs to bring it work.
5420  *
5421  * This adds some complexity to both (5) and (8) but it reduces the total idle
5422  * time.
5423  *
5424  * [XXX more?]
5425  *
5426  *
5427  * CGROUPS
5428  *
5429  * Cgroups make a horror show out of (2), instead of a simple sum we get:
5430  *
5431  *                                s_k,i
5432  *   W_i,0 = \Sum_j \Prod_k w_k * -----                               (9)
5433  *                                 S_k
5434  *
5435  * Where
5436  *
5437  *   s_k,i = \Sum_j w_i,j,k  and  S_k = \Sum_i s_k,i                 (10)
5438  *
5439  * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5440  *
5441  * The big problem is S_k, its a global sum needed to compute a local (W_i)
5442  * property.
5443  *
5444  * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5445  *      rewrite all of this once again.]
5446  */ 
5447 
5448 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5449 
5450 enum fbq_type { regular, remote, all };
5451 
5452 #define LBF_ALL_PINNED  0x01
5453 #define LBF_NEED_BREAK  0x02
5454 #define LBF_DST_PINNED  0x04
5455 #define LBF_SOME_PINNED 0x08
5456 
5457 struct lb_env {
5458         struct sched_domain     *sd;
5459 
5460         struct rq               *src_rq;
5461         int                     src_cpu;
5462 
5463         int                     dst_cpu;
5464         struct rq               *dst_rq;
5465 
5466         struct cpumask          *dst_grpmask;
5467         int                     new_dst_cpu;
5468         enum cpu_idle_type      idle;
5469         long                    imbalance;
5470         /* The set of CPUs under consideration for load-balancing */
5471         struct cpumask          *cpus;
5472 
5473         unsigned int            flags;
5474 
5475         unsigned int            loop;
5476         unsigned int            loop_break;
5477         unsigned int            loop_max;
5478 
5479         enum fbq_type           fbq_type;
5480         struct list_head        tasks;
5481 };
5482 
5483 /*
5484  * Is this task likely cache-hot:
5485  */
5486 static int task_hot(struct task_struct *p, struct lb_env *env)
5487 {
5488         s64 delta;
5489 
5490         lockdep_assert_held(&env->src_rq->lock);
5491 
5492         if (p->sched_class != &fair_sched_class)
5493                 return 0;
5494 
5495         if (unlikely(p->policy == SCHED_IDLE))
5496                 return 0;
5497 
5498         /*
5499          * Buddy candidates are cache hot:
5500          */
5501         if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
5502                         (&p->se == cfs_rq_of(&p->se)->next ||
5503                          &p->se == cfs_rq_of(&p->se)->last))
5504                 return 1;
5505 
5506         if (sysctl_sched_migration_cost == -1)
5507                 return 1;
5508         if (sysctl_sched_migration_cost == 0)
5509                 return 0;
5510 
5511         delta = rq_clock_task(env->src_rq) - p->se.exec_start;
5512 
5513         return delta < (s64)sysctl_sched_migration_cost;
5514 }
5515 
5516 #ifdef CONFIG_NUMA_BALANCING
5517 /*
5518  * Returns 1, if task migration degrades locality
5519  * Returns 0, if task migration improves locality i.e migration preferred.
5520  * Returns -1, if task migration is not affected by locality.
5521  */
5522 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5523 {
5524         struct numa_group *numa_group = rcu_dereference(p->numa_group);
5525         unsigned long src_faults, dst_faults;
5526         int src_nid, dst_nid;
5527 
5528         if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
5529                 return -1;
5530 
5531         if (!sched_feat(NUMA))
5532                 return -1;
5533 
5534         src_nid = cpu_to_node(env->src_cpu);
5535         dst_nid = cpu_to_node(env->dst_cpu);
5536 
5537         if (src_nid == dst_nid)
5538                 return -1;
5539 
5540         /* Migrating away from the preferred node is always bad. */
5541         if (src_nid == p->numa_preferred_nid) {
5542                 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
5543                         return 1;
5544                 else
5545                         return -1;
5546         }
5547 
5548         /* Encourage migration to the preferred node. */
5549         if (dst_nid == p->numa_preferred_nid)
5550                 return 0;
5551 
5552         if (numa_group) {
5553                 src_faults = group_faults(p, src_nid);
5554                 dst_faults = group_faults(p, dst_nid);
5555         } else {
5556                 src_faults = task_faults(p, src_nid);
5557                 dst_faults = task_faults(p, dst_nid);
5558         }
5559 
5560         return dst_faults < src_faults;
5561 }
5562 
5563 #else
5564 static inline int migrate_degrades_locality(struct task_struct *p,
5565                                              struct lb_env *env)
5566 {
5567         return -1;
5568 }
5569 #endif
5570 
5571 /*
5572  * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5573  */
5574 static
5575 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5576 {
5577         int tsk_cache_hot;
5578 
5579         lockdep_assert_held(&env->src_rq->lock);
5580 
5581         /*
5582          * We do not migrate tasks that are:
5583          * 1) throttled_lb_pair, or
5584          * 2) cannot be migrated to this CPU due to cpus_allowed, or
5585          * 3) running (obviously), or
5586          * 4) are cache-hot on their current CPU.
