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

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

Version: ~ [ linux-5.19-rc8 ] ~ [ linux-5.18.14 ] ~ [ linux-5.17.15 ] ~ [ linux-5.16.20 ] ~ [ linux-5.15.57 ] ~ [ linux-5.14.21 ] ~ [ linux-5.13.19 ] ~ [ linux-5.12.19 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.133 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.207 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.253 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.289 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.324 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.302 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.9 ] ~ [ 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 bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
279 {
280         struct rq *rq = rq_of(cfs_rq);
281         int cpu = cpu_of(rq);
282 
283         if (cfs_rq->on_list)
284                 return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
285 
286         cfs_rq->on_list = 1;
287 
288         /*
289          * Ensure we either appear before our parent (if already
290          * enqueued) or force our parent to appear after us when it is
291          * enqueued. The fact that we always enqueue bottom-up
292          * reduces this to two cases and a special case for the root
293          * cfs_rq. Furthermore, it also means that we will always reset
294          * tmp_alone_branch either when the branch is connected
295          * to a tree or when we reach the top of the tree
296          */
297         if (cfs_rq->tg->parent &&
298             cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
299                 /*
300                  * If parent is already on the list, we add the child
301                  * just before. Thanks to circular linked property of
302                  * the list, this means to put the child at the tail
303                  * of the list that starts by parent.
304                  */
305                 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
306                         &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
307                 /*
308                  * The branch is now connected to its tree so we can
309                  * reset tmp_alone_branch to the beginning of the
310                  * list.
311                  */
312                 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
313                 return true;
314         }
315 
316         if (!cfs_rq->tg->parent) {
317                 /*
318                  * cfs rq without parent should be put
319                  * at the tail of the list.
320                  */
321                 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
322                         &rq->leaf_cfs_rq_list);
323                 /*
324                  * We have reach the top of a tree so we can reset
325                  * tmp_alone_branch to the beginning of the list.
326                  */
327                 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
328                 return true;
329         }
330 
331         /*
332          * The parent has not already been added so we want to
333          * make sure that it will be put after us.
334          * tmp_alone_branch points to the begin of the branch
335          * where we will add parent.
336          */
337         list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
338         /*
339          * update tmp_alone_branch to points to the new begin
340          * of the branch
341          */
342         rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
343         return false;
344 }
345 
346 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
347 {
348         if (cfs_rq->on_list) {
349                 struct rq *rq = rq_of(cfs_rq);
350 
351                 /*
352                  * With cfs_rq being unthrottled/throttled during an enqueue,
353                  * it can happen the tmp_alone_branch points the a leaf that
354                  * we finally want to del. In this case, tmp_alone_branch moves
355                  * to the prev element but it will point to rq->leaf_cfs_rq_list
356                  * at the end of the enqueue.
357                  */
358                 if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
359                         rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
360 
361                 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
362                 cfs_rq->on_list = 0;
363         }
364 }
365 
366 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
367 {
368         SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
369 }
370 
371 /* Iterate thr' all leaf cfs_rq's on a runqueue */
372 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos)                      \
373         list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list,    \
374                                  leaf_cfs_rq_list)
375 
376 /* Do the two (enqueued) entities belong to the same group ? */
377 static inline struct cfs_rq *
378 is_same_group(struct sched_entity *se, struct sched_entity *pse)
379 {
380         if (se->cfs_rq == pse->cfs_rq)
381                 return se->cfs_rq;
382 
383         return NULL;
384 }
385 
386 static inline struct sched_entity *parent_entity(struct sched_entity *se)
387 {
388         return se->parent;
389 }
390 
391 static void
392 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
393 {
394         int se_depth, pse_depth;
395 
396         /*
397          * preemption test can be made between sibling entities who are in the
398          * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
399          * both tasks until we find their ancestors who are siblings of common
400          * parent.
401          */
402 
403         /* First walk up until both entities are at same depth */
404         se_depth = (*se)->depth;
405         pse_depth = (*pse)->depth;
406 
407         while (se_depth > pse_depth) {
408                 se_depth--;
409                 *se = parent_entity(*se);
410         }
411 
412         while (pse_depth > se_depth) {
413                 pse_depth--;
414                 *pse = parent_entity(*pse);
415         }
416 
417         while (!is_same_group(*se, *pse)) {
418                 *se = parent_entity(*se);
419                 *pse = parent_entity(*pse);
420         }
421 }
422 
423 #else   /* !CONFIG_FAIR_GROUP_SCHED */
424 
425 static inline struct task_struct *task_of(struct sched_entity *se)
426 {
427         return container_of(se, struct task_struct, se);
428 }
429 
430 #define for_each_sched_entity(se) \
431                 for (; se; se = NULL)
432 
433 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
434 {
435         return &task_rq(p)->cfs;
436 }
437 
438 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
439 {
440         struct task_struct *p = task_of(se);
441         struct rq *rq = task_rq(p);
442 
443         return &rq->cfs;
444 }
445 
446 /* runqueue "owned" by this group */
447 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
448 {
449         return NULL;
450 }
451 
452 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
453 {
454         return true;
455 }
456 
457 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
458 {
459 }
460 
461 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
462 {
463 }
464 
465 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos)      \
466                 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
467 
468 static inline struct sched_entity *parent_entity(struct sched_entity *se)
469 {
470         return NULL;
471 }
472 
473 static inline void
474 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
475 {
476 }
477 
478 #endif  /* CONFIG_FAIR_GROUP_SCHED */
479 
480 static __always_inline
481 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
482 
483 /**************************************************************
484  * Scheduling class tree data structure manipulation methods:
485  */
486 
487 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
488 {
489         s64 delta = (s64)(vruntime - max_vruntime);
490         if (delta > 0)
491                 max_vruntime = vruntime;
492 
493         return max_vruntime;
494 }
495 
496 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
497 {
498         s64 delta = (s64)(vruntime - min_vruntime);
499         if (delta < 0)
500                 min_vruntime = vruntime;
501 
502         return min_vruntime;
503 }
504 
505 static inline int entity_before(struct sched_entity *a,
506                                 struct sched_entity *b)
507 {
508         return (s64)(a->vruntime - b->vruntime) < 0;
509 }
510 
511 static void update_min_vruntime(struct cfs_rq *cfs_rq)
512 {
513         struct sched_entity *curr = cfs_rq->curr;
514         struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
515 
516         u64 vruntime = cfs_rq->min_vruntime;
517 
518         if (curr) {
519                 if (curr->on_rq)
520                         vruntime = curr->vruntime;
521                 else
522                         curr = NULL;
523         }
524 
525         if (leftmost) { /* non-empty tree */
526                 struct sched_entity *se;
527                 se = rb_entry(leftmost, struct sched_entity, run_node);
528 
529                 if (!curr)
530                         vruntime = se->vruntime;
531                 else
532                         vruntime = min_vruntime(vruntime, se->vruntime);
533         }
534 
535         /* ensure we never gain time by being placed backwards. */
536         cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
537 #ifndef CONFIG_64BIT
538         smp_wmb();
539         cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
540 #endif
541 }
542 
543 /*
544  * Enqueue an entity into the rb-tree:
545  */
546 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
547 {
548         struct rb_node **link = &cfs_rq->tasks_timeline.rb_root.rb_node;
549         struct rb_node *parent = NULL;
550         struct sched_entity *entry;
551         bool leftmost = true;
552 
553         /*
554          * Find the right place in the rbtree:
555          */
556         while (*link) {
557                 parent = *link;
558                 entry = rb_entry(parent, struct sched_entity, run_node);
559                 /*
560                  * We dont care about collisions. Nodes with
561                  * the same key stay together.
562                  */
563                 if (entity_before(se, entry)) {
564                         link = &parent->rb_left;
565                 } else {
566                         link = &parent->rb_right;
567                         leftmost = false;
568                 }
569         }
570 
571         rb_link_node(&se->run_node, parent, link);
572         rb_insert_color_cached(&se->run_node,
573                                &cfs_rq->tasks_timeline, leftmost);
574 }
575 
576 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
577 {
578         rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
579 }
580 
581 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
582 {
583         struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
584 
585         if (!left)
586                 return NULL;
587 
588         return rb_entry(left, struct sched_entity, run_node);
589 }
590 
591 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
592 {
593         struct rb_node *next = rb_next(&se->run_node);
594 
595         if (!next)
596                 return NULL;
597 
598         return rb_entry(next, struct sched_entity, run_node);
599 }
600 
601 #ifdef CONFIG_SCHED_DEBUG
602 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
603 {
604         struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
605 
606         if (!last)
607                 return NULL;
608 
609         return rb_entry(last, struct sched_entity, run_node);
610 }
611 
612 /**************************************************************
613  * Scheduling class statistics methods:
614  */
615 
616 int sched_proc_update_handler(struct ctl_table *table, int write,
617                 void __user *buffer, size_t *lenp,
618                 loff_t *ppos)
619 {
620         int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
621         unsigned int factor = get_update_sysctl_factor();
622 
623         if (ret || !write)
624                 return ret;
625 
626         sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
627                                         sysctl_sched_min_granularity);
628 
629 #define WRT_SYSCTL(name) \
630         (normalized_sysctl_##name = sysctl_##name / (factor))
631         WRT_SYSCTL(sched_min_granularity);
632         WRT_SYSCTL(sched_latency);
633         WRT_SYSCTL(sched_wakeup_granularity);
634 #undef WRT_SYSCTL
635 
636         return 0;
637 }
638 #endif
639 
640 /*
641  * delta /= w
642  */
643 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
644 {
645         if (unlikely(se->load.weight != NICE_0_LOAD))
646                 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
647 
648         return delta;
649 }
650 
651 /*
652  * The idea is to set a period in which each task runs once.
653  *
654  * When there are too many tasks (sched_nr_latency) we have to stretch
655  * this period because otherwise the slices get too small.
656  *
657  * p = (nr <= nl) ? l : l*nr/nl
658  */
659 static u64 __sched_period(unsigned long nr_running)
660 {
661         if (unlikely(nr_running > sched_nr_latency))
662                 return nr_running * sysctl_sched_min_granularity;
663         else
664                 return sysctl_sched_latency;
665 }
666 
667 /*
668  * We calculate the wall-time slice from the period by taking a part
669  * proportional to the weight.
670  *
671  * s = p*P[w/rw]
672  */
673 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
674 {
675         u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
676 
677         for_each_sched_entity(se) {
678                 struct load_weight *load;
679                 struct load_weight lw;
680 
681                 cfs_rq = cfs_rq_of(se);
682                 load = &cfs_rq->load;
683 
684                 if (unlikely(!se->on_rq)) {
685                         lw = cfs_rq->load;
686 
687                         update_load_add(&lw, se->load.weight);
688                         load = &lw;
689                 }
690                 slice = __calc_delta(slice, se->load.weight, load);
691         }
692         return slice;
693 }
694 
695 /*
696  * We calculate the vruntime slice of a to-be-inserted task.
697  *
698  * vs = s/w
699  */
700 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
701 {
702         return calc_delta_fair(sched_slice(cfs_rq, se), se);
703 }
704 
705 #include "pelt.h"
706 #ifdef CONFIG_SMP
707 
708 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
709 static unsigned long task_h_load(struct task_struct *p);
710 static unsigned long capacity_of(int cpu);
711 
712 /* Give new sched_entity start runnable values to heavy its load in infant time */
713 void init_entity_runnable_average(struct sched_entity *se)
714 {
715         struct sched_avg *sa = &se->avg;
716 
717         memset(sa, 0, sizeof(*sa));
718 
719         /*
720          * Tasks are initialized with full load to be seen as heavy tasks until
721          * they get a chance to stabilize to their real load level.
722          * Group entities are initialized with zero load to reflect the fact that
723          * nothing has been attached to the task group yet.
724          */
725         if (entity_is_task(se))
726                 sa->runnable_load_avg = sa->load_avg = scale_load_down(se->load.weight);
727 
728         se->runnable_weight = se->load.weight;
729 
730         /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
731 }
732 
733 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
734 static void attach_entity_cfs_rq(struct sched_entity *se);
735 
736 /*
737  * With new tasks being created, their initial util_avgs are extrapolated
738  * based on the cfs_rq's current util_avg:
739  *
740  *   util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
741  *
742  * However, in many cases, the above util_avg does not give a desired
743  * value. Moreover, the sum of the util_avgs may be divergent, such
744  * as when the series is a harmonic series.
745  *
746  * To solve this problem, we also cap the util_avg of successive tasks to
747  * only 1/2 of the left utilization budget:
748  *
749  *   util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
750  *
751  * where n denotes the nth task and cpu_scale the CPU capacity.
752  *
753  * For example, for a CPU with 1024 of capacity, a simplest series from
754  * the beginning would be like:
755  *
756  *  task  util_avg: 512, 256, 128,  64,  32,   16,    8, ...
757  * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
758  *
759  * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
760  * if util_avg > util_avg_cap.
761  */
762 void post_init_entity_util_avg(struct task_struct *p)
763 {
764         struct sched_entity *se = &p->se;
765         struct cfs_rq *cfs_rq = cfs_rq_of(se);
766         struct sched_avg *sa = &se->avg;
767         long cpu_scale = arch_scale_cpu_capacity(NULL, cpu_of(rq_of(cfs_rq)));
768         long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
769 
770         if (cap > 0) {
771                 if (cfs_rq->avg.util_avg != 0) {
772                         sa->util_avg  = cfs_rq->avg.util_avg * se->load.weight;
773                         sa->util_avg /= (cfs_rq->avg.load_avg + 1);
774 
775                         if (sa->util_avg > cap)
776                                 sa->util_avg = cap;
777                 } else {
778                         sa->util_avg = cap;
779                 }
780         }
781 
782         if (p->sched_class != &fair_sched_class) {
783                 /*
784                  * For !fair tasks do:
785                  *
786                 update_cfs_rq_load_avg(now, cfs_rq);
787                 attach_entity_load_avg(cfs_rq, se, 0);
788                 switched_from_fair(rq, p);
789                  *
790                  * such that the next switched_to_fair() has the
791                  * expected state.
792                  */
793                 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
794                 return;
795         }
796 
797         attach_entity_cfs_rq(se);
798 }
799 
800 #else /* !CONFIG_SMP */
801 void init_entity_runnable_average(struct sched_entity *se)
802 {
803 }
804 void post_init_entity_util_avg(struct task_struct *p)
805 {
806 }
807 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
808 {
809 }
810 #endif /* CONFIG_SMP */
811 
812 /*
813  * Update the current task's runtime statistics.
814  */
815 static void update_curr(struct cfs_rq *cfs_rq)
816 {
817         struct sched_entity *curr = cfs_rq->curr;
818         u64 now = rq_clock_task(rq_of(cfs_rq));
819         u64 delta_exec;
820 
821         if (unlikely(!curr))
822                 return;
823 
824         delta_exec = now - curr->exec_start;
825         if (unlikely((s64)delta_exec <= 0))
826                 return;
827 
828         curr->exec_start = now;
829 
830         schedstat_set(curr->statistics.exec_max,
831                       max(delta_exec, curr->statistics.exec_max));
832 
833         curr->sum_exec_runtime += delta_exec;
834         schedstat_add(cfs_rq->exec_clock, delta_exec);
835 
836         curr->vruntime += calc_delta_fair(delta_exec, curr);
837         update_min_vruntime(cfs_rq);
838 
839         if (entity_is_task(curr)) {
840                 struct task_struct *curtask = task_of(curr);
841 
842                 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
843                 cgroup_account_cputime(curtask, delta_exec);
844                 account_group_exec_runtime(curtask, delta_exec);
845         }
846 
847         account_cfs_rq_runtime(cfs_rq, delta_exec);
848 }
849 
850 static void update_curr_fair(struct rq *rq)
851 {
852         update_curr(cfs_rq_of(&rq->curr->se));
853 }
854 
855 static inline void
856 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
857 {
858         u64 wait_start, prev_wait_start;
859 
860         if (!schedstat_enabled())
861                 return;
862 
863         wait_start = rq_clock(rq_of(cfs_rq));
864         prev_wait_start = schedstat_val(se->statistics.wait_start);
865 
866         if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
867             likely(wait_start > prev_wait_start))
868                 wait_start -= prev_wait_start;
869 
870         __schedstat_set(se->statistics.wait_start, wait_start);
871 }
872 
873 static inline void
874 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
875 {
876         struct task_struct *p;
877         u64 delta;
878 
879         if (!schedstat_enabled())
880                 return;
881 
882         delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
883 
884         if (entity_is_task(se)) {
885                 p = task_of(se);
886                 if (task_on_rq_migrating(p)) {
887                         /*
888                          * Preserve migrating task's wait time so wait_start
889                          * time stamp can be adjusted to accumulate wait time
890                          * prior to migration.
