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

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

Version: ~ [ linux-5.12-rc1 ] ~ [ linux-5.11.2 ] ~ [ linux-5.10.19 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.101 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.177 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.222 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.258 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.258 ] ~ [ 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 // SPDX-License-Identifier: GPL-2.0
  2 /*
  3  * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
  4  *
  5  *  Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
  6  *
  7  *  Interactivity improvements by Mike Galbraith
  8  *  (C) 2007 Mike Galbraith <efault@gmx.de>
  9  *
 10  *  Various enhancements by Dmitry Adamushko.
 11  *  (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
 12  *
 13  *  Group scheduling enhancements by Srivatsa Vaddagiri
 14  *  Copyright IBM Corporation, 2007
 15  *  Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
 16  *
 17  *  Scaled math optimizations by Thomas Gleixner
 18  *  Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
 19  *
 20  *  Adaptive scheduling granularity, math enhancements by Peter Zijlstra
 21  *  Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
 22  */
 23 #include "sched.h"
 24 
 25 #include <trace/events/sched.h>
 26 
 27 /*
 28  * Targeted preemption latency for CPU-bound tasks:
 29  *
 30  * NOTE: this latency value is not the same as the concept of
 31  * 'timeslice length' - timeslices in CFS are of variable length
 32  * and have no persistent notion like in traditional, time-slice
 33  * based scheduling concepts.
 34  *
 35  * (to see the precise effective timeslice length of your workload,
 36  *  run vmstat and monitor the context-switches (cs) field)
 37  *
 38  * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
 39  */
 40 unsigned int sysctl_sched_latency                       = 6000000ULL;
 41 static unsigned int normalized_sysctl_sched_latency     = 6000000ULL;
 42 
 43 /*
 44  * The initial- and re-scaling of tunables is configurable
 45  *
 46  * Options are:
 47  *
 48  *   SCHED_TUNABLESCALING_NONE - unscaled, always *1
 49  *   SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
 50  *   SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
 51  *
 52  * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
 53  */
 54 enum sched_tunable_scaling sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
 55 
 56 /*
 57  * Minimal preemption granularity for CPU-bound tasks:
 58  *
 59  * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
 60  */
 61 unsigned int sysctl_sched_min_granularity                       = 750000ULL;
 62 static unsigned int normalized_sysctl_sched_min_granularity     = 750000ULL;
 63 
 64 /*
 65  * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
 66  */
 67 static unsigned int sched_nr_latency = 8;
 68 
 69 /*
 70  * After fork, child runs first. If set to 0 (default) then
 71  * parent will (try to) run first.
 72  */
 73 unsigned int sysctl_sched_child_runs_first __read_mostly;
 74 
 75 /*
 76  * SCHED_OTHER wake-up granularity.
 77  *
 78  * This option delays the preemption effects of decoupled workloads
 79  * and reduces their over-scheduling. Synchronous workloads will still
 80  * have immediate wakeup/sleep latencies.
 81  *
 82  * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
 83  */
 84 unsigned int sysctl_sched_wakeup_granularity                    = 1000000UL;
 85 static unsigned int normalized_sysctl_sched_wakeup_granularity  = 1000000UL;
 86 
 87 const_debug unsigned int sysctl_sched_migration_cost    = 500000UL;
 88 
 89 #ifdef CONFIG_SMP
 90 /*
 91  * For asym packing, by default the lower numbered CPU has higher priority.
 92  */
 93 int __weak arch_asym_cpu_priority(int cpu)
 94 {
 95         return -cpu;
 96 }
 97 
 98 /*
 99  * The margin used when comparing utilization with CPU capacity:
100  * util * margin < capacity * 1024
101  *
102  * (default: ~20%)
103  */
104 static unsigned int capacity_margin                     = 1280;
105 #endif
106 
107 #ifdef CONFIG_CFS_BANDWIDTH
108 /*
109  * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
110  * each time a cfs_rq requests quota.
111  *
112  * Note: in the case that the slice exceeds the runtime remaining (either due
113  * to consumption or the quota being specified to be smaller than the slice)
114  * we will always only issue the remaining available time.
115  *
116  * (default: 5 msec, units: microseconds)
117  */
118 unsigned int sysctl_sched_cfs_bandwidth_slice           = 5000UL;
119 #endif
120 
121 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
122 {
123         lw->weight += inc;
124         lw->inv_weight = 0;
125 }
126 
127 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
128 {
129         lw->weight -= dec;
130         lw->inv_weight = 0;
131 }
132 
133 static inline void update_load_set(struct load_weight *lw, unsigned long w)
134 {
135         lw->weight = w;
136         lw->inv_weight = 0;
137 }
138 
139 /*
140  * Increase the granularity value when there are more CPUs,
141  * because with more CPUs the 'effective latency' as visible
142  * to users decreases. But the relationship is not linear,
143  * so pick a second-best guess by going with the log2 of the
144  * number of CPUs.
145  *
146  * This idea comes from the SD scheduler of Con Kolivas:
147  */
148 static unsigned int get_update_sysctl_factor(void)
149 {
150         unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
151         unsigned int factor;
152 
153         switch (sysctl_sched_tunable_scaling) {
154         case SCHED_TUNABLESCALING_NONE:
155                 factor = 1;
156                 break;
157         case SCHED_TUNABLESCALING_LINEAR:
158                 factor = cpus;
159                 break;
160         case SCHED_TUNABLESCALING_LOG:
161         default:
162                 factor = 1 + ilog2(cpus);
163                 break;
164         }
165 
166         return factor;
167 }
168 
169 static void update_sysctl(void)
170 {
171         unsigned int factor = get_update_sysctl_factor();
172 
173 #define SET_SYSCTL(name) \
174         (sysctl_##name = (factor) * normalized_sysctl_##name)
175         SET_SYSCTL(sched_min_granularity);
176         SET_SYSCTL(sched_latency);
177         SET_SYSCTL(sched_wakeup_granularity);
178 #undef SET_SYSCTL
179 }
180 
181 void sched_init_granularity(void)
182 {
183         update_sysctl();
184 }
185 
186 #define WMULT_CONST     (~0U)
187 #define WMULT_SHIFT     32
188 
189 static void __update_inv_weight(struct load_weight *lw)
190 {
191         unsigned long w;
192 
193         if (likely(lw->inv_weight))
194                 return;
195 
196         w = scale_load_down(lw->weight);
197 
198         if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
199                 lw->inv_weight = 1;
200         else if (unlikely(!w))
201                 lw->inv_weight = WMULT_CONST;
202         else
203                 lw->inv_weight = WMULT_CONST / w;
204 }
205 
206 /*
207  * delta_exec * weight / lw.weight
208  *   OR
209  * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
210  *
211  * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
212  * we're guaranteed shift stays positive because inv_weight is guaranteed to
213  * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
214  *
215  * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
216  * weight/lw.weight <= 1, and therefore our shift will also be positive.
217  */
218 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
219 {
220         u64 fact = scale_load_down(weight);
221         int shift = WMULT_SHIFT;
222 
223         __update_inv_weight(lw);
224 
225         if (unlikely(fact >> 32)) {
226                 while (fact >> 32) {
227                         fact >>= 1;
228                         shift--;
229                 }
230         }
231 
232         /* hint to use a 32x32->64 mul */
233         fact = (u64)(u32)fact * lw->inv_weight;
234 
235         while (fact >> 32) {
236                 fact >>= 1;
237                 shift--;
238         }
239 
240         return mul_u64_u32_shr(delta_exec, fact, shift);
241 }
242 
243 
244 const struct sched_class fair_sched_class;
245 
246 /**************************************************************
247  * CFS operations on generic schedulable entities:
248  */
249 
250 #ifdef CONFIG_FAIR_GROUP_SCHED
251 static inline struct task_struct *task_of(struct sched_entity *se)
252 {
253         SCHED_WARN_ON(!entity_is_task(se));
254         return container_of(se, struct task_struct, se);
255 }
256 
257 /* Walk up scheduling entities hierarchy */
258 #define for_each_sched_entity(se) \
259                 for (; se; se = se->parent)
260 
261 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
262 {
263         return p->se.cfs_rq;
264 }
265 
266 /* runqueue on which this entity is (to be) queued */
267 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
268 {
269         return se->cfs_rq;
270 }
271 
272 /* runqueue "owned" by this group */
273 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
274 {
275         return grp->my_q;
276 }
277 
278 static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
279 {
280         if (!path)
281                 return;
282 
283         if (cfs_rq && task_group_is_autogroup(cfs_rq->tg))
284                 autogroup_path(cfs_rq->tg, path, len);
285         else if (cfs_rq && cfs_rq->tg->css.cgroup)
286                 cgroup_path(cfs_rq->tg->css.cgroup, path, len);
287         else
288                 strlcpy(path, "(null)", len);
289 }
290 
291 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
292 {
293         struct rq *rq = rq_of(cfs_rq);
294         int cpu = cpu_of(rq);
295 
296         if (cfs_rq->on_list)
297                 return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
298 
299         cfs_rq->on_list = 1;
300 
301         /*
302          * Ensure we either appear before our parent (if already
303          * enqueued) or force our parent to appear after us when it is
304          * enqueued. The fact that we always enqueue bottom-up
305          * reduces this to two cases and a special case for the root
306          * cfs_rq. Furthermore, it also means that we will always reset
307          * tmp_alone_branch either when the branch is connected
308          * to a tree or when we reach the top of the tree
309          */
310         if (cfs_rq->tg->parent &&
311             cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
312                 /*
313                  * If parent is already on the list, we add the child
314                  * just before. Thanks to circular linked property of
315                  * the list, this means to put the child at the tail
316                  * of the list that starts by parent.
317                  */
318                 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
319                         &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
320                 /*
321                  * The branch is now connected to its tree so we can
322                  * reset tmp_alone_branch to the beginning of the
323                  * list.
324                  */
325                 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
326                 return true;
327         }
328 
329         if (!cfs_rq->tg->parent) {
330                 /*
331                  * cfs rq without parent should be put
332                  * at the tail of the list.
333                  */
334                 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
335                         &rq->leaf_cfs_rq_list);
336                 /*
337                  * We have reach the top of a tree so we can reset
338                  * tmp_alone_branch to the beginning of the list.
339                  */
340                 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
341                 return true;
342         }
343 
344         /*
345          * The parent has not already been added so we want to
346          * make sure that it will be put after us.
347          * tmp_alone_branch points to the begin of the branch
348          * where we will add parent.
349          */
350         list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
351         /*
352          * update tmp_alone_branch to points to the new begin
353          * of the branch
354          */
355         rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
356         return false;
357 }
358 
359 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
360 {
361         if (cfs_rq->on_list) {
362                 struct rq *rq = rq_of(cfs_rq);
363 
364                 /*
365                  * With cfs_rq being unthrottled/throttled during an enqueue,
366                  * it can happen the tmp_alone_branch points the a leaf that
367                  * we finally want to del. In this case, tmp_alone_branch moves
368                  * to the prev element but it will point to rq->leaf_cfs_rq_list
369                  * at the end of the enqueue.
370                  */
371                 if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
372                         rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
373 
374                 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
375                 cfs_rq->on_list = 0;
376         }
377 }
378 
379 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
380 {
381         SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
382 }
383 
384 /* Iterate thr' all leaf cfs_rq's on a runqueue */
385 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos)                      \
386         list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list,    \
387                                  leaf_cfs_rq_list)
388 
389 /* Do the two (enqueued) entities belong to the same group ? */
390 static inline struct cfs_rq *
391 is_same_group(struct sched_entity *se, struct sched_entity *pse)
392 {
393         if (se->cfs_rq == pse->cfs_rq)
394                 return se->cfs_rq;
395 
396         return NULL;
397 }
398 
399 static inline struct sched_entity *parent_entity(struct sched_entity *se)
400 {
401         return se->parent;
402 }
403 
404 static void
405 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
406 {
407         int se_depth, pse_depth;
408 
409         /*
410          * preemption test can be made between sibling entities who are in the
411          * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
412          * both tasks until we find their ancestors who are siblings of common
413          * parent.
414          */
415 
416         /* First walk up until both entities are at same depth */
417         se_depth = (*se)->depth;
418         pse_depth = (*pse)->depth;
419 
420         while (se_depth > pse_depth) {
421                 se_depth--;
422                 *se = parent_entity(*se);
423         }
424 
425         while (pse_depth > se_depth) {
426                 pse_depth--;
427                 *pse = parent_entity(*pse);
428         }
429 
430         while (!is_same_group(*se, *pse)) {
431                 *se = parent_entity(*se);
432                 *pse = parent_entity(*pse);
433         }
434 }
435 
436 #else   /* !CONFIG_FAIR_GROUP_SCHED */
437 
438 static inline struct task_struct *task_of(struct sched_entity *se)
439 {
440         return container_of(se, struct task_struct, se);
441 }
442 
443 #define for_each_sched_entity(se) \
444                 for (; se; se = NULL)
445 
446 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
447 {
448         return &task_rq(p)->cfs;
449 }
450 
451 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
452 {
453         struct task_struct *p = task_of(se);
454         struct rq *rq = task_rq(p);
455 
456         return &rq->cfs;
457 }
458 
459 /* runqueue "owned" by this group */
460 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
461 {
462         return NULL;
463 }
464 
465 static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
466 {
467         if (path)
468                 strlcpy(path, "(null)", len);
469 }
470 
471 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
472 {
473         return true;
474 }
475 
476 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
477 {
478 }
479 
480 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
481 {
482 }
483 
484 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos)      \
485                 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
486 
487 static inline struct sched_entity *parent_entity(struct sched_entity *se)
488 {
489         return NULL;
490 }
491 
492 static inline void
493 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
494 {
495 }
496 
497 #endif  /* CONFIG_FAIR_GROUP_SCHED */
498 
499 static __always_inline
500 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
501 
502 /**************************************************************
503  * Scheduling class tree data structure manipulation methods:
504  */
505 
506 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
507 {
508         s64 delta = (s64)(vruntime - max_vruntime);
509         if (delta > 0)
510                 max_vruntime = vruntime;
511 
512         return max_vruntime;
513 }
514 
515 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
516 {
517         s64 delta = (s64)(vruntime - min_vruntime);
518         if (delta < 0)
519                 min_vruntime = vruntime;
520 
521         return min_vruntime;
522 }
523 
524 static inline int entity_before(struct sched_entity *a,
525                                 struct sched_entity *b)
526 {
527         return (s64)(a->vruntime - b->vruntime) < 0;
528 }
529 
530 static void update_min_vruntime(struct cfs_rq *cfs_rq)
531 {
532         struct sched_entity *curr = cfs_rq->curr;
533         struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
534 
535         u64 vruntime = cfs_rq->min_vruntime;
536 
537         if (curr) {
538                 if (curr->on_rq)
539                         vruntime = curr->vruntime;
540                 else
541                         curr = NULL;
542         }
543 
544         if (leftmost) { /* non-empty tree */
545                 struct sched_entity *se;
546                 se = rb_entry(leftmost, struct sched_entity, run_node);
547 
548                 if (!curr)
549                         vruntime = se->vruntime;
550                 else
551                         vruntime = min_vruntime(vruntime, se->vruntime);
552         }
553 
554         /* ensure we never gain time by being placed backwards. */
555         cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
556 #ifndef CONFIG_64BIT
557         smp_wmb();
558         cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
559 #endif
560 }
561 
562 /*
563  * Enqueue an entity into the rb-tree:
564  */
565 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
566 {
567         struct rb_node **link = &cfs_rq->tasks_timeline.rb_root.rb_node;
568         struct rb_node *parent = NULL;
569         struct sched_entity *entry;
570         bool leftmost = true;
571 
572         /*
573          * Find the right place in the rbtree:
574          */
575         while (*link) {
576                 parent = *link;
577                 entry = rb_entry(parent, struct sched_entity, run_node);
578                 /*
579                  * We dont care about collisions. Nodes with
580                  * the same key stay together.
581                  */
582                 if (entity_before(se, entry)) {
583                         link = &parent->rb_left;
584                 } else {
585                         link = &parent->rb_right;
586                         leftmost = false;
587                 }
588         }
589 
590         rb_link_node(&se->run_node, parent, link);
591         rb_insert_color_cached(&se->run_node,
592                                &cfs_rq->tasks_timeline, leftmost);
593 }
594 
595 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
596 {
597         rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
598 }
599 
600 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
601 {
602         struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
603 
604         if (!left)
605                 return NULL;
606 
607         return rb_entry(left, struct sched_entity, run_node);
608 }
609 
610 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
611 {
612         struct rb_node *next = rb_next(&se->run_node);
613 
614         if (!next)
615                 return NULL;
616 
617         return rb_entry(next, struct sched_entity, run_node);
618 }
619 
620 #ifdef CONFIG_SCHED_DEBUG
621 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
622 {
623         struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
624 
625         if (!last)
626                 return NULL;
627 
628         return rb_entry(last, struct sched_entity, run_node);
629 }
630 
631 /**************************************************************
632  * Scheduling class statistics methods:
633  */
634 
635 int sched_proc_update_handler(struct ctl_table *table, int write,
636                 void __user *buffer, size_t *lenp,
637                 loff_t *ppos)
638 {
639         int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
640         unsigned int factor = get_update_sysctl_factor();
641 
642         if (ret || !write)
643                 return ret;
644 
645         sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
646                                         sysctl_sched_min_granularity);
647 
648 #define WRT_SYSCTL(name) \
649         (normalized_sysctl_##name = sysctl_##name / (factor))
650         WRT_SYSCTL(sched_min_granularity);
651         WRT_SYSCTL(sched_latency);
652         WRT_SYSCTL(sched_wakeup_granularity);
653 #undef WRT_SYSCTL
654 
655         return 0;
656 }
657 #endif
658 
659 /*
660  * delta /= w
661  */
662 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
663 {
664         if (unlikely(se->load.weight != NICE_0_LOAD))
665                 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
666 
667         return delta;
668 }
669 
670 /*
671  * The idea is to set a period in which each task runs once.
672  *
673  * When there are too many tasks (sched_nr_latency) we have to stretch
674  * this period because otherwise the slices get too small.
675  *
676  * p = (nr <= nl) ? l : l*nr/nl
677  */
678 static u64 __sched_period(unsigned long nr_running)
679 {
680         if (unlikely(nr_running > sched_nr_latency))
681                 return nr_running * sysctl_sched_min_granularity;
682         else
683                 return sysctl_sched_latency;
684 }
685 
686 /*
687  * We calculate the wall-time slice from the period by taking a part
688  * proportional to the weight.
689  *
690  * s = p*P[w/rw]
691  */
692 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
693 {
694         u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
695 
696         for_each_sched_entity(se) {
697                 struct load_weight *load;
698                 struct load_weight lw;
699 
700                 cfs_rq = cfs_rq_of(se);
701                 load = &cfs_rq->load;
702 
703                 if (unlikely(!se->on_rq)) {
704                         lw = cfs_rq->load;
705 
706                         update_load_add(&lw, se->load.weight);
707                         load = &lw;
708                 }
709                 slice = __calc_delta(slice, se->load.weight, load);
710         }
711         return slice;
712 }
713 
714 /*
715  * We calculate the vruntime slice of a to-be-inserted task.
716  *
717  * vs = s/w
718  */
719 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
720 {
721         return calc_delta_fair(sched_slice(cfs_rq, se), se);
722 }
723 
724 #include "pelt.h"
725 #ifdef CONFIG_SMP
726 
727 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
728 static unsigned long task_h_load(struct task_struct *p);
729 static unsigned long capacity_of(int cpu);
730 
731 /* Give new sched_entity start runnable values to heavy its load in infant time */
732 void init_entity_runnable_average(struct sched_entity *se)
733 {
734         struct sched_avg *sa = &se->avg;
735 
736         memset(sa, 0, sizeof(*sa));
737 
738         /*
739          * Tasks are initialized with full load to be seen as heavy tasks until
740          * they get a chance to stabilize to their real load level.
741          * Group entities are initialized with zero load to reflect the fact that
742          * nothing has been attached to the task group yet.
743          */
744         if (entity_is_task(se))
745                 sa->runnable_load_avg = sa->load_avg = scale_load_down(se->load.weight);
746 
747         se->runnable_weight = se->load.weight;
748 
749         /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
750 }
751 
752 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
753 static void attach_entity_cfs_rq(struct sched_entity *se);
754 
755 /*
756  * With new tasks being created, their initial util_avgs are extrapolated
757  * based on the cfs_rq's current util_avg:
758  *
759  *   util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
760  *
761  * However, in many cases, the above util_avg does not give a desired
762  * value. Moreover, the sum of the util_avgs may be divergent, such
763  * as when the series is a harmonic series.
764  *
765  * To solve this problem, we also cap the util_avg of successive tasks to
766  * only 1/2 of the left utilization budget:
767  *
768  *   util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
769  *
770  * where n denotes the nth task and cpu_scale the CPU capacity.
771  *
772  * For example, for a CPU with 1024 of capacity, a simplest series from
773  * the beginning would be like:
774  *
775  *  task  util_avg: 512, 256, 128,  64,  32,   16,    8, ...
776  * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
777  *
778  * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
779  * if util_avg > util_avg_cap.
780  */
781 void post_init_entity_util_avg(struct task_struct *p)
782 {
783         struct sched_entity *se = &p->se;
784         struct cfs_rq *cfs_rq = cfs_rq_of(se);
785         struct sched_avg *sa = &se->avg;
786         long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)));
787         long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
788 
789         if (cap > 0) {
790                 if (cfs_rq->avg.util_avg != 0) {
791                         sa->util_avg  = cfs_rq->avg.util_avg * se->load.weight;
792                         sa->util_avg /= (cfs_rq->avg.load_avg + 1);
793 
794                         if (sa->util_avg > cap)
795                                 sa->util_avg = cap;
796                 } else {
797                         sa->util_avg = cap;
798                 }
799         }
800 
801         if (p->sched_class != &fair_sched_class) {
802                 /*
803                  * For !fair tasks do:
804                  *
805                 update_cfs_rq_load_avg(now, cfs_rq);
806                 attach_entity_load_avg(cfs_rq, se, 0);
807                 switched_from_fair(rq, p);
808                  *
809                  * such that the next switched_to_fair() has the
810                  * expected state.
811                  */
812                 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
813                 return;
814         }
815 
816         attach_entity_cfs_rq(se);
817 }
818 
819 #else /* !CONFIG_SMP */
820 void init_entity_runnable_average(struct sched_entity *se)
821 {
822 }
823 void post_init_entity_util_avg(struct task_struct *p)
824 {
825 }
826 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
827 {
828 }
829 #endif /* CONFIG_SMP */
830 
831 /*
832  * Update the current task's runtime statistics.
833  */
834 static void update_curr(struct cfs_rq *cfs_rq)
835 {
836         struct sched_entity *curr = cfs_rq->curr;
837         u64 now = rq_clock_task(rq_of(cfs_rq));
838         u64 delta_exec;
839 
840         if (unlikely(!curr))
841                 return;
842 
843         delta_exec = now - curr->exec_start;
844         if (unlikely((s64)delta_exec <= 0))
845                 return;
846 
847         curr->exec_start = now;
848 
849         schedstat_set(curr->statistics.exec_max,
850                       max(delta_exec, curr->statistics.exec_max));
851 
852         curr->sum_exec_runtime += delta_exec;
853         schedstat_add(cfs_rq->exec_clock, delta_exec);
854 
855         curr->vruntime += calc_delta_fair(delta_exec, curr);
856         update_min_vruntime(cfs_rq);
857 
858         if (entity_is_task(curr)) {
859                 struct task_struct *curtask = task_of(curr);
860 
861                 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
862                 cgroup_account_cputime(curtask, delta_exec);
863                 account_group_exec_runtime(curtask, delta_exec);
864         }
865 
866         account_cfs_rq_runtime(cfs_rq, delta_exec);
867 }
868 
869 static void update_curr_fair(struct rq *rq)
870 {
871         update_curr(cfs_rq_of(&rq->curr->se));
872 }
873 
874 static inline void
875 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
876 {
877         u64 wait_start, prev_wait_start;
878 
879         if (!schedstat_enabled())
880                 return;
881 
882         wait_start = rq_clock(rq_of(cfs_rq));
883         prev_wait_start = schedstat_val(se->statistics.wait_start);
884 
885         if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
886             likely(wait_start > prev_wait_start))
887                 wait_start -= prev_wait_start;
888 
889         __schedstat_set(se->statistics.wait_start, wait_start);
890 }
891 
892 static inline void
893 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
894 {
895         struct task_struct *p;
896         u64 delta;
897 
898         if (!schedstat_enabled())
899                 return;
900 
901         delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
902 
903         if (entity_is_task(se)) {
904                 p = task_of(se);
905                 if (task_on_rq_migrating(p)) {
906                         /*
907                          * Preserve migrating task's wait time so wait_start
908                          * time stamp can be adjusted to accumulate wait time
909                          * prior to migration.