5587          */
5588         if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5589                 return 0;
5590 
5591         if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5592                 int cpu;
5593 
5594                 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5595 
5596                 env->flags |= LBF_SOME_PINNED;
5597 
5598                 /*
5599                  * Remember if this task can be migrated to any other cpu in
5600                  * our sched_group. We may want to revisit it if we couldn't
5601                  * meet load balance goals by pulling other tasks on src_cpu.
5602                  *
5603                  * Also avoid computing new_dst_cpu if we have already computed
5604                  * one in current iteration.
5605                  */
5606                 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5607                         return 0;
5608 
5609                 /* Prevent to re-select dst_cpu via env's cpus */
5610                 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5611                         if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5612                                 env->flags |= LBF_DST_PINNED;
5613                                 env->new_dst_cpu = cpu;
5614                                 break;
5615                         }
5616                 }
5617 
5618                 return 0;
5619         }
5620 
5621         /* Record that we found atleast one task that could run on dst_cpu */
5622         env->flags &= ~LBF_ALL_PINNED;
5623 
5624         if (task_running(env->src_rq, p)) {
5625                 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5626                 return 0;
5627         }
5628 
5629         /*
5630          * Aggressive migration if:
5631          * 1) destination numa is preferred
5632          * 2) task is cache cold, or
5633          * 3) too many balance attempts have failed.
5634          */
5635         tsk_cache_hot = migrate_degrades_locality(p, env);
5636         if (tsk_cache_hot == -1)
5637                 tsk_cache_hot = task_hot(p, env);
5638 
5639         if (tsk_cache_hot <= 0 ||
5640             env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5641                 if (tsk_cache_hot == 1) {
5642                         schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5643                         schedstat_inc(p, se.statistics.nr_forced_migrations);
5644                 }
5645                 return 1;
5646         }
5647 
5648         schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5649         return 0;
5650 }
5651 
5652 /*
5653  * detach_task() -- detach the task for the migration specified in env
5654  */
5655 static void detach_task(struct task_struct *p, struct lb_env *env)
5656 {
5657         lockdep_assert_held(&env->src_rq->lock);
5658 
5659         deactivate_task(env->src_rq, p, 0);
5660         p->on_rq = TASK_ON_RQ_MIGRATING;
5661         set_task_cpu(p, env->dst_cpu);
5662 }
5663 
5664 /*
5665  * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5666  * part of active balancing operations within "domain".
5667  *
5668  * Returns a task if successful and NULL otherwise.
5669  */
5670 static struct task_struct *detach_one_task(struct lb_env *env)
5671 {
5672         struct task_struct *p, *n;
5673 
5674         lockdep_assert_held(&env->src_rq->lock);
5675 
5676         list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5677                 if (!can_migrate_task(p, env))
5678                         continue;
5679 
5680                 detach_task(p, env);
5681 
5682                 /*
5683                  * Right now, this is only the second place where
5684                  * lb_gained[env->idle] is updated (other is detach_tasks)
5685                  * so we can safely collect stats here rather than
5686                  * inside detach_tasks().
5687                  */
5688                 schedstat_inc(env->sd, lb_gained[env->idle]);
5689                 return p;
5690         }
5691         return NULL;
5692 }
5693 
5694 static const unsigned int sched_nr_migrate_break = 32;
5695 
5696 /*
5697  * detach_tasks() -- tries to detach up to imbalance weighted load from
5698  * busiest_rq, as part of a balancing operation within domain "sd".
5699  *
5700  * Returns number of detached tasks if successful and 0 otherwise.
5701  */
5702 static int detach_tasks(struct lb_env *env)
5703 {
5704         struct list_head *tasks = &env->src_rq->cfs_tasks;
5705         struct task_struct *p;
5706         unsigned long load;
5707         int detached = 0;
5708 
5709         lockdep_assert_held(&env->src_rq->lock);
5710 
5711         if (env->imbalance <= 0)
5712                 return 0;
5713 
5714         while (!list_empty(tasks)) {
5715                 /*
5716                  * We don't want to steal all, otherwise we may be treated likewise,
5717                  * which could at worst lead to a livelock crash.
5718                  */
5719                 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
5720                         break;
5721 
5722                 p = list_first_entry(tasks, struct task_struct, se.group_node);
5723 
5724                 env->loop++;
5725                 /* We've more or less seen every task there is, call it quits */
5726                 if (env->loop > env->loop_max)
5727                         break;
5728 
5729                 /* take a breather every nr_migrate tasks */
5730                 if (env->loop > env->loop_break) {
5731                         env->loop_break += sched_nr_migrate_break;
5732                         env->flags |= LBF_NEED_BREAK;
5733                         break;
5734                 }
5735 
5736                 if (!can_migrate_task(p, env))
5737                         goto next;
5738 
5739                 load = task_h_load(p);
5740 
5741                 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5742                         goto next;
5743 
5744                 if ((load / 2) > env->imbalance)
5745                         goto next;
5746 
5747                 detach_task(p, env);
5748                 list_add(&p->se.group_node, &env->tasks);
5749 
5750                 detached++;
5751                 env->imbalance -= load;
5752 
5753 #ifdef CONFIG_PREEMPT
5754                 /*
5755                  * NEWIDLE balancing is a source of latency, so preemptible
5756                  * kernels will stop after the first task is detached to minimize
5757                  * the critical section.