891                          */
892                         __schedstat_set(se->statistics.wait_start, delta);
893                         return;
894                 }
895                 trace_sched_stat_wait(p, delta);
896         }
897 
898         __schedstat_set(se->statistics.wait_max,
899                       max(schedstat_val(se->statistics.wait_max), delta));
900         __schedstat_inc(se->statistics.wait_count);
901         __schedstat_add(se->statistics.wait_sum, delta);
902         __schedstat_set(se->statistics.wait_start, 0);
903 }
904 
905 static inline void
906 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
907 {
908         struct task_struct *tsk = NULL;
909         u64 sleep_start, block_start;
910 
911         if (!schedstat_enabled())
912                 return;
913 
914         sleep_start = schedstat_val(se->statistics.sleep_start);
915         block_start = schedstat_val(se->statistics.block_start);
916 
917         if (entity_is_task(se))
918                 tsk = task_of(se);
919 
920         if (sleep_start) {
921                 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
922 
923                 if ((s64)delta < 0)
924                         delta = 0;
925 
926                 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
927                         __schedstat_set(se->statistics.sleep_max, delta);
928 
929                 __schedstat_set(se->statistics.sleep_start, 0);
930                 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
931 
932                 if (tsk) {
933                         account_scheduler_latency(tsk, delta >> 10, 1);
934                         trace_sched_stat_sleep(tsk, delta);
935                 }
936         }
937         if (block_start) {
938                 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
939 
940                 if ((s64)delta < 0)
941                         delta = 0;
942 
943                 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
944                         __schedstat_set(se->statistics.block_max, delta);
945 
946                 __schedstat_set(se->statistics.block_start, 0);
947                 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
948 
949                 if (tsk) {
950                         if (tsk->in_iowait) {
951                                 __schedstat_add(se->statistics.iowait_sum, delta);
952                                 __schedstat_inc(se->statistics.iowait_count);
953                                 trace_sched_stat_iowait(tsk, delta);
954                         }
955 
956                         trace_sched_stat_blocked(tsk, delta);
957 
958                         /*
959                          * Blocking time is in units of nanosecs, so shift by
960                          * 20 to get a milliseconds-range estimation of the
961                          * amount of time that the task spent sleeping:
962                          */
963                         if (unlikely(prof_on == SLEEP_PROFILING)) {
964                                 profile_hits(SLEEP_PROFILING,
965                                                 (void *)get_wchan(tsk),
966                                                 delta >> 20);
967                         }
968                         account_scheduler_latency(tsk, delta >> 10, 0);
969                 }
970         }
971 }
972 
973 /*
974  * Task is being enqueued - update stats:
975  */
976 static inline void
977 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
978 {
979         if (!schedstat_enabled())
980                 return;
981 
982         /*
983          * Are we enqueueing a waiting task? (for current tasks
984          * a dequeue/enqueue event is a NOP)
985          */
986         if (se != cfs_rq->curr)
987                 update_stats_wait_start(cfs_rq, se);
988 
989         if (flags & ENQUEUE_WAKEUP)
990                 update_stats_enqueue_sleeper(cfs_rq, se);
991 }
992 
993 static inline void
994 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
995 {
996 
997         if (!schedstat_enabled())
998                 return;
999 
1000         /*
1001          * Mark the end of the wait period if dequeueing a
1002          * waiting task:
1003          */
1004         if (se != cfs_rq->curr)
1005                 update_stats_wait_end(cfs_rq, se);
1006 
1007         if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1008                 struct task_struct *tsk = task_of(se);
1009 
1010                 if (tsk->state & TASK_INTERRUPTIBLE)
1011                         __schedstat_set(se->statistics.sleep_start,
1012                                       rq_clock(rq_of(cfs_rq)));
1013                 if (tsk->state & TASK_UNINTERRUPTIBLE)
1014                         __schedstat_set(se->statistics.block_start,
1015                                       rq_clock(rq_of(cfs_rq)));
1016         }
1017 }
1018 
1019 /*
1020  * We are picking a new current task - update its stats:
1021  */
1022 static inline void
1023 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1024 {
1025         /*
1026          * We are starting a new run period:
1027          */
1028         se->exec_start = rq_clock_task(rq_of(cfs_rq));
1029 }
1030 
1031 /**************************************************
1032  * Scheduling class queueing methods:
1033  */
1034 
1035 #ifdef CONFIG_NUMA_BALANCING
1036 /*
1037  * Approximate time to scan a full NUMA task in ms. The task scan period is
1038  * calculated based on the tasks virtual memory size and
1039  * numa_balancing_scan_size.
1040  */
1041 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1042 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1043 
1044 /* Portion of address space to scan in MB */
1045 unsigned int sysctl_numa_balancing_scan_size = 256;
1046 
1047 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1048 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1049 
1050 struct numa_group {
1051         refcount_t refcount;
1052 
1053         spinlock_t lock; /* nr_tasks, tasks */
1054         int nr_tasks;
1055         pid_t gid;
1056         int active_nodes;
1057 
1058         struct rcu_head rcu;
1059         unsigned long total_faults;
1060         unsigned long max_faults_cpu;
1061         /*
1062          * Faults_cpu is used to decide whether memory should move
1063          * towards the CPU. As a consequence, these stats are weighted
1064          * more by CPU use than by memory faults.
1065          */
1066         unsigned long *faults_cpu;
1067         unsigned long faults[0];
1068 };
1069 
1070 static inline unsigned long group_faults_priv(struct numa_group *ng);
1071 static inline unsigned long group_faults_shared(struct numa_group *ng);
1072 
1073 static unsigned int task_nr_scan_windows(struct task_struct *p)
1074 {
1075         unsigned long rss = 0;
1076         unsigned long nr_scan_pages;
1077 
1078         /*
1079          * Calculations based on RSS as non-present and empty pages are skipped
1080          * by the PTE scanner and NUMA hinting faults should be trapped based
1081          * on resident pages
1082          */
1083         nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1084         rss = get_mm_rss(p->mm);
1085         if (!rss)
1086                 rss = nr_scan_pages;
1087 
1088         rss = round_up(rss, nr_scan_pages);
1089         return rss / nr_scan_pages;
1090 }
1091 
1092 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1093 #define MAX_SCAN_WINDOW 2560
1094 
1095 static unsigned int task_scan_min(struct task_struct *p)
1096 {
1097         unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1098         unsigned int scan, floor;
1099         unsigned int windows = 1;
1100 
1101         if (scan_size < MAX_SCAN_WINDOW)
1102                 windows = MAX_SCAN_WINDOW / scan_size;
1103         floor = 1000 / windows;
1104 
1105         scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1106         return max_t(unsigned int, floor, scan);
1107 }
1108 
1109 static unsigned int task_scan_start(struct task_struct *p)
1110 {
1111         unsigned long smin = task_scan_min(p);
1112         unsigned long period = smin;
1113 
1114         /* Scale the maximum scan period with the amount of shared memory. */
1115         if (p->numa_group) {
1116                 struct numa_group *ng = p->numa_group;
1117                 unsigned long shared = group_faults_shared(ng);
1118                 unsigned long private = group_faults_priv(ng);
1119 
1120                 period *= refcount_read(&ng->refcount);
1121                 period *= shared + 1;
1122                 period /= private + shared + 1;
1123         }
1124 
1125         return max(smin, period);
1126 }
1127 
1128 static unsigned int task_scan_max(struct task_struct *p)
1129 {
1130         unsigned long smin = task_scan_min(p);
1131         unsigned long smax;
1132 
1133         /* Watch for min being lower than max due to floor calculations */
1134         smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1135 
1136         /* Scale the maximum scan period with the amount of shared memory. */
1137         if (p->numa_group) {
1138                 struct numa_group *ng = p->numa_group;
1139                 unsigned long shared = group_faults_shared(ng);
1140                 unsigned long private = group_faults_priv(ng);
1141                 unsigned long period = smax;
1142 
1143                 period *= refcount_read(&ng->refcount);
1144                 period *= shared + 1;
1145                 period /= private + shared + 1;
1146 
1147                 smax = max(smax, period);
1148         }
1149 
1150         return max(smin, smax);
1151 }
1152 
1153 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
1154 {
1155         int mm_users = 0;
1156         struct mm_struct *mm = p->mm;
1157 
1158         if (mm) {
1159                 mm_users = atomic_read(&mm->mm_users);
1160                 if (mm_users == 1) {
1161                         mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1162                         mm->numa_scan_seq = 0;
1163                 }
1164         }
1165         p->node_stamp                   = 0;
1166         p->numa_scan_seq                = mm ? mm->numa_scan_seq : 0;
1167         p->numa_scan_period             = sysctl_numa_balancing_scan_delay;
1168         p->numa_work.next               = &p->numa_work;
1169         p->numa_faults                  = NULL;
1170         p->numa_group                   = NULL;
1171         p->last_task_numa_placement     = 0;
1172         p->last_sum_exec_runtime        = 0;
1173 
1174         /* New address space, reset the preferred nid */
1175         if (!(clone_flags & CLONE_VM)) {
1176                 p->numa_preferred_nid = NUMA_NO_NODE;
1177                 return;
1178         }
1179 
1180         /*
1181          * New thread, keep existing numa_preferred_nid which should be copied
1182          * already by arch_dup_task_struct but stagger when scans start.
1183          */
1184         if (mm) {
1185                 unsigned int delay;
1186 
1187                 delay = min_t(unsigned int, task_scan_max(current),
1188                         current->numa_scan_period * mm_users * NSEC_PER_MSEC);
1189                 delay += 2 * TICK_NSEC;
1190                 p->node_stamp = delay;
1191         }
1192 }
1193 
1194 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1195 {
1196         rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
1197         rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1198 }
1199 
1200 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1201 {
1202         rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1203         rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1204 }
1205 
1206 /* Shared or private faults. */
1207 #define NR_NUMA_HINT_FAULT_TYPES 2
1208 
1209 /* Memory and CPU locality */
1210 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1211 
1212 /* Averaged statistics, and temporary buffers. */
1213 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1214 
1215 pid_t task_numa_group_id(struct task_struct *p)
1216 {
1217         return p->numa_group ? p->numa_group->gid : 0;
1218 }
1219 
1220 /*
1221  * The averaged statistics, shared & private, memory & CPU,
1222  * occupy the first half of the array. The second half of the
1223  * array is for current counters, which are averaged into the
1224  * first set by task_numa_placement.
1225  */
1226 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1227 {
1228         return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1229 }
1230 
1231 static inline unsigned long task_faults(struct task_struct *p, int nid)
1232 {
1233         if (!p->numa_faults)
1234                 return 0;
1235 
1236         return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1237                 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1238 }
1239 
1240 static inline unsigned long group_faults(struct task_struct *p, int nid)
1241 {
1242         if (!p->numa_group)
1243                 return 0;
1244 
1245         return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1246                 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1247 }
1248 
1249 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1250 {
1251         return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1252                 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1253 }
1254 
1255 static inline unsigned long group_faults_priv(struct numa_group *ng)
1256 {
1257         unsigned long faults = 0;
1258         int node;
1259 
1260         for_each_online_node(node) {
1261                 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1262         }
1263 
1264         return faults;
1265 }
1266 
1267 static inline unsigned long group_faults_shared(struct numa_group *ng)
1268 {
1269         unsigned long faults = 0;
1270         int node;
1271 
1272         for_each_online_node(node) {
1273                 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1274         }
1275 
1276         return faults;
1277 }
1278 
1279 /*
1280  * A node triggering more than 1/3 as many NUMA faults as the maximum is
1281  * considered part of a numa group's pseudo-interleaving set. Migrations
1282  * between these nodes are slowed down, to allow things to settle down.
1283  */
1284 #define ACTIVE_NODE_FRACTION 3
1285 
1286 static bool numa_is_active_node(int nid, struct numa_group *ng)
1287 {
1288         return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1289 }
1290 
1291 /* Handle placement on systems where not all nodes are directly connected. */
1292 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1293                                         int maxdist, bool task)
1294 {
1295         unsigned long score = 0;
1296         int node;
1297 
1298         /*
1299          * All nodes are directly connected, and the same distance
1300          * from each other. No need for fancy placement algorithms.
1301          */
1302         if (sched_numa_topology_type == NUMA_DIRECT)
1303                 return 0;
1304 
1305         /*
1306          * This code is called for each node, introducing N^2 complexity,
1307          * which should be ok given the number of nodes rarely exceeds 8.
1308          */
1309         for_each_online_node(node) {
1310                 unsigned long faults;
1311                 int dist = node_distance(nid, node);
1312 
1313                 /*
1314                  * The furthest away nodes in the system are not interesting
1315                  * for placement; nid was already counted.
1316                  */
1317                 if (dist == sched_max_numa_distance || node == nid)
1318                         continue;
1319 
1320                 /*
1321                  * On systems with a backplane NUMA topology, compare groups
1322                  * of nodes, and move tasks towards the group with the most
1323                  * memory accesses. When comparing two nodes at distance
1324                  * "hoplimit", only nodes closer by than "hoplimit" are part
1325                  * of each group. Skip other nodes.
1326                  */
1327                 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1328                                         dist >= maxdist)
1329                         continue;
1330 
1331                 /* Add up the faults from nearby nodes. */
1332                 if (task)
1333                         faults = task_faults(p, node);
1334                 else
1335                         faults = group_faults(p, node);
1336 
1337                 /*
1338                  * On systems with a glueless mesh NUMA topology, there are
1339                  * no fixed "groups of nodes". Instead, nodes that are not
1340                  * directly connected bounce traffic through intermediate
1341                  * nodes; a numa_group can occupy any set of nodes.
1342                  * The further away a node is, the less the faults count.
1343                  * This seems to result in good task placement.
1344                  */
1345                 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1346                         faults *= (sched_max_numa_distance - dist);
1347                         faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1348                 }
1349 
1350                 score += faults;
1351         }
1352 
1353         return score;
1354 }
1355 
1356 /*
1357  * These return the fraction of accesses done by a particular task, or
1358  * task group, on a particular numa node.  The group weight is given a
1359  * larger multiplier, in order to group tasks together that are almost
1360  * evenly spread out between numa nodes.
1361  */
1362 static inline unsigned long task_weight(struct task_struct *p, int nid,
1363                                         int dist)
1364 {
1365         unsigned long faults, total_faults;
1366 
1367         if (!p->numa_faults)
1368                 return 0;
1369 
1370         total_faults = p->total_numa_faults;
1371 
1372         if (!total_faults)
1373                 return 0;
1374 
1375         faults = task_faults(p, nid);
1376         faults += score_nearby_nodes(p, nid, dist, true);
1377 
1378         return 1000 * faults / total_faults;
1379 }
1380 
1381 static inline unsigned long group_weight(struct task_struct *p, int nid,
1382                                          int dist)
1383 {
1384         unsigned long faults, total_faults;
1385 
1386         if (!p->numa_group)
1387                 return 0;
1388 
1389         total_faults = p->numa_group->total_faults;
1390 
1391         if (!total_faults)
1392                 return 0;
1393 
1394         faults = group_faults(p, nid);
1395         faults += score_nearby_nodes(p, nid, dist, false);
1396 
1397         return 1000 * faults / total_faults;
1398 }
1399 
1400 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1401                                 int src_nid, int dst_cpu)
1402 {
1403         struct numa_group *ng = p->numa_group;
1404         int dst_nid = cpu_to_node(dst_cpu);
1405         int last_cpupid, this_cpupid;
1406 
1407         this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1408         last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1409 
1410         /*
1411          * Allow first faults or private faults to migrate immediately early in
1412          * the lifetime of a task. The magic number 4 is based on waiting for
1413          * two full passes of the "multi-stage node selection" test that is
1414          * executed below.
1415          */
1416         if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1417             (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1418                 return true;
1419 
1420         /*
1421          * Multi-stage node selection is used in conjunction with a periodic
1422          * migration fault to build a temporal task<->page relation. By using
1423          * a two-stage filter we remove short/unlikely relations.
1424          *
1425          * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1426          * a task's usage of a particular page (n_p) per total usage of this
1427          * page (n_t) (in a given time-span) to a probability.
1428          *
1429          * Our periodic faults will sample this probability and getting the
1430          * same result twice in a row, given these samples are fully
1431          * independent, is then given by P(n)^2, provided our sample period
1432          * is sufficiently short compared to the usage pattern.
1433          *
1434          * This quadric squishes small probabilities, making it less likely we
1435          * act on an unlikely task<->page relation.
1436          */
1437         if (!cpupid_pid_unset(last_cpupid) &&
1438                                 cpupid_to_nid(last_cpupid) != dst_nid)
1439                 return false;
1440 
1441         /* Always allow migrate on private faults */
1442         if (cpupid_match_pid(p, last_cpupid))
1443                 return true;
1444 
1445         /* A shared fault, but p->numa_group has not been set up yet. */
1446         if (!ng)
1447                 return true;
1448 
1449         /*
1450          * Destination node is much more heavily used than the source
1451          * node? Allow migration.