910                          */
911                         __schedstat_set(se->statistics.wait_start, delta);
912                         return;
913                 }
914                 trace_sched_stat_wait(p, delta);
915         }
916 
917         __schedstat_set(se->statistics.wait_max,
918                       max(schedstat_val(se->statistics.wait_max), delta));
919         __schedstat_inc(se->statistics.wait_count);
920         __schedstat_add(se->statistics.wait_sum, delta);
921         __schedstat_set(se->statistics.wait_start, 0);
922 }
923 
924 static inline void
925 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
926 {
927         struct task_struct *tsk = NULL;
928         u64 sleep_start, block_start;
929 
930         if (!schedstat_enabled())
931                 return;
932 
933         sleep_start = schedstat_val(se->statistics.sleep_start);
934         block_start = schedstat_val(se->statistics.block_start);
935 
936         if (entity_is_task(se))
937                 tsk = task_of(se);
938 
939         if (sleep_start) {
940                 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
941 
942                 if ((s64)delta < 0)
943                         delta = 0;
944 
945                 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
946                         __schedstat_set(se->statistics.sleep_max, delta);
947 
948                 __schedstat_set(se->statistics.sleep_start, 0);
949                 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
950 
951                 if (tsk) {
952                         account_scheduler_latency(tsk, delta >> 10, 1);
953                         trace_sched_stat_sleep(tsk, delta);
954                 }
955         }
956         if (block_start) {
957                 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
958 
959                 if ((s64)delta < 0)
960                         delta = 0;
961 
962                 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
963                         __schedstat_set(se->statistics.block_max, delta);
964 
965                 __schedstat_set(se->statistics.block_start, 0);
966                 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
967 
968                 if (tsk) {
969                         if (tsk->in_iowait) {
970                                 __schedstat_add(se->statistics.iowait_sum, delta);
971                                 __schedstat_inc(se->statistics.iowait_count);
972                                 trace_sched_stat_iowait(tsk, delta);
973                         }
974 
975                         trace_sched_stat_blocked(tsk, delta);
976 
977                         /*
978                          * Blocking time is in units of nanosecs, so shift by
979                          * 20 to get a milliseconds-range estimation of the
980                          * amount of time that the task spent sleeping:
981                          */
982                         if (unlikely(prof_on == SLEEP_PROFILING)) {
983                                 profile_hits(SLEEP_PROFILING,
984                                                 (void *)get_wchan(tsk),
985                                                 delta >> 20);
986                         }
987                         account_scheduler_latency(tsk, delta >> 10, 0);
988                 }
989         }
990 }
991 
992 /*
993  * Task is being enqueued - update stats:
994  */
995 static inline void
996 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
997 {
998         if (!schedstat_enabled())
999                 return;
1000 
1001         /*
1002          * Are we enqueueing a waiting task? (for current tasks
1003          * a dequeue/enqueue event is a NOP)
1004          */
1005         if (se != cfs_rq->curr)
1006                 update_stats_wait_start(cfs_rq, se);
1007 
1008         if (flags & ENQUEUE_WAKEUP)
1009                 update_stats_enqueue_sleeper(cfs_rq, se);
1010 }
1011 
1012 static inline void
1013 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1014 {
1015 
1016         if (!schedstat_enabled())
1017                 return;
1018 
1019         /*
1020          * Mark the end of the wait period if dequeueing a
1021          * waiting task:
1022          */
1023         if (se != cfs_rq->curr)
1024                 update_stats_wait_end(cfs_rq, se);
1025 
1026         if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1027                 struct task_struct *tsk = task_of(se);
1028 
1029                 if (tsk->state & TASK_INTERRUPTIBLE)
1030                         __schedstat_set(se->statistics.sleep_start,
1031                                       rq_clock(rq_of(cfs_rq)));
1032                 if (tsk->state & TASK_UNINTERRUPTIBLE)
1033                         __schedstat_set(se->statistics.block_start,
1034                                       rq_clock(rq_of(cfs_rq)));
1035         }
1036 }
1037 
1038 /*
1039  * We are picking a new current task - update its stats:
1040  */
1041 static inline void
1042 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1043 {
1044         /*
1045          * We are starting a new run period:
1046          */
1047         se->exec_start = rq_clock_task(rq_of(cfs_rq));
1048 }
1049 
1050 /**************************************************
1051  * Scheduling class queueing methods:
1052  */
1053 
1054 #ifdef CONFIG_NUMA_BALANCING
1055 /*
1056  * Approximate time to scan a full NUMA task in ms. The task scan period is
1057  * calculated based on the tasks virtual memory size and
1058  * numa_balancing_scan_size.
1059  */
1060 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1061 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1062 
1063 /* Portion of address space to scan in MB */
1064 unsigned int sysctl_numa_balancing_scan_size = 256;
1065 
1066 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1067 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1068 
1069 struct numa_group {
1070         refcount_t refcount;
1071 
1072         spinlock_t lock; /* nr_tasks, tasks */
1073         int nr_tasks;
1074         pid_t gid;
1075         int active_nodes;
1076 
1077         struct rcu_head rcu;
1078         unsigned long total_faults;
1079         unsigned long max_faults_cpu;
1080         /*
1081          * Faults_cpu is used to decide whether memory should move
1082          * towards the CPU. As a consequence, these stats are weighted
1083          * more by CPU use than by memory faults.
1084          */
1085         unsigned long *faults_cpu;
1086         unsigned long faults[0];
1087 };
1088 
1089 /*
1090  * For functions that can be called in multiple contexts that permit reading
1091  * ->numa_group (see struct task_struct for locking rules).
1092  */
1093 static struct numa_group *deref_task_numa_group(struct task_struct *p)
1094 {
1095         return rcu_dereference_check(p->numa_group, p == current ||
1096                 (lockdep_is_held(&task_rq(p)->lock) && !READ_ONCE(p->on_cpu)));
1097 }
1098 
1099 static struct numa_group *deref_curr_numa_group(struct task_struct *p)
1100 {
1101         return rcu_dereference_protected(p->numa_group, p == current);
1102 }
1103 
1104 static inline unsigned long group_faults_priv(struct numa_group *ng);
1105 static inline unsigned long group_faults_shared(struct numa_group *ng);
1106 
1107 static unsigned int task_nr_scan_windows(struct task_struct *p)
1108 {
1109         unsigned long rss = 0;
1110         unsigned long nr_scan_pages;
1111 
1112         /*
1113          * Calculations based on RSS as non-present and empty pages are skipped
1114          * by the PTE scanner and NUMA hinting faults should be trapped based
1115          * on resident pages
1116          */
1117         nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1118         rss = get_mm_rss(p->mm);
1119         if (!rss)
1120                 rss = nr_scan_pages;
1121 
1122         rss = round_up(rss, nr_scan_pages);
1123         return rss / nr_scan_pages;
1124 }
1125 
1126 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1127 #define MAX_SCAN_WINDOW 2560
1128 
1129 static unsigned int task_scan_min(struct task_struct *p)
1130 {
1131         unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1132         unsigned int scan, floor;
1133         unsigned int windows = 1;
1134 
1135         if (scan_size < MAX_SCAN_WINDOW)
1136                 windows = MAX_SCAN_WINDOW / scan_size;
1137         floor = 1000 / windows;
1138 
1139         scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1140         return max_t(unsigned int, floor, scan);
1141 }
1142 
1143 static unsigned int task_scan_start(struct task_struct *p)
1144 {
1145         unsigned long smin = task_scan_min(p);
1146         unsigned long period = smin;
1147         struct numa_group *ng;
1148 
1149         /* Scale the maximum scan period with the amount of shared memory. */
1150         rcu_read_lock();
1151         ng = rcu_dereference(p->numa_group);
1152         if (ng) {
1153                 unsigned long shared = group_faults_shared(ng);
1154                 unsigned long private = group_faults_priv(ng);
1155 
1156                 period *= refcount_read(&ng->refcount);
1157                 period *= shared + 1;
1158                 period /= private + shared + 1;
1159         }
1160         rcu_read_unlock();
1161 
1162         return max(smin, period);
1163 }
1164 
1165 static unsigned int task_scan_max(struct task_struct *p)
1166 {
1167         unsigned long smin = task_scan_min(p);
1168         unsigned long smax;
1169         struct numa_group *ng;
1170 
1171         /* Watch for min being lower than max due to floor calculations */
1172         smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1173 
1174         /* Scale the maximum scan period with the amount of shared memory. */
1175         ng = deref_curr_numa_group(p);
1176         if (ng) {
1177                 unsigned long shared = group_faults_shared(ng);
1178                 unsigned long private = group_faults_priv(ng);
1179                 unsigned long period = smax;
1180 
1181                 period *= refcount_read(&ng->refcount);
1182                 period *= shared + 1;
1183                 period /= private + shared + 1;
1184 
1185                 smax = max(smax, period);
1186         }
1187 
1188         return max(smin, smax);
1189 }
1190 
1191 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
1192 {
1193         int mm_users = 0;
1194         struct mm_struct *mm = p->mm;
1195 
1196         if (mm) {
1197                 mm_users = atomic_read(&mm->mm_users);
1198                 if (mm_users == 1) {
1199                         mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1200                         mm->numa_scan_seq = 0;
1201                 }
1202         }
1203         p->node_stamp                   = 0;
1204         p->numa_scan_seq                = mm ? mm->numa_scan_seq : 0;
1205         p->numa_scan_period             = sysctl_numa_balancing_scan_delay;
1206         p->numa_work.next               = &p->numa_work;
1207         p->numa_faults                  = NULL;
1208         RCU_INIT_POINTER(p->numa_group, NULL);
1209         p->last_task_numa_placement     = 0;
1210         p->last_sum_exec_runtime        = 0;
1211 
1212         /* New address space, reset the preferred nid */
1213         if (!(clone_flags & CLONE_VM)) {
1214                 p->numa_preferred_nid = NUMA_NO_NODE;
1215                 return;
1216         }
1217 
1218         /*
1219          * New thread, keep existing numa_preferred_nid which should be copied
1220          * already by arch_dup_task_struct but stagger when scans start.
1221          */
1222         if (mm) {
1223                 unsigned int delay;
1224 
1225                 delay = min_t(unsigned int, task_scan_max(current),
1226                         current->numa_scan_period * mm_users * NSEC_PER_MSEC);
1227                 delay += 2 * TICK_NSEC;
1228                 p->node_stamp = delay;
1229         }
1230 }
1231 
1232 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1233 {
1234         rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
1235         rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1236 }
1237 
1238 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1239 {
1240         rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1241         rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1242 }
1243 
1244 /* Shared or private faults. */
1245 #define NR_NUMA_HINT_FAULT_TYPES 2
1246 
1247 /* Memory and CPU locality */
1248 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1249 
1250 /* Averaged statistics, and temporary buffers. */
1251 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1252 
1253 pid_t task_numa_group_id(struct task_struct *p)
1254 {
1255         struct numa_group *ng;
1256         pid_t gid = 0;
1257 
1258         rcu_read_lock();
1259         ng = rcu_dereference(p->numa_group);
1260         if (ng)
1261                 gid = ng->gid;
1262         rcu_read_unlock();
1263 
1264         return gid;
1265 }
1266 
1267 /*
1268  * The averaged statistics, shared & private, memory & CPU,
1269  * occupy the first half of the array. The second half of the
1270  * array is for current counters, which are averaged into the
1271  * first set by task_numa_placement.
1272  */
1273 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1274 {
1275         return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1276 }
1277 
1278 static inline unsigned long task_faults(struct task_struct *p, int nid)
1279 {
1280         if (!p->numa_faults)
1281                 return 0;
1282 
1283         return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1284                 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1285 }
1286 
1287 static inline unsigned long group_faults(struct task_struct *p, int nid)
1288 {
1289         struct numa_group *ng = deref_task_numa_group(p);
1290 
1291         if (!ng)
1292                 return 0;
1293 
1294         return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1295                 ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1296 }
1297 
1298 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1299 {
1300         return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1301                 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1302 }
1303 
1304 static inline unsigned long group_faults_priv(struct numa_group *ng)
1305 {
1306         unsigned long faults = 0;
1307         int node;
1308 
1309         for_each_online_node(node) {
1310                 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1311         }
1312 
1313         return faults;
1314 }
1315 
1316 static inline unsigned long group_faults_shared(struct numa_group *ng)
1317 {
1318         unsigned long faults = 0;
1319         int node;
1320 
1321         for_each_online_node(node) {
1322                 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1323         }
1324 
1325         return faults;
1326 }
1327 
1328 /*
1329  * A node triggering more than 1/3 as many NUMA faults as the maximum is
1330  * considered part of a numa group's pseudo-interleaving set. Migrations
1331  * between these nodes are slowed down, to allow things to settle down.
1332  */
1333 #define ACTIVE_NODE_FRACTION 3
1334 
1335 static bool numa_is_active_node(int nid, struct numa_group *ng)
1336 {
1337         return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1338 }
1339 
1340 /* Handle placement on systems where not all nodes are directly connected. */
1341 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1342                                         int maxdist, bool task)
1343 {
1344         unsigned long score = 0;
1345         int node;
1346 
1347         /*
1348          * All nodes are directly connected, and the same distance
1349          * from each other. No need for fancy placement algorithms.
1350          */
1351         if (sched_numa_topology_type == NUMA_DIRECT)
1352                 return 0;
1353 
1354         /*
1355          * This code is called for each node, introducing N^2 complexity,
1356          * which should be ok given the number of nodes rarely exceeds 8.
1357          */
1358         for_each_online_node(node) {
1359                 unsigned long faults;
1360                 int dist = node_distance(nid, node);
1361 
1362                 /*
1363                  * The furthest away nodes in the system are not interesting
1364                  * for placement; nid was already counted.
1365                  */
1366                 if (dist == sched_max_numa_distance || node == nid)
1367                         continue;
1368 
1369                 /*
1370                  * On systems with a backplane NUMA topology, compare groups
1371                  * of nodes, and move tasks towards the group with the most
1372                  * memory accesses. When comparing two nodes at distance
1373                  * "hoplimit", only nodes closer by than "hoplimit" are part
1374                  * of each group. Skip other nodes.
1375                  */
1376                 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1377                                         dist >= maxdist)
1378                         continue;
1379 
1380                 /* Add up the faults from nearby nodes. */
1381                 if (task)
1382                         faults = task_faults(p, node);
1383                 else
1384                         faults = group_faults(p, node);
1385 
1386                 /*
1387                  * On systems with a glueless mesh NUMA topology, there are
1388                  * no fixed "groups of nodes". Instead, nodes that are not
1389                  * directly connected bounce traffic through intermediate
1390                  * nodes; a numa_group can occupy any set of nodes.
1391                  * The further away a node is, the less the faults count.
1392                  * This seems to result in good task placement.
1393                  */
1394                 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1395                         faults *= (sched_max_numa_distance - dist);
1396                         faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1397                 }
1398 
1399                 score += faults;
1400         }
1401 
1402         return score;
1403 }
1404 
1405 /*
1406  * These return the fraction of accesses done by a particular task, or
1407  * task group, on a particular numa node.  The group weight is given a
1408  * larger multiplier, in order to group tasks together that are almost
1409  * evenly spread out between numa nodes.
1410  */
1411 static inline unsigned long task_weight(struct task_struct *p, int nid,
1412                                         int dist)
1413 {
1414         unsigned long faults, total_faults;
1415 
1416         if (!p->numa_faults)
1417                 return 0;
1418 
1419         total_faults = p->total_numa_faults;
1420 
1421         if (!total_faults)
1422                 return 0;
1423 
1424         faults = task_faults(p, nid);
1425         faults += score_nearby_nodes(p, nid, dist, true);
1426 
1427         return 1000 * faults / total_faults;
1428 }
1429 
1430 static inline unsigned long group_weight(struct task_struct *p, int nid,
1431                                          int dist)
1432 {
1433         struct numa_group *ng = deref_task_numa_group(p);
1434         unsigned long faults, total_faults;
1435 
1436         if (!ng)
1437                 return 0;
1438 
1439         total_faults = ng->total_faults;
1440 
1441         if (!total_faults)
1442                 return 0;
1443 
1444         faults = group_faults(p, nid);
1445         faults += score_nearby_nodes(p, nid, dist, false);
1446 
1447         return 1000 * faults / total_faults;
1448 }
1449 
1450 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1451                                 int src_nid, int dst_cpu)
1452 {
1453         struct numa_group *ng = deref_curr_numa_group(p);
1454         int dst_nid = cpu_to_node(dst_cpu);
1455         int last_cpupid, this_cpupid;
1456 
1457         this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1458         last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1459 
1460         /*
1461          * Allow first faults or private faults to migrate immediately early in
1462          * the lifetime of a task. The magic number 4 is based on waiting for
1463          * two full passes of the "multi-stage node selection" test that is
1464          * executed below.
1465          */
1466         if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1467             (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1468                 return true;
1469 
1470         /*
1471          * Multi-stage node selection is used in conjunction with a periodic
1472          * migration fault to build a temporal task<->page relation. By using
1473          * a two-stage filter we remove short/unlikely relations.
1474          *
1475          * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1476          * a task's usage of a particular page (n_p) per total usage of this
1477          * page (n_t) (in a given time-span) to a probability.
1478          *
1479          * Our periodic faults will sample this probability and getting the
1480          * same result twice in a row, given these samples are fully
1481          * independent, is then given by P(n)^2, provided our sample period
1482          * is sufficiently short compared to the usage pattern.
1483          *
1484          * This quadric squishes small probabilities, making it less likely we
1485          * act on an unlikely task<->page relation.
1486          */
1487         if (!cpupid_pid_unset(last_cpupid) &&
1488                                 cpupid_to_nid(last_cpupid) != dst_nid)
1489                 return false;
1490 
1491         /* Always allow migrate on private faults */
1492         if (cpupid_match_pid(p, last_cpupid))
1493                 return true;
1494 
1495         /* A shared fault, but p->numa_group has not been set up yet. */
1496         if (!ng)
1497                 return true;
1498 
1499         /*
1500          * Destination node is much more heavily used than the source
1501          * node? Allow migration.
1502          */
1503         if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1504                                         ACTIVE_NODE_FRACTION)
1505                 return true;
1506 
1507         /*
1508          * Distribute memory according to CPU & memory use on each node,
1509          * with 3/4 hysteresis to avoid unnecessary memory migrations:
1510          *
1511          * faults_cpu(dst)   3   faults_cpu(src)
1512          * --------------- * - > ---------------
1513          * faults_mem(dst)   4   faults_mem(src)
1514          */
1515         return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1516                group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1517 }
1518 
1519 static unsigned long cpu_runnable_load(struct rq *rq);
1520 
1521 /* Cached statistics for all CPUs within a node */
1522 struct numa_stats {
1523         unsigned long load;
1524 
1525         /* Total compute capacity of CPUs on a node */
1526         unsigned long compute_capacity;
1527 };
1528 
1529 /*
1530  * XXX borrowed from update_sg_lb_stats
1531  */
1532 static void update_numa_stats(struct numa_stats *ns, int nid)
1533 {
1534         int cpu;
1535 
1536         memset(ns, 0, sizeof(*ns));
1537         for_each_cpu(cpu, cpumask_of_node(nid)) {
1538                 struct rq *rq = cpu_rq(cpu);
1539 
1540                 ns->load += cpu_runnable_load(rq);
1541                 ns->compute_capacity += capacity_of(cpu);
1542         }
1543 
1544 }
1545 
1546 struct task_numa_env {
1547         struct task_struct *p;
1548 
1549         int src_cpu, src_nid;
1550         int dst_cpu, dst_nid;
1551 
1552         struct numa_stats src_stats, dst_stats;
1553 
1554         int imbalance_pct;
1555         int dist;
1556 
1557         struct task_struct *best_task;
1558         long best_imp;
1559         int best_cpu;
1560 };
1561 
1562 static void task_numa_assign(struct task_numa_env *env,
1563                              struct task_struct *p, long imp)
1564 {
1565         struct rq *rq = cpu_rq(env->dst_cpu);
1566 
1567         /* Bail out if run-queue part of active NUMA balance. */
1568         if (xchg(&rq->numa_migrate_on, 1))
1569                 return;
1570 
1571         /*
1572          * Clear previous best_cpu/rq numa-migrate flag, since task now
1573          * found a better CPU to move/swap.
1574          */
1575         if (env->best_cpu != -1) {
1576                 rq = cpu_rq(env->best_cpu);
1577                 WRITE_ONCE(rq->numa_migrate_on, 0);
1578         }
1579 
1580         if (env->best_task)
1581                 put_task_struct(env->best_task);
1582         if (p)
1583                 get_task_struct(p);
1584 
1585         env->best_task = p;
1586         env->best_imp = imp;
1587         env->best_cpu = env->dst_cpu;
1588 }
1589 
1590 static bool load_too_imbalanced(long src_load, long dst_load,
1591                                 struct task_numa_env *env)
1592 {
1593         long imb, old_imb;
1594         long orig_src_load, orig_dst_load;
1595         long src_capacity, dst_capacity;
1596 
1597         /*
1598          * The load is corrected for the CPU capacity available on each node.
1599          *
1600          * src_load        dst_load
1601          * ------------ vs ---------
1602          * src_capacity    dst_capacity
1603          */
1604         src_capacity = env->src_stats.compute_capacity;
1605         dst_capacity = env->dst_stats.compute_capacity;
1606 
1607         imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1608 
1609         orig_src_load = env->src_stats.load;
1610         orig_dst_load = env->dst_stats.load;
1611 
1612         old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1613 
1614         /* Would this change make things worse? */
1615         return (imb > old_imb);
1616 }
1617 
1618 /*
1619  * Maximum NUMA importance can be 1998 (2*999);
1620  * SMALLIMP @ 30 would be close to 1998/64.
1621  * Used to deter task migration.
1622  */
1623 #define SMALLIMP        30
1624 
1625 /*
1626  * This checks if the overall compute and NUMA accesses of the system would
1627  * be improved if the source tasks was migrated to the target dst_cpu taking
1628  * into account that it might be best if task running on the dst_cpu should
1629  * be exchanged with the source task
1630  */
1631 static void task_numa_compare(struct task_numa_env *env,
1632                               long taskimp, long groupimp, bool maymove)
1633 {
1634         struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
1635         struct rq *dst_rq = cpu_rq(env->dst_cpu);
1636         long imp = p_ng ? groupimp : taskimp;
1637         struct task_struct *cur;
1638         long src_load, dst_load;
1639         int dist = env->dist;
1640         long moveimp = imp;
1641         long load;
1642 
1643         if (READ_ONCE(dst_rq->numa_migrate_on))
1644                 return;
1645 
1646         rcu_read_lock();
1647         cur = task_rcu_dereference(&dst_rq->curr);
1648         if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1649                 cur = NULL;
1650 
1651         /*
1652          * Because we have preemption enabled we can get migrated around and
1653          * end try selecting ourselves (current == env->p) as a swap candidate.
1654          */
1655         if (cur == env->p)
1656                 goto unlock;
1657 
1658         if (!cur) {
1659                 if (maymove && moveimp >= env->best_imp)
1660                         goto assign;
1661                 else
1662                         goto unlock;
1663         }
1664 
1665         /*
1666          * "imp" is the fault differential for the source task between the
1667          * source and destination node. Calculate the total differential for
1668          * the source task and potential destination task. The more negative
1669          * the value is, the more remote accesses that would be expected to
1670          * be incurred if the tasks were swapped.
1671          */
1672         /* Skip this swap candidate if cannot move to the source cpu */
1673         if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
1674                 goto unlock;
1675 
1676         /*
1677          * If dst and source tasks are in the same NUMA group, or not
1678          * in any group then look only at task weights.
1679          */
1680         cur_ng = rcu_dereference(cur->numa_group);
1681         if (cur_ng == p_ng) {
1682                 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1683                       task_weight(cur, env->dst_nid, dist);
1684                 /*
1685                  * Add some hysteresis to prevent swapping the
1686                  * tasks within a group over tiny differences.
1687                  */
1688                 if (cur_ng)
1689                         imp -= imp / 16;
1690         } else {
1691                 /*
1692                  * Compare the group weights. If a task is all by itself
1693                  * (not part of a group), use the task weight instead.
1694                  */
1695                 if (cur_ng && p_ng)
1696                         imp += group_weight(cur, env->src_nid, dist) -
1697                                group_weight(cur, env->dst_nid, dist);
1698                 else
1699                         imp += task_weight(cur, env->src_nid, dist) -
1700                                task_weight(cur, env->dst_nid, dist);
1701         }
1702 
1703         if (maymove && moveimp > imp && moveimp > env->best_imp) {
1704                 imp = moveimp;
1705                 cur = NULL;
1706                 goto assign;
1707         }
1708 
1709         /*
1710          * If the NUMA importance is less than SMALLIMP,
1711          * task migration might only result in ping pong
1712          * of tasks and also hurt performance due to cache
1713          * misses.
1714          */
1715         if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
1716                 goto unlock;
1717 
1718         /*
1719          * In the overloaded case, try and keep the load balanced.
1720          */
1721         load = task_h_load(env->p) - task_h_load(cur);
1722         if (!load)
1723                 goto assign;
1724 
1725         dst_load = env->dst_stats.load + load;
1726         src_load = env->src_stats.load - load;
1727 
1728         if (load_too_imbalanced(src_load, dst_load, env))
1729                 goto unlock;
1730 
1731 assign:
1732         /*
1733          * One idle CPU per node is evaluated for a task numa move.
1734          * Call select_idle_sibling to maybe find a better one.
1735          */
1736         if (!cur) {
1737                 /*
1738                  * select_idle_siblings() uses an per-CPU cpumask that
1739                  * can be used from IRQ context.