5758                  */
5759                 if (env->idle == CPU_NEWLY_IDLE)
5760                         break;
5761 #endif
5762 
5763                 /*
5764                  * We only want to steal up to the prescribed amount of
5765                  * weighted load.
5766                  */
5767                 if (env->imbalance <= 0)
5768                         break;
5769 
5770                 continue;
5771 next:
5772                 list_move_tail(&p->se.group_node, tasks);
5773         }
5774 
5775         /*
5776          * Right now, this is one of only two places we collect this stat
5777          * so we can safely collect detach_one_task() stats here rather
5778          * than inside detach_one_task().
5779          */
5780         schedstat_add(env->sd, lb_gained[env->idle], detached);
5781 
5782         return detached;
5783 }
5784 
5785 /*
5786  * attach_task() -- attach the task detached by detach_task() to its new rq.
5787  */
5788 static void attach_task(struct rq *rq, struct task_struct *p)
5789 {
5790         lockdep_assert_held(&rq->lock);
5791 
5792         BUG_ON(task_rq(p) != rq);
5793         p->on_rq = TASK_ON_RQ_QUEUED;
5794         activate_task(rq, p, 0);
5795         check_preempt_curr(rq, p, 0);
5796 }
5797 
5798 /*
5799  * attach_one_task() -- attaches the task returned from detach_one_task() to
5800  * its new rq.
5801  */
5802 static void attach_one_task(struct rq *rq, struct task_struct *p)
5803 {
5804         raw_spin_lock(&rq->lock);
5805         attach_task(rq, p);
5806         raw_spin_unlock(&rq->lock);
5807 }
5808 
5809 /*
5810  * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
5811  * new rq.
5812  */
5813 static void attach_tasks(struct lb_env *env)
5814 {
5815         struct list_head *tasks = &env->tasks;
5816         struct task_struct *p;
5817 
5818         raw_spin_lock(&env->dst_rq->lock);
5819 
5820         while (!list_empty(tasks)) {
5821                 p = list_first_entry(tasks, struct task_struct, se.group_node);
5822                 list_del_init(&p->se.group_node);
5823 
5824                 attach_task(env->dst_rq, p);
5825         }
5826 
5827         raw_spin_unlock(&env->dst_rq->lock);
5828 }
5829 
5830 #ifdef CONFIG_FAIR_GROUP_SCHED
5831 static void update_blocked_averages(int cpu)
5832 {
5833         struct rq *rq = cpu_rq(cpu);
5834         struct cfs_rq *cfs_rq;
5835         unsigned long flags;
5836 
5837         raw_spin_lock_irqsave(&rq->lock, flags);
5838         update_rq_clock(rq);
5839 
5840         /*
5841          * Iterates the task_group tree in a bottom up fashion, see
5842          * list_add_leaf_cfs_rq() for details.
5843          */
5844         for_each_leaf_cfs_rq(rq, cfs_rq) {
5845                 /* throttled entities do not contribute to load */
5846                 if (throttled_hierarchy(cfs_rq))
5847                         continue;
5848 
5849                 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
5850                         update_tg_load_avg(cfs_rq, 0);
5851         }
5852         raw_spin_unlock_irqrestore(&rq->lock, flags);
5853 }
5854 
5855 /*
5856  * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5857  * This needs to be done in a top-down fashion because the load of a child
5858  * group is a fraction of its parents load.
5859  */
5860 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5861 {
5862         struct rq *rq = rq_of(cfs_rq);
5863         struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5864         unsigned long now = jiffies;
5865         unsigned long load;
5866 
5867         if (cfs_rq->last_h_load_update == now)
5868                 return;
5869 
5870         cfs_rq->h_load_next = NULL;
5871         for_each_sched_entity(se) {
5872                 cfs_rq = cfs_rq_of(se);
5873                 cfs_rq->h_load_next = se;
5874                 if (cfs_rq->last_h_load_update == now)
5875                         break;
5876         }
5877 
5878         if (!se) {
5879                 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
5880                 cfs_rq->last_h_load_update = now;
5881         }
5882 
5883         while ((se = cfs_rq->h_load_next) != NULL) {
5884                 load = cfs_rq->h_load;
5885                 load = div64_ul(load * se->avg.load_avg,
5886                         cfs_rq_load_avg(cfs_rq) + 1);
5887                 cfs_rq = group_cfs_rq(se);
5888                 cfs_rq->h_load = load;
5889                 cfs_rq->last_h_load_update = now;
5890         }
5891 }
5892 
5893 static unsigned long task_h_load(struct task_struct *p)
5894 {
5895         struct cfs_rq *cfs_rq = task_cfs_rq(p);
5896 
5897         update_cfs_rq_h_load(cfs_rq);
5898         return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
5899                         cfs_rq_load_avg(cfs_rq) + 1);
5900 }
5901 #else
5902 static inline void update_blocked_averages(int cpu)
5903 {
5904         struct rq *rq = cpu_rq(cpu);
5905         struct cfs_rq *cfs_rq = &rq->cfs;
5906         unsigned long flags;
5907 
5908         raw_spin_lock_irqsave(&rq->lock, flags);
5909         update_rq_clock(rq);
5910         update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
5911         raw_spin_unlock_irqrestore(&rq->lock, flags);
5912 }
5913 
5914 static unsigned long task_h_load(struct task_struct *p)
5915 {
5916         return p->se.avg.load_avg;
5917 }
5918 #endif
5919 
5920 /********** Helpers for find_busiest_group ************************/
5921 
5922 enum group_type {
5923         group_other = 0,
5924         group_imbalanced,
5925         group_overloaded,
5926 };
5927 
5928 /*
5929  * sg_lb_stats - stats of a sched_group required for load_balancing
5930  */
5931 struct sg_lb_stats {
5932         unsigned long avg_load; /*Avg load across the CPUs of the group */
5933         unsigned long group_load; /* Total load over the CPUs of the group */
5934         unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5935         unsigned long load_per_task;
5936         unsigned long group_capacity;
5937         unsigned long group_usage; /* Total usage of the group */
5938         unsigned int sum_nr_running; /* Nr tasks running in the group */
5939         unsigned int idle_cpus;
5940         unsigned int group_weight;
5941         enum group_type group_type;
5942         int group_no_capacity;
5943 #ifdef CONFIG_NUMA_BALANCING
5944         unsigned int nr_numa_running;
5945         unsigned int nr_preferred_running;
5946 #endif
5947 };
5948 
5949 /*
5950  * sd_lb_stats - Structure to store the statistics of a sched_domain
5951  *               during load balancing.