1452          */
1453         if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1454                                         ACTIVE_NODE_FRACTION)
1455                 return true;
1456 
1457         /*
1458          * Distribute memory according to CPU & memory use on each node,
1459          * with 3/4 hysteresis to avoid unnecessary memory migrations:
1460          *
1461          * faults_cpu(dst)   3   faults_cpu(src)
1462          * --------------- * - > ---------------
1463          * faults_mem(dst)   4   faults_mem(src)
1464          */
1465         return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1466                group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1467 }
1468 
1469 static unsigned long weighted_cpuload(struct rq *rq);
1470 static unsigned long source_load(int cpu, int type);
1471 static unsigned long target_load(int cpu, int type);
1472 
1473 /* Cached statistics for all CPUs within a node */
1474 struct numa_stats {
1475         unsigned long load;
1476 
1477         /* Total compute capacity of CPUs on a node */
1478         unsigned long compute_capacity;
1479 };
1480 
1481 /*
1482  * XXX borrowed from update_sg_lb_stats
1483  */
1484 static void update_numa_stats(struct numa_stats *ns, int nid)
1485 {
1486         int cpu;
1487 
1488         memset(ns, 0, sizeof(*ns));
1489         for_each_cpu(cpu, cpumask_of_node(nid)) {
1490                 struct rq *rq = cpu_rq(cpu);
1491 
1492                 ns->load += weighted_cpuload(rq);
1493                 ns->compute_capacity += capacity_of(cpu);
1494         }
1495 
1496 }
1497 
1498 struct task_numa_env {
1499         struct task_struct *p;
1500 
1501         int src_cpu, src_nid;
1502         int dst_cpu, dst_nid;
1503 
1504         struct numa_stats src_stats, dst_stats;
1505 
1506         int imbalance_pct;
1507         int dist;
1508 
1509         struct task_struct *best_task;
1510         long best_imp;
1511         int best_cpu;
1512 };
1513 
1514 static void task_numa_assign(struct task_numa_env *env,
1515                              struct task_struct *p, long imp)
1516 {
1517         struct rq *rq = cpu_rq(env->dst_cpu);
1518 
1519         /* Bail out if run-queue part of active NUMA balance. */
1520         if (xchg(&rq->numa_migrate_on, 1))
1521                 return;
1522 
1523         /*
1524          * Clear previous best_cpu/rq numa-migrate flag, since task now
1525          * found a better CPU to move/swap.
1526          */
1527         if (env->best_cpu != -1) {
1528                 rq = cpu_rq(env->best_cpu);
1529                 WRITE_ONCE(rq->numa_migrate_on, 0);
1530         }
1531 
1532         if (env->best_task)
1533                 put_task_struct(env->best_task);
1534         if (p)
1535                 get_task_struct(p);
1536 
1537         env->best_task = p;
1538         env->best_imp = imp;
1539         env->best_cpu = env->dst_cpu;
1540 }
1541 
1542 static bool load_too_imbalanced(long src_load, long dst_load,
1543                                 struct task_numa_env *env)
1544 {
1545         long imb, old_imb;
1546         long orig_src_load, orig_dst_load;
1547         long src_capacity, dst_capacity;
1548 
1549         /*
1550          * The load is corrected for the CPU capacity available on each node.
1551          *
1552          * src_load        dst_load
1553          * ------------ vs ---------
1554          * src_capacity    dst_capacity
1555          */
1556         src_capacity = env->src_stats.compute_capacity;
1557         dst_capacity = env->dst_stats.compute_capacity;
1558 
1559         imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1560 
1561         orig_src_load = env->src_stats.load;
1562         orig_dst_load = env->dst_stats.load;
1563 
1564         old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1565 
1566         /* Would this change make things worse? */
1567         return (imb > old_imb);
1568 }
1569 
1570 /*
1571  * Maximum NUMA importance can be 1998 (2*999);
1572  * SMALLIMP @ 30 would be close to 1998/64.
1573  * Used to deter task migration.
1574  */
1575 #define SMALLIMP        30
1576 
1577 /*
1578  * This checks if the overall compute and NUMA accesses of the system would
1579  * be improved if the source tasks was migrated to the target dst_cpu taking
1580  * into account that it might be best if task running on the dst_cpu should
1581  * be exchanged with the source task
1582  */
1583 static void task_numa_compare(struct task_numa_env *env,
1584                               long taskimp, long groupimp, bool maymove)
1585 {
1586         struct rq *dst_rq = cpu_rq(env->dst_cpu);
1587         struct task_struct *cur;
1588         long src_load, dst_load;
1589         long load;
1590         long imp = env->p->numa_group ? groupimp : taskimp;
1591         long moveimp = imp;
1592         int dist = env->dist;
1593 
1594         if (READ_ONCE(dst_rq->numa_migrate_on))
1595                 return;
1596 
1597         rcu_read_lock();
1598         cur = task_rcu_dereference(&dst_rq->curr);
1599         if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1600                 cur = NULL;
1601 
1602         /*
1603          * Because we have preemption enabled we can get migrated around and
1604          * end try selecting ourselves (current == env->p) as a swap candidate.
1605          */
1606         if (cur == env->p)
1607                 goto unlock;
1608 
1609         if (!cur) {
1610                 if (maymove && moveimp >= env->best_imp)
1611                         goto assign;
1612                 else
1613                         goto unlock;
1614         }
1615 
1616         /*
1617          * "imp" is the fault differential for the source task between the
1618          * source and destination node. Calculate the total differential for
1619          * the source task and potential destination task. The more negative
1620          * the value is, the more remote accesses that would be expected to
1621          * be incurred if the tasks were swapped.
1622          */
1623         /* Skip this swap candidate if cannot move to the source cpu */
1624         if (!cpumask_test_cpu(env->src_cpu, &cur->cpus_allowed))
1625                 goto unlock;
1626 
1627         /*
1628          * If dst and source tasks are in the same NUMA group, or not
1629          * in any group then look only at task weights.
1630          */
1631         if (cur->numa_group == env->p->numa_group) {
1632                 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1633                       task_weight(cur, env->dst_nid, dist);
1634                 /*
1635                  * Add some hysteresis to prevent swapping the
1636                  * tasks within a group over tiny differences.
1637                  */
1638                 if (cur->numa_group)
1639                         imp -= imp / 16;
1640         } else {
1641                 /*
1642                  * Compare the group weights. If a task is all by itself
1643                  * (not part of a group), use the task weight instead.
1644                  */
1645                 if (cur->numa_group && env->p->numa_group)
1646                         imp += group_weight(cur, env->src_nid, dist) -
1647                                group_weight(cur, env->dst_nid, dist);
1648                 else
1649                         imp += task_weight(cur, env->src_nid, dist) -
1650                                task_weight(cur, env->dst_nid, dist);
1651         }
1652 
1653         if (maymove && moveimp > imp && moveimp > env->best_imp) {
1654                 imp = moveimp;
1655                 cur = NULL;
1656                 goto assign;
1657         }
1658 
1659         /*
1660          * If the NUMA importance is less than SMALLIMP,
1661          * task migration might only result in ping pong
1662          * of tasks and also hurt performance due to cache
1663          * misses.
1664          */
1665         if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
1666                 goto unlock;
1667 
1668         /*
1669          * In the overloaded case, try and keep the load balanced.
1670          */
1671         load = task_h_load(env->p) - task_h_load(cur);
1672         if (!load)
1673                 goto assign;
1674 
1675         dst_load = env->dst_stats.load + load;
1676         src_load = env->src_stats.load - load;
1677 
1678         if (load_too_imbalanced(src_load, dst_load, env))
1679                 goto unlock;
1680 
1681 assign:
1682         /*
1683          * One idle CPU per node is evaluated for a task numa move.
1684          * Call select_idle_sibling to maybe find a better one.
1685          */
1686         if (!cur) {
1687                 /*
1688                  * select_idle_siblings() uses an per-CPU cpumask that
1689                  * can be used from IRQ context.
1690                  */
1691                 local_irq_disable();
1692                 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1693                                                    env->dst_cpu);
1694                 local_irq_enable();
1695         }
1696 
1697         task_numa_assign(env, cur, imp);
1698 unlock:
1699         rcu_read_unlock();
1700 }
1701 
1702 static void task_numa_find_cpu(struct task_numa_env *env,
1703                                 long taskimp, long groupimp)
1704 {
1705         long src_load, dst_load, load;
1706         bool maymove = false;
1707         int cpu;
1708 
1709         load = task_h_load(env->p);
1710         dst_load = env->dst_stats.load + load;
1711         src_load = env->src_stats.load - load;
1712 
1713         /*
1714          * If the improvement from just moving env->p direction is better
1715          * than swapping tasks around, check if a move is possible.
1716          */
1717         maymove = !load_too_imbalanced(src_load, dst_load, env);
1718 
1719         for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1720                 /* Skip this CPU if the source task cannot migrate */
1721                 if (!cpumask_test_cpu(cpu, &env->p->cpus_allowed))
1722                         continue;
1723 
1724                 env->dst_cpu = cpu;
1725                 task_numa_compare(env, taskimp, groupimp, maymove);
1726         }
1727 }
1728 
1729 static int task_numa_migrate(struct task_struct *p)
1730 {
1731         struct task_numa_env env = {
1732                 .p = p,
1733 
1734                 .src_cpu = task_cpu(p),
1735                 .src_nid = task_node(p),
1736 
1737                 .imbalance_pct = 112,
1738 
1739                 .best_task = NULL,
1740                 .best_imp = 0,
1741                 .best_cpu = -1,
1742         };
1743         struct sched_domain *sd;
1744         struct rq *best_rq;
1745         unsigned long taskweight, groupweight;
1746         int nid, ret, dist;
1747         long taskimp, groupimp;
1748 
1749         /*
1750          * Pick the lowest SD_NUMA domain, as that would have the smallest
1751          * imbalance and would be the first to start moving tasks about.
1752          *
1753          * And we want to avoid any moving of tasks about, as that would create
1754          * random movement of tasks -- counter the numa conditions we're trying
1755          * to satisfy here.
1756          */
1757         rcu_read_lock();
1758         sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1759         if (sd)
1760                 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1761         rcu_read_unlock();
1762 
1763         /*
1764          * Cpusets can break the scheduler domain tree into smaller
1765          * balance domains, some of which do not cross NUMA boundaries.
1766          * Tasks that are "trapped" in such domains cannot be migrated
1767          * elsewhere, so there is no point in (re)trying.
1768          */
1769         if (unlikely(!sd)) {
1770                 sched_setnuma(p, task_node(p));
1771                 return -EINVAL;
1772         }
1773 
1774         env.dst_nid = p->numa_preferred_nid;
1775         dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1776         taskweight = task_weight(p, env.src_nid, dist);
1777         groupweight = group_weight(p, env.src_nid, dist);
1778         update_numa_stats(&env.src_stats, env.src_nid);
1779         taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1780         groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1781         update_numa_stats(&env.dst_stats, env.dst_nid);
1782 
1783         /* Try to find a spot on the preferred nid. */
1784         task_numa_find_cpu(&env, taskimp, groupimp);
1785 
1786         /*
1787          * Look at other nodes in these cases:
1788          * - there is no space available on the preferred_nid
1789          * - the task is part of a numa_group that is interleaved across
1790          *   multiple NUMA nodes; in order to better consolidate the group,
1791          *   we need to check other locations.
1792          */
1793         if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1794                 for_each_online_node(nid) {
1795                         if (nid == env.src_nid || nid == p->numa_preferred_nid)
1796                                 continue;
1797 
1798                         dist = node_distance(env.src_nid, env.dst_nid);
1799                         if (sched_numa_topology_type == NUMA_BACKPLANE &&
1800                                                 dist != env.dist) {
1801                                 taskweight = task_weight(p, env.src_nid, dist);
1802                                 groupweight = group_weight(p, env.src_nid, dist);
1803                         }
1804 
1805                         /* Only consider nodes where both task and groups benefit */
1806                         taskimp = task_weight(p, nid, dist) - taskweight;
1807                         groupimp = group_weight(p, nid, dist) - groupweight;
1808                         if (taskimp < 0 && groupimp < 0)
1809                                 continue;
1810 
1811                         env.dist = dist;
1812                         env.dst_nid = nid;
1813                         update_numa_stats(&env.dst_stats, env.dst_nid);
1814                         task_numa_find_cpu(&env, taskimp, groupimp);
1815                 }
1816         }
1817 
1818         /*
1819          * If the task is part of a workload that spans multiple NUMA nodes,
1820          * and is migrating into one of the workload's active nodes, remember
1821          * this node as the task's preferred numa node, so the workload can
1822          * settle down.
1823          * A task that migrated to a second choice node will be better off
1824          * trying for a better one later. Do not set the preferred node here.
1825          */
1826         if (p->numa_group) {
1827                 if (env.best_cpu == -1)
1828                         nid = env.src_nid;
1829                 else
1830                         nid = cpu_to_node(env.best_cpu);
1831 
1832                 if (nid != p->numa_preferred_nid)
1833                         sched_setnuma(p, nid);
1834         }
1835 
1836         /* No better CPU than the current one was found. */
1837         if (env.best_cpu == -1)
1838                 return -EAGAIN;
1839 
1840         best_rq = cpu_rq(env.best_cpu);
1841         if (env.best_task == NULL) {
1842                 ret = migrate_task_to(p, env.best_cpu);
1843                 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1844                 if (ret != 0)
1845                         trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1846                 return ret;
1847         }
1848 
1849         ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
1850         WRITE_ONCE(best_rq->numa_migrate_on, 0);
1851 
1852         if (ret != 0)
1853                 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1854         put_task_struct(env.best_task);
1855         return ret;
1856 }
1857 
1858 /* Attempt to migrate a task to a CPU on the preferred node. */
1859 static void numa_migrate_preferred(struct task_struct *p)
1860 {
1861         unsigned long interval = HZ;
1862 
1863         /* This task has no NUMA fault statistics yet */
1864         if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
1865                 return;
1866 
1867         /* Periodically retry migrating the task to the preferred node */
1868         interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1869         p->numa_migrate_retry = jiffies + interval;
1870 
1871         /* Success if task is already running on preferred CPU */
1872         if (task_node(p) == p->numa_preferred_nid)
1873                 return;
1874 
1875         /* Otherwise, try migrate to a CPU on the preferred node */
1876         task_numa_migrate(p);
1877 }
1878 
1879 /*
1880  * Find out how many nodes on the workload is actively running on. Do this by
1881  * tracking the nodes from which NUMA hinting faults are triggered. This can
1882  * be different from the set of nodes where the workload's memory is currently
1883  * located.
1884  */
1885 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1886 {
1887         unsigned long faults, max_faults = 0;
1888         int nid, active_nodes = 0;
1889 
1890         for_each_online_node(nid) {
1891                 faults = group_faults_cpu(numa_group, nid);
1892                 if (faults > max_faults)
1893                         max_faults = faults;
1894         }
1895 
1896         for_each_online_node(nid) {
1897                 faults = group_faults_cpu(numa_group, nid);
1898                 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1899                         active_nodes++;
1900         }
1901 
1902         numa_group->max_faults_cpu = max_faults;
1903         numa_group->active_nodes = active_nodes;
1904 }
1905 
1906 /*
1907  * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1908  * increments. The more local the fault statistics are, the higher the scan
1909  * period will be for the next scan window. If local/(local+remote) ratio is
1910  * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1911  * the scan period will decrease. Aim for 70% local accesses.
1912  */
1913 #define NUMA_PERIOD_SLOTS 10
1914 #define NUMA_PERIOD_THRESHOLD 7
1915 
1916 /*
1917  * Increase the scan period (slow down scanning) if the majority of
1918  * our memory is already on our local node, or if the majority of
1919  * the page accesses are shared with other processes.
1920  * Otherwise, decrease the scan period.
1921  */
1922 static void update_task_scan_period(struct task_struct *p,
1923                         unsigned long shared, unsigned long private)
1924 {
1925         unsigned int period_slot;
1926         int lr_ratio, ps_ratio;
1927         int diff;
1928 
1929         unsigned long remote = p->numa_faults_locality[0];
1930         unsigned long local = p->numa_faults_locality[1];
1931 
1932         /*
1933          * If there were no record hinting faults then either the task is
1934          * completely idle or all activity is areas that are not of interest
1935          * to automatic numa balancing. Related to that, if there were failed
1936          * migration then it implies we are migrating too quickly or the local
1937          * node is overloaded. In either case, scan slower
1938          */
1939         if (local + shared == 0 || p->numa_faults_locality[2]) {
1940                 p->numa_scan_period = min(p->numa_scan_period_max,
1941                         p->numa_scan_period << 1);
1942 
1943                 p->mm->numa_next_scan = jiffies +
1944                         msecs_to_jiffies(p->numa_scan_period);
1945 
1946                 return;
1947         }
1948 
1949         /*
1950          * Prepare to scale scan period relative to the current period.
1951          *       == NUMA_PERIOD_THRESHOLD scan period stays the same
1952          *       <  NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1953          *       >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1954          */
1955         period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1956         lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1957         ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
1958 
1959         if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
1960                 /*
1961                  * Most memory accesses are local. There is no need to
1962                  * do fast NUMA scanning, since memory is already local.
1963                  */
1964                 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
1965                 if (!slot)
1966                         slot = 1;
1967                 diff = slot * period_slot;
1968         } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
1969                 /*
1970                  * Most memory accesses are shared with other tasks.
1971                  * There is no point in continuing fast NUMA scanning,
1972                  * since other tasks may just move the memory elsewhere.