1740                  */
1741                 local_irq_disable();
1742                 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1743                                                    env->dst_cpu);
1744                 local_irq_enable();
1745         }
1746 
1747         task_numa_assign(env, cur, imp);
1748 unlock:
1749         rcu_read_unlock();
1750 }
1751 
1752 static void task_numa_find_cpu(struct task_numa_env *env,
1753                                 long taskimp, long groupimp)
1754 {
1755         long src_load, dst_load, load;
1756         bool maymove = false;
1757         int cpu;
1758 
1759         load = task_h_load(env->p);
1760         dst_load = env->dst_stats.load + load;
1761         src_load = env->src_stats.load - load;
1762 
1763         /*
1764          * If the improvement from just moving env->p direction is better
1765          * than swapping tasks around, check if a move is possible.
1766          */
1767         maymove = !load_too_imbalanced(src_load, dst_load, env);
1768 
1769         for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1770                 /* Skip this CPU if the source task cannot migrate */
1771                 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
1772                         continue;
1773 
1774                 env->dst_cpu = cpu;
1775                 task_numa_compare(env, taskimp, groupimp, maymove);
1776         }
1777 }
1778 
1779 static int task_numa_migrate(struct task_struct *p)
1780 {
1781         struct task_numa_env env = {
1782                 .p = p,
1783 
1784                 .src_cpu = task_cpu(p),
1785                 .src_nid = task_node(p),
1786 
1787                 .imbalance_pct = 112,
1788 
1789                 .best_task = NULL,
1790                 .best_imp = 0,
1791                 .best_cpu = -1,
1792         };
1793         unsigned long taskweight, groupweight;
1794         struct sched_domain *sd;
1795         long taskimp, groupimp;
1796         struct numa_group *ng;
1797         struct rq *best_rq;
1798         int nid, ret, dist;
1799 
1800         /*
1801          * Pick the lowest SD_NUMA domain, as that would have the smallest
1802          * imbalance and would be the first to start moving tasks about.
1803          *
1804          * And we want to avoid any moving of tasks about, as that would create
1805          * random movement of tasks -- counter the numa conditions we're trying
1806          * to satisfy here.
1807          */
1808         rcu_read_lock();
1809         sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1810         if (sd)
1811                 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1812         rcu_read_unlock();
1813 
1814         /*
1815          * Cpusets can break the scheduler domain tree into smaller
1816          * balance domains, some of which do not cross NUMA boundaries.
1817          * Tasks that are "trapped" in such domains cannot be migrated
1818          * elsewhere, so there is no point in (re)trying.
1819          */
1820         if (unlikely(!sd)) {
1821                 sched_setnuma(p, task_node(p));
1822                 return -EINVAL;
1823         }
1824 
1825         env.dst_nid = p->numa_preferred_nid;
1826         dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1827         taskweight = task_weight(p, env.src_nid, dist);
1828         groupweight = group_weight(p, env.src_nid, dist);
1829         update_numa_stats(&env.src_stats, env.src_nid);
1830         taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1831         groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1832         update_numa_stats(&env.dst_stats, env.dst_nid);
1833 
1834         /* Try to find a spot on the preferred nid. */
1835         task_numa_find_cpu(&env, taskimp, groupimp);
1836 
1837         /*
1838          * Look at other nodes in these cases:
1839          * - there is no space available on the preferred_nid
1840          * - the task is part of a numa_group that is interleaved across
1841          *   multiple NUMA nodes; in order to better consolidate the group,
1842          *   we need to check other locations.
1843          */
1844         ng = deref_curr_numa_group(p);
1845         if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
1846                 for_each_online_node(nid) {
1847                         if (nid == env.src_nid || nid == p->numa_preferred_nid)
1848                                 continue;
1849 
1850                         dist = node_distance(env.src_nid, env.dst_nid);
1851                         if (sched_numa_topology_type == NUMA_BACKPLANE &&
1852                                                 dist != env.dist) {
1853                                 taskweight = task_weight(p, env.src_nid, dist);
1854                                 groupweight = group_weight(p, env.src_nid, dist);
1855                         }
1856 
1857                         /* Only consider nodes where both task and groups benefit */
1858                         taskimp = task_weight(p, nid, dist) - taskweight;
1859                         groupimp = group_weight(p, nid, dist) - groupweight;
1860                         if (taskimp < 0 && groupimp < 0)
1861                                 continue;
1862 
1863                         env.dist = dist;
1864                         env.dst_nid = nid;
1865                         update_numa_stats(&env.dst_stats, env.dst_nid);
1866                         task_numa_find_cpu(&env, taskimp, groupimp);
1867                 }
1868         }
1869 
1870         /*
1871          * If the task is part of a workload that spans multiple NUMA nodes,
1872          * and is migrating into one of the workload's active nodes, remember
1873          * this node as the task's preferred numa node, so the workload can
1874          * settle down.
1875          * A task that migrated to a second choice node will be better off
1876          * trying for a better one later. Do not set the preferred node here.
1877          */
1878         if (ng) {
1879                 if (env.best_cpu == -1)
1880                         nid = env.src_nid;
1881                 else
1882                         nid = cpu_to_node(env.best_cpu);
1883 
1884                 if (nid != p->numa_preferred_nid)
1885                         sched_setnuma(p, nid);
1886         }
1887 
1888         /* No better CPU than the current one was found. */
1889         if (env.best_cpu == -1)
1890                 return -EAGAIN;
1891 
1892         best_rq = cpu_rq(env.best_cpu);
1893         if (env.best_task == NULL) {
1894                 ret = migrate_task_to(p, env.best_cpu);
1895                 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1896                 if (ret != 0)
1897                         trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1898                 return ret;
1899         }
1900 
1901         ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
1902         WRITE_ONCE(best_rq->numa_migrate_on, 0);
1903 
1904         if (ret != 0)
1905                 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1906         put_task_struct(env.best_task);
1907         return ret;
1908 }
1909 
1910 /* Attempt to migrate a task to a CPU on the preferred node. */
1911 static void numa_migrate_preferred(struct task_struct *p)
1912 {
1913         unsigned long interval = HZ;
1914 
1915         /* This task has no NUMA fault statistics yet */
1916         if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
1917                 return;
1918 
1919         /* Periodically retry migrating the task to the preferred node */
1920         interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1921         p->numa_migrate_retry = jiffies + interval;
1922 
1923         /* Success if task is already running on preferred CPU */
1924         if (task_node(p) == p->numa_preferred_nid)
1925                 return;
1926 
1927         /* Otherwise, try migrate to a CPU on the preferred node */
1928         task_numa_migrate(p);
1929 }
1930 
1931 /*
1932  * Find out how many nodes on the workload is actively running on. Do this by
1933  * tracking the nodes from which NUMA hinting faults are triggered. This can
1934  * be different from the set of nodes where the workload's memory is currently
1935  * located.
1936  */
1937 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1938 {
1939         unsigned long faults, max_faults = 0;
1940         int nid, active_nodes = 0;
1941 
1942         for_each_online_node(nid) {
1943                 faults = group_faults_cpu(numa_group, nid);
1944                 if (faults > max_faults)
1945                         max_faults = faults;
1946         }
1947 
1948         for_each_online_node(nid) {
1949                 faults = group_faults_cpu(numa_group, nid);
1950                 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1951                         active_nodes++;
1952         }
1953 
1954         numa_group->max_faults_cpu = max_faults;
1955         numa_group->active_nodes = active_nodes;
1956 }
1957 
1958 /*
1959  * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1960  * increments. The more local the fault statistics are, the higher the scan
1961  * period will be for the next scan window. If local/(local+remote) ratio is
1962  * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1963  * the scan period will decrease. Aim for 70% local accesses.
1964  */
1965 #define NUMA_PERIOD_SLOTS 10
1966 #define NUMA_PERIOD_THRESHOLD 7
1967 
1968 /*
1969  * Increase the scan period (slow down scanning) if the majority of
1970  * our memory is already on our local node, or if the majority of
1971  * the page accesses are shared with other processes.
1972  * Otherwise, decrease the scan period.
1973  */
1974 static void update_task_scan_period(struct task_struct *p,
1975                         unsigned long shared, unsigned long private)
1976 {
1977         unsigned int period_slot;
1978         int lr_ratio, ps_ratio;
1979         int diff;
1980 
1981         unsigned long remote = p->numa_faults_locality[0];
1982         unsigned long local = p->numa_faults_locality[1];
1983 
1984         /*
1985          * If there were no record hinting faults then either the task is
1986          * completely idle or all activity is areas that are not of interest
1987          * to automatic numa balancing. Related to that, if there were failed
1988          * migration then it implies we are migrating too quickly or the local
1989          * node is overloaded. In either case, scan slower
1990          */
1991         if (local + shared == 0 || p->numa_faults_locality[2]) {
1992                 p->numa_scan_period = min(p->numa_scan_period_max,
1993                         p->numa_scan_period << 1);
1994 
1995                 p->mm->numa_next_scan = jiffies +
1996                         msecs_to_jiffies(p->numa_scan_period);
1997 
1998                 return;
1999         }
2000 
2001         /*
2002          * Prepare to scale scan period relative to the current period.
2003          *       == NUMA_PERIOD_THRESHOLD scan period stays the same
2004          *       <  NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
2005          *       >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
2006          */
2007         period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
2008         lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
2009         ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
2010 
2011         if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
2012                 /*
2013                  * Most memory accesses are local. There is no need to
2014                  * do fast NUMA scanning, since memory is already local.
2015                  */
2016                 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
2017                 if (!slot)
2018                         slot = 1;
2019                 diff = slot * period_slot;
2020         } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
2021                 /*
2022                  * Most memory accesses are shared with other tasks.
2023                  * There is no point in continuing fast NUMA scanning,
2024                  * since other tasks may just move the memory elsewhere.
2025                  */
2026                 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
2027                 if (!slot)
2028                         slot = 1;
2029                 diff = slot * period_slot;
2030         } else {
2031                 /*
2032                  * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2033                  * yet they are not on the local NUMA node. Speed up
2034                  * NUMA scanning to get the memory moved over.
2035                  */
2036                 int ratio = max(lr_ratio, ps_ratio);
2037                 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2038         }
2039 
2040         p->numa_scan_period = clamp(p->numa_scan_period + diff,
2041                         task_scan_min(p), task_scan_max(p));
2042         memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2043 }
2044 
2045 /*
2046  * Get the fraction of time the task has been running since the last
2047  * NUMA placement cycle. The scheduler keeps similar statistics, but
2048  * decays those on a 32ms period, which is orders of magnitude off
2049  * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2050  * stats only if the task is so new there are no NUMA statistics yet.
2051  */
2052 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2053 {
2054         u64 runtime, delta, now;
2055         /* Use the start of this time slice to avoid calculations. */
2056         now = p->se.exec_start;
2057         runtime = p->se.sum_exec_runtime;
2058 
2059         if (p->last_task_numa_placement) {
2060                 delta = runtime - p->last_sum_exec_runtime;
2061                 *period = now - p->last_task_numa_placement;
2062 
2063                 /* Avoid time going backwards, prevent potential divide error: */
2064                 if (unlikely((s64)*period < 0))
2065                         *period = 0;
2066         } else {
2067                 delta = p->se.avg.load_sum;
2068                 *period = LOAD_AVG_MAX;
2069         }
2070 
2071         p->last_sum_exec_runtime = runtime;
2072         p->last_task_numa_placement = now;
2073 
2074         return delta;
2075 }
2076 
2077 /*
2078  * Determine the preferred nid for a task in a numa_group. This needs to
2079  * be done in a way that produces consistent results with group_weight,
2080  * otherwise workloads might not converge.
2081  */
2082 static int preferred_group_nid(struct task_struct *p, int nid)
2083 {
2084         nodemask_t nodes;
2085         int dist;
2086 
2087         /* Direct connections between all NUMA nodes. */
2088         if (sched_numa_topology_type == NUMA_DIRECT)
2089                 return nid;
2090 
2091         /*
2092          * On a system with glueless mesh NUMA topology, group_weight
2093          * scores nodes according to the number of NUMA hinting faults on
2094          * both the node itself, and on nearby nodes.
2095          */
2096         if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2097                 unsigned long score, max_score = 0;
2098                 int node, max_node = nid;
2099 
2100                 dist = sched_max_numa_distance;
2101 
2102                 for_each_online_node(node) {
2103                         score = group_weight(p, node, dist);
2104                         if (score > max_score) {
2105                                 max_score = score;
2106                                 max_node = node;
2107                         }
2108                 }
2109                 return max_node;
2110         }
2111 
2112         /*
2113          * Finding the preferred nid in a system with NUMA backplane
2114          * interconnect topology is more involved. The goal is to locate
2115          * tasks from numa_groups near each other in the system, and
2116          * untangle workloads from different sides of the system. This requires
2117          * searching down the hierarchy of node groups, recursively searching
2118          * inside the highest scoring group of nodes. The nodemask tricks
2119          * keep the complexity of the search down.
2120          */
2121         nodes = node_online_map;
2122         for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2123                 unsigned long max_faults = 0;
2124                 nodemask_t max_group = NODE_MASK_NONE;
2125                 int a, b;
2126 
2127                 /* Are there nodes at this distance from each other? */
2128                 if (!find_numa_distance(dist))
2129                         continue;
2130 
2131                 for_each_node_mask(a, nodes) {
2132                         unsigned long faults = 0;
2133                         nodemask_t this_group;
2134                         nodes_clear(this_group);
2135 
2136                         /* Sum group's NUMA faults; includes a==b case. */
2137                         for_each_node_mask(b, nodes) {
2138                                 if (node_distance(a, b) < dist) {
2139                                         faults += group_faults(p, b);
2140                                         node_set(b, this_group);
2141                                         node_clear(b, nodes);
2142                                 }
2143                         }
2144 
2145                         /* Remember the top group. */
2146                         if (faults > max_faults) {
2147                                 max_faults = faults;
2148                                 max_group = this_group;
2149                                 /*
2150                                  * subtle: at the smallest distance there is
2151                                  * just one node left in each "group", the
2152                                  * winner is the preferred nid.
2153                                  */
2154                                 nid = a;
2155                         }
2156                 }
2157                 /* Next round, evaluate the nodes within max_group. */
2158                 if (!max_faults)
2159                         break;
2160                 nodes = max_group;
2161         }
2162         return nid;
2163 }
2164 
2165 static void task_numa_placement(struct task_struct *p)
2166 {
2167         int seq, nid, max_nid = NUMA_NO_NODE;
2168         unsigned long max_faults = 0;
2169         unsigned long fault_types[2] = { 0, 0 };
2170         unsigned long total_faults;
2171         u64 runtime, period;
2172         spinlock_t *group_lock = NULL;
2173         struct numa_group *ng;
2174 
2175         /*
2176          * The p->mm->numa_scan_seq field gets updated without
2177          * exclusive access. Use READ_ONCE() here to ensure
2178          * that the field is read in a single access:
2179          */
2180         seq = READ_ONCE(p->mm->numa_scan_seq);
2181         if (p->numa_scan_seq == seq)
2182                 return;
2183         p->numa_scan_seq = seq;
2184         p->numa_scan_period_max = task_scan_max(p);
2185 
2186         total_faults = p->numa_faults_locality[0] +
2187                        p->numa_faults_locality[1];
2188         runtime = numa_get_avg_runtime(p, &period);
2189 
2190         /* If the task is part of a group prevent parallel updates to group stats */
2191         ng = deref_curr_numa_group(p);
2192         if (ng) {
2193                 group_lock = &ng->lock;
2194                 spin_lock_irq(group_lock);
2195         }
2196 
2197         /* Find the node with the highest number of faults */
2198         for_each_online_node(nid) {
2199                 /* Keep track of the offsets in numa_faults array */
2200                 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2201                 unsigned long faults = 0, group_faults = 0;
2202                 int priv;
2203 
2204                 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2205                         long diff, f_diff, f_weight;
2206 
2207                         mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2208                         membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2209                         cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2210                         cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2211 
2212                         /* Decay existing window, copy faults since last scan */
2213                         diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2214                         fault_types[priv] += p->numa_faults[membuf_idx];
2215                         p->numa_faults[membuf_idx] = 0;
2216 
2217                         /*
2218                          * Normalize the faults_from, so all tasks in a group
2219                          * count according to CPU use, instead of by the raw
2220                          * number of faults. Tasks with little runtime have
2221                          * little over-all impact on throughput, and thus their
2222                          * faults are less important.
2223                          */
2224                         f_weight = div64_u64(runtime << 16, period + 1);
2225                         f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2226                                    (total_faults + 1);
2227                         f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2228                         p->numa_faults[cpubuf_idx] = 0;
2229 
2230                         p->numa_faults[mem_idx] += diff;
2231                         p->numa_faults[cpu_idx] += f_diff;
2232                         faults += p->numa_faults[mem_idx];
2233                         p->total_numa_faults += diff;
2234                         if (ng) {
2235                                 /*
2236                                  * safe because we can only change our own group
2237                                  *
2238                                  * mem_idx represents the offset for a given
2239                                  * nid and priv in a specific region because it
2240                                  * is at the beginning of the numa_faults array.
2241                                  */
2242                                 ng->faults[mem_idx] += diff;
2243                                 ng->faults_cpu[mem_idx] += f_diff;
2244                                 ng->total_faults += diff;
2245                                 group_faults += ng->faults[mem_idx];
2246                         }
2247                 }
2248 
2249                 if (!ng) {
2250                         if (faults > max_faults) {
2251                                 max_faults = faults;
2252                                 max_nid = nid;
2253                         }
2254                 } else if (group_faults > max_faults) {
2255                         max_faults = group_faults;
2256                         max_nid = nid;
2257                 }
2258         }
2259 
2260         if (ng) {
2261                 numa_group_count_active_nodes(ng);
2262                 spin_unlock_irq(group_lock);
2263                 max_nid = preferred_group_nid(p, max_nid);
2264         }
2265 
2266         if (max_faults) {
2267                 /* Set the new preferred node */
2268                 if (max_nid != p->numa_preferred_nid)
2269                         sched_setnuma(p, max_nid);
2270         }
2271 
2272         update_task_scan_period(p, fault_types[0], fault_types[1]);
2273 }
2274 
2275 static inline int get_numa_group(struct numa_group *grp)
2276 {
2277         return refcount_inc_not_zero(&grp->refcount);
2278 }
2279 
2280 static inline void put_numa_group(struct numa_group *grp)
2281 {
2282         if (refcount_dec_and_test(&grp->refcount))
2283                 kfree_rcu(grp, rcu);
2284 }
2285 
2286 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2287                         int *priv)
2288 {
2289         struct numa_group *grp, *my_grp;
2290         struct task_struct *tsk;
2291         bool join = false;
2292         int cpu = cpupid_to_cpu(cpupid);
2293         int i;
2294 
2295         if (unlikely(!deref_curr_numa_group(p))) {
2296                 unsigned int size = sizeof(struct numa_group) +
2297                                     4*nr_node_ids*sizeof(unsigned long);
2298 
2299                 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2300                 if (!grp)
2301                         return;
2302 
2303                 refcount_set(&grp->refcount, 1);
2304                 grp->active_nodes = 1;
2305                 grp->max_faults_cpu = 0;
2306                 spin_lock_init(&grp->lock);
2307                 grp->gid = p->pid;
2308                 /* Second half of the array tracks nids where faults happen */
2309                 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2310                                                 nr_node_ids;
2311 
2312                 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2313                         grp->faults[i] = p->numa_faults[i];
2314 
2315                 grp->total_faults = p->total_numa_faults;
2316 
2317                 grp->nr_tasks++;
2318                 rcu_assign_pointer(p->numa_group, grp);
2319         }
2320 
2321         rcu_read_lock();
2322         tsk = READ_ONCE(cpu_rq(cpu)->curr);
2323 
2324         if (!cpupid_match_pid(tsk, cpupid))
2325                 goto no_join;
2326 
2327         grp = rcu_dereference(tsk->numa_group);
2328         if (!grp)
2329                 goto no_join;
2330 
2331         my_grp = deref_curr_numa_group(p);
2332         if (grp == my_grp)
2333                 goto no_join;
2334 
2335         /*
2336          * Only join the other group if its bigger; if we're the bigger group,
2337          * the other task will join us.
2338          */
2339         if (my_grp->nr_tasks > grp->nr_tasks)
2340                 goto no_join;
2341 
2342         /*
2343          * Tie-break on the grp address.
2344          */
2345         if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2346                 goto no_join;
2347 
2348         /* Always join threads in the same process. */
2349         if (tsk->mm == current->mm)
2350                 join = true;
2351 
2352         /* Simple filter to avoid false positives due to PID collisions */
2353         if (flags & TNF_SHARED)
2354                 join = true;
2355 
2356         /* Update priv based on whether false sharing was detected */
2357         *priv = !join;
2358 
2359         if (join && !get_numa_group(grp))
2360                 goto no_join;
2361 
2362         rcu_read_unlock();
2363 
2364         if (!join)
2365                 return;
2366 
2367         BUG_ON(irqs_disabled());
2368         double_lock_irq(&my_grp->lock, &grp->lock);
2369 
2370         for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2371                 my_grp->faults[i] -= p->numa_faults[i];
2372                 grp->faults[i] += p->numa_faults[i];
2373         }
2374         my_grp->total_faults -= p->total_numa_faults;
2375         grp->total_faults += p->total_numa_faults;
2376 
2377         my_grp->nr_tasks--;
2378         grp->nr_tasks++;
2379 
2380         spin_unlock(&my_grp->lock);
2381         spin_unlock_irq(&grp->lock);
2382 
2383         rcu_assign_pointer(p->numa_group, grp);
2384 
2385         put_numa_group(my_grp);
2386         return;
2387 
2388 no_join:
2389         rcu_read_unlock();
2390         return;
2391 }
2392 
2393 /*
2394  * Get rid of NUMA staticstics associated with a task (either current or dead).
2395  * If @final is set, the task is dead and has reached refcount zero, so we can
2396  * safely free all relevant data structures. Otherwise, there might be
2397  * concurrent reads from places like load balancing and procfs, and we should
2398  * reset the data back to default state without freeing ->numa_faults.
2399  */
2400 void task_numa_free(struct task_struct *p, bool final)
2401 {
2402         /* safe: p either is current or is being freed by current */
2403         struct numa_group *grp = rcu_dereference_raw(p->numa_group);
2404         unsigned long *numa_faults = p->numa_faults;
2405         unsigned long flags;
2406         int i;
2407 
2408         if (!numa_faults)
2409                 return;
2410 
2411         if (grp) {
2412                 spin_lock_irqsave(&grp->lock, flags);
2413                 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2414                         grp->faults[i] -= p->numa_faults[i];
2415                 grp->total_faults -= p->total_numa_faults;
2416 
2417                 grp->nr_tasks--;
2418                 spin_unlock_irqrestore(&grp->lock, flags);
2419                 RCU_INIT_POINTER(p->numa_group, NULL);
2420                 put_numa_group(grp);
2421         }
2422 
2423         if (final) {
2424                 p->numa_faults = NULL;
2425                 kfree(numa_faults);
2426         } else {
2427                 p->total_numa_faults = 0;
2428                 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2429                         numa_faults[i] = 0;
2430         }
2431 }
2432 
2433 /*
2434  * Got a PROT_NONE fault for a page on @node.
2435  */
2436 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2437 {
2438         struct task_struct *p = current;
2439         bool migrated = flags & TNF_MIGRATED;
2440         int cpu_node = task_node(current);
2441         int local = !!(flags & TNF_FAULT_LOCAL);
2442         struct numa_group *ng;
2443         int priv;
2444 
2445         if (!static_branch_likely(&sched_numa_balancing))
2446                 return;
2447 
2448         /* for example, ksmd faulting in a user's mm */
2449         if (!p->mm)
2450                 return;
2451 
2452         /* Allocate buffer to track faults on a per-node basis */
2453         if (unlikely(!p->numa_faults)) {
2454                 int size = sizeof(*p->numa_faults) *
2455                            NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2456 
2457                 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2458                 if (!p->numa_faults)
2459                         return;
2460 
2461                 p->total_numa_faults = 0;
2462                 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2463         }
2464 
2465         /*
2466          * First accesses are treated as private, otherwise consider accesses
2467          * to be private if the accessing pid has not changed
2468          */
2469         if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2470                 priv = 1;
2471         } else {
2472                 priv = cpupid_match_pid(p, last_cpupid);
2473                 if (!priv && !(flags & TNF_NO_GROUP))
2474                         task_numa_group(p, last_cpupid, flags, &priv);
2475         }
2476 
2477         /*
2478          * If a workload spans multiple NUMA nodes, a shared fault that
2479          * occurs wholly within the set of nodes that the workload is
2480          * actively using should be counted as local. This allows the
2481          * scan rate to slow down when a workload has settled down.
2482          */
2483         ng = deref_curr_numa_group(p);
2484         if (!priv && !local && ng && ng->active_nodes > 1 &&
2485                                 numa_is_active_node(cpu_node, ng) &&
2486                                 numa_is_active_node(mem_node, ng))
2487                 local = 1;
2488 
2489         /*
2490          * Retry to migrate task to preferred node periodically, in case it
2491          * previously failed, or the scheduler moved us.