5952  */
5953 struct sd_lb_stats {
5954         struct sched_group *busiest;    /* Busiest group in this sd */
5955         struct sched_group *local;      /* Local group in this sd */
5956         unsigned long total_load;       /* Total load of all groups in sd */
5957         unsigned long total_capacity;   /* Total capacity of all groups in sd */
5958         unsigned long avg_load; /* Average load across all groups in sd */
5959 
5960         struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5961         struct sg_lb_stats local_stat;  /* Statistics of the local group */
5962 };
5963 
5964 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5965 {
5966         /*
5967          * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5968          * local_stat because update_sg_lb_stats() does a full clear/assignment.
5969          * We must however clear busiest_stat::avg_load because
5970          * update_sd_pick_busiest() reads this before assignment.
5971          */
5972         *sds = (struct sd_lb_stats){
5973                 .busiest = NULL,
5974                 .local = NULL,
5975                 .total_load = 0UL,
5976                 .total_capacity = 0UL,
5977                 .busiest_stat = {
5978                         .avg_load = 0UL,
5979                         .sum_nr_running = 0,
5980                         .group_type = group_other,
5981                 },
5982         };
5983 }
5984 
5985 /**
5986  * get_sd_load_idx - Obtain the load index for a given sched domain.
5987  * @sd: The sched_domain whose load_idx is to be obtained.
5988  * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5989  *
5990  * Return: The load index.
5991  */
5992 static inline int get_sd_load_idx(struct sched_domain *sd,
5993                                         enum cpu_idle_type idle)
5994 {
5995         int load_idx;
5996 
5997         switch (idle) {
5998         case CPU_NOT_IDLE:
5999                 load_idx = sd->busy_idx;
6000                 break;
6001 
6002         case CPU_NEWLY_IDLE:
6003                 load_idx = sd->newidle_idx;
6004                 break;
6005         default:
6006                 load_idx = sd->idle_idx;
6007                 break;
6008         }
6009 
6010         return load_idx;
6011 }
6012 
6013 static unsigned long default_scale_cpu_capacity(struct sched_domain *sd, int cpu)
6014 {
6015         if ((sd->flags & SD_SHARE_CPUCAPACITY) && (sd->span_weight > 1))
6016                 return sd->smt_gain / sd->span_weight;
6017 
6018         return SCHED_CAPACITY_SCALE;
6019 }
6020 
6021 unsigned long __weak arch_scale_cpu_capacity(struct sched_domain *sd, int cpu)
6022 {
6023         return default_scale_cpu_capacity(sd, cpu);
6024 }
6025 
6026 static unsigned long scale_rt_capacity(int cpu)
6027 {
6028         struct rq *rq = cpu_rq(cpu);
6029         u64 total, used, age_stamp, avg;
6030         s64 delta;
6031 
6032         /*
6033          * Since we're reading these variables without serialization make sure
6034          * we read them once before doing sanity checks on them.