1973                  */
1974                 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
1975                 if (!slot)
1976                         slot = 1;
1977                 diff = slot * period_slot;
1978         } else {
1979                 /*
1980                  * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
1981                  * yet they are not on the local NUMA node. Speed up
1982                  * NUMA scanning to get the memory moved over.
1983                  */
1984                 int ratio = max(lr_ratio, ps_ratio);
1985                 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1986         }
1987 
1988         p->numa_scan_period = clamp(p->numa_scan_period + diff,
1989                         task_scan_min(p), task_scan_max(p));
1990         memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1991 }
1992 
1993 /*
1994  * Get the fraction of time the task has been running since the last
1995  * NUMA placement cycle. The scheduler keeps similar statistics, but
1996  * decays those on a 32ms period, which is orders of magnitude off
1997  * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1998  * stats only if the task is so new there are no NUMA statistics yet.
1999  */
2000 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2001 {
2002         u64 runtime, delta, now;
2003         /* Use the start of this time slice to avoid calculations. */
2004         now = p->se.exec_start;
2005         runtime = p->se.sum_exec_runtime;
2006 
2007         if (p->last_task_numa_placement) {
2008                 delta = runtime - p->last_sum_exec_runtime;
2009                 *period = now - p->last_task_numa_placement;
2010 
2011                 /* Avoid time going backwards, prevent potential divide error: */
2012                 if (unlikely((s64)*period < 0))
2013                         *period = 0;
2014         } else {
2015                 delta = p->se.avg.load_sum;
2016                 *period = LOAD_AVG_MAX;
2017         }
2018 
2019         p->last_sum_exec_runtime = runtime;
2020         p->last_task_numa_placement = now;
2021 
2022         return delta;
2023 }
2024 
2025 /*
2026  * Determine the preferred nid for a task in a numa_group. This needs to
2027  * be done in a way that produces consistent results with group_weight,
2028  * otherwise workloads might not converge.
2029  */
2030 static int preferred_group_nid(struct task_struct *p, int nid)
2031 {
2032         nodemask_t nodes;
2033         int dist;
2034 
2035         /* Direct connections between all NUMA nodes. */
2036         if (sched_numa_topology_type == NUMA_DIRECT)
2037                 return nid;
2038 
2039         /*
2040          * On a system with glueless mesh NUMA topology, group_weight
2041          * scores nodes according to the number of NUMA hinting faults on
2042          * both the node itself, and on nearby nodes.
2043          */
2044         if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2045                 unsigned long score, max_score = 0;
2046                 int node, max_node = nid;
2047 
2048                 dist = sched_max_numa_distance;
2049 
2050                 for_each_online_node(node) {
2051                         score = group_weight(p, node, dist);
2052                         if (score > max_score) {
2053                                 max_score = score;
2054                                 max_node = node;
2055                         }
2056                 }
2057                 return max_node;
2058         }
2059 
2060         /*
2061          * Finding the preferred nid in a system with NUMA backplane
2062          * interconnect topology is more involved. The goal is to locate
2063          * tasks from numa_groups near each other in the system, and
2064          * untangle workloads from different sides of the system. This requires
2065          * searching down the hierarchy of node groups, recursively searching
2066          * inside the highest scoring group of nodes. The nodemask tricks
2067          * keep the complexity of the search down.
2068          */
2069         nodes = node_online_map;
2070         for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2071                 unsigned long max_faults = 0;
2072                 nodemask_t max_group = NODE_MASK_NONE;
2073                 int a, b;
2074 
2075                 /* Are there nodes at this distance from each other? */
2076                 if (!find_numa_distance(dist))
2077                         continue;
2078 
2079                 for_each_node_mask(a, nodes) {
2080                         unsigned long faults = 0;
2081                         nodemask_t this_group;
2082                         nodes_clear(this_group);
2083 
2084                         /* Sum group's NUMA faults; includes a==b case. */
2085                         for_each_node_mask(b, nodes) {
2086                                 if (node_distance(a, b) < dist) {
2087                                         faults += group_faults(p, b);
2088                                         node_set(b, this_group);
2089                                         node_clear(b, nodes);
2090                                 }
2091                         }
2092 
2093                         /* Remember the top group. */
2094                         if (faults > max_faults) {
2095                                 max_faults = faults;
2096                                 max_group = this_group;
2097                                 /*
2098                                  * subtle: at the smallest distance there is
2099                                  * just one node left in each "group", the
2100                                  * winner is the preferred nid.
2101                                  */
2102                                 nid = a;
2103                         }
2104                 }
2105                 /* Next round, evaluate the nodes within max_group. */
2106                 if (!max_faults)
2107                         break;
2108                 nodes = max_group;
2109         }
2110         return nid;
2111 }
2112 
2113 static void task_numa_placement(struct task_struct *p)
2114 {
2115         int seq, nid, max_nid = NUMA_NO_NODE;
2116         unsigned long max_faults = 0;
2117         unsigned long fault_types[2] = { 0, 0 };
2118         unsigned long total_faults;
2119         u64 runtime, period;
2120         spinlock_t *group_lock = NULL;
2121 
2122         /*
2123          * The p->mm->numa_scan_seq field gets updated without
2124          * exclusive access. Use READ_ONCE() here to ensure
2125          * that the field is read in a single access:
2126          */
2127         seq = READ_ONCE(p->mm->numa_scan_seq);
2128         if (p->numa_scan_seq == seq)
2129                 return;
2130         p->numa_scan_seq = seq;
2131         p->numa_scan_period_max = task_scan_max(p);
2132 
2133         total_faults = p->numa_faults_locality[0] +
2134                        p->numa_faults_locality[1];
2135         runtime = numa_get_avg_runtime(p, &period);
2136 
2137         /* If the task is part of a group prevent parallel updates to group stats */
2138         if (p->numa_group) {
2139                 group_lock = &p->numa_group->lock;
2140                 spin_lock_irq(group_lock);
2141         }
2142 
2143         /* Find the node with the highest number of faults */
2144         for_each_online_node(nid) {
2145                 /* Keep track of the offsets in numa_faults array */
2146                 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2147                 unsigned long faults = 0, group_faults = 0;
2148                 int priv;
2149 
2150                 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2151                         long diff, f_diff, f_weight;
2152 
2153                         mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2154                         membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2155                         cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2156                         cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2157 
2158                         /* Decay existing window, copy faults since last scan */
2159                         diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2160                         fault_types[priv] += p->numa_faults[membuf_idx];
2161                         p->numa_faults[membuf_idx] = 0;
2162 
2163                         /*
2164                          * Normalize the faults_from, so all tasks in a group
2165                          * count according to CPU use, instead of by the raw
2166                          * number of faults. Tasks with little runtime have
2167                          * little over-all impact on throughput, and thus their
2168                          * faults are less important.
2169                          */
2170                         f_weight = div64_u64(runtime << 16, period + 1);
2171                         f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2172                                    (total_faults + 1);
2173                         f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2174                         p->numa_faults[cpubuf_idx] = 0;
2175 
2176                         p->numa_faults[mem_idx] += diff;
2177                         p->numa_faults[cpu_idx] += f_diff;
2178                         faults += p->numa_faults[mem_idx];
2179                         p->total_numa_faults += diff;
2180                         if (p->numa_group) {
2181                                 /*
2182                                  * safe because we can only change our own group
2183                                  *
2184                                  * mem_idx represents the offset for a given
2185                                  * nid and priv in a specific region because it
2186                                  * is at the beginning of the numa_faults array.
2187                                  */
2188                                 p->numa_group->faults[mem_idx] += diff;
2189                                 p->numa_group->faults_cpu[mem_idx] += f_diff;
2190                                 p->numa_group->total_faults += diff;
2191                                 group_faults += p->numa_group->faults[mem_idx];
2192                         }
2193                 }
2194 
2195                 if (!p->numa_group) {
2196                         if (faults > max_faults) {
2197                                 max_faults = faults;
2198                                 max_nid = nid;
2199                         }
2200                 } else if (group_faults > max_faults) {
2201                         max_faults = group_faults;
2202                         max_nid = nid;
2203                 }
2204         }
2205 
2206         if (p->numa_group) {
2207                 numa_group_count_active_nodes(p->numa_group);
2208                 spin_unlock_irq(group_lock);
2209                 max_nid = preferred_group_nid(p, max_nid);
2210         }
2211 
2212         if (max_faults) {
2213                 /* Set the new preferred node */
2214                 if (max_nid != p->numa_preferred_nid)
2215                         sched_setnuma(p, max_nid);
2216         }
2217 
2218         update_task_scan_period(p, fault_types[0], fault_types[1]);
2219 }
2220 
2221 static inline int get_numa_group(struct numa_group *grp)
2222 {
2223         return refcount_inc_not_zero(&grp->refcount);
2224 }
2225 
2226 static inline void put_numa_group(struct numa_group *grp)
2227 {
2228         if (refcount_dec_and_test(&grp->refcount))
2229                 kfree_rcu(grp, rcu);
2230 }
2231 
2232 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2233                         int *priv)
2234 {
2235         struct numa_group *grp, *my_grp;
2236         struct task_struct *tsk;
2237         bool join = false;
2238         int cpu = cpupid_to_cpu(cpupid);
2239         int i;
2240 
2241         if (unlikely(!p->numa_group)) {
2242                 unsigned int size = sizeof(struct numa_group) +
2243                                     4*nr_node_ids*sizeof(unsigned long);
2244 
2245                 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2246                 if (!grp)
2247                         return;
2248 
2249                 refcount_set(&grp->refcount, 1);
2250                 grp->active_nodes = 1;
2251                 grp->max_faults_cpu = 0;
2252                 spin_lock_init(&grp->lock);
2253                 grp->gid = p->pid;
2254                 /* Second half of the array tracks nids where faults happen */
2255                 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2256                                                 nr_node_ids;
2257 
2258                 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2259                         grp->faults[i] = p->numa_faults[i];
2260 
2261                 grp->total_faults = p->total_numa_faults;
2262 
2263                 grp->nr_tasks++;
2264                 rcu_assign_pointer(p->numa_group, grp);
2265         }
2266 
2267         rcu_read_lock();
2268         tsk = READ_ONCE(cpu_rq(cpu)->curr);
2269 
2270         if (!cpupid_match_pid(tsk, cpupid))
2271                 goto no_join;
2272 
2273         grp = rcu_dereference(tsk->numa_group);
2274         if (!grp)
2275                 goto no_join;
2276 
2277         my_grp = p->numa_group;
2278         if (grp == my_grp)
2279                 goto no_join;
2280 
2281         /*
2282          * Only join the other group if its bigger; if we're the bigger group,
2283          * the other task will join us.
2284          */
2285         if (my_grp->nr_tasks > grp->nr_tasks)
2286                 goto no_join;
2287 
2288         /*
2289          * Tie-break on the grp address.
2290          */
2291         if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2292                 goto no_join;
2293 
2294         /* Always join threads in the same process. */
2295         if (tsk->mm == current->mm)
2296                 join = true;
2297 
2298         /* Simple filter to avoid false positives due to PID collisions */
2299         if (flags & TNF_SHARED)
2300                 join = true;
2301 
2302         /* Update priv based on whether false sharing was detected */
2303         *priv = !join;
2304 
2305         if (join && !get_numa_group(grp))
2306                 goto no_join;
2307 
2308         rcu_read_unlock();
2309 
2310         if (!join)
2311                 return;
2312 
2313         BUG_ON(irqs_disabled());
2314         double_lock_irq(&my_grp->lock, &grp->lock);
2315 
2316         for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2317                 my_grp->faults[i] -= p->numa_faults[i];
2318                 grp->faults[i] += p->numa_faults[i];
2319         }
2320         my_grp->total_faults -= p->total_numa_faults;
2321         grp->total_faults += p->total_numa_faults;
2322 
2323         my_grp->nr_tasks--;
2324         grp->nr_tasks++;
2325 
2326         spin_unlock(&my_grp->lock);
2327         spin_unlock_irq(&grp->lock);
2328 
2329         rcu_assign_pointer(p->numa_group, grp);
2330 
2331         put_numa_group(my_grp);
2332         return;
2333 
2334 no_join:
2335         rcu_read_unlock();
2336         return;
2337 }
2338 
2339 void task_numa_free(struct task_struct *p)
2340 {
2341         struct numa_group *grp = p->numa_group;
2342         void *numa_faults = p->numa_faults;
2343         unsigned long flags;
2344         int i;
2345 
2346         if (grp) {
2347                 spin_lock_irqsave(&grp->lock, flags);
2348                 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2349                         grp->faults[i] -= p->numa_faults[i];
2350                 grp->total_faults -= p->total_numa_faults;
2351 
2352                 grp->nr_tasks--;
2353                 spin_unlock_irqrestore(&grp->lock, flags);
2354                 RCU_INIT_POINTER(p->numa_group, NULL);
2355                 put_numa_group(grp);
2356         }
2357 
2358         p->numa_faults = NULL;
2359         kfree(numa_faults);
2360 }
2361 
2362 /*
2363  * Got a PROT_NONE fault for a page on @node.
2364  */
2365 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2366 {
2367         struct task_struct *p = current;
2368         bool migrated = flags & TNF_MIGRATED;
2369         int cpu_node = task_node(current);
2370         int local = !!(flags & TNF_FAULT_LOCAL);
2371         struct numa_group *ng;
2372         int priv;
2373 
2374         if (!static_branch_likely(&sched_numa_balancing))
2375                 return;
2376 
2377         /* for example, ksmd faulting in a user's mm */
2378         if (!p->mm)
2379                 return;
2380 
2381         /* Allocate buffer to track faults on a per-node basis */
2382         if (unlikely(!p->numa_faults)) {
2383                 int size = sizeof(*p->numa_faults) *
2384                            NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2385 
2386                 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2387                 if (!p->numa_faults)
2388                         return;
2389 
2390                 p->total_numa_faults = 0;
2391                 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2392         }
2393 
2394         /*
2395          * First accesses are treated as private, otherwise consider accesses
2396          * to be private if the accessing pid has not changed
2397          */
2398         if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2399                 priv = 1;
2400         } else {
2401                 priv = cpupid_match_pid(p, last_cpupid);
2402                 if (!priv && !(flags & TNF_NO_GROUP))
2403                         task_numa_group(p, last_cpupid, flags, &priv);
2404         }
2405 
2406         /*
2407          * If a workload spans multiple NUMA nodes, a shared fault that
2408          * occurs wholly within the set of nodes that the workload is
2409          * actively using should be counted as local. This allows the
2410          * scan rate to slow down when a workload has settled down.
2411          */
2412         ng = p->numa_group;
2413         if (!priv && !local && ng && ng->active_nodes > 1 &&
2414                                 numa_is_active_node(cpu_node, ng) &&
2415                                 numa_is_active_node(mem_node, ng))
2416                 local = 1;
2417 
2418         /*
2419          * Retry to migrate task to preferred node periodically, in case it
2420          * previously failed, or the scheduler moved us.
2421          */
2422         if (time_after(jiffies, p->numa_migrate_retry)) {
2423                 task_numa_placement(p);
2424                 numa_migrate_preferred(p);
2425         }
2426 
2427         if (migrated)
2428                 p->numa_pages_migrated += pages;
2429         if (flags & TNF_MIGRATE_FAIL)
2430                 p->numa_faults_locality[2] += pages;
2431 
2432         p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2433         p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2434         p->numa_faults_locality[local] += pages;
2435 }
2436 
2437 static void reset_ptenuma_scan(struct task_struct *p)
2438 {
2439         /*
2440          * We only did a read acquisition of the mmap sem, so
2441          * p->mm->numa_scan_seq is written to without exclusive access
2442          * and the update is not guaranteed to be atomic. That's not
2443          * much of an issue though, since this is just used for
2444          * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2445          * expensive, to avoid any form of compiler optimizations:
2446          */
2447         WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2448         p->mm->numa_scan_offset = 0;
2449 }
2450 
2451 /*
2452  * The expensive part of numa migration is done from task_work context.
2453  * Triggered from task_tick_numa().
2454  */
2455 void task_numa_work(struct callback_head *work)
2456 {
2457         unsigned long migrate, next_scan, now = jiffies;
2458         struct task_struct *p = current;
2459         struct mm_struct *mm = p->mm;
2460         u64 runtime = p->se.sum_exec_runtime;
2461         struct vm_area_struct *vma;
2462         unsigned long start, end;
2463         unsigned long nr_pte_updates = 0;
2464         long pages, virtpages;
2465 
2466         SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2467 
2468         work->next = work; /* protect against double add */
2469         /*
2470          * Who cares about NUMA placement when they're dying.
2471          *
2472          * NOTE: make sure not to dereference p->mm before this check,
2473          * exit_task_work() happens _after_ exit_mm() so we could be called
2474          * without p->mm even though we still had it when we enqueued this
2475          * work.
2476          */
2477         if (p->flags & PF_EXITING)
2478                 return;
2479 
2480         if (!mm->numa_next_scan) {
2481                 mm->numa_next_scan = now +
2482                         msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2483         }
2484 
2485         /*
2486          * Enforce maximal scan/migration frequency..