2492          */
2493         if (time_after(jiffies, p->numa_migrate_retry)) {
2494                 task_numa_placement(p);
2495                 numa_migrate_preferred(p);
2496         }
2497 
2498         if (migrated)
2499                 p->numa_pages_migrated += pages;
2500         if (flags & TNF_MIGRATE_FAIL)
2501                 p->numa_faults_locality[2] += pages;
2502 
2503         p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2504         p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2505         p->numa_faults_locality[local] += pages;
2506 }
2507 
2508 static void reset_ptenuma_scan(struct task_struct *p)
2509 {
2510         /*
2511          * We only did a read acquisition of the mmap sem, so
2512          * p->mm->numa_scan_seq is written to without exclusive access
2513          * and the update is not guaranteed to be atomic. That's not
2514          * much of an issue though, since this is just used for
2515          * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2516          * expensive, to avoid any form of compiler optimizations:
2517          */
2518         WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2519         p->mm->numa_scan_offset = 0;
2520 }
2521 
2522 /*
2523  * The expensive part of numa migration is done from task_work context.
2524  * Triggered from task_tick_numa().
2525  */
2526 void task_numa_work(struct callback_head *work)
2527 {
2528         unsigned long migrate, next_scan, now = jiffies;
2529         struct task_struct *p = current;
2530         struct mm_struct *mm = p->mm;
2531         u64 runtime = p->se.sum_exec_runtime;
2532         struct vm_area_struct *vma;
2533         unsigned long start, end;
2534         unsigned long nr_pte_updates = 0;
2535         long pages, virtpages;
2536 
2537         SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2538 
2539         work->next = work; /* protect against double add */
2540         /*
2541          * Who cares about NUMA placement when they're dying.
2542          *
2543          * NOTE: make sure not to dereference p->mm before this check,
2544          * exit_task_work() happens _after_ exit_mm() so we could be called
2545          * without p->mm even though we still had it when we enqueued this
2546          * work.
2547          */
2548         if (p->flags & PF_EXITING)
2549                 return;
2550 
2551         if (!mm->numa_next_scan) {
2552                 mm->numa_next_scan = now +
2553                         msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2554         }
2555 
2556         /*
2557          * Enforce maximal scan/migration frequency..
2558          */
2559         migrate = mm->numa_next_scan;
2560         if (time_before(now, migrate))
2561                 return;
2562 
2563         if (p->numa_scan_period == 0) {
2564                 p->numa_scan_period_max = task_scan_max(p);
2565                 p->numa_scan_period = task_scan_start(p);
2566         }
2567 
2568         next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2569         if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2570                 return;
2571 
2572         /*
2573          * Delay this task enough that another task of this mm will likely win
2574          * the next time around.
2575          */
2576         p->node_stamp += 2 * TICK_NSEC;
2577 
2578         start = mm->numa_scan_offset;
2579         pages = sysctl_numa_balancing_scan_size;
2580         pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2581         virtpages = pages * 8;     /* Scan up to this much virtual space */
2582         if (!pages)
2583                 return;
2584 
2585 
2586         if (!down_read_trylock(&mm->mmap_sem))
2587                 return;
2588         vma = find_vma(mm, start);
2589         if (!vma) {
2590                 reset_ptenuma_scan(p);
2591                 start = 0;
2592                 vma = mm->mmap;
2593         }
2594         for (; vma; vma = vma->vm_next) {
2595                 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2596                         is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2597                         continue;
2598                 }
2599 
2600                 /*
2601                  * Shared library pages mapped by multiple processes are not
2602                  * migrated as it is expected they are cache replicated. Avoid
2603                  * hinting faults in read-only file-backed mappings or the vdso
2604                  * as migrating the pages will be of marginal benefit.
2605                  */
2606                 if (!vma->vm_mm ||
2607                     (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2608                         continue;
2609 
2610                 /*
2611                  * Skip inaccessible VMAs to avoid any confusion between
2612                  * PROT_NONE and NUMA hinting ptes
2613                  */
2614                 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2615                         continue;
2616 
2617                 do {
2618                         start = max(start, vma->vm_start);
2619                         end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2620                         end = min(end, vma->vm_end);
2621                         nr_pte_updates = change_prot_numa(vma, start, end);
2622 
2623                         /*
2624                          * Try to scan sysctl_numa_balancing_size worth of
2625                          * hpages that have at least one present PTE that
2626                          * is not already pte-numa. If the VMA contains
2627                          * areas that are unused or already full of prot_numa
2628                          * PTEs, scan up to virtpages, to skip through those
2629                          * areas faster.
2630                          */
2631                         if (nr_pte_updates)
2632                                 pages -= (end - start) >> PAGE_SHIFT;
2633                         virtpages -= (end - start) >> PAGE_SHIFT;
2634 
2635                         start = end;
2636                         if (pages <= 0 || virtpages <= 0)
2637                                 goto out;
2638 
2639                         cond_resched();
2640                 } while (end != vma->vm_end);
2641         }
2642 
2643 out:
2644         /*
2645          * It is possible to reach the end of the VMA list but the last few
2646          * VMAs are not guaranteed to the vma_migratable. If they are not, we
2647          * would find the !migratable VMA on the next scan but not reset the
2648          * scanner to the start so check it now.
2649          */
2650         if (vma)
2651                 mm->numa_scan_offset = start;
2652         else
2653                 reset_ptenuma_scan(p);
2654         up_read(&mm->mmap_sem);
2655 
2656         /*
2657          * Make sure tasks use at least 32x as much time to run other code
2658          * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2659          * Usually update_task_scan_period slows down scanning enough; on an
2660          * overloaded system we need to limit overhead on a per task basis.
2661          */
2662         if (unlikely(p->se.sum_exec_runtime != runtime)) {
2663                 u64 diff = p->se.sum_exec_runtime - runtime;
2664                 p->node_stamp += 32 * diff;
2665         }
2666 }
2667 
2668 /*
2669  * Drive the periodic memory faults..
2670  */
2671 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2672 {
2673         struct callback_head *work = &curr->numa_work;
2674         u64 period, now;
2675 
2676         /*
2677          * We don't care about NUMA placement if we don't have memory.
2678          */
2679         if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2680                 return;
2681 
2682         /*
2683          * Using runtime rather than walltime has the dual advantage that
2684          * we (mostly) drive the selection from busy threads and that the
2685          * task needs to have done some actual work before we bother with
2686          * NUMA placement.
2687          */
2688         now = curr->se.sum_exec_runtime;
2689         period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2690 
2691         if (now > curr->node_stamp + period) {
2692                 if (!curr->node_stamp)
2693                         curr->numa_scan_period = task_scan_start(curr);
2694                 curr->node_stamp += period;
2695 
2696                 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2697                         init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2698                         task_work_add(curr, work, true);
2699                 }
2700         }
2701 }
2702 
2703 static void update_scan_period(struct task_struct *p, int new_cpu)
2704 {
2705         int src_nid = cpu_to_node(task_cpu(p));
2706         int dst_nid = cpu_to_node(new_cpu);
2707 
2708         if (!static_branch_likely(&sched_numa_balancing))
2709                 return;
2710 
2711         if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
2712                 return;
2713 
2714         if (src_nid == dst_nid)
2715                 return;
2716 
2717         /*
2718          * Allow resets if faults have been trapped before one scan
2719          * has completed. This is most likely due to a new task that
2720          * is pulled cross-node due to wakeups or load balancing.
2721          */
2722         if (p->numa_scan_seq) {
2723                 /*
2724                  * Avoid scan adjustments if moving to the preferred
2725                  * node or if the task was not previously running on
2726                  * the preferred node.
2727                  */
2728                 if (dst_nid == p->numa_preferred_nid ||
2729                     (p->numa_preferred_nid != NUMA_NO_NODE &&
2730                         src_nid != p->numa_preferred_nid))
2731                         return;
2732         }
2733 
2734         p->numa_scan_period = task_scan_start(p);
2735 }
2736 
2737 #else
2738 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2739 {
2740 }
2741 
2742 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2743 {
2744 }
2745 
2746 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2747 {
2748 }
2749 
2750 static inline void update_scan_period(struct task_struct *p, int new_cpu)
2751 {
2752 }
2753 
2754 #endif /* CONFIG_NUMA_BALANCING */
2755 
2756 static void
2757 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2758 {
2759         update_load_add(&cfs_rq->load, se->load.weight);
2760 #ifdef CONFIG_SMP
2761         if (entity_is_task(se)) {
2762                 struct rq *rq = rq_of(cfs_rq);
2763 
2764                 account_numa_enqueue(rq, task_of(se));
2765                 list_add(&se->group_node, &rq->cfs_tasks);
2766         }
2767 #endif
2768         cfs_rq->nr_running++;
2769 }
2770 
2771 static void
2772 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2773 {
2774         update_load_sub(&cfs_rq->load, se->load.weight);
2775 #ifdef CONFIG_SMP
2776         if (entity_is_task(se)) {
2777                 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2778                 list_del_init(&se->group_node);
2779         }
2780 #endif
2781         cfs_rq->nr_running--;
2782 }
2783 
2784 /*
2785  * Signed add and clamp on underflow.
2786  *
2787  * Explicitly do a load-store to ensure the intermediate value never hits
2788  * memory. This allows lockless observations without ever seeing the negative
2789  * values.
2790  */
2791 #define add_positive(_ptr, _val) do {                           \
2792         typeof(_ptr) ptr = (_ptr);                              \
2793         typeof(_val) val = (_val);                              \
2794         typeof(*ptr) res, var = READ_ONCE(*ptr);                \
2795                                                                 \
2796         res = var + val;                                        \
2797                                                                 \
2798         if (val < 0 && res > var)                               \
2799                 res = 0;                                        \
2800                                                                 \
2801         WRITE_ONCE(*ptr, res);                                  \
2802 } while (0)
2803 
2804 /*
2805  * Unsigned subtract and clamp on underflow.
2806  *
2807  * Explicitly do a load-store to ensure the intermediate value never hits
2808  * memory. This allows lockless observations without ever seeing the negative
2809  * values.
2810  */
2811 #define sub_positive(_ptr, _val) do {                           \
2812         typeof(_ptr) ptr = (_ptr);                              \
2813         typeof(*ptr) val = (_val);                              \
2814         typeof(*ptr) res, var = READ_ONCE(*ptr);                \
2815         res = var - val;                                        \
2816         if (res > var)                                          \
2817                 res = 0;                                        \
2818         WRITE_ONCE(*ptr, res);                                  \
2819 } while (0)
2820 
2821 /*
2822  * Remove and clamp on negative, from a local variable.
2823  *
2824  * A variant of sub_positive(), which does not use explicit load-store
2825  * and is thus optimized for local variable updates.
2826  */
2827 #define lsub_positive(_ptr, _val) do {                          \
2828         typeof(_ptr) ptr = (_ptr);                              \
2829         *ptr -= min_t(typeof(*ptr), *ptr, _val);                \
2830 } while (0)
2831 
2832 #ifdef CONFIG_SMP
2833 static inline void
2834 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2835 {
2836         cfs_rq->runnable_weight += se->runnable_weight;
2837 
2838         cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg;
2839         cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum;
2840 }
2841 
2842 static inline void
2843 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2844 {
2845         cfs_rq->runnable_weight -= se->runnable_weight;
2846 
2847         sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg);
2848         sub_positive(&cfs_rq->avg.runnable_load_sum,
2849                      se_runnable(se) * se->avg.runnable_load_sum);
2850 }
2851 
2852 static inline void
2853 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2854 {
2855         cfs_rq->avg.load_avg += se->avg.load_avg;
2856         cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
2857 }
2858 
2859 static inline void
2860 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2861 {
2862         sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2863         sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
2864 }
2865 #else
2866 static inline void
2867 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2868 static inline void
2869 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2870 static inline void
2871 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2872 static inline void
2873 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2874 #endif
2875 
2876 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2877                             unsigned long weight, unsigned long runnable)
2878 {
2879         if (se->on_rq) {
2880                 /* commit outstanding execution time */
2881                 if (cfs_rq->curr == se)
2882                         update_curr(cfs_rq);
2883                 account_entity_dequeue(cfs_rq, se);
2884                 dequeue_runnable_load_avg(cfs_rq, se);
2885         }
2886         dequeue_load_avg(cfs_rq, se);
2887 
2888         se->runnable_weight = runnable;
2889         update_load_set(&se->load, weight);
2890 
2891 #ifdef CONFIG_SMP
2892         do {
2893                 u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
2894 
2895                 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
2896                 se->avg.runnable_load_avg =
2897                         div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider);
2898         } while (0);
2899 #endif
2900 
2901         enqueue_load_avg(cfs_rq, se);
2902         if (se->on_rq) {
2903                 account_entity_enqueue(cfs_rq, se);
2904                 enqueue_runnable_load_avg(cfs_rq, se);
2905         }
2906 }
2907 
2908 void reweight_task(struct task_struct *p, int prio)
2909 {
2910         struct sched_entity *se = &p->se;
2911         struct cfs_rq *cfs_rq = cfs_rq_of(se);
2912         struct load_weight *load = &se->load;
2913         unsigned long weight = scale_load(sched_prio_to_weight[prio]);
2914 
2915         reweight_entity(cfs_rq, se, weight, weight);
2916         load->inv_weight = sched_prio_to_wmult[prio];
2917 }
2918 
2919 #ifdef CONFIG_FAIR_GROUP_SCHED
2920 #ifdef CONFIG_SMP
2921 /*
2922  * All this does is approximate the hierarchical proportion which includes that
2923  * global sum we all love to hate.
2924  *
2925  * That is, the weight of a group entity, is the proportional share of the
2926  * group weight based on the group runqueue weights. That is:
2927  *
2928  *                     tg->weight * grq->load.weight
2929  *   ge->load.weight = -----------------------------               (1)
2930  *                        \Sum grq->load.weight
2931  *
2932  * Now, because computing that sum is prohibitively expensive to compute (been
2933  * there, done that) we approximate it with this average stuff. The average
2934  * moves slower and therefore the approximation is cheaper and more stable.
2935  *
2936  * So instead of the above, we substitute:
2937  *
2938  *   grq->load.weight -> grq->avg.load_avg                         (2)
2939  *
2940  * which yields the following:
2941  *
2942  *                     tg->weight * grq->avg.load_avg
2943  *   ge->load.weight = ------------------------------              (3)
2944  *                              tg->load_avg
2945  *
2946  * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2947  *
2948  * That is shares_avg, and it is right (given the approximation (2)).
2949  *
2950  * The problem with it is that because the average is slow -- it was designed
2951  * to be exactly that of course -- this leads to transients in boundary
2952  * conditions. In specific, the case where the group was idle and we start the
2953  * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2954  * yielding bad latency etc..
2955  *
2956  * Now, in that special case (1) reduces to:
2957  *
2958  *                     tg->weight * grq->load.weight
2959  *   ge->load.weight = ----------------------------- = tg->weight   (4)
2960  *                          grp->load.weight
2961  *
2962  * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2963  *
2964  * So what we do is modify our approximation (3) to approach (4) in the (near)
2965  * UP case, like:
2966  *
2967  *   ge->load.weight =
2968  *
2969  *              tg->weight * grq->load.weight
2970  *     ---------------------------------------------------         (5)
2971  *     tg->load_avg - grq->avg.load_avg + grq->load.weight
2972  *
2973  * But because grq->load.weight can drop to 0, resulting in a divide by zero,
2974  * we need to use grq->avg.load_avg as its lower bound, which then gives:
2975  *
2976  *
2977  *                     tg->weight * grq->load.weight
2978  *   ge->load.weight = -----------------------------               (6)
2979  *                              tg_load_avg'
2980  *
2981  * Where:
2982  *
2983  *   tg_load_avg' = tg->load_avg - grq->avg.load_avg +
2984  *                  max(grq->load.weight, grq->avg.load_avg)
2985  *
2986  * And that is shares_weight and is icky. In the (near) UP case it approaches
2987  * (4) while in the normal case it approaches (3). It consistently
2988  * overestimates the ge->load.weight and therefore:
2989  *
2990  *   \Sum ge->load.weight >= tg->weight
2991  *
2992  * hence icky!
2993  */
2994 static long calc_group_shares(struct cfs_rq *cfs_rq)
2995 {
2996         long tg_weight, tg_shares, load, shares;
2997         struct task_group *tg = cfs_rq->tg;
2998 
2999         tg_shares = READ_ONCE(tg->shares);
3000 
3001         load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3002 
3003         tg_weight = atomic_long_read(&tg->load_avg);
3004 
3005         /* Ensure tg_weight >= load */
3006         tg_weight -= cfs_rq->tg_load_avg_contrib;
3007         tg_weight += load;
3008 
3009         shares = (tg_shares * load);
3010         if (tg_weight)
3011                 shares /= tg_weight;
3012 
3013         /*
3014          * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3015          * of a group with small tg->shares value. It is a floor value which is
3016          * assigned as a minimum load.weight to the sched_entity representing
3017          * the group on a CPU.
3018          *
3019          * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3020          * on an 8-core system with 8 tasks each runnable on one CPU shares has
3021          * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3022          * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3023          * instead of 0.
3024          */
3025         return clamp_t(long, shares, MIN_SHARES, tg_shares);
3026 }
3027 
3028 /*
3029  * This calculates the effective runnable weight for a group entity based on
3030  * the group entity weight calculated above.
3031  *
3032  * Because of the above approximation (2), our group entity weight is
3033  * an load_avg based ratio (3). This means that it includes blocked load and
3034  * does not represent the runnable weight.
3035  *
3036  * Approximate the group entity's runnable weight per ratio from the group
3037  * runqueue:
3038  *
3039  *                                           grq->avg.runnable_load_avg
3040  *   ge->runnable_weight = ge->load.weight * -------------------------- (7)
3041  *                                               grq->avg.load_avg
3042  *
3043  * However, analogous to above, since the avg numbers are slow, this leads to
3044  * transients in the from-idle case. Instead we use:
3045  *
3046  *   ge->runnable_weight = ge->load.weight *
3047  *
3048  *              max(grq->avg.runnable_load_avg, grq->runnable_weight)
3049  *              -----------------------------------------------------   (8)
3050  *                    max(grq->avg.load_avg, grq->load.weight)
3051  *
3052  * Where these max() serve both to use the 'instant' values to fix the slow
3053  * from-idle and avoid the /0 on to-idle, similar to (6).
3054  */
3055 static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares)
3056 {
3057         long runnable, load_avg;
3058 
3059         load_avg = max(cfs_rq->avg.load_avg,
3060                        scale_load_down(cfs_rq->load.weight));
3061 
3062         runnable = max(cfs_rq->avg.runnable_load_avg,
3063                        scale_load_down(cfs_rq->runnable_weight));
3064 
3065         runnable *= shares;
3066         if (load_avg)
3067                 runnable /= load_avg;
3068 
3069         return clamp_t(long, runnable, MIN_SHARES, shares);
3070 }
3071 #endif /* CONFIG_SMP */
3072 
3073 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3074 
3075 /*
3076  * Recomputes the group entity based on the current state of its group
3077  * runqueue.
3078  */
3079 static void update_cfs_group(struct sched_entity *se)
3080 {
3081         struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3082         long shares, runnable;
3083 
3084         if (!gcfs_rq)
3085                 return;
3086 
3087         if (throttled_hierarchy(gcfs_rq))
3088                 return;
3089 
3090 #ifndef CONFIG_SMP
3091         runnable = shares = READ_ONCE(gcfs_rq->tg->shares);
3092 
3093         if (likely(se->load.weight == shares))
3094                 return;
3095 #else
3096         shares   = calc_group_shares(gcfs_rq);
3097         runnable = calc_group_runnable(gcfs_rq, shares);
3098 #endif
3099 
3100         reweight_entity(cfs_rq_of(se), se, shares, runnable);
3101 }
3102 
3103 #else /* CONFIG_FAIR_GROUP_SCHED */
3104 static inline void update_cfs_group(struct sched_entity *se)
3105 {
3106 }
3107 #endif /* CONFIG_FAIR_GROUP_SCHED */
3108 
3109 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3110 {
3111         struct rq *rq = rq_of(cfs_rq);
3112 
3113         if (&rq->cfs == cfs_rq || (flags & SCHED_CPUFREQ_MIGRATION)) {
3114                 /*
3115                  * There are a few boundary cases this might miss but it should
3116                  * get called often enough that that should (hopefully) not be
3117                  * a real problem.
3118                  *
3119                  * It will not get called when we go idle, because the idle
3120                  * thread is a different class (!fair), nor will the utilization
3121                  * number include things like RT tasks.
3122                  *
3123                  * As is, the util number is not freq-invariant (we'd have to
3124                  * implement arch_scale_freq_capacity() for that).
3125                  *
3126                  * See cpu_util().
3127                  */
3128                 cpufreq_update_util(rq, flags);
3129         }
3130 }
3131 
3132 #ifdef CONFIG_SMP
3133 #ifdef CONFIG_FAIR_GROUP_SCHED
3134 /**
3135  * update_tg_load_avg - update the tg's load avg
3136  * @cfs_rq: the cfs_rq whose avg changed
3137  * @force: update regardless of how small the difference
3138  *
3139  * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3140  * However, because tg->load_avg is a global value there are performance
3141  * considerations.
3142  *
3143  * In order to avoid having to look at the other cfs_rq's, we use a
3144  * differential update where we store the last value we propagated. This in
3145  * turn allows skipping updates if the differential is 'small'.
3146  *
3147  * Updating tg's load_avg is necessary before update_cfs_share().
3148  */
3149 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3150 {
3151         long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3152 
3153         /*
3154          * No need to update load_avg for root_task_group as it is not used.
3155          */
3156         if (cfs_rq->tg == &root_task_group)
3157                 return;
3158 
3159         if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3160                 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3161                 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3162         }
3163 }
3164 
3165 /*
3166  * Called within set_task_rq() right before setting a task's CPU. The
3167  * caller only guarantees p->pi_lock is held; no other assumptions,
3168  * including the state of rq->lock, should be made.
3169  */
3170 void set_task_rq_fair(struct sched_entity *se,
3171                       struct cfs_rq *prev, struct cfs_rq *next)
3172 {
3173         u64 p_last_update_time;
3174         u64 n_last_update_time;
3175 
3176         if (!sched_feat(ATTACH_AGE_LOAD))
3177                 return;
3178 
3179         /*
3180          * We are supposed to update the task to "current" time, then its up to
3181          * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3182          * getting what current time is, so simply throw away the out-of-date
3183          * time. This will result in the wakee task is less decayed, but giving
3184          * the wakee more load sounds not bad.
3185          */
3186         if (!(se->avg.last_update_time && prev))
3187                 return;
3188 
3189 #ifndef CONFIG_64BIT
3190         {
3191                 u64 p_last_update_time_copy;
3192                 u64 n_last_update_time_copy;
3193 
3194                 do {
3195                         p_last_update_time_copy = prev->load_last_update_time_copy;
3196                         n_last_update_time_copy = next->load_last_update_time_copy;
3197 
3198                         smp_rmb();
3199 
3200                         p_last_update_time = prev->avg.last_update_time;
3201                         n_last_update_time = next->avg.last_update_time;
3202 
3203                 } while (p_last_update_time != p_last_update_time_copy ||
3204                          n_last_update_time != n_last_update_time_copy);
3205         }
3206 #else
3207         p_last_update_time = prev->avg.last_update_time;
3208         n_last_update_time = next->avg.last_update_time;
3209 #endif
3210         __update_load_avg_blocked_se(p_last_update_time, se);
3211         se->avg.last_update_time = n_last_update_time;
3212 }
3213 
3214 
3215 /*
3216  * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3217  * propagate its contribution. The key to this propagation is the invariant
3218  * that for each group:
3219  *
3220  *   ge->avg == grq->avg                                                (1)
3221  *
3222  * _IFF_ we look at the pure running and runnable sums. Because they
3223  * represent the very same entity, just at different points in the hierarchy.
3224  *
3225  * Per the above update_tg_cfs_util() is trivial and simply copies the running
3226  * sum over (but still wrong, because the group entity and group rq do not have
3227  * their PELT windows aligned).
3228  *
3229  * However, update_tg_cfs_runnable() is more complex. So we have:
3230  *
3231  *   ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg          (2)
3232  *
3233  * And since, like util, the runnable part should be directly transferable,
3234  * the following would _appear_ to be the straight forward approach:
3235  *
3236  *   grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg       (3)
3237  *
3238  * And per (1) we have:
3239  *
3240  *   ge->avg.runnable_avg == grq->avg.runnable_avg
3241  *
3242  * Which gives:
3243  *
3244  *                      ge->load.weight * grq->avg.load_avg
3245  *   ge->avg.load_avg = -----------------------------------             (4)
3246  *                               grq->load.weight
3247  *
3248  * Except that is wrong!
3249  *
3250  * Because while for entities historical weight is not important and we
3251  * really only care about our future and therefore can consider a pure
3252  * runnable sum, runqueues can NOT do this.
3253  *
3254  * We specifically want runqueues to have a load_avg that includes
3255  * historical weights. Those represent the blocked load, the load we expect
3256  * to (shortly) return to us. This only works by keeping the weights as
3257  * integral part of the sum. We therefore cannot decompose as per (3).