6035          */
6036         age_stamp = READ_ONCE(rq->age_stamp);
6037         avg = READ_ONCE(rq->rt_avg);
6038         delta = __rq_clock_broken(rq) - age_stamp;
6039 
6040         if (unlikely(delta < 0))
6041                 delta = 0;
6042 
6043         total = sched_avg_period() + delta;
6044 
6045         used = div_u64(avg, total);
6046 
6047         if (likely(used < SCHED_CAPACITY_SCALE))
6048                 return SCHED_CAPACITY_SCALE - used;
6049 
6050         return 1;
6051 }
6052 
6053 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6054 {
6055         unsigned long capacity = SCHED_CAPACITY_SCALE;
6056         struct sched_group *sdg = sd->groups;
6057 
6058         if (sched_feat(ARCH_CAPACITY))
6059                 capacity *= arch_scale_cpu_capacity(sd, cpu);
6060         else
6061                 capacity *= default_scale_cpu_capacity(sd, cpu);
6062 
6063         capacity >>= SCHED_CAPACITY_SHIFT;
6064 
6065         cpu_rq(cpu)->cpu_capacity_orig = capacity;
6066 
6067         capacity *= scale_rt_capacity(cpu);
6068         capacity >>= SCHED_CAPACITY_SHIFT;
6069 
6070         if (!capacity)
6071                 capacity = 1;
6072 
6073         cpu_rq(cpu)->cpu_capacity = capacity;
6074         sdg->sgc->capacity = capacity;
6075 }
6076 
6077 void update_group_capacity(struct sched_domain *sd, int cpu)
6078 {
6079         struct sched_domain *child = sd->child;
6080         struct sched_group *group, *sdg = sd->groups;
6081         unsigned long capacity;
6082         unsigned long interval;
6083 
6084         interval = msecs_to_jiffies(sd->balance_interval);
6085         interval = clamp(interval, 1UL, max_load_balance_interval);
6086         sdg->sgc->next_update = jiffies + interval;
6087 
6088         if (!child) {
6089                 update_cpu_capacity(sd, cpu);
6090                 return;
6091         }
6092 
6093         capacity = 0;
6094 
6095         if (child->flags & SD_OVERLAP) {
6096                 /*
6097                  * SD_OVERLAP domains cannot assume that child groups
6098                  * span the current group.
6099                  */
6100 
6101                 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6102                         struct sched_group_capacity *sgc;
6103                         struct rq *rq = cpu_rq(cpu);
6104 
6105                         /*
6106                          * build_sched_domains() -> init_sched_groups_capacity()
6107                          * gets here before we've attached the domains to the
6108                          * runqueues.
6109                          *
6110                          * Use capacity_of(), which is set irrespective of domains
6111                          * in update_cpu_capacity().
6112                          *
6113                          * This avoids capacity from being 0 and
6114                          * causing divide-by-zero issues on boot.
6115                          */
6116                         if (unlikely(!rq->sd)) {
6117                                 capacity += capacity_of(cpu);
6118                                 continue;
6119                         }
6120 
6121                         sgc = rq->sd->groups->sgc;
6122                         capacity += sgc->capacity;
6123                 }
6124         } else  {
6125                 /*
6126                  * !SD_OVERLAP domains can assume that child groups
6127                  * span the current group.
6128                  */ 
6129 
6130                 group = child->groups;
6131                 do {
6132                         capacity += group->sgc->capacity;
6133                         group = group->next;
6134                 } while (group != child->groups);
6135         }
6136 
6137         sdg->sgc->capacity = capacity;
6138 }
6139 
6140 /*
6141  * Check whether the capacity of the rq has been noticeably reduced by side
6142  * activity. The imbalance_pct is used for the threshold.
6143  * Return true is the capacity is reduced
6144  */
6145 static inline int
6146 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6147 {
6148         return ((rq->cpu_capacity * sd->imbalance_pct) <
6149                                 (rq->cpu_capacity_orig * 100));
6150 }
6151 
6152 /*
6153  * Group imbalance indicates (and tries to solve) the problem where balancing
6154  * groups is inadequate due to tsk_cpus_allowed() constraints.
6155  *
6156  * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6157  * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6158  * Something like:
6159  *
6160  *      { 0 1 2 3 } { 4 5 6 7 }
6161  *              *     * * *
6162  *
6163  * If we were to balance group-wise we'd place two tasks in the first group and
6164  * two tasks in the second group. Clearly this is undesired as it will overload
6165  * cpu 3 and leave one of the cpus in the second group unused.
6166  *
6167  * The current solution to this issue is detecting the skew in the first group
6168  * by noticing the lower domain failed to reach balance and had difficulty
6169  * moving tasks due to affinity constraints.
6170  *
6171  * When this is so detected; this group becomes a candidate for busiest; see
6172  * update_sd_pick_busiest(). And calculate_imbalance() and
6173  * find_busiest_group() avoid some of the usual balance conditions to allow it
6174  * to create an effective group imbalance.
6175  *
6176  * This is a somewhat tricky proposition since the next run might not find the
6177  * group imbalance and decide the groups need to be balanced again. A most
6178  * subtle and fragile situation.
6179  */
6180 
6181 static inline int sg_imbalanced(struct sched_group *group)
6182 {
6183         return group->sgc->imbalance;
6184 }
6185 
6186 /*
6187  * group_has_capacity returns true if the group has spare capacity that could
6188  * be used by some tasks.
6189  * We consider that a group has spare capacity if the  * number of task is
6190  * smaller than the number of CPUs or if the usage is lower than the available
6191  * capacity for CFS tasks.
6192  * For the latter, we use a threshold to stabilize the state, to take into
6193  * account the variance of the tasks' load and to return true if the available
6194  * capacity in meaningful for the load balancer.
6195  * As an example, an available capacity of 1% can appear but it doesn't make
6196  * any benefit for the load balance.