2487          */
2488         migrate = mm->numa_next_scan;
2489         if (time_before(now, migrate))
2490                 return;
2491 
2492         if (p->numa_scan_period == 0) {
2493                 p->numa_scan_period_max = task_scan_max(p);
2494                 p->numa_scan_period = task_scan_start(p);
2495         }
2496 
2497         next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2498         if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2499                 return;
2500 
2501         /*
2502          * Delay this task enough that another task of this mm will likely win
2503          * the next time around.
2504          */
2505         p->node_stamp += 2 * TICK_NSEC;
2506 
2507         start = mm->numa_scan_offset;
2508         pages = sysctl_numa_balancing_scan_size;
2509         pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2510         virtpages = pages * 8;     /* Scan up to this much virtual space */
2511         if (!pages)
2512                 return;
2513 
2514 
2515         if (!down_read_trylock(&mm->mmap_sem))
2516                 return;
2517         vma = find_vma(mm, start);
2518         if (!vma) {
2519                 reset_ptenuma_scan(p);
2520                 start = 0;
2521                 vma = mm->mmap;
2522         }
2523         for (; vma; vma = vma->vm_next) {
2524                 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2525                         is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2526                         continue;
2527                 }
2528 
2529                 /*
2530                  * Shared library pages mapped by multiple processes are not
2531                  * migrated as it is expected they are cache replicated. Avoid
2532                  * hinting faults in read-only file-backed mappings or the vdso
2533                  * as migrating the pages will be of marginal benefit.
2534                  */
2535                 if (!vma->vm_mm ||
2536                     (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2537                         continue;
2538 
2539                 /*
2540                  * Skip inaccessible VMAs to avoid any confusion between
2541                  * PROT_NONE and NUMA hinting ptes
2542                  */
2543                 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2544                         continue;
2545 
2546                 do {
2547                         start = max(start, vma->vm_start);
2548                         end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2549                         end = min(end, vma->vm_end);
2550                         nr_pte_updates = change_prot_numa(vma, start, end);
2551 
2552                         /*
2553                          * Try to scan sysctl_numa_balancing_size worth of
2554                          * hpages that have at least one present PTE that
2555                          * is not already pte-numa. If the VMA contains
2556                          * areas that are unused or already full of prot_numa
2557                          * PTEs, scan up to virtpages, to skip through those
2558                          * areas faster.
2559                          */
2560                         if (nr_pte_updates)
2561                                 pages -= (end - start) >> PAGE_SHIFT;
2562                         virtpages -= (end - start) >> PAGE_SHIFT;
2563 
2564                         start = end;
2565                         if (pages <= 0 || virtpages <= 0)
2566                                 goto out;
2567 
2568                         cond_resched();
2569                 } while (end != vma->vm_end);
2570         }
2571 
2572 out:
2573         /*
2574          * It is possible to reach the end of the VMA list but the last few
2575          * VMAs are not guaranteed to the vma_migratable. If they are not, we
2576          * would find the !migratable VMA on the next scan but not reset the
2577          * scanner to the start so check it now.
2578          */
2579         if (vma)
2580                 mm->numa_scan_offset = start;
2581         else
2582                 reset_ptenuma_scan(p);
2583         up_read(&mm->mmap_sem);
2584 
2585         /*
2586          * Make sure tasks use at least 32x as much time to run other code
2587          * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2588          * Usually update_task_scan_period slows down scanning enough; on an
2589          * overloaded system we need to limit overhead on a per task basis.
2590          */
2591         if (unlikely(p->se.sum_exec_runtime != runtime)) {
2592                 u64 diff = p->se.sum_exec_runtime - runtime;
2593                 p->node_stamp += 32 * diff;
2594         }
2595 }
2596 
2597 /*
2598  * Drive the periodic memory faults..
2599  */
2600 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2601 {
2602         struct callback_head *work = &curr->numa_work;
2603         u64 period, now;
2604 
2605         /*
2606          * We don't care about NUMA placement if we don't have memory.
2607          */
2608         if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2609                 return;
2610 
2611         /*
2612          * Using runtime rather than walltime has the dual advantage that
2613          * we (mostly) drive the selection from busy threads and that the
2614          * task needs to have done some actual work before we bother with
2615          * NUMA placement.
2616          */
2617         now = curr->se.sum_exec_runtime;
2618         period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2619 
2620         if (now > curr->node_stamp + period) {
2621                 if (!curr->node_stamp)
2622                         curr->numa_scan_period = task_scan_start(curr);
2623                 curr->node_stamp += period;
2624 
2625                 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2626                         init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2627                         task_work_add(curr, work, true);
2628                 }
2629         }
2630 }
2631 
2632 static void update_scan_period(struct task_struct *p, int new_cpu)
2633 {
2634         int src_nid = cpu_to_node(task_cpu(p));
2635         int dst_nid = cpu_to_node(new_cpu);
2636 
2637         if (!static_branch_likely(&sched_numa_balancing))
2638                 return;
2639 
2640         if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
2641                 return;
2642 
2643         if (src_nid == dst_nid)
2644                 return;
2645 
2646         /*
2647          * Allow resets if faults have been trapped before one scan
2648          * has completed. This is most likely due to a new task that
2649          * is pulled cross-node due to wakeups or load balancing.
2650          */
2651         if (p->numa_scan_seq) {
2652                 /*
2653                  * Avoid scan adjustments if moving to the preferred
2654                  * node or if the task was not previously running on
2655                  * the preferred node.
2656                  */
2657                 if (dst_nid == p->numa_preferred_nid ||
2658                     (p->numa_preferred_nid != NUMA_NO_NODE &&
2659                         src_nid != p->numa_preferred_nid))
2660                         return;
2661         }
2662 
2663         p->numa_scan_period = task_scan_start(p);
2664 }
2665 
2666 #else
2667 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2668 {
2669 }
2670 
2671 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2672 {
2673 }
2674 
2675 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2676 {
2677 }
2678 
2679 static inline void update_scan_period(struct task_struct *p, int new_cpu)
2680 {
2681 }
2682 
2683 #endif /* CONFIG_NUMA_BALANCING */
2684 
2685 static void
2686 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2687 {
2688         update_load_add(&cfs_rq->load, se->load.weight);
2689         if (!parent_entity(se))
2690                 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2691 #ifdef CONFIG_SMP
2692         if (entity_is_task(se)) {
2693                 struct rq *rq = rq_of(cfs_rq);
2694 
2695                 account_numa_enqueue(rq, task_of(se));
2696                 list_add(&se->group_node, &rq->cfs_tasks);
2697         }
2698 #endif
2699         cfs_rq->nr_running++;
2700 }
2701 
2702 static void
2703 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2704 {
2705         update_load_sub(&cfs_rq->load, se->load.weight);
2706         if (!parent_entity(se))
2707                 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2708 #ifdef CONFIG_SMP
2709         if (entity_is_task(se)) {
2710                 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2711                 list_del_init(&se->group_node);
2712         }
2713 #endif
2714         cfs_rq->nr_running--;
2715 }
2716 
2717 /*
2718  * Signed add and clamp on underflow.
2719  *
2720  * Explicitly do a load-store to ensure the intermediate value never hits
2721  * memory. This allows lockless observations without ever seeing the negative
2722  * values.
2723  */
2724 #define add_positive(_ptr, _val) do {                           \
2725         typeof(_ptr) ptr = (_ptr);                              \
2726         typeof(_val) val = (_val);                              \
2727         typeof(*ptr) res, var = READ_ONCE(*ptr);                \
2728                                                                 \
2729         res = var + val;                                        \
2730                                                                 \
2731         if (val < 0 && res > var)                               \
2732                 res = 0;                                        \
2733                                                                 \
2734         WRITE_ONCE(*ptr, res);                                  \
2735 } while (0)
2736 
2737 /*
2738  * Unsigned subtract and clamp on underflow.
2739  *
2740  * Explicitly do a load-store to ensure the intermediate value never hits
2741  * memory. This allows lockless observations without ever seeing the negative
2742  * values.
2743  */
2744 #define sub_positive(_ptr, _val) do {                           \
2745         typeof(_ptr) ptr = (_ptr);                              \
2746         typeof(*ptr) val = (_val);                              \
2747         typeof(*ptr) res, var = READ_ONCE(*ptr);                \
2748         res = var - val;                                        \
2749         if (res > var)                                          \
2750                 res = 0;                                        \
2751         WRITE_ONCE(*ptr, res);                                  \
2752 } while (0)
2753 
2754 /*
2755  * Remove and clamp on negative, from a local variable.
2756  *
2757  * A variant of sub_positive(), which does not use explicit load-store
2758  * and is thus optimized for local variable updates.
2759  */
2760 #define lsub_positive(_ptr, _val) do {                          \
2761         typeof(_ptr) ptr = (_ptr);                              \
2762         *ptr -= min_t(typeof(*ptr), *ptr, _val);                \
2763 } while (0)
2764 
2765 #ifdef CONFIG_SMP
2766 static inline void
2767 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2768 {
2769         cfs_rq->runnable_weight += se->runnable_weight;
2770 
2771         cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg;
2772         cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum;
2773 }
2774 
2775 static inline void
2776 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2777 {
2778         cfs_rq->runnable_weight -= se->runnable_weight;
2779 
2780         sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg);
2781         sub_positive(&cfs_rq->avg.runnable_load_sum,
2782                      se_runnable(se) * se->avg.runnable_load_sum);
2783 }
2784 
2785 static inline void
2786 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2787 {
2788         cfs_rq->avg.load_avg += se->avg.load_avg;
2789         cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
2790 }
2791 
2792 static inline void
2793 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2794 {
2795         sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2796         sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
2797 }
2798 #else
2799 static inline void
2800 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2801 static inline void
2802 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2803 static inline void
2804 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2805 static inline void
2806 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2807 #endif
2808 
2809 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2810                             unsigned long weight, unsigned long runnable)
2811 {
2812         if (se->on_rq) {
2813                 /* commit outstanding execution time */
2814                 if (cfs_rq->curr == se)
2815                         update_curr(cfs_rq);
2816                 account_entity_dequeue(cfs_rq, se);
2817                 dequeue_runnable_load_avg(cfs_rq, se);
2818         }
2819         dequeue_load_avg(cfs_rq, se);
2820 
2821         se->runnable_weight = runnable;
2822         update_load_set(&se->load, weight);
2823 
2824 #ifdef CONFIG_SMP
2825         do {
2826                 u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
2827 
2828                 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
2829                 se->avg.runnable_load_avg =
2830                         div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider);
2831         } while (0);
2832 #endif
2833 
2834         enqueue_load_avg(cfs_rq, se);
2835         if (se->on_rq) {
2836                 account_entity_enqueue(cfs_rq, se);
2837                 enqueue_runnable_load_avg(cfs_rq, se);
2838         }
2839 }
2840 
2841 void reweight_task(struct task_struct *p, int prio)
2842 {
2843         struct sched_entity *se = &p->se;
2844         struct cfs_rq *cfs_rq = cfs_rq_of(se);
2845         struct load_weight *load = &se->load;
2846         unsigned long weight = scale_load(sched_prio_to_weight[prio]);
2847 
2848         reweight_entity(cfs_rq, se, weight, weight);
2849         load->inv_weight = sched_prio_to_wmult[prio];
2850 }
2851 
2852 #ifdef CONFIG_FAIR_GROUP_SCHED
2853 #ifdef CONFIG_SMP
2854 /*
2855  * All this does is approximate the hierarchical proportion which includes that
2856  * global sum we all love to hate.
2857  *
2858  * That is, the weight of a group entity, is the proportional share of the
2859  * group weight based on the group runqueue weights. That is:
2860  *
2861  *                     tg->weight * grq->load.weight
2862  *   ge->load.weight = -----------------------------               (1)
2863  *                        \Sum grq->load.weight
2864  *
2865  * Now, because computing that sum is prohibitively expensive to compute (been
2866  * there, done that) we approximate it with this average stuff. The average
2867  * moves slower and therefore the approximation is cheaper and more stable.
2868  *
2869  * So instead of the above, we substitute:
2870  *
2871  *   grq->load.weight -> grq->avg.load_avg                         (2)
2872  *
2873  * which yields the following:
2874  *
2875  *                     tg->weight * grq->avg.load_avg
2876  *   ge->load.weight = ------------------------------              (3)
2877  *                              tg->load_avg
2878  *
2879  * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2880  *
2881  * That is shares_avg, and it is right (given the approximation (2)).
2882  *
2883  * The problem with it is that because the average is slow -- it was designed
2884  * to be exactly that of course -- this leads to transients in boundary
2885  * conditions. In specific, the case where the group was idle and we start the
2886  * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2887  * yielding bad latency etc..
2888  *
2889  * Now, in that special case (1) reduces to:
2890  *
2891  *                     tg->weight * grq->load.weight
2892  *   ge->load.weight = ----------------------------- = tg->weight   (4)
2893  *                          grp->load.weight
2894  *
2895  * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2896  *
2897  * So what we do is modify our approximation (3) to approach (4) in the (near)
2898  * UP case, like:
2899  *
2900  *   ge->load.weight =
2901  *
2902  *              tg->weight * grq->load.weight
2903  *     ---------------------------------------------------         (5)
2904  *     tg->load_avg - grq->avg.load_avg + grq->load.weight
2905  *
2906  * But because grq->load.weight can drop to 0, resulting in a divide by zero,
2907  * we need to use grq->avg.load_avg as its lower bound, which then gives:
2908  *
2909  *
2910  *                     tg->weight * grq->load.weight
2911  *   ge->load.weight = -----------------------------               (6)
2912  *                              tg_load_avg'
2913  *
2914  * Where:
2915  *
2916  *   tg_load_avg' = tg->load_avg - grq->avg.load_avg +
2917  *                  max(grq->load.weight, grq->avg.load_avg)
2918  *
2919  * And that is shares_weight and is icky. In the (near) UP case it approaches
2920  * (4) while in the normal case it approaches (3). It consistently
2921  * overestimates the ge->load.weight and therefore:
2922  *
2923  *   \Sum ge->load.weight >= tg->weight
2924  *
2925  * hence icky!
2926  */
2927 static long calc_group_shares(struct cfs_rq *cfs_rq)
2928 {
2929         long tg_weight, tg_shares, load, shares;
2930         struct task_group *tg = cfs_rq->tg;
2931 
2932         tg_shares = READ_ONCE(tg->shares);
2933 
2934         load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
2935 
2936         tg_weight = atomic_long_read(&tg->load_avg);
2937 
2938         /* Ensure tg_weight >= load */
2939         tg_weight -= cfs_rq->tg_load_avg_contrib;
2940         tg_weight += load;
2941 
2942         shares = (tg_shares * load);
2943         if (tg_weight)
2944                 shares /= tg_weight;
2945 
2946         /*
2947          * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2948          * of a group with small tg->shares value. It is a floor value which is
2949          * assigned as a minimum load.weight to the sched_entity representing
2950          * the group on a CPU.
2951          *
2952          * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2953          * on an 8-core system with 8 tasks each runnable on one CPU shares has
2954          * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2955          * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2956          * instead of 0.
2957          */
2958         return clamp_t(long, shares, MIN_SHARES, tg_shares);
2959 }
2960 
2961 /*
2962  * This calculates the effective runnable weight for a group entity based on
2963  * the group entity weight calculated above.
2964  *
2965  * Because of the above approximation (2), our group entity weight is
2966  * an load_avg based ratio (3). This means that it includes blocked load and
2967  * does not represent the runnable weight.
2968  *
2969  * Approximate the group entity's runnable weight per ratio from the group
2970  * runqueue:
2971  *
2972  *                                           grq->avg.runnable_load_avg
2973  *   ge->runnable_weight = ge->load.weight * -------------------------- (7)
2974  *                                               grq->avg.load_avg
2975  *
2976  * However, analogous to above, since the avg numbers are slow, this leads to
2977  * transients in the from-idle case. Instead we use:
2978  *
2979  *   ge->runnable_weight = ge->load.weight *
2980  *
2981  *              max(grq->avg.runnable_load_avg, grq->runnable_weight)
2982  *              -----------------------------------------------------   (8)
2983  *                    max(grq->avg.load_avg, grq->load.weight)
2984  *
2985  * Where these max() serve both to use the 'instant' values to fix the slow
2986  * from-idle and avoid the /0 on to-idle, similar to (6).
2987  */
2988 static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares)
2989 {
2990         long runnable, load_avg;
2991 
2992         load_avg = max(cfs_rq->avg.load_avg,
2993                        scale_load_down(cfs_rq->load.weight));
2994 
2995         runnable = max(cfs_rq->avg.runnable_load_avg,
2996                        scale_load_down(cfs_rq->runnable_weight));
2997 
2998         runnable *= shares;
2999         if (load_avg)
3000                 runnable /= load_avg;
3001 
3002         return clamp_t(long, runnable, MIN_SHARES, shares);
3003 }
3004 #endif /* CONFIG_SMP */
3005 
3006 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3007 
3008 /*
3009  * Recomputes the group entity based on the current state of its group
3010  * runqueue.