3258  *
3259  * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3260  * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3261  * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3262  * runnable section of these tasks overlap (or not). If they were to perfectly
3263  * align the rq as a whole would be runnable 2/3 of the time. If however we
3264  * always have at least 1 runnable task, the rq as a whole is always runnable.
3265  *
3266  * So we'll have to approximate.. :/
3267  *
3268  * Given the constraint:
3269  *
3270  *   ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3271  *
3272  * We can construct a rule that adds runnable to a rq by assuming minimal
3273  * overlap.
3274  *
3275  * On removal, we'll assume each task is equally runnable; which yields:
3276  *
3277  *   grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3278  *
3279  * XXX: only do this for the part of runnable > running ?
3280  *
3281  */
3282 
3283 static inline void
3284 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3285 {
3286         long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3287 
3288         /* Nothing to update */
3289         if (!delta)
3290                 return;
3291 
3292         /*
3293          * The relation between sum and avg is:
3294          *
3295          *   LOAD_AVG_MAX - 1024 + sa->period_contrib
3296          *
3297          * however, the PELT windows are not aligned between grq and gse.
3298          */
3299 
3300         /* Set new sched_entity's utilization */
3301         se->avg.util_avg = gcfs_rq->avg.util_avg;
3302         se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3303 
3304         /* Update parent cfs_rq utilization */
3305         add_positive(&cfs_rq->avg.util_avg, delta);
3306         cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3307 }
3308 
3309 static inline void
3310 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3311 {
3312         long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3313         unsigned long runnable_load_avg, load_avg;
3314         u64 runnable_load_sum, load_sum = 0;
3315         s64 delta_sum;
3316 
3317         if (!runnable_sum)
3318                 return;
3319 
3320         gcfs_rq->prop_runnable_sum = 0;
3321 
3322         if (runnable_sum >= 0) {
3323                 /*
3324                  * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3325                  * the CPU is saturated running == runnable.
3326                  */
3327                 runnable_sum += se->avg.load_sum;
3328                 runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
3329         } else {
3330                 /*
3331                  * Estimate the new unweighted runnable_sum of the gcfs_rq by
3332                  * assuming all tasks are equally runnable.
3333                  */
3334                 if (scale_load_down(gcfs_rq->load.weight)) {
3335                         load_sum = div_s64(gcfs_rq->avg.load_sum,
3336                                 scale_load_down(gcfs_rq->load.weight));
3337                 }
3338 
3339                 /* But make sure to not inflate se's runnable */
3340                 runnable_sum = min(se->avg.load_sum, load_sum);
3341         }
3342 
3343         /*
3344          * runnable_sum can't be lower than running_sum
3345          * Rescale running sum to be in the same range as runnable sum
3346          * running_sum is in [0 : LOAD_AVG_MAX <<  SCHED_CAPACITY_SHIFT]
3347          * runnable_sum is in [0 : LOAD_AVG_MAX]
3348          */
3349         running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
3350         runnable_sum = max(runnable_sum, running_sum);
3351 
3352         load_sum = (s64)se_weight(se) * runnable_sum;
3353         load_avg = div_s64(load_sum, LOAD_AVG_MAX);
3354 
3355         delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3356         delta_avg = load_avg - se->avg.load_avg;
3357 
3358         se->avg.load_sum = runnable_sum;
3359         se->avg.load_avg = load_avg;
3360         add_positive(&cfs_rq->avg.load_avg, delta_avg);
3361         add_positive(&cfs_rq->avg.load_sum, delta_sum);
3362 
3363         runnable_load_sum = (s64)se_runnable(se) * runnable_sum;
3364         runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX);
3365         delta_sum = runnable_load_sum - se_weight(se) * se->avg.runnable_load_sum;
3366         delta_avg = runnable_load_avg - se->avg.runnable_load_avg;
3367 
3368         se->avg.runnable_load_sum = runnable_sum;
3369         se->avg.runnable_load_avg = runnable_load_avg;
3370 
3371         if (se->on_rq) {
3372                 add_positive(&cfs_rq->avg.runnable_load_avg, delta_avg);
3373                 add_positive(&cfs_rq->avg.runnable_load_sum, delta_sum);
3374         }
3375 }
3376 
3377 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3378 {
3379         cfs_rq->propagate = 1;
3380         cfs_rq->prop_runnable_sum += runnable_sum;
3381 }
3382 
3383 /* Update task and its cfs_rq load average */
3384 static inline int propagate_entity_load_avg(struct sched_entity *se)
3385 {
3386         struct cfs_rq *cfs_rq, *gcfs_rq;
3387 
3388         if (entity_is_task(se))
3389                 return 0;
3390 
3391         gcfs_rq = group_cfs_rq(se);
3392         if (!gcfs_rq->propagate)
3393                 return 0;
3394 
3395         gcfs_rq->propagate = 0;
3396 
3397         cfs_rq = cfs_rq_of(se);
3398 
3399         add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3400 
3401         update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3402         update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3403 
3404         trace_pelt_cfs_tp(cfs_rq);
3405         trace_pelt_se_tp(se);
3406 
3407         return 1;
3408 }
3409 
3410 /*
3411  * Check if we need to update the load and the utilization of a blocked
3412  * group_entity:
3413  */
3414 static inline bool skip_blocked_update(struct sched_entity *se)
3415 {
3416         struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3417 
3418         /*
3419          * If sched_entity still have not zero load or utilization, we have to
3420          * decay it:
3421          */
3422         if (se->avg.load_avg || se->avg.util_avg)
3423                 return false;
3424 
3425         /*
3426          * If there is a pending propagation, we have to update the load and
3427          * the utilization of the sched_entity:
3428          */
3429         if (gcfs_rq->propagate)
3430                 return false;
3431 
3432         /*
3433          * Otherwise, the load and the utilization of the sched_entity is
3434          * already zero and there is no pending propagation, so it will be a
3435          * waste of time to try to decay it:
3436          */
3437         return true;
3438 }
3439 
3440 #else /* CONFIG_FAIR_GROUP_SCHED */
3441 
3442 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3443 
3444 static inline int propagate_entity_load_avg(struct sched_entity *se)
3445 {
3446         return 0;
3447 }
3448 
3449 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3450 
3451 #endif /* CONFIG_FAIR_GROUP_SCHED */
3452 
3453 /**
3454  * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3455  * @now: current time, as per cfs_rq_clock_pelt()
3456  * @cfs_rq: cfs_rq to update
3457  *
3458  * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3459  * avg. The immediate corollary is that all (fair) tasks must be attached, see
3460  * post_init_entity_util_avg().
3461  *
3462  * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3463  *
3464  * Returns true if the load decayed or we removed load.
3465  *
3466  * Since both these conditions indicate a changed cfs_rq->avg.load we should
3467  * call update_tg_load_avg() when this function returns true.
3468  */
3469 static inline int
3470 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3471 {
3472         unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0;
3473         struct sched_avg *sa = &cfs_rq->avg;
3474         int decayed = 0;
3475 
3476         if (cfs_rq->removed.nr) {
3477                 unsigned long r;
3478                 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3479 
3480                 raw_spin_lock(&cfs_rq->removed.lock);
3481                 swap(cfs_rq->removed.util_avg, removed_util);
3482                 swap(cfs_rq->removed.load_avg, removed_load);
3483                 swap(cfs_rq->removed.runnable_sum, removed_runnable_sum);
3484                 cfs_rq->removed.nr = 0;
3485                 raw_spin_unlock(&cfs_rq->removed.lock);
3486 
3487                 r = removed_load;
3488                 sub_positive(&sa->load_avg, r);
3489                 sub_positive(&sa->load_sum, r * divider);
3490 
3491                 r = removed_util;
3492                 sub_positive(&sa->util_avg, r);
3493                 sub_positive(&sa->util_sum, r * divider);
3494 
3495                 add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum);
3496 
3497                 decayed = 1;
3498         }
3499 
3500         decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
3501 
3502 #ifndef CONFIG_64BIT
3503         smp_wmb();
3504         cfs_rq->load_last_update_time_copy = sa->last_update_time;
3505 #endif
3506 
3507         if (decayed)
3508                 cfs_rq_util_change(cfs_rq, 0);
3509 
3510         return decayed;
3511 }
3512 
3513 /**
3514  * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3515  * @cfs_rq: cfs_rq to attach to
3516  * @se: sched_entity to attach
3517  * @flags: migration hints
3518  *
3519  * Must call update_cfs_rq_load_avg() before this, since we rely on
3520  * cfs_rq->avg.last_update_time being current.
3521  */
3522 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3523 {
3524         u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3525 
3526         /*
3527          * When we attach the @se to the @cfs_rq, we must align the decay
3528          * window because without that, really weird and wonderful things can
3529          * happen.
3530          *
3531          * XXX illustrate
3532          */
3533         se->avg.last_update_time = cfs_rq->avg.last_update_time;
3534         se->avg.period_contrib = cfs_rq->avg.period_contrib;
3535 
3536         /*
3537          * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3538          * period_contrib. This isn't strictly correct, but since we're
3539          * entirely outside of the PELT hierarchy, nobody cares if we truncate
3540          * _sum a little.
3541          */
3542         se->avg.util_sum = se->avg.util_avg * divider;
3543 
3544         se->avg.load_sum = divider;
3545         if (se_weight(se)) {
3546                 se->avg.load_sum =
3547                         div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3548         }
3549 
3550         se->avg.runnable_load_sum = se->avg.load_sum;
3551 
3552         enqueue_load_avg(cfs_rq, se);
3553         cfs_rq->avg.util_avg += se->avg.util_avg;
3554         cfs_rq->avg.util_sum += se->avg.util_sum;
3555 
3556         add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3557 
3558         cfs_rq_util_change(cfs_rq, flags);
3559 
3560         trace_pelt_cfs_tp(cfs_rq);
3561 }
3562 
3563 /**
3564  * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3565  * @cfs_rq: cfs_rq to detach from
3566  * @se: sched_entity to detach
3567  *
3568  * Must call update_cfs_rq_load_avg() before this, since we rely on
3569  * cfs_rq->avg.last_update_time being current.
3570  */
3571 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3572 {
3573         dequeue_load_avg(cfs_rq, se);
3574         sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3575         sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3576 
3577         add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3578 
3579         cfs_rq_util_change(cfs_rq, 0);
3580 
3581         trace_pelt_cfs_tp(cfs_rq);
3582 }
3583 
3584 /*
3585  * Optional action to be done while updating the load average
3586  */
3587 #define UPDATE_TG       0x1
3588 #define SKIP_AGE_LOAD   0x2
3589 #define DO_ATTACH       0x4
3590 
3591 /* Update task and its cfs_rq load average */
3592 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3593 {
3594         u64 now = cfs_rq_clock_pelt(cfs_rq);
3595         int decayed;
3596 
3597         /*
3598          * Track task load average for carrying it to new CPU after migrated, and
3599          * track group sched_entity load average for task_h_load calc in migration
3600          */
3601         if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3602                 __update_load_avg_se(now, cfs_rq, se);
3603 
3604         decayed  = update_cfs_rq_load_avg(now, cfs_rq);
3605         decayed |= propagate_entity_load_avg(se);
3606 
3607         if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3608 
3609                 /*
3610                  * DO_ATTACH means we're here from enqueue_entity().
3611                  * !last_update_time means we've passed through
3612                  * migrate_task_rq_fair() indicating we migrated.
3613                  *
3614                  * IOW we're enqueueing a task on a new CPU.
3615                  */
3616                 attach_entity_load_avg(cfs_rq, se, SCHED_CPUFREQ_MIGRATION);
3617                 update_tg_load_avg(cfs_rq, 0);
3618 
3619         } else if (decayed && (flags & UPDATE_TG))
3620                 update_tg_load_avg(cfs_rq, 0);
3621 }
3622 
3623 #ifndef CONFIG_64BIT
3624 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3625 {
3626         u64 last_update_time_copy;
3627         u64 last_update_time;
3628 
3629         do {
3630                 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3631                 smp_rmb();
3632                 last_update_time = cfs_rq->avg.last_update_time;
3633         } while (last_update_time != last_update_time_copy);
3634 
3635         return last_update_time;
3636 }
3637 #else
3638 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3639 {
3640         return cfs_rq->avg.last_update_time;
3641 }
3642 #endif
3643 
3644 /*
3645  * Synchronize entity load avg of dequeued entity without locking
3646  * the previous rq.
3647  */
3648 static void sync_entity_load_avg(struct sched_entity *se)
3649 {
3650         struct cfs_rq *cfs_rq = cfs_rq_of(se);
3651         u64 last_update_time;
3652 
3653         last_update_time = cfs_rq_last_update_time(cfs_rq);
3654         __update_load_avg_blocked_se(last_update_time, se);
3655 }
3656 
3657 /*
3658  * Task first catches up with cfs_rq, and then subtract
3659  * itself from the cfs_rq (task must be off the queue now).
3660  */
3661 static void remove_entity_load_avg(struct sched_entity *se)
3662 {
3663         struct cfs_rq *cfs_rq = cfs_rq_of(se);
3664         unsigned long flags;
3665 
3666         /*
3667          * tasks cannot exit without having gone through wake_up_new_task() ->
3668          * post_init_entity_util_avg() which will have added things to the
3669          * cfs_rq, so we can remove unconditionally.
3670          */
3671 
3672         sync_entity_load_avg(se);
3673 
3674         raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3675         ++cfs_rq->removed.nr;
3676         cfs_rq->removed.util_avg        += se->avg.util_avg;
3677         cfs_rq->removed.load_avg        += se->avg.load_avg;
3678         cfs_rq->removed.runnable_sum    += se->avg.load_sum; /* == runnable_sum */
3679         raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3680 }
3681 
3682 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3683 {
3684         return cfs_rq->avg.runnable_load_avg;
3685 }
3686 
3687 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3688 {
3689         return cfs_rq->avg.load_avg;
3690 }
3691 
3692 static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3693 
3694 static inline unsigned long task_util(struct task_struct *p)
3695 {
3696         return READ_ONCE(p->se.avg.util_avg);
3697 }
3698 
3699 static inline unsigned long _task_util_est(struct task_struct *p)
3700 {
3701         struct util_est ue = READ_ONCE(p->se.avg.util_est);
3702 
3703         return (max(ue.ewma, ue.enqueued) | UTIL_AVG_UNCHANGED);
3704 }
3705 
3706 static inline unsigned long task_util_est(struct task_struct *p)
3707 {
3708         return max(task_util(p), _task_util_est(p));
3709 }
3710 
3711 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3712                                     struct task_struct *p)
3713 {
3714         unsigned int enqueued;
3715 
3716         if (!sched_feat(UTIL_EST))
3717                 return;
3718 
3719         /* Update root cfs_rq's estimated utilization */
3720         enqueued  = cfs_rq->avg.util_est.enqueued;
3721         enqueued += _task_util_est(p);
3722         WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3723 }
3724 
3725 /*
3726  * Check if a (signed) value is within a specified (unsigned) margin,
3727  * based on the observation that:
3728  *
3729  *     abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3730  *
3731  * NOTE: this only works when value + maring < INT_MAX.
3732  */
3733 static inline bool within_margin(int value, int margin)
3734 {
3735         return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3736 }
3737 
3738 static void
3739 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep)
3740 {
3741         long last_ewma_diff;
3742         struct util_est ue;
3743         int cpu;
3744 
3745         if (!sched_feat(UTIL_EST))
3746                 return;
3747 
3748         /* Update root cfs_rq's estimated utilization */
3749         ue.enqueued  = cfs_rq->avg.util_est.enqueued;
3750         ue.enqueued -= min_t(unsigned int, ue.enqueued, _task_util_est(p));
3751         WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued);
3752 
3753         /*
3754          * Skip update of task's estimated utilization when the task has not
3755          * yet completed an activation, e.g. being migrated.
3756          */
3757         if (!task_sleep)
3758                 return;
3759 
3760         /*
3761          * If the PELT values haven't changed since enqueue time,
3762          * skip the util_est update.
3763          */
3764         ue = p->se.avg.util_est;
3765         if (ue.enqueued & UTIL_AVG_UNCHANGED)
3766                 return;
3767 
3768         /*
3769          * Skip update of task's estimated utilization when its EWMA is
3770          * already ~1% close to its last activation value.
3771          */
3772         ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
3773         last_ewma_diff = ue.enqueued - ue.ewma;
3774         if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
3775                 return;
3776 
3777         /*
3778          * To avoid overestimation of actual task utilization, skip updates if
3779          * we cannot grant there is idle time in this CPU.
3780          */
3781         cpu = cpu_of(rq_of(cfs_rq));
3782         if (task_util(p) > capacity_orig_of(cpu))
3783                 return;
3784 
3785         /*
3786          * Update Task's estimated utilization
3787          *
3788          * When *p completes an activation we can consolidate another sample
3789          * of the task size. This is done by storing the current PELT value
3790          * as ue.enqueued and by using this value to update the Exponential
3791          * Weighted Moving Average (EWMA):
3792          *
3793          *  ewma(t) = w *  task_util(p) + (1-w) * ewma(t-1)
3794          *          = w *  task_util(p) +         ewma(t-1)  - w * ewma(t-1)
3795          *          = w * (task_util(p) -         ewma(t-1)) +     ewma(t-1)
3796          *          = w * (      last_ewma_diff            ) +     ewma(t-1)
3797          *          = w * (last_ewma_diff  +  ewma(t-1) / w)
3798          *
3799          * Where 'w' is the weight of new samples, which is configured to be
3800          * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
3801          */
3802         ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
3803         ue.ewma  += last_ewma_diff;
3804         ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
3805         WRITE_ONCE(p->se.avg.util_est, ue);
3806 }
3807 
3808 static inline int task_fits_capacity(struct task_struct *p, long capacity)
3809 {
3810         return capacity * 1024 > task_util_est(p) * capacity_margin;
3811 }
3812 
3813 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
3814 {
3815         if (!static_branch_unlikely(&sched_asym_cpucapacity))
3816                 return;
3817 
3818         if (!p) {
3819                 rq->misfit_task_load = 0;
3820                 return;
3821         }
3822 
3823         if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
3824                 rq->misfit_task_load = 0;
3825                 return;
3826         }
3827 
3828         rq->misfit_task_load = task_h_load(p);
3829 }
3830 
3831 #else /* CONFIG_SMP */
3832 
3833 #define UPDATE_TG       0x0
3834 #define SKIP_AGE_LOAD   0x0
3835 #define DO_ATTACH       0x0
3836 
3837 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
3838 {
3839         cfs_rq_util_change(cfs_rq, 0);
3840 }
3841 
3842 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3843 
3844 static inline void
3845 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) {}
3846 static inline void
3847 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3848 
3849 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3850 {
3851         return 0;
3852 }
3853 
3854 static inline void
3855 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
3856 
3857 static inline void
3858 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p,
3859                  bool task_sleep) {}
3860 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
3861 
3862 #endif /* CONFIG_SMP */
3863 
3864 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3865 {
3866 #ifdef CONFIG_SCHED_DEBUG
3867         s64 d = se->vruntime - cfs_rq->min_vruntime;
3868 
3869         if (d < 0)
3870                 d = -d;
3871 
3872         if (d > 3*sysctl_sched_latency)
3873                 schedstat_inc(cfs_rq->nr_spread_over);
3874 #endif
3875 }
3876 
3877 static void
3878 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3879 {
3880         u64 vruntime = cfs_rq->min_vruntime;
3881 
3882         /*
3883          * The 'current' period is already promised to the current tasks,
3884          * however the extra weight of the new task will slow them down a
3885          * little, place the new task so that it fits in the slot that
3886          * stays open at the end.
3887          */
3888         if (initial && sched_feat(START_DEBIT))
3889                 vruntime += sched_vslice(cfs_rq, se);
3890 
3891         /* sleeps up to a single latency don't count. */
3892         if (!initial) {
3893                 unsigned long thresh = sysctl_sched_latency;
3894 
3895                 /*
3896                  * Halve their sleep time's effect, to allow
3897                  * for a gentler effect of sleepers:
3898                  */
3899                 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3900                         thresh >>= 1;
3901 
3902                 vruntime -= thresh;
3903         }
3904 
3905         /* ensure we never gain time by being placed backwards. */
3906         se->vruntime = max_vruntime(se->vruntime, vruntime);
3907 }
3908 
3909 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3910 
3911 static inline void check_schedstat_required(void)
3912 {
3913 #ifdef CONFIG_SCHEDSTATS
3914         if (schedstat_enabled())
3915                 return;
3916 
3917         /* Force schedstat enabled if a dependent tracepoint is active */
3918         if (trace_sched_stat_wait_enabled()    ||
3919                         trace_sched_stat_sleep_enabled()   ||
3920                         trace_sched_stat_iowait_enabled()  ||
3921                         trace_sched_stat_blocked_enabled() ||
3922                         trace_sched_stat_runtime_enabled())  {
3923                 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3924                              "stat_blocked and stat_runtime require the "
3925                              "kernel parameter schedstats=enable or "
3926                              "kernel.sched_schedstats=1\n");
3927         }
3928 #endif
3929 }
3930 
3931 
3932 /*
3933  * MIGRATION
3934  *
3935  *      dequeue
3936  *        update_curr()
3937  *          update_min_vruntime()
3938  *        vruntime -= min_vruntime
3939  *
3940  *      enqueue
3941  *        update_curr()
3942  *          update_min_vruntime()
3943  *        vruntime += min_vruntime
3944  *
3945  * this way the vruntime transition between RQs is done when both
3946  * min_vruntime are up-to-date.
3947  *
3948  * WAKEUP (remote)
3949  *
3950  *      ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3951  *        vruntime -= min_vruntime
3952  *
3953  *      enqueue
3954  *        update_curr()
3955  *          update_min_vruntime()
3956  *        vruntime += min_vruntime
3957  *
3958  * this way we don't have the most up-to-date min_vruntime on the originating
3959  * CPU and an up-to-date min_vruntime on the destination CPU.
3960  */
3961 
3962 static void
3963 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3964 {
3965         bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3966         bool curr = cfs_rq->curr == se;
3967 
3968         /*
3969          * If we're the current task, we must renormalise before calling
3970          * update_curr().
3971          */
3972         if (renorm && curr)
3973                 se->vruntime += cfs_rq->min_vruntime;
3974 
3975         update_curr(cfs_rq);
3976 
3977         /*
3978          * Otherwise, renormalise after, such that we're placed at the current
3979          * moment in time, instead of some random moment in the past. Being
3980          * placed in the past could significantly boost this task to the
3981          * fairness detriment of existing tasks.
3982          */
3983         if (renorm && !curr)
3984                 se->vruntime += cfs_rq->min_vruntime;
3985 
3986         /*
3987          * When enqueuing a sched_entity, we must:
3988          *   - Update loads to have both entity and cfs_rq synced with now.
3989          *   - Add its load to cfs_rq->runnable_avg
3990          *   - For group_entity, update its weight to reflect the new share of
3991          *     its group cfs_rq
3992          *   - Add its new weight to cfs_rq->load.weight
3993          */
3994         update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
3995         update_cfs_group(se);
3996         enqueue_runnable_load_avg(cfs_rq, se);
3997         account_entity_enqueue(cfs_rq, se);
3998 
3999         if (flags & ENQUEUE_WAKEUP)
4000                 place_entity(cfs_rq, se, 0);
4001 
4002         check_schedstat_required();
4003         update_stats_enqueue(cfs_rq, se, flags);
4004         check_spread(cfs_rq, se);
4005         if (!curr)
4006                 __enqueue_entity(cfs_rq, se);
4007         se->on_rq = 1;
4008 
4009         if (cfs_rq->nr_running == 1) {
4010                 list_add_leaf_cfs_rq(cfs_rq);
4011                 check_enqueue_throttle(cfs_rq);
4012         }
4013 }
4014 
4015 static void __clear_buddies_last(struct sched_entity *se)
4016 {
4017         for_each_sched_entity(se) {
4018                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4019                 if (cfs_rq->last != se)
4020                         break;
4021 
4022                 cfs_rq->last = NULL;
4023         }
4024 }
4025 
4026 static void __clear_buddies_next(struct sched_entity *se)
4027 {
4028         for_each_sched_entity(se) {
4029                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4030                 if (cfs_rq->next != se)
4031                         break;
4032 
4033                 cfs_rq->next = NULL;
4034         }
4035 }
4036 
4037 static void __clear_buddies_skip(struct sched_entity *se)
4038 {
4039         for_each_sched_entity(se) {
4040                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4041                 if (cfs_rq->skip != se)
4042                         break;
4043 
4044                 cfs_rq->skip = NULL;
4045         }
4046 }
4047 
4048 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4049 {
4050         if (cfs_rq->last == se)
4051                 __clear_buddies_last(se);
4052 
4053         if (cfs_rq->next == se)
4054                 __clear_buddies_next(se);
4055 
4056         if (cfs_rq->skip == se)
4057                 __clear_buddies_skip(se);
4058 }
4059 
4060 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4061 
4062 static void
4063 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4064 {
4065         /*
4066          * Update run-time statistics of the 'current'.
4067          */
4068         update_curr(cfs_rq);
4069 
4070         /*
4071          * When dequeuing a sched_entity, we must:
4072          *   - Update loads to have both entity and cfs_rq synced with now.
4073          *   - Subtract its load from the cfs_rq->runnable_avg.
4074          *   - Subtract its previous weight from cfs_rq->load.weight.