6197  */
6198 static inline bool
6199 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6200 {
6201         if (sgs->sum_nr_running < sgs->group_weight)
6202                 return true;
6203 
6204         if ((sgs->group_capacity * 100) >
6205                         (sgs->group_usage * env->sd->imbalance_pct))
6206                 return true;
6207 
6208         return false;
6209 }
6210 
6211 /*
6212  *  group_is_overloaded returns true if the group has more tasks than it can
6213  *  handle.
6214  *  group_is_overloaded is not equals to !group_has_capacity because a group
6215  *  with the exact right number of tasks, has no more spare capacity but is not
6216  *  overloaded so both group_has_capacity and group_is_overloaded return
6217  *  false.
6218  */
6219 static inline bool
6220 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6221 {
6222         if (sgs->sum_nr_running <= sgs->group_weight)
6223                 return false;
6224 
6225         if ((sgs->group_capacity * 100) <
6226                         (sgs->group_usage * env->sd->imbalance_pct))
6227                 return true;
6228 
6229         return false;
6230 }
6231 
6232 static enum group_type group_classify(struct lb_env *env,
6233                 struct sched_group *group,
6234                 struct sg_lb_stats *sgs)
6235 {
6236         if (sgs->group_no_capacity)
6237                 return group_overloaded;
6238 
6239         if (sg_imbalanced(group))
6240                 return group_imbalanced;
6241 
6242         return group_other;
6243 }
6244 
6245 /**
6246  * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6247  * @env: The load balancing environment.
6248  * @group: sched_group whose statistics are to be updated.
6249  * @load_idx: Load index of sched_domain of this_cpu for load calc.
6250  * @local_group: Does group contain this_cpu.
6251  * @sgs: variable to hold the statistics for this group.
6252  * @overload: Indicate more than one runnable task for any CPU.
6253  */
6254 static inline void update_sg_lb_stats(struct lb_env *env,
6255                         struct sched_group *group, int load_idx,
6256                         int local_group, struct sg_lb_stats *sgs,
6257                         bool *overload)
6258 {
6259         unsigned long load;
6260         int i;
6261 
6262         memset(sgs, 0, sizeof(*sgs));
6263 
6264         for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6265                 struct rq *rq = cpu_rq(i);
6266 
6267                 /* Bias balancing toward cpus of our domain */
6268                 if (local_group)
6269                         load = target_load(i, load_idx);
6270                 else
6271                         load = source_load(i, load_idx);
6272 
6273                 sgs->group_load += load;
6274                 sgs->group_usage += get_cpu_usage(i);
6275                 sgs->sum_nr_running += rq->cfs.h_nr_running;
6276 
6277                 if (rq->nr_running > 1)
6278                         *overload = true;
6279 
6280 #ifdef CONFIG_NUMA_BALANCING
6281                 sgs->nr_numa_running += rq->nr_numa_running;
6282                 sgs->nr_preferred_running += rq->nr_preferred_running;
6283 #endif
6284                 sgs->sum_weighted_load += weighted_cpuload(i);
6285                 if (idle_cpu(i))
6286                         sgs->idle_cpus++;
6287         }
6288 
6289         /* Adjust by relative CPU capacity of the group */
6290         sgs->group_capacity = group->sgc->capacity;
6291         sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6292 
6293         if (sgs->sum_nr_running)
6294                 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6295 
6296         sgs->group_weight = group->group_weight;
6297 
6298         sgs->group_no_capacity = group_is_overloaded(env, sgs);
6299         sgs->group_type = group_classify(env, group, sgs);
6300 }
6301 
6302 /**
6303  * update_sd_pick_busiest - return 1 on busiest group
6304  * @env: The load balancing environment.
6305  * @sds: sched_domain statistics
6306  * @sg: sched_group candidate to be checked for being the busiest
6307  * @sgs: sched_group statistics
6308  *
6309  * Determine if @sg is a busier group than the previously selected
6310  * busiest group.
6311  *
6312  * Return: %true if @sg is a busier group than the previously selected
6313  * busiest group. %false otherwise.
6314  */
6315 static bool update_sd_pick_busiest(struct lb_env *env,
6316                                    struct sd_lb_stats *sds,
6317                                    struct sched_group *sg,
6318                                    struct sg_lb_stats *sgs)
6319 {
6320         struct sg_lb_stats *busiest = &sds->busiest_stat;
6321 
6322         if (sgs->group_type > busiest->group_type)
6323                 return true;
6324 
6325         if (sgs->group_type < busiest->group_type)
6326                 return false;
6327 
6328         if (sgs->avg_load <= busiest->avg_load)
6329                 return false;
6330 
6331         /* This is the busiest node in its class. */
6332         if (!(env->sd->flags & SD_ASYM_PACKING))
6333                 return true;
6334 
6335         /*
6336          * ASYM_PACKING needs to move all the work to the lowest
6337          * numbered CPUs in the group, therefore mark all groups
6338          * higher than ourself as busy.