3011  */
3012 static void update_cfs_group(struct sched_entity *se)
3013 {
3014         struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3015         long shares, runnable;
3016 
3017         if (!gcfs_rq)
3018                 return;
3019 
3020         if (throttled_hierarchy(gcfs_rq))
3021                 return;
3022 
3023 #ifndef CONFIG_SMP
3024         runnable = shares = READ_ONCE(gcfs_rq->tg->shares);
3025 
3026         if (likely(se->load.weight == shares))
3027                 return;
3028 #else
3029         shares   = calc_group_shares(gcfs_rq);
3030         runnable = calc_group_runnable(gcfs_rq, shares);
3031 #endif
3032 
3033         reweight_entity(cfs_rq_of(se), se, shares, runnable);
3034 }
3035 
3036 #else /* CONFIG_FAIR_GROUP_SCHED */
3037 static inline void update_cfs_group(struct sched_entity *se)
3038 {
3039 }
3040 #endif /* CONFIG_FAIR_GROUP_SCHED */
3041 
3042 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3043 {
3044         struct rq *rq = rq_of(cfs_rq);
3045 
3046         if (&rq->cfs == cfs_rq || (flags & SCHED_CPUFREQ_MIGRATION)) {
3047                 /*
3048                  * There are a few boundary cases this might miss but it should
3049                  * get called often enough that that should (hopefully) not be
3050                  * a real problem.
3051                  *
3052                  * It will not get called when we go idle, because the idle
3053                  * thread is a different class (!fair), nor will the utilization
3054                  * number include things like RT tasks.
3055                  *
3056                  * As is, the util number is not freq-invariant (we'd have to
3057                  * implement arch_scale_freq_capacity() for that).
3058                  *
3059                  * See cpu_util().
3060                  */
3061                 cpufreq_update_util(rq, flags);
3062         }
3063 }
3064 
3065 #ifdef CONFIG_SMP
3066 #ifdef CONFIG_FAIR_GROUP_SCHED
3067 /**
3068  * update_tg_load_avg - update the tg's load avg
3069  * @cfs_rq: the cfs_rq whose avg changed
3070  * @force: update regardless of how small the difference
3071  *
3072  * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3073  * However, because tg->load_avg is a global value there are performance
3074  * considerations.
3075  *
3076  * In order to avoid having to look at the other cfs_rq's, we use a
3077  * differential update where we store the last value we propagated. This in
3078  * turn allows skipping updates if the differential is 'small'.
3079  *
3080  * Updating tg's load_avg is necessary before update_cfs_share().
3081  */
3082 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3083 {
3084         long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3085 
3086         /*
3087          * No need to update load_avg for root_task_group as it is not used.
3088          */
3089         if (cfs_rq->tg == &root_task_group)
3090                 return;
3091 
3092         if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3093                 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3094                 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3095         }
3096 }
3097 
3098 /*
3099  * Called within set_task_rq() right before setting a task's CPU. The
3100  * caller only guarantees p->pi_lock is held; no other assumptions,
3101  * including the state of rq->lock, should be made.
3102  */
3103 void set_task_rq_fair(struct sched_entity *se,
3104                       struct cfs_rq *prev, struct cfs_rq *next)
3105 {
3106         u64 p_last_update_time;
3107         u64 n_last_update_time;
3108 
3109         if (!sched_feat(ATTACH_AGE_LOAD))
3110                 return;
3111 
3112         /*
3113          * We are supposed to update the task to "current" time, then its up to
3114          * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3115          * getting what current time is, so simply throw away the out-of-date
3116          * time. This will result in the wakee task is less decayed, but giving
3117          * the wakee more load sounds not bad.
3118          */
3119         if (!(se->avg.last_update_time && prev))
3120                 return;
3121 
3122 #ifndef CONFIG_64BIT
3123         {
3124                 u64 p_last_update_time_copy;
3125                 u64 n_last_update_time_copy;
3126 
3127                 do {
3128                         p_last_update_time_copy = prev->load_last_update_time_copy;
3129                         n_last_update_time_copy = next->load_last_update_time_copy;
3130 
3131                         smp_rmb();
3132 
3133                         p_last_update_time = prev->avg.last_update_time;
3134                         n_last_update_time = next->avg.last_update_time;
3135 
3136                 } while (p_last_update_time != p_last_update_time_copy ||
3137                          n_last_update_time != n_last_update_time_copy);
3138         }
3139 #else
3140         p_last_update_time = prev->avg.last_update_time;
3141         n_last_update_time = next->avg.last_update_time;
3142 #endif
3143         __update_load_avg_blocked_se(p_last_update_time, se);
3144         se->avg.last_update_time = n_last_update_time;
3145 }
3146 
3147 
3148 /*
3149  * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3150  * propagate its contribution. The key to this propagation is the invariant
3151  * that for each group:
3152  *
3153  *   ge->avg == grq->avg                                                (1)
3154  *
3155  * _IFF_ we look at the pure running and runnable sums. Because they
3156  * represent the very same entity, just at different points in the hierarchy.
3157  *
3158  * Per the above update_tg_cfs_util() is trivial and simply copies the running
3159  * sum over (but still wrong, because the group entity and group rq do not have
3160  * their PELT windows aligned).
3161  *
3162  * However, update_tg_cfs_runnable() is more complex. So we have:
3163  *
3164  *   ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg          (2)
3165  *
3166  * And since, like util, the runnable part should be directly transferable,
3167  * the following would _appear_ to be the straight forward approach:
3168  *
3169  *   grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg       (3)
3170  *
3171  * And per (1) we have:
3172  *
3173  *   ge->avg.runnable_avg == grq->avg.runnable_avg
3174  *
3175  * Which gives:
3176  *
3177  *                      ge->load.weight * grq->avg.load_avg
3178  *   ge->avg.load_avg = -----------------------------------             (4)
3179  *                               grq->load.weight
3180  *
3181  * Except that is wrong!
3182  *
3183  * Because while for entities historical weight is not important and we
3184  * really only care about our future and therefore can consider a pure
3185  * runnable sum, runqueues can NOT do this.
3186  *
3187  * We specifically want runqueues to have a load_avg that includes
3188  * historical weights. Those represent the blocked load, the load we expect
3189  * to (shortly) return to us. This only works by keeping the weights as
3190  * integral part of the sum. We therefore cannot decompose as per (3).
3191  *
3192  * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3193  * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3194  * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3195  * runnable section of these tasks overlap (or not). If they were to perfectly
3196  * align the rq as a whole would be runnable 2/3 of the time. If however we
3197  * always have at least 1 runnable task, the rq as a whole is always runnable.
3198  *
3199  * So we'll have to approximate.. :/
3200  *
3201  * Given the constraint:
3202  *
3203  *   ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3204  *
3205  * We can construct a rule that adds runnable to a rq by assuming minimal
3206  * overlap.
3207  *
3208  * On removal, we'll assume each task is equally runnable; which yields:
3209  *
3210  *   grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3211  *
3212  * XXX: only do this for the part of runnable > running ?
3213  *
3214  */
3215 
3216 static inline void
3217 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3218 {
3219         long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3220 
3221         /* Nothing to update */
3222         if (!delta)
3223                 return;
3224 
3225         /*
3226          * The relation between sum and avg is:
3227          *
3228          *   LOAD_AVG_MAX - 1024 + sa->period_contrib
3229          *
3230          * however, the PELT windows are not aligned between grq and gse.
3231          */
3232 
3233         /* Set new sched_entity's utilization */
3234         se->avg.util_avg = gcfs_rq->avg.util_avg;
3235         se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3236 
3237         /* Update parent cfs_rq utilization */
3238         add_positive(&cfs_rq->avg.util_avg, delta);
3239         cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3240 }
3241 
3242 static inline void
3243 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3244 {
3245         long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3246         unsigned long runnable_load_avg, load_avg;
3247         u64 runnable_load_sum, load_sum = 0;
3248         s64 delta_sum;
3249 
3250         if (!runnable_sum)
3251                 return;
3252 
3253         gcfs_rq->prop_runnable_sum = 0;
3254 
3255         if (runnable_sum >= 0) {
3256                 /*
3257                  * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3258                  * the CPU is saturated running == runnable.
3259                  */
3260                 runnable_sum += se->avg.load_sum;
3261                 runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
3262         } else {
3263                 /*
3264                  * Estimate the new unweighted runnable_sum of the gcfs_rq by
3265                  * assuming all tasks are equally runnable.
3266                  */
3267                 if (scale_load_down(gcfs_rq->load.weight)) {
3268                         load_sum = div_s64(gcfs_rq->avg.load_sum,
3269                                 scale_load_down(gcfs_rq->load.weight));
3270                 }
3271 
3272                 /* But make sure to not inflate se's runnable */
3273                 runnable_sum = min(se->avg.load_sum, load_sum);
3274         }
3275 
3276         /*
3277          * runnable_sum can't be lower than running_sum
3278          * Rescale running sum to be in the same range as runnable sum
3279          * running_sum is in [0 : LOAD_AVG_MAX <<  SCHED_CAPACITY_SHIFT]
3280          * runnable_sum is in [0 : LOAD_AVG_MAX]
3281          */
3282         running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
3283         runnable_sum = max(runnable_sum, running_sum);
3284 
3285         load_sum = (s64)se_weight(se) * runnable_sum;
3286         load_avg = div_s64(load_sum, LOAD_AVG_MAX);
3287 
3288         delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3289         delta_avg = load_avg - se->avg.load_avg;
3290 
3291         se->avg.load_sum = runnable_sum;
3292         se->avg.load_avg = load_avg;
3293         add_positive(&cfs_rq->avg.load_avg, delta_avg);
3294         add_positive(&cfs_rq->avg.load_sum, delta_sum);
3295 
3296         runnable_load_sum = (s64)se_runnable(se) * runnable_sum;
3297         runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX);
3298         delta_sum = runnable_load_sum - se_weight(se) * se->avg.runnable_load_sum;
3299         delta_avg = runnable_load_avg - se->avg.runnable_load_avg;
3300 
3301         se->avg.runnable_load_sum = runnable_sum;
3302         se->avg.runnable_load_avg = runnable_load_avg;
3303 
3304         if (se->on_rq) {
3305                 add_positive(&cfs_rq->avg.runnable_load_avg, delta_avg);
3306                 add_positive(&cfs_rq->avg.runnable_load_sum, delta_sum);
3307         }
3308 }
3309 
3310 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3311 {
3312         cfs_rq->propagate = 1;
3313         cfs_rq->prop_runnable_sum += runnable_sum;
3314 }
3315 
3316 /* Update task and its cfs_rq load average */
3317 static inline int propagate_entity_load_avg(struct sched_entity *se)
3318 {
3319         struct cfs_rq *cfs_rq, *gcfs_rq;
3320 
3321         if (entity_is_task(se))
3322                 return 0;
3323 
3324         gcfs_rq = group_cfs_rq(se);
3325         if (!gcfs_rq->propagate)
3326                 return 0;
3327 
3328         gcfs_rq->propagate = 0;
3329 
3330         cfs_rq = cfs_rq_of(se);
3331 
3332         add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3333 
3334         update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3335         update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3336 
3337         return 1;
3338 }
3339 
3340 /*
3341  * Check if we need to update the load and the utilization of a blocked
3342  * group_entity:
3343  */
3344 static inline bool skip_blocked_update(struct sched_entity *se)
3345 {
3346         struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3347 
3348         /*
3349          * If sched_entity still have not zero load or utilization, we have to
3350          * decay it:
3351          */
3352         if (se->avg.load_avg || se->avg.util_avg)
3353                 return false;
3354 
3355         /*
3356          * If there is a pending propagation, we have to update the load and
3357          * the utilization of the sched_entity:
3358          */
3359         if (gcfs_rq->propagate)
3360                 return false;
3361 
3362         /*
3363          * Otherwise, the load and the utilization of the sched_entity is
3364          * already zero and there is no pending propagation, so it will be a
3365          * waste of time to try to decay it:
3366          */
3367         return true;
3368 }
3369 
3370 #else /* CONFIG_FAIR_GROUP_SCHED */
3371 
3372 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3373 
3374 static inline int propagate_entity_load_avg(struct sched_entity *se)
3375 {
3376         return 0;
3377 }
3378 
3379 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3380 
3381 #endif /* CONFIG_FAIR_GROUP_SCHED */
3382 
3383 /**
3384  * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3385  * @now: current time, as per cfs_rq_clock_pelt()
3386  * @cfs_rq: cfs_rq to update
3387  *
3388  * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3389  * avg. The immediate corollary is that all (fair) tasks must be attached, see
3390  * post_init_entity_util_avg().
3391  *
3392  * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3393  *
3394  * Returns true if the load decayed or we removed load.
3395  *
3396  * Since both these conditions indicate a changed cfs_rq->avg.load we should
3397  * call update_tg_load_avg() when this function returns true.
3398  */
3399 static inline int
3400 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3401 {
3402         unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0;
3403         struct sched_avg *sa = &cfs_rq->avg;
3404         int decayed = 0;
3405 
3406         if (cfs_rq->removed.nr) {
3407                 unsigned long r;
3408                 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3409 
3410                 raw_spin_lock(&cfs_rq->removed.lock);
3411                 swap(cfs_rq->removed.util_avg, removed_util);
3412                 swap(cfs_rq->removed.load_avg, removed_load);
3413                 swap(cfs_rq->removed.runnable_sum, removed_runnable_sum);
3414                 cfs_rq->removed.nr = 0;
3415                 raw_spin_unlock(&cfs_rq->removed.lock);
3416 
3417                 r = removed_load;
3418                 sub_positive(&sa->load_avg, r);
3419                 sub_positive(&sa->load_sum, r * divider);
3420 
3421                 r = removed_util;
3422                 sub_positive(&sa->util_avg, r);
3423                 sub_positive(&sa->util_sum, r * divider);
3424 
3425                 add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum);
3426 
3427                 decayed = 1;
3428         }
3429 
3430         decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
3431 
3432 #ifndef CONFIG_64BIT
3433         smp_wmb();
3434         cfs_rq->load_last_update_time_copy = sa->last_update_time;
3435 #endif
3436 
3437         if (decayed)
3438                 cfs_rq_util_change(cfs_rq, 0);
3439 
3440         return decayed;
3441 }
3442 
3443 /**
3444  * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3445  * @cfs_rq: cfs_rq to attach to
3446  * @se: sched_entity to attach
3447  * @flags: migration hints
3448  *
3449  * Must call update_cfs_rq_load_avg() before this, since we rely on
3450  * cfs_rq->avg.last_update_time being current.
3451  */
3452 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3453 {
3454         u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3455 
3456         /*
3457          * When we attach the @se to the @cfs_rq, we must align the decay
3458          * window because without that, really weird and wonderful things can
3459          * happen.
3460          *
3461          * XXX illustrate
3462          */
3463         se->avg.last_update_time = cfs_rq->avg.last_update_time;
3464         se->avg.period_contrib = cfs_rq->avg.period_contrib;
3465 
3466         /*
3467          * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3468          * period_contrib. This isn't strictly correct, but since we're
3469          * entirely outside of the PELT hierarchy, nobody cares if we truncate
3470          * _sum a little.
3471          */
3472         se->avg.util_sum = se->avg.util_avg * divider;
3473 
3474         se->avg.load_sum = divider;
3475         if (se_weight(se)) {
3476                 se->avg.load_sum =
3477                         div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3478         }
3479 
3480         se->avg.runnable_load_sum = se->avg.load_sum;
3481 
3482         enqueue_load_avg(cfs_rq, se);
3483         cfs_rq->avg.util_avg += se->avg.util_avg;
3484         cfs_rq->avg.util_sum += se->avg.util_sum;
3485 
3486         add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3487 
3488         cfs_rq_util_change(cfs_rq, flags);
3489 }
3490 
3491 /**
3492  * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3493  * @cfs_rq: cfs_rq to detach from
3494  * @se: sched_entity to detach
3495  *
3496  * Must call update_cfs_rq_load_avg() before this, since we rely on
3497  * cfs_rq->avg.last_update_time being current.
3498  */
3499 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3500 {
3501         dequeue_load_avg(cfs_rq, se);
3502         sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3503         sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3504 
3505         add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3506 
3507         cfs_rq_util_change(cfs_rq, 0);
3508 }
3509 
3510 /*
3511  * Optional action to be done while updating the load average
3512  */
3513 #define UPDATE_TG       0x1
3514 #define SKIP_AGE_LOAD   0x2
3515 #define DO_ATTACH       0x4
3516 
3517 /* Update task and its cfs_rq load average */
3518 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3519 {
3520         u64 now = cfs_rq_clock_pelt(cfs_rq);
3521         int decayed;
3522 
3523         /*
3524          * Track task load average for carrying it to new CPU after migrated, and
3525          * track group sched_entity load average for task_h_load calc in migration
3526          */
3527         if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3528                 __update_load_avg_se(now, cfs_rq, se);
3529 
3530         decayed  = update_cfs_rq_load_avg(now, cfs_rq);
3531         decayed |= propagate_entity_load_avg(se);
3532 
3533         if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3534 
3535                 /*
3536                  * DO_ATTACH means we're here from enqueue_entity().
3537                  * !last_update_time means we've passed through
3538                  * migrate_task_rq_fair() indicating we migrated.
3539                  *
3540                  * IOW we're enqueueing a task on a new CPU.