4075          *   - For group entity, update its weight to reflect the new share
4076          *     of its group cfs_rq.
4077          */
4078         update_load_avg(cfs_rq, se, UPDATE_TG);
4079         dequeue_runnable_load_avg(cfs_rq, se);
4080 
4081         update_stats_dequeue(cfs_rq, se, flags);
4082 
4083         clear_buddies(cfs_rq, se);
4084 
4085         if (se != cfs_rq->curr)
4086                 __dequeue_entity(cfs_rq, se);
4087         se->on_rq = 0;
4088         account_entity_dequeue(cfs_rq, se);
4089 
4090         /*
4091          * Normalize after update_curr(); which will also have moved
4092          * min_vruntime if @se is the one holding it back. But before doing
4093          * update_min_vruntime() again, which will discount @se's position and
4094          * can move min_vruntime forward still more.
4095          */
4096         if (!(flags & DEQUEUE_SLEEP))
4097                 se->vruntime -= cfs_rq->min_vruntime;
4098 
4099         /* return excess runtime on last dequeue */
4100         return_cfs_rq_runtime(cfs_rq);
4101 
4102         update_cfs_group(se);
4103 
4104         /*
4105          * Now advance min_vruntime if @se was the entity holding it back,
4106          * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4107          * put back on, and if we advance min_vruntime, we'll be placed back
4108          * further than we started -- ie. we'll be penalized.
4109          */
4110         if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4111                 update_min_vruntime(cfs_rq);
4112 }
4113 
4114 /*
4115  * Preempt the current task with a newly woken task if needed:
4116  */
4117 static void
4118 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4119 {
4120         unsigned long ideal_runtime, delta_exec;
4121         struct sched_entity *se;
4122         s64 delta;
4123 
4124         ideal_runtime = sched_slice(cfs_rq, curr);
4125         delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4126         if (delta_exec > ideal_runtime) {
4127                 resched_curr(rq_of(cfs_rq));
4128                 /*
4129                  * The current task ran long enough, ensure it doesn't get
4130                  * re-elected due to buddy favours.
4131                  */
4132                 clear_buddies(cfs_rq, curr);
4133                 return;
4134         }
4135 
4136         /*
4137          * Ensure that a task that missed wakeup preemption by a
4138          * narrow margin doesn't have to wait for a full slice.
4139          * This also mitigates buddy induced latencies under load.
4140          */
4141         if (delta_exec < sysctl_sched_min_granularity)
4142                 return;
4143 
4144         se = __pick_first_entity(cfs_rq);
4145         delta = curr->vruntime - se->vruntime;
4146 
4147         if (delta < 0)
4148                 return;
4149 
4150         if (delta > ideal_runtime)
4151                 resched_curr(rq_of(cfs_rq));
4152 }
4153 
4154 static void
4155 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4156 {
4157         /* 'current' is not kept within the tree. */
4158         if (se->on_rq) {
4159                 /*
4160                  * Any task has to be enqueued before it get to execute on
4161                  * a CPU. So account for the time it spent waiting on the
4162                  * runqueue.
4163                  */
4164                 update_stats_wait_end(cfs_rq, se);
4165                 __dequeue_entity(cfs_rq, se);
4166                 update_load_avg(cfs_rq, se, UPDATE_TG);
4167         }
4168 
4169         update_stats_curr_start(cfs_rq, se);
4170         cfs_rq->curr = se;
4171 
4172         /*
4173          * Track our maximum slice length, if the CPU's load is at
4174          * least twice that of our own weight (i.e. dont track it
4175          * when there are only lesser-weight tasks around):
4176          */
4177         if (schedstat_enabled() &&
4178             rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
4179                 schedstat_set(se->statistics.slice_max,
4180                         max((u64)schedstat_val(se->statistics.slice_max),
4181                             se->sum_exec_runtime - se->prev_sum_exec_runtime));
4182         }
4183 
4184         se->prev_sum_exec_runtime = se->sum_exec_runtime;
4185 }
4186 
4187 static int
4188 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4189 
4190 /*
4191  * Pick the next process, keeping these things in mind, in this order:
4192  * 1) keep things fair between processes/task groups
4193  * 2) pick the "next" process, since someone really wants that to run
4194  * 3) pick the "last" process, for cache locality
4195  * 4) do not run the "skip" process, if something else is available
4196  */
4197 static struct sched_entity *
4198 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4199 {
4200         struct sched_entity *left = __pick_first_entity(cfs_rq);
4201         struct sched_entity *se;
4202 
4203         /*
4204          * If curr is set we have to see if its left of the leftmost entity
4205          * still in the tree, provided there was anything in the tree at all.
4206          */
4207         if (!left || (curr && entity_before(curr, left)))
4208                 left = curr;
4209 
4210         se = left; /* ideally we run the leftmost entity */
4211 
4212         /*
4213          * Avoid running the skip buddy, if running something else can
4214          * be done without getting too unfair.
4215          */
4216         if (cfs_rq->skip == se) {
4217                 struct sched_entity *second;
4218 
4219                 if (se == curr) {
4220                         second = __pick_first_entity(cfs_rq);
4221                 } else {
4222                         second = __pick_next_entity(se);
4223                         if (!second || (curr && entity_before(curr, second)))
4224                                 second = curr;
4225                 }
4226 
4227                 if (second && wakeup_preempt_entity(second, left) < 1)
4228                         se = second;
4229         }
4230 
4231         /*
4232          * Prefer last buddy, try to return the CPU to a preempted task.
4233          */
4234         if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4235                 se = cfs_rq->last;
4236 
4237         /*
4238          * Someone really wants this to run. If it's not unfair, run it.
4239          */
4240         if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4241                 se = cfs_rq->next;
4242 
4243         clear_buddies(cfs_rq, se);
4244 
4245         return se;
4246 }
4247 
4248 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4249 
4250 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4251 {
4252         /*
4253          * If still on the runqueue then deactivate_task()
4254          * was not called and update_curr() has to be done:
4255          */
4256         if (prev->on_rq)
4257                 update_curr(cfs_rq);
4258 
4259         /* throttle cfs_rqs exceeding runtime */
4260         check_cfs_rq_runtime(cfs_rq);
4261 
4262         check_spread(cfs_rq, prev);
4263 
4264         if (prev->on_rq) {
4265                 update_stats_wait_start(cfs_rq, prev);
4266                 /* Put 'current' back into the tree. */
4267                 __enqueue_entity(cfs_rq, prev);
4268                 /* in !on_rq case, update occurred at dequeue */
4269                 update_load_avg(cfs_rq, prev, 0);
4270         }
4271         cfs_rq->curr = NULL;
4272 }
4273 
4274 static void
4275 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4276 {
4277         /*
4278          * Update run-time statistics of the 'current'.
4279          */
4280         update_curr(cfs_rq);
4281 
4282         /*
4283          * Ensure that runnable average is periodically updated.
4284          */
4285         update_load_avg(cfs_rq, curr, UPDATE_TG);
4286         update_cfs_group(curr);
4287 
4288 #ifdef CONFIG_SCHED_HRTICK
4289         /*
4290          * queued ticks are scheduled to match the slice, so don't bother
4291          * validating it and just reschedule.
4292          */
4293         if (queued) {
4294                 resched_curr(rq_of(cfs_rq));
4295                 return;
4296         }
4297         /*
4298          * don't let the period tick interfere with the hrtick preemption
4299          */
4300         if (!sched_feat(DOUBLE_TICK) &&
4301                         hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4302                 return;
4303 #endif
4304 
4305         if (cfs_rq->nr_running > 1)
4306                 check_preempt_tick(cfs_rq, curr);
4307 }
4308 
4309 
4310 /**************************************************
4311  * CFS bandwidth control machinery
4312  */
4313 
4314 #ifdef CONFIG_CFS_BANDWIDTH
4315 
4316 #ifdef CONFIG_JUMP_LABEL
4317 static struct static_key __cfs_bandwidth_used;
4318 
4319 static inline bool cfs_bandwidth_used(void)
4320 {
4321         return static_key_false(&__cfs_bandwidth_used);
4322 }
4323 
4324 void cfs_bandwidth_usage_inc(void)
4325 {
4326         static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4327 }
4328 
4329 void cfs_bandwidth_usage_dec(void)
4330 {
4331         static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4332 }
4333 #else /* CONFIG_JUMP_LABEL */
4334 static bool cfs_bandwidth_used(void)
4335 {
4336         return true;
4337 }
4338 
4339 void cfs_bandwidth_usage_inc(void) {}
4340 void cfs_bandwidth_usage_dec(void) {}
4341 #endif /* CONFIG_JUMP_LABEL */
4342 
4343 /*
4344  * default period for cfs group bandwidth.
4345  * default: 0.1s, units: nanoseconds
4346  */
4347 static inline u64 default_cfs_period(void)
4348 {
4349         return 100000000ULL;
4350 }
4351 
4352 static inline u64 sched_cfs_bandwidth_slice(void)
4353 {
4354         return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4355 }
4356 
4357 /*
4358  * Replenish runtime according to assigned quota. We use sched_clock_cpu
4359  * directly instead of rq->clock to avoid adding additional synchronization
4360  * around rq->lock.
4361  *
4362  * requires cfs_b->lock
4363  */
4364 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4365 {
4366         if (cfs_b->quota != RUNTIME_INF)
4367                 cfs_b->runtime = cfs_b->quota;
4368 }
4369 
4370 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4371 {
4372         return &tg->cfs_bandwidth;
4373 }
4374 
4375 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4376 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4377 {
4378         if (unlikely(cfs_rq->throttle_count))
4379                 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4380 
4381         return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4382 }
4383 
4384 /* returns 0 on failure to allocate runtime */
4385 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4386 {
4387         struct task_group *tg = cfs_rq->tg;
4388         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4389         u64 amount = 0, min_amount;
4390 
4391         /* note: this is a positive sum as runtime_remaining <= 0 */
4392         min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4393 
4394         raw_spin_lock(&cfs_b->lock);
4395         if (cfs_b->quota == RUNTIME_INF)
4396                 amount = min_amount;
4397         else {
4398                 start_cfs_bandwidth(cfs_b);
4399 
4400                 if (cfs_b->runtime > 0) {
4401                         amount = min(cfs_b->runtime, min_amount);
4402                         cfs_b->runtime -= amount;
4403                         cfs_b->idle = 0;
4404                 }
4405         }
4406         raw_spin_unlock(&cfs_b->lock);
4407 
4408         cfs_rq->runtime_remaining += amount;
4409 
4410         return cfs_rq->runtime_remaining > 0;
4411 }
4412 
4413 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4414 {
4415         /* dock delta_exec before expiring quota (as it could span periods) */
4416         cfs_rq->runtime_remaining -= delta_exec;
4417 
4418         if (likely(cfs_rq->runtime_remaining > 0))
4419                 return;
4420 
4421         if (cfs_rq->throttled)
4422                 return;
4423         /*
4424          * if we're unable to extend our runtime we resched so that the active
4425          * hierarchy can be throttled
4426          */
4427         if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4428                 resched_curr(rq_of(cfs_rq));
4429 }
4430 
4431 static __always_inline
4432 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4433 {
4434         if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4435                 return;
4436 
4437         __account_cfs_rq_runtime(cfs_rq, delta_exec);
4438 }
4439 
4440 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4441 {
4442         return cfs_bandwidth_used() && cfs_rq->throttled;
4443 }
4444 
4445 /* check whether cfs_rq, or any parent, is throttled */
4446 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4447 {
4448         return cfs_bandwidth_used() && cfs_rq->throttle_count;
4449 }
4450 
4451 /*
4452  * Ensure that neither of the group entities corresponding to src_cpu or
4453  * dest_cpu are members of a throttled hierarchy when performing group
4454  * load-balance operations.
4455  */
4456 static inline int throttled_lb_pair(struct task_group *tg,
4457                                     int src_cpu, int dest_cpu)
4458 {
4459         struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4460 
4461         src_cfs_rq = tg->cfs_rq[src_cpu];
4462         dest_cfs_rq = tg->cfs_rq[dest_cpu];
4463 
4464         return throttled_hierarchy(src_cfs_rq) ||
4465                throttled_hierarchy(dest_cfs_rq);
4466 }
4467 
4468 static int tg_unthrottle_up(struct task_group *tg, void *data)
4469 {
4470         struct rq *rq = data;
4471         struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4472 
4473         cfs_rq->throttle_count--;
4474         if (!cfs_rq->throttle_count) {
4475                 /* adjust cfs_rq_clock_task() */
4476                 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4477                                              cfs_rq->throttled_clock_task;
4478 
4479                 /* Add cfs_rq with already running entity in the list */
4480                 if (cfs_rq->nr_running >= 1)
4481                         list_add_leaf_cfs_rq(cfs_rq);
4482         }
4483 
4484         return 0;
4485 }
4486 
4487 static int tg_throttle_down(struct task_group *tg, void *data)
4488 {
4489         struct rq *rq = data;
4490         struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4491 
4492         /* group is entering throttled state, stop time */
4493         if (!cfs_rq->throttle_count) {
4494                 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4495                 list_del_leaf_cfs_rq(cfs_rq);
4496         }
4497         cfs_rq->throttle_count++;
4498 
4499         return 0;
4500 }
4501 
4502 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4503 {
4504         struct rq *rq = rq_of(cfs_rq);
4505         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4506         struct sched_entity *se;
4507         long task_delta, dequeue = 1;
4508         bool empty;
4509 
4510         se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4511 
4512         /* freeze hierarchy runnable averages while throttled */
4513         rcu_read_lock();
4514         walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4515         rcu_read_unlock();
4516 
4517         task_delta = cfs_rq->h_nr_running;
4518         for_each_sched_entity(se) {
4519                 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4520                 /* throttled entity or throttle-on-deactivate */
4521                 if (!se->on_rq)
4522                         break;
4523 
4524                 if (dequeue)
4525                         dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4526                 qcfs_rq->h_nr_running -= task_delta;
4527 
4528                 if (qcfs_rq->load.weight)
4529                         dequeue = 0;
4530         }
4531 
4532         if (!se)
4533                 sub_nr_running(rq, task_delta);
4534 
4535         cfs_rq->throttled = 1;
4536         cfs_rq->throttled_clock = rq_clock(rq);
4537         raw_spin_lock(&cfs_b->lock);
4538         empty = list_empty(&cfs_b->throttled_cfs_rq);
4539 
4540         /*
4541          * Add to the _head_ of the list, so that an already-started
4542          * distribute_cfs_runtime will not see us. If disribute_cfs_runtime is
4543          * not running add to the tail so that later runqueues don't get starved.
4544          */
4545         if (cfs_b->distribute_running)
4546                 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4547         else
4548                 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4549 
4550         /*
4551          * If we're the first throttled task, make sure the bandwidth
4552          * timer is running.
4553          */
4554         if (empty)
4555                 start_cfs_bandwidth(cfs_b);
4556 
4557         raw_spin_unlock(&cfs_b->lock);
4558 }
4559 
4560 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4561 {
4562         struct rq *rq = rq_of(cfs_rq);
4563         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4564         struct sched_entity *se;
4565         int enqueue = 1;
4566         long task_delta;
4567 
4568         se = cfs_rq->tg->se[cpu_of(rq)];
4569 
4570         cfs_rq->throttled = 0;
4571 
4572         update_rq_clock(rq);
4573 
4574         raw_spin_lock(&cfs_b->lock);
4575         cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4576         list_del_rcu(&cfs_rq->throttled_list);
4577         raw_spin_unlock(&cfs_b->lock);
4578 
4579         /* update hierarchical throttle state */
4580         walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4581 
4582         if (!cfs_rq->load.weight)
4583                 return;
4584 
4585         task_delta = cfs_rq->h_nr_running;
4586         for_each_sched_entity(se) {
4587                 if (se->on_rq)
4588                         enqueue = 0;
4589 
4590                 cfs_rq = cfs_rq_of(se);
4591                 if (enqueue)
4592                         enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4593                 cfs_rq->h_nr_running += task_delta;
4594 
4595                 if (cfs_rq_throttled(cfs_rq))
4596                         break;
4597         }
4598 
4599         assert_list_leaf_cfs_rq(rq);
4600 
4601         if (!se)
4602                 add_nr_running(rq, task_delta);
4603 
4604         /* Determine whether we need to wake up potentially idle CPU: */
4605         if (rq->curr == rq->idle && rq->cfs.nr_running)
4606                 resched_curr(rq);
4607 }
4608 
4609 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, u64 remaining)
4610 {
4611         struct cfs_rq *cfs_rq;
4612         u64 runtime;
4613         u64 starting_runtime = remaining;
4614 
4615         rcu_read_lock();
4616         list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4617                                 throttled_list) {
4618                 struct rq *rq = rq_of(cfs_rq);
4619                 struct rq_flags rf;
4620 
4621                 rq_lock_irqsave(rq, &rf);
4622                 if (!cfs_rq_throttled(cfs_rq))
4623                         goto next;
4624 
4625                 /* By the above check, this should never be true */
4626                 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
4627 
4628                 runtime = -cfs_rq->runtime_remaining + 1;
4629                 if (runtime > remaining)
4630                         runtime = remaining;
4631                 remaining -= runtime;
4632 
4633                 cfs_rq->runtime_remaining += runtime;
4634 
4635                 /* we check whether we're throttled above */
4636                 if (cfs_rq->runtime_remaining > 0)
4637                         unthrottle_cfs_rq(cfs_rq);
4638 
4639 next:
4640                 rq_unlock_irqrestore(rq, &rf);
4641 
4642                 if (!remaining)
4643                         break;
4644         }
4645         rcu_read_unlock();
4646 
4647         return starting_runtime - remaining;
4648 }
4649 
4650 /*
4651  * Responsible for refilling a task_group's bandwidth and unthrottling its
4652  * cfs_rqs as appropriate. If there has been no activity within the last
4653  * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4654  * used to track this state.
4655  */
4656 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
4657 {
4658         u64 runtime;
4659         int throttled;
4660 
4661         /* no need to continue the timer with no bandwidth constraint */
4662         if (cfs_b->quota == RUNTIME_INF)
4663                 goto out_deactivate;
4664 
4665         throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4666         cfs_b->nr_periods += overrun;
4667 
4668         /*
4669          * idle depends on !throttled (for the case of a large deficit), and if
4670          * we're going inactive then everything else can be deferred
4671          */
4672         if (cfs_b->idle && !throttled)
4673                 goto out_deactivate;
4674 
4675         __refill_cfs_bandwidth_runtime(cfs_b);
4676 
4677         if (!throttled) {
4678                 /* mark as potentially idle for the upcoming period */
4679                 cfs_b->idle = 1;
4680                 return 0;
4681         }
4682 
4683         /* account preceding periods in which throttling occurred */
4684         cfs_b->nr_throttled += overrun;
4685 
4686         /*
4687          * This check is repeated as we are holding onto the new bandwidth while
4688          * we unthrottle. This can potentially race with an unthrottled group
4689          * trying to acquire new bandwidth from the global pool. This can result
4690          * in us over-using our runtime if it is all used during this loop, but
4691          * only by limited amounts in that extreme case.
4692          */
4693         while (throttled && cfs_b->runtime > 0 && !cfs_b->distribute_running) {
4694                 runtime = cfs_b->runtime;
4695                 cfs_b->distribute_running = 1;
4696                 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4697                 /* we can't nest cfs_b->lock while distributing bandwidth */
4698                 runtime = distribute_cfs_runtime(cfs_b, runtime);
4699                 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4700 
4701                 cfs_b->distribute_running = 0;
4702                 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4703 
4704                 lsub_positive(&cfs_b->runtime, runtime);
4705         }
4706 
4707         /*
4708          * While we are ensured activity in the period following an
4709          * unthrottle, this also covers the case in which the new bandwidth is
4710          * insufficient to cover the existing bandwidth deficit.  (Forcing the
4711          * timer to remain active while there are any throttled entities.)
4712          */
4713         cfs_b->idle = 0;
4714 
4715         return 0;
4716 
4717 out_deactivate:
4718         return 1;
4719 }
4720 
4721 /* a cfs_rq won't donate quota below this amount */
4722 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4723 /* minimum remaining period time to redistribute slack quota */
4724 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4725 /* how long we wait to gather additional slack before distributing */
4726 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4727 
4728 /*
4729  * Are we near the end of the current quota period?
4730  *
4731  * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4732  * hrtimer base being cleared by hrtimer_start. In the case of
4733  * migrate_hrtimers, base is never cleared, so we are fine.
4734  */
4735 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4736 {
4737         struct hrtimer *refresh_timer = &cfs_b->period_timer;
4738         u64 remaining;
4739 
4740         /* if the call-back is running a quota refresh is already occurring */
4741         if (hrtimer_callback_running(refresh_timer))
4742                 return 1;
4743 
4744         /* is a quota refresh about to occur? */
4745         remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4746         if (remaining < min_expire)
4747                 return 1;
4748 
4749         return 0;
4750 }
4751 
4752 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4753 {
4754         u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4755 
4756         /* if there's a quota refresh soon don't bother with slack */
4757         if (runtime_refresh_within(cfs_b, min_left))
4758                 return;
4759 
4760         /* don't push forwards an existing deferred unthrottle */
4761         if (cfs_b->slack_started)
4762                 return;
4763         cfs_b->slack_started = true;
4764 
4765         hrtimer_start(&cfs_b->slack_timer,
4766                         ns_to_ktime(cfs_bandwidth_slack_period),
4767                         HRTIMER_MODE_REL);
4768 }
4769 
4770 /* we know any runtime found here is valid as update_curr() precedes return */
4771 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4772 {
4773         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4774         s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4775 
4776         if (slack_runtime <= 0)
4777                 return;
4778 
4779         raw_spin_lock(&cfs_b->lock);
4780         if (cfs_b->quota != RUNTIME_INF) {
4781                 cfs_b->runtime += slack_runtime;
4782 
4783                 /* we are under rq->lock, defer unthrottling using a timer */
4784                 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4785                     !list_empty(&cfs_b->throttled_cfs_rq))
4786                         start_cfs_slack_bandwidth(cfs_b);
4787         }
4788         raw_spin_unlock(&cfs_b->lock);
4789 
4790         /* even if it's not valid for return we don't want to try again */
4791         cfs_rq->runtime_remaining -= slack_runtime;
4792 }
4793 
4794 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4795 {
4796         if (!cfs_bandwidth_used())
4797                 return;
4798 
4799         if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4800                 return;
4801 
4802         __return_cfs_rq_runtime(cfs_rq);
4803 }
4804 
4805 /*
4806  * This is done with a timer (instead of inline with bandwidth return) since
4807  * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4808  */
4809 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4810 {
4811         u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4812         unsigned long flags;
4813 
4814         /* confirm we're still not at a refresh boundary */
4815         raw_spin_lock_irqsave(&cfs_b->lock, flags);
4816         cfs_b->slack_started = false;
4817         if (cfs_b->distribute_running) {
4818                 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4819                 return;
4820         }
4821 
4822         if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4823                 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4824                 return;
4825         }
4826 
4827         if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4828                 runtime = cfs_b->runtime;
4829 
4830         if (runtime)
4831                 cfs_b->distribute_running = 1;
4832 
4833         raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4834 
4835         if (!runtime)
4836                 return;
4837 
4838         runtime = distribute_cfs_runtime(cfs_b, runtime);
4839 
4840         raw_spin_lock_irqsave(&cfs_b->lock, flags);
4841         lsub_positive(&cfs_b->runtime, runtime);
4842         cfs_b->distribute_running = 0;
4843         raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4844 }
4845 
4846 /*
4847  * When a group wakes up we want to make sure that its quota is not already
4848  * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4849  * runtime as update_curr() throttling can not not trigger until it's on-rq.
4850  */
4851 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4852 {
4853         if (!cfs_bandwidth_used())
4854                 return;
4855 
4856         /* an active group must be handled by the update_curr()->put() path */
4857         if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4858                 return;
4859 
4860         /* ensure the group is not already throttled */
4861         if (cfs_rq_throttled(cfs_rq))
4862                 return;
4863 
4864         /* update runtime allocation */
4865         account_cfs_rq_runtime(cfs_rq, 0);
4866         if (cfs_rq->runtime_remaining <= 0)
4867                 throttle_cfs_rq(cfs_rq);
4868 }
4869 
4870 static void sync_throttle(struct task_group *tg, int cpu)
4871 {
4872         struct cfs_rq *pcfs_rq, *cfs_rq;
4873 
4874         if (!cfs_bandwidth_used())
4875                 return;
4876 
4877         if (!tg->parent)
4878                 return;
4879 
4880         cfs_rq = tg->cfs_rq[cpu];
4881         pcfs_rq = tg->parent->cfs_rq[cpu];
4882 
4883         cfs_rq->throttle_count = pcfs_rq->throttle_count;
4884         cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4885 }
4886 
4887 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4888 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4889 {
4890         if (!cfs_bandwidth_used())
4891                 return false;
4892 
4893         if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4894                 return false;
4895 
4896         /*
4897          * it's possible for a throttled entity to be forced into a running
4898          * state (e.g. set_curr_task), in this case we're finished.