6339          */
6340         if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6341                 if (!sds->busiest)
6342                         return true;
6343 
6344                 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
6345                         return true;
6346         }
6347 
6348         return false;
6349 }
6350 
6351 #ifdef CONFIG_NUMA_BALANCING
6352 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6353 {
6354         if (sgs->sum_nr_running > sgs->nr_numa_running)
6355                 return regular;
6356         if (sgs->sum_nr_running > sgs->nr_preferred_running)
6357                 return remote;
6358         return all;
6359 }
6360 
6361 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6362 {
6363         if (rq->nr_running > rq->nr_numa_running)
6364                 return regular;
6365         if (rq->nr_running > rq->nr_preferred_running)
6366                 return remote;
6367         return all;
6368 }
6369 #else
6370 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6371 {
6372         return all;
6373 }
6374 
6375 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6376 {
6377         return regular;
6378 }
6379 #endif /* CONFIG_NUMA_BALANCING */
6380 
6381 /**
6382  * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6383  * @env: The load balancing environment.
6384  * @sds: variable to hold the statistics for this sched_domain.
6385  */
6386 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6387 {
6388         struct sched_domain *child = env->sd->child;
6389         struct sched_group *sg = env->sd->groups;
6390         struct sg_lb_stats tmp_sgs;
6391         int load_idx, prefer_sibling = 0;
6392         bool overload = false;
6393 
6394         if (child && child->flags & SD_PREFER_SIBLING)
6395                 prefer_sibling = 1;
6396 
6397         load_idx = get_sd_load_idx(env->sd, env->idle);
6398 
6399         do {
6400                 struct sg_lb_stats *sgs = &tmp_sgs;
6401                 int local_group;
6402 
6403                 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6404                 if (local_group) {
6405                         sds->local = sg;
6406                         sgs = &sds->local_stat;
6407 
6408                         if (env->idle != CPU_NEWLY_IDLE ||
6409                             time_after_eq(jiffies, sg->sgc->next_update))
6410                                 update_group_capacity(env->sd, env->dst_cpu);
6411                 }
6412 
6413                 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6414                                                 &overload);
6415 
6416                 if (local_group)
6417                         goto next_group;
6418 
6419                 /*
6420                  * In case the child domain prefers tasks go to siblings
6421                  * first, lower the sg capacity so that we'll try
6422                  * and move all the excess tasks away. We lower the capacity
6423                  * of a group only if the local group has the capacity to fit
6424                  * these excess tasks. The extra check prevents the case where
6425                  * you always pull from the heaviest group when it is already
6426                  * under-utilized (possible with a large weight task outweighs
6427                  * the tasks on the system).
6428                  */
6429                 if (prefer_sibling && sds->local &&
6430                     group_has_capacity(env, &sds->local_stat) &&
6431                     (sgs->sum_nr_running > 1)) {
6432                         sgs->group_no_capacity = 1;
6433                         sgs->group_type = group_overloaded;
6434                 }
6435 
6436                 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6437                         sds->busiest = sg;
6438                         sds->busiest_stat = *sgs;
6439                 }
6440 
6441 next_group:
6442                 /* Now, start updating sd_lb_stats */
6443                 sds->total_load += sgs->group_load;
6444                 sds->total_capacity += sgs->group_capacity;
6445 
6446                 sg = sg->next;
6447         } while (sg != env->sd->groups);
6448 
6449         if (env->sd->flags & SD_NUMA)
6450                 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6451 
6452         if (!env->sd->parent) {
6453                 /* update overload indicator if we are at root domain */
6454                 if (env->dst_rq->rd->overload != overload)
6455                         env->dst_rq->rd->overload = overload;
6456         }
6457 
6458 }
6459 
6460 /**
6461  * check_asym_packing - Check to see if the group is packed into the
6462  *                      sched doman.
6463  *
6464  * This is primarily intended to used at the sibling level.  Some
6465  * cores like POWER7 prefer to use lower numbered SMT threads.  In the
6466  * case of POWER7, it can move to lower SMT modes only when higher
6467  * threads are idle.  When in lower SMT modes, the threads will
6468  * perform better since they share less core resources.  Hence when we
6469  * have idle threads, we want them to be the higher ones.
6470  *
6471  * This packing function is run on idle threads.  It checks to see if
6472  * the busiest CPU in this domain (core in the P7 case) has a higher
6473  * CPU number than the packing function is being run on.  Here we are
6474  * assuming lower CPU number will be equivalent to lower a SMT thread
6475  * number.
6476  *
6477  * Return: 1 when packing is required and a task should be moved to
6478  * this CPU.  The amount of the imbalance is returned in *imbalance.
6479  *
6480  * @env: The load balancing environment.
6481  * @sds: Statistics of the sched_domain which is to be packed
6482  */
6483 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6484 {
6485         int busiest_cpu;
6486 
6487         if (!(env->sd->flags & SD_ASYM_PACKING))
6488                 return 0;
6489 
6490         if (!sds->busiest)
6491                 return 0;
6492 
6493         busiest_cpu = group_first_cpu(sds->busiest);
6494         if (env->dst_cpu > busiest_cpu)
6495                 return 0;
6496 
6497         env->imbalance = DIV_ROUND_CLOSEST(
6498                 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6499                 SCHED_CAPACITY_SCALE);
6500 
6501         return 1;
6502 }
6503 
6504 /**
6505  * fix_small_imbalance - Calculate the minor imbalance that exists
6506  *                      amongst the groups of a sched_domain, during
6507  *                      load balancing.
6508  * @env: The load balancing environment.