3541                  */
3542                 attach_entity_load_avg(cfs_rq, se, SCHED_CPUFREQ_MIGRATION);
3543                 update_tg_load_avg(cfs_rq, 0);
3544 
3545         } else if (decayed && (flags & UPDATE_TG))
3546                 update_tg_load_avg(cfs_rq, 0);
3547 }
3548 
3549 #ifndef CONFIG_64BIT
3550 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3551 {
3552         u64 last_update_time_copy;
3553         u64 last_update_time;
3554 
3555         do {
3556                 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3557                 smp_rmb();
3558                 last_update_time = cfs_rq->avg.last_update_time;
3559         } while (last_update_time != last_update_time_copy);
3560 
3561         return last_update_time;
3562 }
3563 #else
3564 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3565 {
3566         return cfs_rq->avg.last_update_time;
3567 }
3568 #endif
3569 
3570 /*
3571  * Synchronize entity load avg of dequeued entity without locking
3572  * the previous rq.
3573  */
3574 void sync_entity_load_avg(struct sched_entity *se)
3575 {
3576         struct cfs_rq *cfs_rq = cfs_rq_of(se);
3577         u64 last_update_time;
3578 
3579         last_update_time = cfs_rq_last_update_time(cfs_rq);
3580         __update_load_avg_blocked_se(last_update_time, se);
3581 }
3582 
3583 /*
3584  * Task first catches up with cfs_rq, and then subtract
3585  * itself from the cfs_rq (task must be off the queue now).
3586  */
3587 void remove_entity_load_avg(struct sched_entity *se)
3588 {
3589         struct cfs_rq *cfs_rq = cfs_rq_of(se);
3590         unsigned long flags;
3591 
3592         /*
3593          * tasks cannot exit without having gone through wake_up_new_task() ->
3594          * post_init_entity_util_avg() which will have added things to the
3595          * cfs_rq, so we can remove unconditionally.
3596          */
3597 
3598         sync_entity_load_avg(se);
3599 
3600         raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3601         ++cfs_rq->removed.nr;
3602         cfs_rq->removed.util_avg        += se->avg.util_avg;
3603         cfs_rq->removed.load_avg        += se->avg.load_avg;
3604         cfs_rq->removed.runnable_sum    += se->avg.load_sum; /* == runnable_sum */
3605         raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3606 }
3607 
3608 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3609 {
3610         return cfs_rq->avg.runnable_load_avg;
3611 }
3612 
3613 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3614 {
3615         return cfs_rq->avg.load_avg;
3616 }
3617 
3618 static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3619 
3620 static inline unsigned long task_util(struct task_struct *p)
3621 {
3622         return READ_ONCE(p->se.avg.util_avg);
3623 }
3624 
3625 static inline unsigned long _task_util_est(struct task_struct *p)
3626 {
3627         struct util_est ue = READ_ONCE(p->se.avg.util_est);
3628 
3629         return (max(ue.ewma, ue.enqueued) | UTIL_AVG_UNCHANGED);
3630 }
3631 
3632 static inline unsigned long task_util_est(struct task_struct *p)
3633 {
3634         return max(task_util(p), _task_util_est(p));
3635 }
3636 
3637 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3638                                     struct task_struct *p)
3639 {
3640         unsigned int enqueued;
3641 
3642         if (!sched_feat(UTIL_EST))
3643                 return;
3644 
3645         /* Update root cfs_rq's estimated utilization */
3646         enqueued  = cfs_rq->avg.util_est.enqueued;
3647         enqueued += _task_util_est(p);
3648         WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3649 }
3650 
3651 /*
3652  * Check if a (signed) value is within a specified (unsigned) margin,
3653  * based on the observation that:
3654  *
3655  *     abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3656  *
3657  * NOTE: this only works when value + maring < INT_MAX.
3658  */
3659 static inline bool within_margin(int value, int margin)
3660 {
3661         return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3662 }
3663 
3664 static void
3665 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep)
3666 {
3667         long last_ewma_diff;
3668         struct util_est ue;
3669         int cpu;
3670 
3671         if (!sched_feat(UTIL_EST))
3672                 return;
3673 
3674         /* Update root cfs_rq's estimated utilization */
3675         ue.enqueued  = cfs_rq->avg.util_est.enqueued;
3676         ue.enqueued -= min_t(unsigned int, ue.enqueued, _task_util_est(p));
3677         WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued);
3678 
3679         /*
3680          * Skip update of task's estimated utilization when the task has not
3681          * yet completed an activation, e.g. being migrated.
3682          */
3683         if (!task_sleep)
3684                 return;
3685 
3686         /*
3687          * If the PELT values haven't changed since enqueue time,
3688          * skip the util_est update.
3689          */
3690         ue = p->se.avg.util_est;
3691         if (ue.enqueued & UTIL_AVG_UNCHANGED)
3692                 return;
3693 
3694         /*
3695          * Skip update of task's estimated utilization when its EWMA is
3696          * already ~1% close to its last activation value.
3697          */
3698         ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
3699         last_ewma_diff = ue.enqueued - ue.ewma;
3700         if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
3701                 return;
3702 
3703         /*
3704          * To avoid overestimation of actual task utilization, skip updates if
3705          * we cannot grant there is idle time in this CPU.
3706          */
3707         cpu = cpu_of(rq_of(cfs_rq));
3708         if (task_util(p) > capacity_orig_of(cpu))
3709                 return;
3710 
3711         /*
3712          * Update Task's estimated utilization
3713          *
3714          * When *p completes an activation we can consolidate another sample
3715          * of the task size. This is done by storing the current PELT value
3716          * as ue.enqueued and by using this value to update the Exponential
3717          * Weighted Moving Average (EWMA):
3718          *
3719          *  ewma(t) = w *  task_util(p) + (1-w) * ewma(t-1)
3720          *          = w *  task_util(p) +         ewma(t-1)  - w * ewma(t-1)
3721          *          = w * (task_util(p) -         ewma(t-1)) +     ewma(t-1)
3722          *          = w * (      last_ewma_diff            ) +     ewma(t-1)
3723          *          = w * (last_ewma_diff  +  ewma(t-1) / w)
3724          *
3725          * Where 'w' is the weight of new samples, which is configured to be
3726          * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
3727          */
3728         ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
3729         ue.ewma  += last_ewma_diff;
3730         ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
3731         WRITE_ONCE(p->se.avg.util_est, ue);
3732 }
3733 
3734 static inline int task_fits_capacity(struct task_struct *p, long capacity)
3735 {
3736         return capacity * 1024 > task_util_est(p) * capacity_margin;
3737 }
3738 
3739 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
3740 {
3741         if (!static_branch_unlikely(&sched_asym_cpucapacity))
3742                 return;
3743 
3744         if (!p) {
3745                 rq->misfit_task_load = 0;
3746                 return;
3747         }
3748 
3749         if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
3750                 rq->misfit_task_load = 0;
3751                 return;
3752         }
3753 
3754         rq->misfit_task_load = task_h_load(p);
3755 }
3756 
3757 #else /* CONFIG_SMP */
3758 
3759 #define UPDATE_TG       0x0
3760 #define SKIP_AGE_LOAD   0x0
3761 #define DO_ATTACH       0x0
3762 
3763 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
3764 {
3765         cfs_rq_util_change(cfs_rq, 0);
3766 }
3767 
3768 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3769 
3770 static inline void
3771 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) {}
3772 static inline void
3773 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3774 
3775 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3776 {
3777         return 0;
3778 }
3779 
3780 static inline void
3781 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
3782 
3783 static inline void
3784 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p,
3785                  bool task_sleep) {}
3786 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
3787 
3788 #endif /* CONFIG_SMP */
3789 
3790 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3791 {
3792 #ifdef CONFIG_SCHED_DEBUG
3793         s64 d = se->vruntime - cfs_rq->min_vruntime;
3794 
3795         if (d < 0)
3796                 d = -d;
3797 
3798         if (d > 3*sysctl_sched_latency)
3799                 schedstat_inc(cfs_rq->nr_spread_over);
3800 #endif
3801 }
3802 
3803 static void
3804 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3805 {
3806         u64 vruntime = cfs_rq->min_vruntime;
3807 
3808         /*
3809          * The 'current' period is already promised to the current tasks,
3810          * however the extra weight of the new task will slow them down a
3811          * little, place the new task so that it fits in the slot that
3812          * stays open at the end.
3813          */
3814         if (initial && sched_feat(START_DEBIT))
3815                 vruntime += sched_vslice(cfs_rq, se);
3816 
3817         /* sleeps up to a single latency don't count. */
3818         if (!initial) {
3819                 unsigned long thresh = sysctl_sched_latency;
3820 
3821                 /*
3822                  * Halve their sleep time's effect, to allow
3823                  * for a gentler effect of sleepers:
3824                  */
3825                 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3826                         thresh >>= 1;
3827 
3828                 vruntime -= thresh;
3829         }
3830 
3831         /* ensure we never gain time by being placed backwards. */
3832         se->vruntime = max_vruntime(se->vruntime, vruntime);
3833 }
3834 
3835 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3836 
3837 static inline void check_schedstat_required(void)
3838 {
3839 #ifdef CONFIG_SCHEDSTATS
3840         if (schedstat_enabled())
3841                 return;
3842 
3843         /* Force schedstat enabled if a dependent tracepoint is active */
3844         if (trace_sched_stat_wait_enabled()    ||
3845                         trace_sched_stat_sleep_enabled()   ||
3846                         trace_sched_stat_iowait_enabled()  ||
3847                         trace_sched_stat_blocked_enabled() ||
3848                         trace_sched_stat_runtime_enabled())  {
3849                 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3850                              "stat_blocked and stat_runtime require the "
3851                              "kernel parameter schedstats=enable or "
3852                              "kernel.sched_schedstats=1\n");
3853         }
3854 #endif
3855 }
3856 
3857 
3858 /*
3859  * MIGRATION
3860  *
3861  *      dequeue
3862  *        update_curr()
3863  *          update_min_vruntime()
3864  *        vruntime -= min_vruntime
3865  *
3866  *      enqueue
3867  *        update_curr()
3868  *          update_min_vruntime()
3869  *        vruntime += min_vruntime
3870  *
3871  * this way the vruntime transition between RQs is done when both
3872  * min_vruntime are up-to-date.
3873  *
3874  * WAKEUP (remote)
3875  *
3876  *      ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3877  *        vruntime -= min_vruntime
3878  *
3879  *      enqueue
3880  *        update_curr()
3881  *          update_min_vruntime()
3882  *        vruntime += min_vruntime
3883  *
3884  * this way we don't have the most up-to-date min_vruntime on the originating
3885  * CPU and an up-to-date min_vruntime on the destination CPU.
3886  */
3887 
3888 static void
3889 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3890 {
3891         bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3892         bool curr = cfs_rq->curr == se;
3893 
3894         /*
3895          * If we're the current task, we must renormalise before calling
3896          * update_curr().
3897          */
3898         if (renorm && curr)
3899                 se->vruntime += cfs_rq->min_vruntime;
3900 
3901         update_curr(cfs_rq);
3902 
3903         /*
3904          * Otherwise, renormalise after, such that we're placed at the current
3905          * moment in time, instead of some random moment in the past. Being
3906          * placed in the past could significantly boost this task to the
3907          * fairness detriment of existing tasks.
3908          */
3909         if (renorm && !curr)
3910                 se->vruntime += cfs_rq->min_vruntime;
3911 
3912         /*
3913          * When enqueuing a sched_entity, we must:
3914          *   - Update loads to have both entity and cfs_rq synced with now.
3915          *   - Add its load to cfs_rq->runnable_avg
3916          *   - For group_entity, update its weight to reflect the new share of
3917          *     its group cfs_rq
3918          *   - Add its new weight to cfs_rq->load.weight
3919          */
3920         update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
3921         update_cfs_group(se);
3922         enqueue_runnable_load_avg(cfs_rq, se);
3923         account_entity_enqueue(cfs_rq, se);
3924 
3925         if (flags & ENQUEUE_WAKEUP)
3926                 place_entity(cfs_rq, se, 0);
3927 
3928         check_schedstat_required();
3929         update_stats_enqueue(cfs_rq, se, flags);
3930         check_spread(cfs_rq, se);
3931         if (!curr)
3932                 __enqueue_entity(cfs_rq, se);
3933         se->on_rq = 1;
3934 
3935         if (cfs_rq->nr_running == 1) {
3936                 list_add_leaf_cfs_rq(cfs_rq);
3937                 check_enqueue_throttle(cfs_rq);
3938         }
3939 }
3940 
3941 static void __clear_buddies_last(struct sched_entity *se)
3942 {
3943         for_each_sched_entity(se) {
3944                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3945                 if (cfs_rq->last != se)
3946                         break;
3947 
3948                 cfs_rq->last = NULL;
3949         }
3950 }
3951 
3952 static void __clear_buddies_next(struct sched_entity *se)
3953 {
3954         for_each_sched_entity(se) {
3955                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3956                 if (cfs_rq->next != se)
3957                         break;
3958 
3959                 cfs_rq->next = NULL;
3960         }
3961 }
3962 
3963 static void __clear_buddies_skip(struct sched_entity *se)
3964 {
3965         for_each_sched_entity(se) {
3966                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3967                 if (cfs_rq->skip != se)
3968                         break;
3969 
3970                 cfs_rq->skip = NULL;
3971         }
3972 }
3973 
3974 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3975 {
3976         if (cfs_rq->last == se)
3977                 __clear_buddies_last(se);
3978 
3979         if (cfs_rq->next == se)
3980                 __clear_buddies_next(se);
3981 
3982         if (cfs_rq->skip == se)
3983                 __clear_buddies_skip(se);
3984 }
3985 
3986 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3987 
3988 static void
3989 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3990 {
3991         /*
3992          * Update run-time statistics of the 'current'.
3993          */
3994         update_curr(cfs_rq);
3995 
3996         /*
3997          * When dequeuing a sched_entity, we must:
3998          *   - Update loads to have both entity and cfs_rq synced with now.
3999          *   - Subtract its load from the cfs_rq->runnable_avg.
4000          *   - Subtract its previous weight from cfs_rq->load.weight.
4001          *   - For group entity, update its weight to reflect the new share
4002          *     of its group cfs_rq.
4003          */
4004         update_load_avg(cfs_rq, se, UPDATE_TG);
4005         dequeue_runnable_load_avg(cfs_rq, se);
4006 
4007         update_stats_dequeue(cfs_rq, se, flags);
4008 
4009         clear_buddies(cfs_rq, se);
4010 
4011         if (se != cfs_rq->curr)
4012                 __dequeue_entity(cfs_rq, se);
4013         se->on_rq = 0;
4014         account_entity_dequeue(cfs_rq, se);
4015 
4016         /*
4017          * Normalize after update_curr(); which will also have moved
4018          * min_vruntime if @se is the one holding it back. But before doing
4019          * update_min_vruntime() again, which will discount @se's position and
4020          * can move min_vruntime forward still more.
4021          */
4022         if (!(flags & DEQUEUE_SLEEP))
4023                 se->vruntime -= cfs_rq->min_vruntime;
4024 
4025         /* return excess runtime on last dequeue */
4026         return_cfs_rq_runtime(cfs_rq);
4027 
4028         update_cfs_group(se);
4029 
4030         /*
4031          * Now advance min_vruntime if @se was the entity holding it back,
4032          * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4033          * put back on, and if we advance min_vruntime, we'll be placed back
4034          * further than we started -- ie. we'll be penalized.
4035          */
4036         if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4037                 update_min_vruntime(cfs_rq);
4038 }
4039 
4040 /*
4041  * Preempt the current task with a newly woken task if needed:
4042  */
4043 static void
4044 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4045 {
4046         unsigned long ideal_runtime, delta_exec;
4047         struct sched_entity *se;
4048         s64 delta;
4049 
4050         ideal_runtime = sched_slice(cfs_rq, curr);
4051         delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4052         if (delta_exec > ideal_runtime) {
4053                 resched_curr(rq_of(cfs_rq));
4054                 /*
4055                  * The current task ran long enough, ensure it doesn't get
4056                  * re-elected due to buddy favours.
4057                  */
4058                 clear_buddies(cfs_rq, curr);
4059                 return;
4060         }
4061 
4062         /*
4063          * Ensure that a task that missed wakeup preemption by a
4064          * narrow margin doesn't have to wait for a full slice.
4065          * This also mitigates buddy induced latencies under load.
4066          */
4067         if (delta_exec < sysctl_sched_min_granularity)
4068                 return;
4069 
4070         se = __pick_first_entity(cfs_rq);
4071         delta = curr->vruntime - se->vruntime;
4072 
4073         if (delta < 0)
4074                 return;
4075 
4076         if (delta > ideal_runtime)
4077                 resched_curr(rq_of(cfs_rq));
4078 }
4079 
4080 static void
4081 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4082 {
4083         /* 'current' is not kept within the tree. */
4084         if (se->on_rq) {
4085                 /*
4086                  * Any task has to be enqueued before it get to execute on
4087                  * a CPU. So account for the time it spent waiting on the
4088                  * runqueue.