4899          */
4900         if (cfs_rq_throttled(cfs_rq))
4901                 return true;
4902 
4903         throttle_cfs_rq(cfs_rq);
4904         return true;
4905 }
4906 
4907 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4908 {
4909         struct cfs_bandwidth *cfs_b =
4910                 container_of(timer, struct cfs_bandwidth, slack_timer);
4911 
4912         do_sched_cfs_slack_timer(cfs_b);
4913 
4914         return HRTIMER_NORESTART;
4915 }
4916 
4917 extern const u64 max_cfs_quota_period;
4918 
4919 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4920 {
4921         struct cfs_bandwidth *cfs_b =
4922                 container_of(timer, struct cfs_bandwidth, period_timer);
4923         unsigned long flags;
4924         int overrun;
4925         int idle = 0;
4926         int count = 0;
4927 
4928         raw_spin_lock_irqsave(&cfs_b->lock, flags);
4929         for (;;) {
4930                 overrun = hrtimer_forward_now(timer, cfs_b->period);
4931                 if (!overrun)
4932                         break;
4933 
4934                 if (++count > 3) {
4935                         u64 new, old = ktime_to_ns(cfs_b->period);
4936 
4937                         /*
4938                          * Grow period by a factor of 2 to avoid losing precision.
4939                          * Precision loss in the quota/period ratio can cause __cfs_schedulable
4940                          * to fail.
4941                          */
4942                         new = old * 2;
4943                         if (new < max_cfs_quota_period) {
4944                                 cfs_b->period = ns_to_ktime(new);
4945                                 cfs_b->quota *= 2;
4946 
4947                                 pr_warn_ratelimited(
4948         "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
4949                                         smp_processor_id(),
4950                                         div_u64(new, NSEC_PER_USEC),
4951                                         div_u64(cfs_b->quota, NSEC_PER_USEC));
4952                         } else {
4953                                 pr_warn_ratelimited(
4954         "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
4955                                         smp_processor_id(),
4956                                         div_u64(old, NSEC_PER_USEC),
4957                                         div_u64(cfs_b->quota, NSEC_PER_USEC));
4958                         }
4959 
4960                         /* reset count so we don't come right back in here */
4961                         count = 0;
4962                 }
4963 
4964                 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
4965         }
4966         if (idle)
4967                 cfs_b->period_active = 0;
4968         raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4969 
4970         return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4971 }
4972 
4973 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4974 {
4975         raw_spin_lock_init(&cfs_b->lock);
4976         cfs_b->runtime = 0;
4977         cfs_b->quota = RUNTIME_INF;
4978         cfs_b->period = ns_to_ktime(default_cfs_period());
4979 
4980         INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4981         hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4982         cfs_b->period_timer.function = sched_cfs_period_timer;
4983         hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4984         cfs_b->slack_timer.function = sched_cfs_slack_timer;
4985         cfs_b->distribute_running = 0;
4986         cfs_b->slack_started = false;
4987 }
4988 
4989 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4990 {
4991         cfs_rq->runtime_enabled = 0;
4992         INIT_LIST_HEAD(&cfs_rq->throttled_list);
4993 }
4994 
4995 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4996 {
4997         lockdep_assert_held(&cfs_b->lock);
4998 
4999         if (cfs_b->period_active)
5000                 return;
5001 
5002         cfs_b->period_active = 1;
5003         hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
5004         hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
5005 }
5006 
5007 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5008 {
5009         /* init_cfs_bandwidth() was not called */
5010         if (!cfs_b->throttled_cfs_rq.next)
5011                 return;
5012 
5013         hrtimer_cancel(&cfs_b->period_timer);
5014         hrtimer_cancel(&cfs_b->slack_timer);
5015 }
5016 
5017 /*
5018  * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5019  *
5020  * The race is harmless, since modifying bandwidth settings of unhooked group
5021  * bits doesn't do much.
5022  */
5023 
5024 /* cpu online calback */
5025 static void __maybe_unused update_runtime_enabled(struct rq *rq)
5026 {
5027         struct task_group *tg;
5028 
5029         lockdep_assert_held(&rq->lock);
5030 
5031         rcu_read_lock();
5032         list_for_each_entry_rcu(tg, &task_groups, list) {
5033                 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5034                 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5035 
5036                 raw_spin_lock(&cfs_b->lock);
5037                 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5038                 raw_spin_unlock(&cfs_b->lock);
5039         }
5040         rcu_read_unlock();
5041 }
5042 
5043 /* cpu offline callback */
5044 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5045 {
5046         struct task_group *tg;
5047 
5048         lockdep_assert_held(&rq->lock);
5049 
5050         rcu_read_lock();
5051         list_for_each_entry_rcu(tg, &task_groups, list) {
5052                 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5053 
5054                 if (!cfs_rq->runtime_enabled)
5055                         continue;
5056 
5057                 /*
5058                  * clock_task is not advancing so we just need to make sure
5059                  * there's some valid quota amount
5060                  */
5061                 cfs_rq->runtime_remaining = 1;
5062                 /*
5063                  * Offline rq is schedulable till CPU is completely disabled
5064                  * in take_cpu_down(), so we prevent new cfs throttling here.
5065                  */
5066                 cfs_rq->runtime_enabled = 0;
5067 
5068                 if (cfs_rq_throttled(cfs_rq))
5069                         unthrottle_cfs_rq(cfs_rq);
5070         }
5071         rcu_read_unlock();
5072 }
5073 
5074 #else /* CONFIG_CFS_BANDWIDTH */
5075 
5076 static inline bool cfs_bandwidth_used(void)
5077 {
5078         return false;
5079 }
5080 
5081 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
5082 {
5083         return rq_clock_task(rq_of(cfs_rq));
5084 }
5085 
5086 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5087 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5088 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5089 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5090 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5091 
5092 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5093 {
5094         return 0;
5095 }
5096 
5097 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5098 {
5099         return 0;
5100 }
5101 
5102 static inline int throttled_lb_pair(struct task_group *tg,
5103                                     int src_cpu, int dest_cpu)
5104 {
5105         return 0;
5106 }
5107 
5108 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5109 
5110 #ifdef CONFIG_FAIR_GROUP_SCHED
5111 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5112 #endif
5113 
5114 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5115 {
5116         return NULL;
5117 }
5118 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5119 static inline void update_runtime_enabled(struct rq *rq) {}
5120 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5121 
5122 #endif /* CONFIG_CFS_BANDWIDTH */
5123 
5124 /**************************************************
5125  * CFS operations on tasks:
5126  */
5127 
5128 #ifdef CONFIG_SCHED_HRTICK
5129 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5130 {
5131         struct sched_entity *se = &p->se;
5132         struct cfs_rq *cfs_rq = cfs_rq_of(se);
5133 
5134         SCHED_WARN_ON(task_rq(p) != rq);
5135 
5136         if (rq->cfs.h_nr_running > 1) {
5137                 u64 slice = sched_slice(cfs_rq, se);
5138                 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5139                 s64 delta = slice - ran;
5140 
5141                 if (delta < 0) {
5142                         if (rq->curr == p)
5143                                 resched_curr(rq);
5144                         return;
5145                 }
5146                 hrtick_start(rq, delta);
5147         }
5148 }
5149 
5150 /*
5151  * called from enqueue/dequeue and updates the hrtick when the
5152  * current task is from our class and nr_running is low enough
5153  * to matter.
5154  */
5155 static void hrtick_update(struct rq *rq)
5156 {
5157         struct task_struct *curr = rq->curr;
5158 
5159         if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
5160                 return;
5161 
5162         if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5163                 hrtick_start_fair(rq, curr);
5164 }
5165 #else /* !CONFIG_SCHED_HRTICK */
5166 static inline void
5167 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5168 {
5169 }
5170 
5171 static inline void hrtick_update(struct rq *rq)
5172 {
5173 }
5174 #endif
5175 
5176 #ifdef CONFIG_SMP
5177 static inline unsigned long cpu_util(int cpu);
5178 
5179 static inline bool cpu_overutilized(int cpu)
5180 {
5181         return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5182 }
5183 
5184 static inline void update_overutilized_status(struct rq *rq)
5185 {
5186         if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
5187                 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
5188                 trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
5189         }
5190 }
5191 #else
5192 static inline void update_overutilized_status(struct rq *rq) { }
5193 #endif
5194 
5195 /*
5196  * The enqueue_task method is called before nr_running is
5197  * increased. Here we update the fair scheduling stats and
5198  * then put the task into the rbtree:
5199  */
5200 static void
5201 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5202 {
5203         struct cfs_rq *cfs_rq;
5204         struct sched_entity *se = &p->se;
5205 
5206         /*
5207          * The code below (indirectly) updates schedutil which looks at
5208          * the cfs_rq utilization to select a frequency.
5209          * Let's add the task's estimated utilization to the cfs_rq's
5210          * estimated utilization, before we update schedutil.
5211          */
5212         util_est_enqueue(&rq->cfs, p);
5213 
5214         /*
5215          * If in_iowait is set, the code below may not trigger any cpufreq
5216          * utilization updates, so do it here explicitly with the IOWAIT flag
5217          * passed.
5218          */
5219         if (p->in_iowait)
5220                 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5221 
5222         for_each_sched_entity(se) {
5223                 if (se->on_rq)
5224                         break;
5225                 cfs_rq = cfs_rq_of(se);
5226                 enqueue_entity(cfs_rq, se, flags);
5227 
5228                 /*
5229                  * end evaluation on encountering a throttled cfs_rq
5230                  *
5231                  * note: in the case of encountering a throttled cfs_rq we will
5232                  * post the final h_nr_running increment below.
5233                  */
5234                 if (cfs_rq_throttled(cfs_rq))
5235                         break;
5236                 cfs_rq->h_nr_running++;
5237 
5238                 flags = ENQUEUE_WAKEUP;
5239         }
5240 
5241         for_each_sched_entity(se) {
5242                 cfs_rq = cfs_rq_of(se);
5243                 cfs_rq->h_nr_running++;
5244 
5245                 if (cfs_rq_throttled(cfs_rq))
5246                         break;
5247 
5248                 update_load_avg(cfs_rq, se, UPDATE_TG);
5249                 update_cfs_group(se);
5250         }
5251 
5252         if (!se) {
5253                 add_nr_running(rq, 1);
5254                 /*
5255                  * Since new tasks are assigned an initial util_avg equal to
5256                  * half of the spare capacity of their CPU, tiny tasks have the
5257                  * ability to cross the overutilized threshold, which will
5258                  * result in the load balancer ruining all the task placement
5259                  * done by EAS. As a way to mitigate that effect, do not account
5260                  * for the first enqueue operation of new tasks during the
5261                  * overutilized flag detection.
5262                  *
5263                  * A better way of solving this problem would be to wait for
5264                  * the PELT signals of tasks to converge before taking them
5265                  * into account, but that is not straightforward to implement,
5266                  * and the following generally works well enough in practice.
5267                  */
5268                 if (flags & ENQUEUE_WAKEUP)
5269                         update_overutilized_status(rq);
5270 
5271         }
5272 
5273         if (cfs_bandwidth_used()) {
5274                 /*
5275                  * When bandwidth control is enabled; the cfs_rq_throttled()
5276                  * breaks in the above iteration can result in incomplete
5277                  * leaf list maintenance, resulting in triggering the assertion
5278                  * below.
5279                  */
5280                 for_each_sched_entity(se) {
5281                         cfs_rq = cfs_rq_of(se);
5282 
5283                         if (list_add_leaf_cfs_rq(cfs_rq))
5284                                 break;
5285                 }
5286         }
5287 
5288         assert_list_leaf_cfs_rq(rq);
5289 
5290         hrtick_update(rq);
5291 }
5292 
5293 static void set_next_buddy(struct sched_entity *se);
5294 
5295 /*
5296  * The dequeue_task method is called before nr_running is
5297  * decreased. We remove the task from the rbtree and
5298  * update the fair scheduling stats:
5299  */
5300 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5301 {
5302         struct cfs_rq *cfs_rq;
5303         struct sched_entity *se = &p->se;
5304         int task_sleep = flags & DEQUEUE_SLEEP;
5305 
5306         for_each_sched_entity(se) {
5307                 cfs_rq = cfs_rq_of(se);
5308                 dequeue_entity(cfs_rq, se, flags);
5309 
5310                 /*
5311                  * end evaluation on encountering a throttled cfs_rq
5312                  *
5313                  * note: in the case of encountering a throttled cfs_rq we will
5314                  * post the final h_nr_running decrement below.
5315                 */
5316                 if (cfs_rq_throttled(cfs_rq))
5317                         break;
5318                 cfs_rq->h_nr_running--;
5319 
5320                 /* Don't dequeue parent if it has other entities besides us */
5321                 if (cfs_rq->load.weight) {
5322                         /* Avoid re-evaluating load for this entity: */
5323                         se = parent_entity(se);
5324                         /*
5325                          * Bias pick_next to pick a task from this cfs_rq, as
5326                          * p is sleeping when it is within its sched_slice.
5327                          */
5328                         if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5329                                 set_next_buddy(se);
5330                         break;
5331                 }
5332                 flags |= DEQUEUE_SLEEP;
5333         }
5334 
5335         for_each_sched_entity(se) {
5336                 cfs_rq = cfs_rq_of(se);
5337                 cfs_rq->h_nr_running--;
5338 
5339                 if (cfs_rq_throttled(cfs_rq))
5340                         break;
5341 
5342                 update_load_avg(cfs_rq, se, UPDATE_TG);
5343                 update_cfs_group(se);
5344         }
5345 
5346         if (!se)
5347                 sub_nr_running(rq, 1);
5348 
5349         util_est_dequeue(&rq->cfs, p, task_sleep);
5350         hrtick_update(rq);
5351 }
5352 
5353 #ifdef CONFIG_SMP
5354 
5355 /* Working cpumask for: load_balance, load_balance_newidle. */
5356 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5357 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5358 
5359 #ifdef CONFIG_NO_HZ_COMMON
5360 
5361 static struct {
5362         cpumask_var_t idle_cpus_mask;
5363         atomic_t nr_cpus;
5364         int has_blocked;                /* Idle CPUS has blocked load */
5365         unsigned long next_balance;     /* in jiffy units */
5366         unsigned long next_blocked;     /* Next update of blocked load in jiffies */
5367 } nohz ____cacheline_aligned;
5368 
5369 #endif /* CONFIG_NO_HZ_COMMON */
5370 
5371 static unsigned long cpu_runnable_load(struct rq *rq)
5372 {
5373         return cfs_rq_runnable_load_avg(&rq->cfs);
5374 }
5375 
5376 static unsigned long capacity_of(int cpu)
5377 {
5378         return cpu_rq(cpu)->cpu_capacity;
5379 }
5380 
5381 static unsigned long cpu_avg_load_per_task(int cpu)
5382 {
5383         struct rq *rq = cpu_rq(cpu);
5384         unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5385         unsigned long load_avg = cpu_runnable_load(rq);
5386 
5387         if (nr_running)
5388                 return load_avg / nr_running;
5389 
5390         return 0;
5391 }
5392 
5393 static void record_wakee(struct task_struct *p)
5394 {
5395         /*
5396          * Only decay a single time; tasks that have less then 1 wakeup per
5397          * jiffy will not have built up many flips.
5398          */
5399         if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5400                 current->wakee_flips >>= 1;
5401                 current->wakee_flip_decay_ts = jiffies;
5402         }
5403 
5404         if (current->last_wakee != p) {
5405                 current->last_wakee = p;
5406                 current->wakee_flips++;
5407         }
5408 }
5409 
5410 /*
5411  * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5412  *
5413  * A waker of many should wake a different task than the one last awakened
5414  * at a frequency roughly N times higher than one of its wakees.
5415  *
5416  * In order to determine whether we should let the load spread vs consolidating
5417  * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5418  * partner, and a factor of lls_size higher frequency in the other.
5419  *
5420  * With both conditions met, we can be relatively sure that the relationship is
5421  * non-monogamous, with partner count exceeding socket size.
5422  *
5423  * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5424  * whatever is irrelevant, spread criteria is apparent partner count exceeds
5425  * socket size.
5426  */
5427 static int wake_wide(struct task_struct *p)
5428 {
5429         unsigned int master = current->wakee_flips;
5430         unsigned int slave = p->wakee_flips;
5431         int factor = this_cpu_read(sd_llc_size);
5432 
5433         if (master < slave)
5434                 swap(master, slave);
5435         if (slave < factor || master < slave * factor)
5436                 return 0;
5437         return 1;
5438 }
5439 
5440 /*
5441  * The purpose of wake_affine() is to quickly determine on which CPU we can run
5442  * soonest. For the purpose of speed we only consider the waking and previous
5443  * CPU.
5444  *
5445  * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5446  *                      cache-affine and is (or will be) idle.
5447  *
5448  * wake_affine_weight() - considers the weight to reflect the average
5449  *                        scheduling latency of the CPUs. This seems to work
5450  *                        for the overloaded case.
5451  */
5452 static int
5453 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5454 {
5455         /*
5456          * If this_cpu is idle, it implies the wakeup is from interrupt
5457          * context. Only allow the move if cache is shared. Otherwise an
5458          * interrupt intensive workload could force all tasks onto one
5459          * node depending on the IO topology or IRQ affinity settings.
5460          *
5461          * If the prev_cpu is idle and cache affine then avoid a migration.
5462          * There is no guarantee that the cache hot data from an interrupt
5463          * is more important than cache hot data on the prev_cpu and from
5464          * a cpufreq perspective, it's better to have higher utilisation
5465          * on one CPU.
5466          */
5467         if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5468                 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5469 
5470         if (sync && cpu_rq(this_cpu)->nr_running == 1)
5471                 return this_cpu;
5472 
5473         return nr_cpumask_bits;
5474 }
5475 
5476 static int
5477 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5478                    int this_cpu, int prev_cpu, int sync)
5479 {
5480         s64 this_eff_load, prev_eff_load;
5481         unsigned long task_load;
5482 
5483         this_eff_load = cpu_runnable_load(cpu_rq(this_cpu));
5484 
5485         if (sync) {
5486                 unsigned long current_load = task_h_load(current);
5487 
5488                 if (current_load > this_eff_load)
5489                         return this_cpu;
5490 
5491                 this_eff_load -= current_load;
5492         }
5493 
5494         task_load = task_h_load(p);
5495 
5496         this_eff_load += task_load;
5497         if (sched_feat(WA_BIAS))
5498                 this_eff_load *= 100;
5499         this_eff_load *= capacity_of(prev_cpu);
5500 
5501         prev_eff_load = cpu_runnable_load(cpu_rq(prev_cpu));
5502         prev_eff_load -= task_load;
5503         if (sched_feat(WA_BIAS))
5504                 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5505         prev_eff_load *= capacity_of(this_cpu);
5506 
5507         /*
5508          * If sync, adjust the weight of prev_eff_load such that if
5509          * prev_eff == this_eff that select_idle_sibling() will consider
5510          * stacking the wakee on top of the waker if no other CPU is
5511          * idle.
5512          */
5513         if (sync)
5514                 prev_eff_load += 1;
5515 
5516         return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5517 }
5518 
5519 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5520                        int this_cpu, int prev_cpu, int sync)
5521 {
5522         int target = nr_cpumask_bits;
5523 
5524         if (sched_feat(WA_IDLE))
5525                 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5526 
5527         if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5528                 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5529 
5530         schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5531         if (target == nr_cpumask_bits)
5532                 return prev_cpu;
5533 
5534         schedstat_inc(sd->ttwu_move_affine);
5535         schedstat_inc(p->se.statistics.nr_wakeups_affine);
5536         return target;
5537 }
5538 
5539 static unsigned long cpu_util_without(int cpu, struct task_struct *p);
5540 
5541 static unsigned long capacity_spare_without(int cpu, struct task_struct *p)
5542 {
5543         return max_t(long, capacity_of(cpu) - cpu_util_without(cpu, p), 0);
5544 }
5545 
5546 /*
5547  * find_idlest_group finds and returns the least busy CPU group within the
5548  * domain.
5549  *
5550  * Assumes p is allowed on at least one CPU in sd.
5551  */
5552 static struct sched_group *
5553 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5554                   int this_cpu, int sd_flag)
5555 {
5556         struct sched_group *idlest = NULL, *group = sd->groups;
5557         struct sched_group *most_spare_sg = NULL;
5558         unsigned long min_runnable_load = ULONG_MAX;
5559         unsigned long this_runnable_load = ULONG_MAX;
5560         unsigned long min_avg_load = ULONG_MAX, this_avg_load = ULONG_MAX;
5561         unsigned long most_spare = 0, this_spare = 0;
5562         int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5563         unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5564                                 (sd->imbalance_pct-100) / 100;
5565 
5566         do {
5567                 unsigned long load, avg_load, runnable_load;
5568                 unsigned long spare_cap, max_spare_cap;
5569                 int local_group;
5570                 int i;
5571 
5572                 /* Skip over this group if it has no CPUs allowed */
5573                 if (!cpumask_intersects(sched_group_span(group),
5574                                         p->cpus_ptr))
5575                         continue;
5576 
5577                 local_group = cpumask_test_cpu(this_cpu,
5578                                                sched_group_span(group));
5579 
5580                 /*
5581                  * Tally up the load of all CPUs in the group and find
5582                  * the group containing the CPU with most spare capacity.
5583                  */
5584                 avg_load = 0;
5585                 runnable_load = 0;
5586                 max_spare_cap = 0;
5587 
5588                 for_each_cpu(i, sched_group_span(group)) {
5589                         load = cpu_runnable_load(cpu_rq(i));
5590                         runnable_load += load;
5591 
5592                         avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5593 
5594                         spare_cap = capacity_spare_without(i, p);
5595 
5596                         if (spare_cap > max_spare_cap)
5597                                 max_spare_cap = spare_cap;
5598                 }
5599 
5600                 /* Adjust by relative CPU capacity of the group */
5601                 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5602                                         group->sgc->capacity;
5603                 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5604                                         group->sgc->capacity;
5605 
5606                 if (local_group) {
5607                         this_runnable_load = runnable_load;
5608                         this_avg_load = avg_load;
5609                         this_spare = max_spare_cap;
5610                 } else {
5611                         if (min_runnable_load > (runnable_load + imbalance)) {
5612                                 /*
5613                                  * The runnable load is significantly smaller
5614                                  * so we can pick this new CPU:
5615                                  */
5616                                 min_runnable_load = runnable_load;
5617                                 min_avg_load = avg_load;
5618                                 idlest = group;
5619                         } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5620                                    (100*min_avg_load > imbalance_scale*avg_load)) {
5621                                 /*
5622                                  * The runnable loads are close so take the
5623                                  * blocked load into account through avg_load:
5624                                  */
5625                                 min_avg_load = avg_load;
5626                                 idlest = group;
5627                         }
5628 
5629                         if (most_spare < max_spare_cap) {
5630                                 most_spare = max_spare_cap;
5631                                 most_spare_sg = group;
5632                         }
5633                 }
5634         } while (group = group->next, group != sd->groups);
5635 
5636         /*
5637          * The cross-over point between using spare capacity or least load
5638          * is too conservative for high utilization tasks on partially
5639          * utilized systems if we require spare_capacity > task_util(p),
5640          * so we allow for some task stuffing by using
5641          * spare_capacity > task_util(p)/2.
5642          *
5643          * Spare capacity can't be used for fork because the utilization has
5644          * not been set yet, we must first select a rq to compute the initial
5645          * utilization.
5646          */
5647         if (sd_flag & SD_BALANCE_FORK)
5648                 goto skip_spare;
5649 
5650         if (this_spare > task_util(p) / 2 &&
5651             imbalance_scale*this_spare > 100*most_spare)
5652                 return NULL;
5653 
5654         if (most_spare > task_util(p) / 2)
5655                 return most_spare_sg;
5656 
5657 skip_spare:
5658         if (!idlest)
5659                 return NULL;
5660 
5661         /*
5662          * When comparing groups across NUMA domains, it's possible for the
5663          * local domain to be very lightly loaded relative to the remote
5664          * domains but "imbalance" skews the comparison making remote CPUs
5665          * look much more favourable. When considering cross-domain, add
5666          * imbalance to the runnable load on the remote node and consider
5667          * staying local.
5668          */
5669         if ((sd->flags & SD_NUMA) &&
5670             min_runnable_load + imbalance >= this_runnable_load)
5671                 return NULL;
5672 
5673         if (min_runnable_load > (this_runnable_load + imbalance))
5674                 return NULL;
5675 
5676         if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5677              (100*this_avg_load < imbalance_scale*min_avg_load))
5678                 return NULL;
5679 
5680         return idlest;
5681 }
5682 
5683 /*
5684  * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5685  */
5686 static int
5687 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5688 {
5689         unsigned long load, min_load = ULONG_MAX;
5690         unsigned int min_exit_latency = UINT_MAX;
5691         u64 latest_idle_timestamp = 0;
5692         int least_loaded_cpu = this_cpu;
5693         int shallowest_idle_cpu = -1;
5694         int i;
5695 
5696         /* Check if we have any choice: */
5697         if (group->group_weight == 1)
5698                 return cpumask_first(sched_group_span(group));
5699 
5700         /* Traverse only the allowed CPUs */
5701         for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
5702                 if (available_idle_cpu(i)) {
5703                         struct rq *rq = cpu_rq(i);
5704                         struct cpuidle_state *idle = idle_get_state(rq);
5705                         if (idle && idle->exit_latency < min_exit_latency) {
5706                                 /*
5707                                  * We give priority to a CPU whose idle state
5708                                  * has the smallest exit latency irrespective
5709                                  * of any idle timestamp.