6509  * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6510  */
6511 static inline
6512 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6513 {
6514         unsigned long tmp, capa_now = 0, capa_move = 0;
6515         unsigned int imbn = 2;
6516         unsigned long scaled_busy_load_per_task;
6517         struct sg_lb_stats *local, *busiest;
6518 
6519         local = &sds->local_stat;
6520         busiest = &sds->busiest_stat;
6521 
6522         if (!local->sum_nr_running)
6523                 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6524         else if (busiest->load_per_task > local->load_per_task)
6525                 imbn = 1;
6526 
6527         scaled_busy_load_per_task =
6528                 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6529                 busiest->group_capacity;
6530 
6531         if (busiest->avg_load + scaled_busy_load_per_task >=
6532             local->avg_load + (scaled_busy_load_per_task * imbn)) {
6533                 env->imbalance = busiest->load_per_task;
6534                 return;
6535         }
6536 
6537         /*
6538          * OK, we don't have enough imbalance to justify moving tasks,
6539          * however we may be able to increase total CPU capacity used by
6540          * moving them.
6541          */
6542 
6543         capa_now += busiest->group_capacity *
6544                         min(busiest->load_per_task, busiest->avg_load);
6545         capa_now += local->group_capacity *
6546                         min(local->load_per_task, local->avg_load);
6547         capa_now /= SCHED_CAPACITY_SCALE;
6548 
6549         /* Amount of load we'd subtract */
6550         if (busiest->avg_load > scaled_busy_load_per_task) {
6551                 capa_move += busiest->group_capacity *
6552                             min(busiest->load_per_task,
6553                                 busiest->avg_load - scaled_busy_load_per_task);
6554         }
6555 
6556         /* Amount of load we'd add */
6557         if (busiest->avg_load * busiest->group_capacity <
6558             busiest->load_per_task * SCHED_CAPACITY_SCALE) {
6559                 tmp = (busiest->avg_load * busiest->group_capacity) /
6560                       local->group_capacity;
6561         } else {
6562                 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6563                       local->group_capacity;
6564         }
6565         capa_move += local->group_capacity *
6566                     min(local->load_per_task, local->avg_load + tmp);
6567         capa_move /= SCHED_CAPACITY_SCALE;
6568 
6569         /* Move if we gain throughput */
6570         if (capa_move > capa_now)
6571                 env->imbalance = busiest->load_per_task;
6572 }
6573 
6574 /**
6575  * calculate_imbalance - Calculate the amount of imbalance present within the
6576  *                       groups of a given sched_domain during load balance.
6577  * @env: load balance environment
6578  * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6579  */
6580 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6581 {
6582         unsigned long max_pull, load_above_capacity = ~0UL;
6583         struct sg_lb_stats *local, *busiest;
6584 
6585         local = &sds->local_stat;
6586         busiest = &sds->busiest_stat;
6587 
6588         if (busiest->group_type == group_imbalanced) {
6589                 /*
6590                  * In the group_imb case we cannot rely on group-wide averages
6591                  * to ensure cpu-load equilibrium, look at wider averages. XXX
6592                  */
6593                 busiest->load_per_task =
6594                         min(busiest->load_per_task, sds->avg_load);
6595         }
6596 
6597         /*
6598          * In the presence of smp nice balancing, certain scenarios can have
6599          * max load less than avg load(as we skip the groups at or below
6600          * its cpu_capacity, while calculating max_load..)
6601          */
6602         if (busiest->avg_load <= sds->avg_load ||
6603             local->avg_load >= sds->avg_load) {
6604                 env->imbalance = 0;
6605                 return fix_small_imbalance(env, sds);
6606         }
6607 
6608         /*
6609          * If there aren't any idle cpus, avoid creating some.
6610          */
6611         if (busiest->group_type == group_overloaded &&
6612             local->group_type   == group_overloaded) {
6613                 load_above_capacity = busiest->sum_nr_running *
6614                                         SCHED_LOAD_SCALE;
6615                 if (load_above_capacity > busiest->group_capacity)
6616                         load_above_capacity -= busiest->group_capacity;
6617                 else
6618                         load_above_capacity = ~0UL;
6619         }
6620 
6621         /*
6622          * We're trying to get all the cpus to the average_load, so we don't
6623          * want to push ourselves above the average load, nor do we wish to
6624          * reduce the max loaded cpu below the average load. At the same time,
6625          * we also don't want to reduce the group load below the group capacity
6626          * (so that we can implement power-savings policies etc). Thus we look
6627          * for the minimum possible imbalance.
6628          */
6629         max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6630 
6631         /* How much load to actually move to equalise the imbalance */
6632         env->imbalance = min(
6633                 max_pull * busiest->group_capacity,
6634                 (sds->avg_load - local->avg_load) * local->group_capacity
6635         ) / SCHED_CAPACITY_SCALE;
6636 
6637         /*
6638          * if *imbalance is less than the average load per runnable task
6639          * there is no guarantee that any tasks will be moved so we'll have
6640          * a think about bumping its value to force at least one task to be
6641          * moved
6642          */
6643         if (env->imbalance < busiest->load_per_task)
6644                 return fix_small_imbalance(env, sds);
6645 }
6646 
6647 /******* find_busiest_group() helpers end here *********************/
6648 
6649 /**
6650  * find_busiest_group - Returns the busiest group within the sched_domain
6651  * if there is an imbalance. If there isn't an imbalance, and