5710                                  */
5711                                 min_exit_latency = idle->exit_latency;
5712                                 latest_idle_timestamp = rq->idle_stamp;
5713                                 shallowest_idle_cpu = i;
5714                         } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5715                                    rq->idle_stamp > latest_idle_timestamp) {
5716                                 /*
5717                                  * If equal or no active idle state, then
5718                                  * the most recently idled CPU might have
5719                                  * a warmer cache.
5720                                  */
5721                                 latest_idle_timestamp = rq->idle_stamp;
5722                                 shallowest_idle_cpu = i;
5723                         }
5724                 } else if (shallowest_idle_cpu == -1) {
5725                         load = cpu_runnable_load(cpu_rq(i));
5726                         if (load < min_load) {
5727                                 min_load = load;
5728                                 least_loaded_cpu = i;
5729                         }
5730                 }
5731         }
5732 
5733         return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5734 }
5735 
5736 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
5737                                   int cpu, int prev_cpu, int sd_flag)
5738 {
5739         int new_cpu = cpu;
5740 
5741         if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
5742                 return prev_cpu;
5743 
5744         /*
5745          * We need task's util for capacity_spare_without, sync it up to
5746          * prev_cpu's last_update_time.
5747          */
5748         if (!(sd_flag & SD_BALANCE_FORK))
5749                 sync_entity_load_avg(&p->se);
5750 
5751         while (sd) {
5752                 struct sched_group *group;
5753                 struct sched_domain *tmp;
5754                 int weight;
5755 
5756                 if (!(sd->flags & sd_flag)) {
5757                         sd = sd->child;
5758                         continue;
5759                 }
5760 
5761                 group = find_idlest_group(sd, p, cpu, sd_flag);
5762                 if (!group) {
5763                         sd = sd->child;
5764                         continue;
5765                 }
5766 
5767                 new_cpu = find_idlest_group_cpu(group, p, cpu);
5768                 if (new_cpu == cpu) {
5769                         /* Now try balancing at a lower domain level of 'cpu': */
5770                         sd = sd->child;
5771                         continue;
5772                 }
5773 
5774                 /* Now try balancing at a lower domain level of 'new_cpu': */
5775                 cpu = new_cpu;
5776                 weight = sd->span_weight;
5777                 sd = NULL;
5778                 for_each_domain(cpu, tmp) {
5779                         if (weight <= tmp->span_weight)
5780                                 break;
5781                         if (tmp->flags & sd_flag)
5782                                 sd = tmp;
5783                 }
5784         }
5785 
5786         return new_cpu;
5787 }
5788 
5789 #ifdef CONFIG_SCHED_SMT
5790 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5791 EXPORT_SYMBOL_GPL(sched_smt_present);
5792 
5793 static inline void set_idle_cores(int cpu, int val)
5794 {
5795         struct sched_domain_shared *sds;
5796 
5797         sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5798         if (sds)
5799                 WRITE_ONCE(sds->has_idle_cores, val);
5800 }
5801 
5802 static inline bool test_idle_cores(int cpu, bool def)
5803 {
5804         struct sched_domain_shared *sds;
5805 
5806         sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5807         if (sds)
5808                 return READ_ONCE(sds->has_idle_cores);
5809 
5810         return def;
5811 }
5812 
5813 /*
5814  * Scans the local SMT mask to see if the entire core is idle, and records this
5815  * information in sd_llc_shared->has_idle_cores.
5816  *
5817  * Since SMT siblings share all cache levels, inspecting this limited remote
5818  * state should be fairly cheap.
5819  */
5820 void __update_idle_core(struct rq *rq)
5821 {
5822         int core = cpu_of(rq);
5823         int cpu;
5824 
5825         rcu_read_lock();
5826         if (test_idle_cores(core, true))
5827                 goto unlock;
5828 
5829         for_each_cpu(cpu, cpu_smt_mask(core)) {
5830                 if (cpu == core)
5831                         continue;
5832 
5833                 if (!available_idle_cpu(cpu))
5834                         goto unlock;
5835         }
5836 
5837         set_idle_cores(core, 1);
5838 unlock:
5839         rcu_read_unlock();
5840 }
5841 
5842 /*
5843  * Scan the entire LLC domain for idle cores; this dynamically switches off if
5844  * there are no idle cores left in the system; tracked through
5845  * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5846  */
5847 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5848 {
5849         struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5850         int core, cpu;
5851 
5852         if (!static_branch_likely(&sched_smt_present))
5853                 return -1;
5854 
5855         if (!test_idle_cores(target, false))
5856                 return -1;
5857 
5858         cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
5859 
5860         for_each_cpu_wrap(core, cpus, target) {
5861                 bool idle = true;
5862 
5863                 for_each_cpu(cpu, cpu_smt_mask(core)) {
5864                         __cpumask_clear_cpu(cpu, cpus);
5865                         if (!available_idle_cpu(cpu))
5866                                 idle = false;
5867                 }
5868 
5869                 if (idle)
5870                         return core;
5871         }
5872 
5873         /*
5874          * Failed to find an idle core; stop looking for one.
5875          */
5876         set_idle_cores(target, 0);
5877 
5878         return -1;
5879 }
5880 
5881 /*
5882  * Scan the local SMT mask for idle CPUs.
5883  */
5884 static int select_idle_smt(struct task_struct *p, int target)
5885 {
5886         int cpu;
5887 
5888         if (!static_branch_likely(&sched_smt_present))
5889                 return -1;
5890 
5891         for_each_cpu(cpu, cpu_smt_mask(target)) {
5892                 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
5893                         continue;
5894                 if (available_idle_cpu(cpu))
5895                         return cpu;
5896         }
5897 
5898         return -1;
5899 }
5900 
5901 #else /* CONFIG_SCHED_SMT */
5902 
5903 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5904 {
5905         return -1;
5906 }
5907 
5908 static inline int select_idle_smt(struct task_struct *p, int target)
5909 {
5910         return -1;
5911 }
5912 
5913 #endif /* CONFIG_SCHED_SMT */
5914 
5915 /*
5916  * Scan the LLC domain for idle CPUs; this is dynamically regulated by
5917  * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
5918  * average idle time for this rq (as found in rq->avg_idle).
5919  */
5920 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
5921 {
5922         struct sched_domain *this_sd;
5923         u64 avg_cost, avg_idle;
5924         u64 time, cost;
5925         s64 delta;
5926         int cpu, nr = INT_MAX;
5927         int this = smp_processor_id();
5928 
5929         this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
5930         if (!this_sd)
5931                 return -1;
5932 
5933         /*
5934          * Due to large variance we need a large fuzz factor; hackbench in
5935          * particularly is sensitive here.
5936          */
5937         avg_idle = this_rq()->avg_idle / 512;
5938         avg_cost = this_sd->avg_scan_cost + 1;
5939 
5940         if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
5941                 return -1;
5942 
5943         if (sched_feat(SIS_PROP)) {
5944                 u64 span_avg = sd->span_weight * avg_idle;
5945                 if (span_avg > 4*avg_cost)
5946                         nr = div_u64(span_avg, avg_cost);
5947                 else
5948                         nr = 4;
5949         }
5950 
5951         time = cpu_clock(this);
5952 
5953         for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
5954                 if (!--nr)
5955                         return -1;
5956                 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
5957                         continue;
5958                 if (available_idle_cpu(cpu))
5959                         break;
5960         }
5961 
5962         time = cpu_clock(this) - time;
5963         cost = this_sd->avg_scan_cost;
5964         delta = (s64)(time - cost) / 8;
5965         this_sd->avg_scan_cost += delta;
5966 
5967         return cpu;
5968 }
5969 
5970 /*
5971  * Try and locate an idle core/thread in the LLC cache domain.
5972  */
5973 static int select_idle_sibling(struct task_struct *p, int prev, int target)
5974 {
5975         struct sched_domain *sd;
5976         int i, recent_used_cpu;
5977 
5978         if (available_idle_cpu(target))
5979                 return target;
5980 
5981         /*
5982          * If the previous CPU is cache affine and idle, don't be stupid:
5983          */
5984         if (prev != target && cpus_share_cache(prev, target) && available_idle_cpu(prev))
5985                 return prev;
5986 
5987         /* Check a recently used CPU as a potential idle candidate: */
5988         recent_used_cpu = p->recent_used_cpu;
5989         if (recent_used_cpu != prev &&
5990             recent_used_cpu != target &&
5991             cpus_share_cache(recent_used_cpu, target) &&
5992             available_idle_cpu(recent_used_cpu) &&
5993             cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr)) {
5994                 /*
5995                  * Replace recent_used_cpu with prev as it is a potential
5996                  * candidate for the next wake:
5997                  */
5998                 p->recent_used_cpu = prev;
5999                 return recent_used_cpu;
6000         }
6001 
6002         sd = rcu_dereference(per_cpu(sd_llc, target));
6003         if (!sd)
6004                 return target;
6005 
6006         i = select_idle_core(p, sd, target);
6007         if ((unsigned)i < nr_cpumask_bits)
6008                 return i;
6009 
6010         i = select_idle_cpu(p, sd, target);
6011         if ((unsigned)i < nr_cpumask_bits)
6012                 return i;
6013 
6014         i = select_idle_smt(p, target);
6015         if ((unsigned)i < nr_cpumask_bits)
6016                 return i;
6017 
6018         return target;
6019 }
6020 
6021 /**
6022  * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
6023  * @cpu: the CPU to get the utilization of
6024  *
6025  * The unit of the return value must be the one of capacity so we can compare
6026  * the utilization with the capacity of the CPU that is available for CFS task
6027  * (ie cpu_capacity).
6028  *
6029  * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6030  * recent utilization of currently non-runnable tasks on a CPU. It represents
6031  * the amount of utilization of a CPU in the range [0..capacity_orig] where
6032  * capacity_orig is the cpu_capacity available at the highest frequency
6033  * (arch_scale_freq_capacity()).
6034  * The utilization of a CPU converges towards a sum equal to or less than the
6035  * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6036  * the running time on this CPU scaled by capacity_curr.
6037  *
6038  * The estimated utilization of a CPU is defined to be the maximum between its
6039  * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6040  * currently RUNNABLE on that CPU.
6041  * This allows to properly represent the expected utilization of a CPU which
6042  * has just got a big task running since a long sleep period. At the same time
6043  * however it preserves the benefits of the "blocked utilization" in
6044  * describing the potential for other tasks waking up on the same CPU.
6045  *
6046  * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6047  * higher than capacity_orig because of unfortunate rounding in
6048  * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6049  * the average stabilizes with the new running time. We need to check that the
6050  * utilization stays within the range of [0..capacity_orig] and cap it if
6051  * necessary. Without utilization capping, a group could be seen as overloaded
6052  * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6053  * available capacity. We allow utilization to overshoot capacity_curr (but not
6054  * capacity_orig) as it useful for predicting the capacity required after task
6055  * migrations (scheduler-driven DVFS).
6056  *
6057  * Return: the (estimated) utilization for the specified CPU
6058  */
6059 static inline unsigned long cpu_util(int cpu)
6060 {
6061         struct cfs_rq *cfs_rq;
6062         unsigned int util;
6063 
6064         cfs_rq = &cpu_rq(cpu)->cfs;
6065         util = READ_ONCE(cfs_rq->avg.util_avg);
6066 
6067         if (sched_feat(UTIL_EST))
6068                 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6069 
6070         return min_t(unsigned long, util, capacity_orig_of(cpu));
6071 }
6072 
6073 /*
6074  * cpu_util_without: compute cpu utilization without any contributions from *p
6075  * @cpu: the CPU which utilization is requested
6076  * @p: the task which utilization should be discounted
6077  *
6078  * The utilization of a CPU is defined by the utilization of tasks currently
6079  * enqueued on that CPU as well as tasks which are currently sleeping after an
6080  * execution on that CPU.
6081  *
6082  * This method returns the utilization of the specified CPU by discounting the
6083  * utilization of the specified task, whenever the task is currently
6084  * contributing to the CPU utilization.
6085  */
6086 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
6087 {
6088         struct cfs_rq *cfs_rq;
6089         unsigned int util;
6090 
6091         /* Task has no contribution or is new */
6092         if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6093                 return cpu_util(cpu);
6094 
6095         cfs_rq = &cpu_rq(cpu)->cfs;
6096         util = READ_ONCE(cfs_rq->avg.util_avg);
6097 
6098         /* Discount task's util from CPU's util */
6099         lsub_positive(&util, task_util(p));
6100 
6101         /*
6102          * Covered cases:
6103          *
6104          * a) if *p is the only task sleeping on this CPU, then:
6105          *      cpu_util (== task_util) > util_est (== 0)
6106          *    and thus we return:
6107          *      cpu_util_without = (cpu_util - task_util) = 0
6108          *
6109          * b) if other tasks are SLEEPING on this CPU, which is now exiting
6110          *    IDLE, then:
6111          *      cpu_util >= task_util
6112          *      cpu_util > util_est (== 0)
6113          *    and thus we discount *p's blocked utilization to return:
6114          *      cpu_util_without = (cpu_util - task_util) >= 0
6115          *
6116          * c) if other tasks are RUNNABLE on that CPU and
6117          *      util_est > cpu_util
6118          *    then we use util_est since it returns a more restrictive
6119          *    estimation of the spare capacity on that CPU, by just
6120          *    considering the expected utilization of tasks already
6121          *    runnable on that CPU.
6122          *
6123          * Cases a) and b) are covered by the above code, while case c) is
6124          * covered by the following code when estimated utilization is
6125          * enabled.
6126          */
6127         if (sched_feat(UTIL_EST)) {
6128                 unsigned int estimated =
6129                         READ_ONCE(cfs_rq->avg.util_est.enqueued);
6130 
6131                 /*
6132                  * Despite the following checks we still have a small window
6133                  * for a possible race, when an execl's select_task_rq_fair()
6134                  * races with LB's detach_task():
6135                  *
6136                  *   detach_task()
6137                  *     p->on_rq = TASK_ON_RQ_MIGRATING;
6138                  *     ---------------------------------- A
6139                  *     deactivate_task()                   \
6140                  *       dequeue_task()                     + RaceTime
6141                  *         util_est_dequeue()              /
6142                  *     ---------------------------------- B
6143                  *
6144                  * The additional check on "current == p" it's required to
6145                  * properly fix the execl regression and it helps in further
6146                  * reducing the chances for the above race.
6147                  */
6148                 if (unlikely(task_on_rq_queued(p) || current == p))
6149                         lsub_positive(&estimated, _task_util_est(p));
6150 
6151                 util = max(util, estimated);
6152         }
6153 
6154         /*
6155          * Utilization (estimated) can exceed the CPU capacity, thus let's
6156          * clamp to the maximum CPU capacity to ensure consistency with
6157          * the cpu_util call.
6158          */
6159         return min_t(unsigned long, util, capacity_orig_of(cpu));
6160 }
6161 
6162 /*
6163  * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6164  * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6165  *
6166  * In that case WAKE_AFFINE doesn't make sense and we'll let
6167  * BALANCE_WAKE sort things out.
6168  */
6169 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6170 {
6171         long min_cap, max_cap;
6172 
6173         if (!static_branch_unlikely(&sched_asym_cpucapacity))
6174                 return 0;
6175 
6176         min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6177         max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
6178 
6179         /* Minimum capacity is close to max, no need to abort wake_affine */
6180         if (max_cap - min_cap < max_cap >> 3)
6181                 return 0;
6182 
6183         /* Bring task utilization in sync with prev_cpu */
6184         sync_entity_load_avg(&p->se);
6185 
6186         return !task_fits_capacity(p, min_cap);
6187 }
6188 
6189 /*
6190  * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
6191  * to @dst_cpu.
6192  */
6193 static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
6194 {
6195         struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6196         unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg);
6197 
6198         /*
6199          * If @p migrates from @cpu to another, remove its contribution. Or,
6200          * if @p migrates from another CPU to @cpu, add its contribution. In
6201          * the other cases, @cpu is not impacted by the migration, so the
6202          * util_avg should already be correct.
6203          */
6204         if (task_cpu(p) == cpu && dst_cpu != cpu)
6205                 sub_positive(&util, task_util(p));
6206         else if (task_cpu(p) != cpu && dst_cpu == cpu)
6207                 util += task_util(p);
6208 
6209         if (sched_feat(UTIL_EST)) {
6210                 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
6211 
6212                 /*
6213                  * During wake-up, the task isn't enqueued yet and doesn't
6214                  * appear in the cfs_rq->avg.util_est.enqueued of any rq,
6215                  * so just add it (if needed) to "simulate" what will be
6216                  * cpu_util() after the task has been enqueued.
6217                  */
6218                 if (dst_cpu == cpu)
6219                         util_est += _task_util_est(p);
6220 
6221                 util = max(util, util_est);
6222         }
6223 
6224         return min(util, capacity_orig_of(cpu));
6225 }
6226 
6227 /*
6228  * compute_energy(): Estimates the energy that would be consumed if @p was
6229  * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
6230  * landscape of the * CPUs after the task migration, and uses the Energy Model
6231  * to compute what would be the energy if we decided to actually migrate that
6232  * task.
6233  */
6234 static long
6235 compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd)
6236 {
6237         unsigned int max_util, util_cfs, cpu_util, cpu_cap;
6238         unsigned long sum_util, energy = 0;
6239         struct task_struct *tsk;
6240         int cpu;
6241 
6242         for (; pd; pd = pd->next) {
6243                 struct cpumask *pd_mask = perf_domain_span(pd);
6244 
6245                 /*
6246                  * The energy model mandates all the CPUs of a performance
6247                  * domain have the same capacity.
6248                  */
6249                 cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask));
6250                 max_util = sum_util = 0;
6251 
6252                 /*
6253                  * The capacity state of CPUs of the current rd can be driven by
6254                  * CPUs of another rd if they belong to the same performance
6255                  * domain. So, account for the utilization of these CPUs too
6256                  * by masking pd with cpu_online_mask instead of the rd span.
6257                  *
6258                  * If an entire performance domain is outside of the current rd,
6259                  * it will not appear in its pd list and will not be accounted
6260                  * by compute_energy().
6261                  */
6262                 for_each_cpu_and(cpu, pd_mask, cpu_online_mask) {
6263                         util_cfs = cpu_util_next(cpu, p, dst_cpu);
6264 
6265                         /*
6266                          * Busy time computation: utilization clamping is not
6267                          * required since the ratio (sum_util / cpu_capacity)
6268                          * is already enough to scale the EM reported power
6269                          * consumption at the (eventually clamped) cpu_capacity.
6270                          */
6271                         sum_util += schedutil_cpu_util(cpu, util_cfs, cpu_cap,
6272                                                        ENERGY_UTIL, NULL);
6273 
6274                         /*
6275                          * Performance domain frequency: utilization clamping
6276                          * must be considered since it affects the selection
6277                          * of the performance domain frequency.
6278                          * NOTE: in case RT tasks are running, by default the
6279                          * FREQUENCY_UTIL's utilization can be max OPP.
6280                          */
6281                         tsk = cpu == dst_cpu ? p : NULL;
6282                         cpu_util = schedutil_cpu_util(cpu, util_cfs, cpu_cap,
6283                                                       FREQUENCY_UTIL, tsk);
6284                         max_util = max(max_util, cpu_util);
6285                 }
6286 
6287                 energy += em_pd_energy(pd->em_pd, max_util, sum_util);
6288         }
6289 
6290         return energy;
6291 }
6292 
6293 /*
6294  * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6295  * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6296  * spare capacity in each performance domain and uses it as a potential
6297  * candidate to execute the task. Then, it uses the Energy Model to figure
6298  * out which of the CPU candidates is the most energy-efficient.
6299  *
6300  * The rationale for this heuristic is as follows. In a performance domain,
6301  * all the most energy efficient CPU candidates (according to the Energy
6302  * Model) are those for which we'll request a low frequency. When there are
6303  * several CPUs for which the frequency request will be the same, we don't
6304  * have enough data to break the tie between them, because the Energy Model
6305  * only includes active power costs. With this model, if we assume that
6306  * frequency requests follow utilization (e.g. using schedutil), the CPU with
6307  * the maximum spare capacity in a performance domain is guaranteed to be among
6308  * the best candidates of the performance domain.
6309  *
6310  * In practice, it could be preferable from an energy standpoint to pack
6311  * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6312  * but that could also hurt our chances to go cluster idle, and we have no
6313  * ways to tell with the current Energy Model if this is actually a good
6314  * idea or not. So, find_energy_efficient_cpu() basically favors
6315  * cluster-packing, and spreading inside a cluster. That should at least be
6316  * a good thing for latency, and this is consistent with the idea that most
6317  * of the energy savings of EAS come from the asymmetry of the system, and
6318  * not so much from breaking the tie between identical CPUs. That's also the
6319  * reason why EAS is enabled in the topology code only for systems where
6320  * SD_ASYM_CPUCAPACITY is set.
6321  *
6322  * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6323  * they don't have any useful utilization data yet and it's not possible to
6324  * forecast their impact on energy consumption. Consequently, they will be
6325  * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6326  * to be energy-inefficient in some use-cases. The alternative would be to
6327  * bias new tasks towards specific types of CPUs first, or to try to infer
6328  * their util_avg from the parent task, but those heuristics could hurt
6329  * other use-cases too. So, until someone finds a better way to solve this,
6330  * let's keep things simple by re-using the existing slow path.
6331  */
6332 
6333 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
6334 {
6335         unsigned long prev_energy = ULONG_MAX, best_energy = ULONG_MAX;
6336         struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
6337         int cpu, best_energy_cpu = prev_cpu;
6338         struct perf_domain *head, *pd;
6339         unsigned long cpu_cap, util;
6340         struct sched_domain *sd;
6341 
6342         rcu_read_lock();
6343         pd = rcu_dereference(rd->pd);
6344         if (!pd || READ_ONCE(rd->overutilized))
6345                 goto fail;
6346         head = pd;
6347 
6348         /*
6349          * Energy-aware wake-up happens on the lowest sched_domain starting
6350          * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6351          */
6352         sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
6353         while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
6354                 sd = sd->parent;
6355         if (!sd)
6356                 goto fail;
6357 
6358         sync_entity_load_avg(&p->se);
6359         if (!task_util_est(p))
6360                 goto unlock;
6361 
6362         for (; pd; pd = pd->next) {
6363                 unsigned long cur_energy, spare_cap, max_spare_cap = 0;
6364                 int max_spare_cap_cpu = -1;
6365 
6366                 for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) {
6367                         if (!cpumask_test_cpu(cpu, p->cpus_ptr))
6368                                 continue;
6369 
6370                         /* Skip CPUs that will be overutilized. */
6371                         util = cpu_util_next(cpu, p, cpu);
6372                         cpu_cap = capacity_of(cpu);
6373                         if (cpu_cap * 1024 < util * capacity_margin)
6374                                 continue;
6375 
6376                         /* Always use prev_cpu as a candidate. */
6377                         if (cpu == prev_cpu) {
6378                                 prev_energy = compute_energy(p, prev_cpu, head);
6379                                 best_energy = min(best_energy, prev_energy);
6380                                 continue;
6381                         }
6382 
6383                         /*
6384                          * Find the CPU with the maximum spare capacity in
6385                          * the performance domain
6386                          */
6387                         spare_cap = cpu_cap - util;
6388                         if (spare_cap > max_spare_cap) {
6389                                 max_spare_cap = spare_cap;
6390                                 max_spare_cap_cpu = cpu;
6391                         }
6392                 }
6393 
6394                 /* Evaluate the energy impact of using this CPU. */
6395                 if (max_spare_cap_cpu >= 0) {
6396                         cur_energy = compute_energy(p, max_spare_cap_cpu, head);
6397                         if (cur_energy < best_energy) {
6398                                 best_energy = cur_energy;
6399                                 best_energy_cpu = max_spare_cap_cpu;
6400                         }
6401                 }
6402         }
6403 unlock:
6404         rcu_read_unlock();
6405 
6406         /*
6407          * Pick the best CPU if prev_cpu cannot be used, or if it saves at
6408          * least 6% of the energy used by prev_cpu.
6409          */
6410         if (prev_energy == ULONG_MAX)
6411                 return best_energy_cpu;
6412 
6413         if ((prev_energy - best_energy) > (prev_energy >> 4))
6414                 return best_energy_cpu;
6415 
6416         return prev_cpu;
6417 
6418 fail:
6419         rcu_read_unlock();
6420 
6421         return -1;
6422 }
6423 
6424 /*
6425  * select_task_rq_fair: Select target runqueue for the waking task in domains
6426  * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6427  * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6428  *
6429  * Balances load by selecting the idlest CPU in the idlest group, or under
6430  * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6431  *
6432  * Returns the target CPU number.
6433  *
6434  * preempt must be disabled.
6435  */
6436 static int
6437 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6438 {
6439         struct sched_domain *tmp, *sd = NULL;
6440         int cpu = smp_processor_id();
6441         int new_cpu = prev_cpu;
6442         int want_affine = 0;
6443         int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6444 
6445         if (sd_flag & SD_BALANCE_WAKE) {
6446                 record_wakee(p);
6447 
6448                 if (sched_energy_enabled()) {
6449                         new_cpu =