4089                  */
4090                 update_stats_wait_end(cfs_rq, se);
4091                 __dequeue_entity(cfs_rq, se);
4092                 update_load_avg(cfs_rq, se, UPDATE_TG);
4093         }
4094 
4095         update_stats_curr_start(cfs_rq, se);
4096         cfs_rq->curr = se;
4097 
4098         /*
4099          * Track our maximum slice length, if the CPU's load is at
4100          * least twice that of our own weight (i.e. dont track it
4101          * when there are only lesser-weight tasks around):
4102          */
4103         if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
4104                 schedstat_set(se->statistics.slice_max,
4105                         max((u64)schedstat_val(se->statistics.slice_max),
4106                             se->sum_exec_runtime - se->prev_sum_exec_runtime));
4107         }
4108 
4109         se->prev_sum_exec_runtime = se->sum_exec_runtime;
4110 }
4111 
4112 static int
4113 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4114 
4115 /*
4116  * Pick the next process, keeping these things in mind, in this order:
4117  * 1) keep things fair between processes/task groups
4118  * 2) pick the "next" process, since someone really wants that to run
4119  * 3) pick the "last" process, for cache locality
4120  * 4) do not run the "skip" process, if something else is available
4121  */
4122 static struct sched_entity *
4123 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4124 {
4125         struct sched_entity *left = __pick_first_entity(cfs_rq);
4126         struct sched_entity *se;
4127 
4128         /*
4129          * If curr is set we have to see if its left of the leftmost entity
4130          * still in the tree, provided there was anything in the tree at all.
4131          */
4132         if (!left || (curr && entity_before(curr, left)))
4133                 left = curr;
4134 
4135         se = left; /* ideally we run the leftmost entity */
4136 
4137         /*
4138          * Avoid running the skip buddy, if running something else can
4139          * be done without getting too unfair.
4140          */
4141         if (cfs_rq->skip == se) {
4142                 struct sched_entity *second;
4143 
4144                 if (se == curr) {
4145                         second = __pick_first_entity(cfs_rq);
4146                 } else {
4147                         second = __pick_next_entity(se);
4148                         if (!second || (curr && entity_before(curr, second)))
4149                                 second = curr;
4150                 }
4151 
4152                 if (second && wakeup_preempt_entity(second, left) < 1)
4153                         se = second;
4154         }
4155 
4156         /*
4157          * Prefer last buddy, try to return the CPU to a preempted task.
4158          */
4159         if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4160                 se = cfs_rq->last;
4161 
4162         /*
4163          * Someone really wants this to run. If it's not unfair, run it.
4164          */
4165         if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4166                 se = cfs_rq->next;
4167 
4168         clear_buddies(cfs_rq, se);
4169 
4170         return se;
4171 }
4172 
4173 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4174 
4175 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4176 {
4177         /*
4178          * If still on the runqueue then deactivate_task()
4179          * was not called and update_curr() has to be done:
4180          */
4181         if (prev->on_rq)
4182                 update_curr(cfs_rq);
4183 
4184         /* throttle cfs_rqs exceeding runtime */
4185         check_cfs_rq_runtime(cfs_rq);
4186 
4187         check_spread(cfs_rq, prev);
4188 
4189         if (prev->on_rq) {
4190                 update_stats_wait_start(cfs_rq, prev);
4191                 /* Put 'current' back into the tree. */
4192                 __enqueue_entity(cfs_rq, prev);
4193                 /* in !on_rq case, update occurred at dequeue */
4194                 update_load_avg(cfs_rq, prev, 0);
4195         }
4196         cfs_rq->curr = NULL;
4197 }
4198 
4199 static void
4200 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4201 {
4202         /*
4203          * Update run-time statistics of the 'current'.
4204          */
4205         update_curr(cfs_rq);
4206 
4207         /*
4208          * Ensure that runnable average is periodically updated.
4209          */
4210         update_load_avg(cfs_rq, curr, UPDATE_TG);
4211         update_cfs_group(curr);
4212 
4213 #ifdef CONFIG_SCHED_HRTICK
4214         /*
4215          * queued ticks are scheduled to match the slice, so don't bother
4216          * validating it and just reschedule.
4217          */
4218         if (queued) {
4219                 resched_curr(rq_of(cfs_rq));
4220                 return;
4221         }
4222         /*
4223          * don't let the period tick interfere with the hrtick preemption
4224          */
4225         if (!sched_feat(DOUBLE_TICK) &&
4226                         hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4227                 return;
4228 #endif
4229 
4230         if (cfs_rq->nr_running > 1)
4231                 check_preempt_tick(cfs_rq, curr);
4232 }
4233 
4234 
4235 /**************************************************
4236  * CFS bandwidth control machinery
4237  */
4238 
4239 #ifdef CONFIG_CFS_BANDWIDTH
4240 
4241 #ifdef CONFIG_JUMP_LABEL
4242 static struct static_key __cfs_bandwidth_used;
4243 
4244 static inline bool cfs_bandwidth_used(void)
4245 {
4246         return static_key_false(&__cfs_bandwidth_used);
4247 }
4248 
4249 void cfs_bandwidth_usage_inc(void)
4250 {
4251         static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4252 }
4253 
4254 void cfs_bandwidth_usage_dec(void)
4255 {
4256         static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4257 }
4258 #else /* CONFIG_JUMP_LABEL */
4259 static bool cfs_bandwidth_used(void)
4260 {
4261         return true;
4262 }
4263 
4264 void cfs_bandwidth_usage_inc(void) {}
4265 void cfs_bandwidth_usage_dec(void) {}
4266 #endif /* CONFIG_JUMP_LABEL */
4267 
4268 /*
4269  * default period for cfs group bandwidth.
4270  * default: 0.1s, units: nanoseconds
4271  */
4272 static inline u64 default_cfs_period(void)
4273 {
4274         return 100000000ULL;
4275 }
4276 
4277 static inline u64 sched_cfs_bandwidth_slice(void)
4278 {
4279         return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4280 }
4281 
4282 /*
4283  * Replenish runtime according to assigned quota and update expiration time.
4284  * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4285  * additional synchronization around rq->lock.
4286  *
4287  * requires cfs_b->lock
4288  */
4289 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4290 {
4291         u64 now;
4292 
4293         if (cfs_b->quota == RUNTIME_INF)
4294                 return;
4295 
4296         now = sched_clock_cpu(smp_processor_id());
4297         cfs_b->runtime = cfs_b->quota;
4298         cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
4299         cfs_b->expires_seq++;
4300 }
4301 
4302 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4303 {
4304         return &tg->cfs_bandwidth;
4305 }
4306 
4307 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4308 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4309 {
4310         if (unlikely(cfs_rq->throttle_count))
4311                 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4312 
4313         return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4314 }
4315 
4316 /* returns 0 on failure to allocate runtime */
4317 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4318 {
4319         struct task_group *tg = cfs_rq->tg;
4320         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4321         u64 amount = 0, min_amount, expires;
4322         int expires_seq;
4323 
4324         /* note: this is a positive sum as runtime_remaining <= 0 */
4325         min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4326 
4327         raw_spin_lock(&cfs_b->lock);
4328         if (cfs_b->quota == RUNTIME_INF)
4329                 amount = min_amount;
4330         else {
4331                 start_cfs_bandwidth(cfs_b);
4332 
4333                 if (cfs_b->runtime > 0) {
4334                         amount = min(cfs_b->runtime, min_amount);
4335                         cfs_b->runtime -= amount;
4336                         cfs_b->idle = 0;
4337                 }
4338         }
4339         expires_seq = cfs_b->expires_seq;
4340         expires = cfs_b->runtime_expires;
4341         raw_spin_unlock(&cfs_b->lock);
4342 
4343         cfs_rq->runtime_remaining += amount;
4344         /*
4345          * we may have advanced our local expiration to account for allowed
4346          * spread between our sched_clock and the one on which runtime was
4347          * issued.
4348          */
4349         if (cfs_rq->expires_seq != expires_seq) {
4350                 cfs_rq->expires_seq = expires_seq;
4351                 cfs_rq->runtime_expires = expires;
4352         }
4353 
4354         return cfs_rq->runtime_remaining > 0;
4355 }
4356 
4357 /*
4358  * Note: This depends on the synchronization provided by sched_clock and the
4359  * fact that rq->clock snapshots this value.
4360  */
4361 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4362 {
4363         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4364 
4365         /* if the deadline is ahead of our clock, nothing to do */
4366         if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
4367                 return;
4368 
4369         if (cfs_rq->runtime_remaining < 0)
4370                 return;
4371 
4372         /*
4373          * If the local deadline has passed we have to consider the
4374          * possibility that our sched_clock is 'fast' and the global deadline
4375          * has not truly expired.
4376          *
4377          * Fortunately we can check determine whether this the case by checking
4378          * whether the global deadline(cfs_b->expires_seq) has advanced.
4379          */
4380         if (cfs_rq->expires_seq == cfs_b->expires_seq) {
4381                 /* extend local deadline, drift is bounded above by 2 ticks */
4382                 cfs_rq->runtime_expires += TICK_NSEC;
4383         } else {
4384                 /* global deadline is ahead, expiration has passed */
4385                 cfs_rq->runtime_remaining = 0;
4386         }
4387 }
4388 
4389 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4390 {
4391         /* dock delta_exec before expiring quota (as it could span periods) */
4392         cfs_rq->runtime_remaining -= delta_exec;
4393         expire_cfs_rq_runtime(cfs_rq);
4394 
4395         if (likely(cfs_rq->runtime_remaining > 0))
4396                 return;
4397 
4398         /*
4399          * if we're unable to extend our runtime we resched so that the active
4400          * hierarchy can be throttled
4401          */
4402         if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4403                 resched_curr(rq_of(cfs_rq));
4404 }
4405 
4406 static __always_inline
4407 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4408 {
4409         if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4410                 return;
4411 
4412         __account_cfs_rq_runtime(cfs_rq, delta_exec);
4413 }
4414 
4415 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4416 {
4417         return cfs_bandwidth_used() && cfs_rq->throttled;
4418 }
4419 
4420 /* check whether cfs_rq, or any parent, is throttled */
4421 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4422 {
4423         return cfs_bandwidth_used() && cfs_rq->throttle_count;
4424 }
4425 
4426 /*
4427  * Ensure that neither of the group entities corresponding to src_cpu or
4428  * dest_cpu are members of a throttled hierarchy when performing group
4429  * load-balance operations.
4430  */
4431 static inline int throttled_lb_pair(struct task_group *tg,
4432                                     int src_cpu, int dest_cpu)
4433 {
4434         struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4435 
4436         src_cfs_rq = tg->cfs_rq[src_cpu];
4437         dest_cfs_rq = tg->cfs_rq[dest_cpu];
4438 
4439         return throttled_hierarchy(src_cfs_rq) ||
4440                throttled_hierarchy(dest_cfs_rq);
4441 }
4442 
4443 static int tg_unthrottle_up(struct task_group *tg, void *data)
4444 {
4445         struct rq *rq = data;
4446         struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4447 
4448         cfs_rq->throttle_count--;
4449         if (!cfs_rq->throttle_count) {
4450                 /* adjust cfs_rq_clock_task() */
4451                 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4452                                              cfs_rq->throttled_clock_task;
4453 
4454                 /* Add cfs_rq with already running entity in the list */
4455                 if (cfs_rq->nr_running >= 1)
4456                         list_add_leaf_cfs_rq(cfs_rq);
4457         }
4458 
4459         return 0;
4460 }
4461 
4462 static int tg_throttle_down(struct task_group *tg, void *data)
4463 {
4464         struct rq *rq = data;
4465         struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4466 
4467         /* group is entering throttled state, stop time */
4468         if (!cfs_rq->throttle_count) {
4469                 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4470                 list_del_leaf_cfs_rq(cfs_rq);
4471         }
4472         cfs_rq->throttle_count++;
4473 
4474         return 0;
4475 }
4476 
4477 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4478 {
4479         struct rq *rq = rq_of(cfs_rq);
4480         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4481         struct sched_entity *se;
4482         long task_delta, dequeue = 1;
4483         bool empty;
4484 
4485         se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4486 
4487         /* freeze hierarchy runnable averages while throttled */
4488         rcu_read_lock();
4489         walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4490         rcu_read_unlock();
4491 
4492         task_delta = cfs_rq->h_nr_running;
4493         for_each_sched_entity(se) {
4494                 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4495                 /* throttled entity or throttle-on-deactivate */
4496                 if (!se->on_rq)
4497                         break;
4498 
4499                 if (dequeue)
4500                         dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4501                 qcfs_rq->h_nr_running -= task_delta;
4502 
4503                 if (qcfs_rq->load.weight)
4504                         dequeue = 0;
4505         }
4506 
4507         if (!se)
4508                 sub_nr_running(rq, task_delta);
4509 
4510         cfs_rq->throttled = 1;
4511         cfs_rq->throttled_clock = rq_clock(rq);
4512         raw_spin_lock(&cfs_b->lock);
4513         empty = list_empty(&cfs_b->throttled_cfs_rq);
4514 
4515         /*
4516          * Add to the _head_ of the list, so that an already-started
4517          * distribute_cfs_runtime will not see us. If disribute_cfs_runtime is
4518          * not running add to the tail so that later runqueues don't get starved.
4519          */
4520         if (cfs_b->distribute_running)
4521                 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4522         else
4523                 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4524 
4525         /*
4526          * If we're the first throttled task, make sure the bandwidth
4527          * timer is running.
4528          */
4529         if (empty)
4530                 start_cfs_bandwidth(cfs_b);
4531 
4532         raw_spin_unlock(&cfs_b->lock);
4533 }
4534 
4535 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4536 {
4537         struct rq *rq = rq_of(cfs_rq);
4538         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4539         struct sched_entity *se;
4540         int enqueue = 1;
4541         long task_delta;
4542 
4543         se = cfs_rq->tg->se[cpu_of(rq)];
4544 
4545         cfs_rq->throttled = 0;
4546 
4547         update_rq_clock(rq);
4548 
4549         raw_spin_lock(&cfs_b->lock);
4550         cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4551         list_del_rcu(&cfs_rq->throttled_list);
4552         raw_spin_unlock(&cfs_b->lock);
4553 
4554         /* update hierarchical throttle state */
4555         walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4556 
4557         if (!cfs_rq->load.weight)
4558                 return;
4559 
4560         task_delta = cfs_rq->h_nr_running;
4561         for_each_sched_entity(se) {
4562                 if (se->on_rq)
4563                         enqueue = 0;
4564 
4565                 cfs_rq = cfs_rq_of(se);
4566                 if (enqueue)
4567                         enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4568                 cfs_rq->h_nr_running += task_delta;
4569 
4570                 if (cfs_rq_throttled(cfs_rq))
4571                         break;
4572         }
4573 
4574         assert_list_leaf_cfs_rq(rq);
4575 
4576         if (!se)
4577                 add_nr_running(rq, task_delta);
4578 
4579         /* Determine whether we need to wake up potentially idle CPU: */
4580         if (rq->curr == rq->idle && rq->cfs.nr_running)
4581                 resched_curr(rq);
4582 }
4583 
4584 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4585                 u64 remaining, u64 expires)
4586 {
4587         struct cfs_rq *cfs_rq;
4588         u64 runtime;
4589         u64 starting_runtime = remaining;
4590 
4591         rcu_read_lock();
4592         list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4593                                 throttled_list) {
4594                 struct rq *rq = rq_of(cfs_rq);
4595                 struct rq_flags rf;
4596 
4597                 rq_lock_irqsave(rq, &rf);
4598                 if (!cfs_rq_throttled(cfs_rq))
4599                         goto next;
4600 
4601                 runtime = -cfs_rq->runtime_remaining + 1;
4602                 if (runtime > remaining)
4603                         runtime = remaining;
4604                 remaining -= runtime;
4605 
4606                 cfs_rq->runtime_remaining += runtime;
4607                 cfs_rq->runtime_expires = expires;
4608 
4609                 /* we check whether we're throttled above */
4610                 if (cfs_rq->runtime_remaining > 0)
4611                         unthrottle_cfs_rq(cfs_rq);
4612 
4613 next:
4614                 rq_unlock_irqrestore(rq, &rf);
4615 
4616                 if (!remaining)
4617                         break;
4618         }
4619         rcu_read_unlock();
4620 
4621         return starting_runtime - remaining;
4622 }
4623 
4624 /*
4625  * Responsible for refilling a task_group's bandwidth and unthrottling its
4626  * cfs_rqs as appropriate. If there has been no activity within the last
4627  * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4628  * used to track this state.
4629  */
4630 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
4631 {
4632         u64 runtime, runtime_expires;
4633         int throttled;
4634 
4635         /* no need to continue the timer with no bandwidth constraint */
4636         if (cfs_b->quota == RUNTIME_INF)
4637                 goto out_deactivate;
4638 
4639         throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4640         cfs_b->nr_periods += overrun;
4641 
4642         /*
4643          * idle depends on !throttled (for the case of a large deficit), and if
4644          * we're going inactive then everything else can be deferred
4645          */
4646         if (cfs_b->idle && !throttled)
4647                 goto out_deactivate;
4648 
4649         __refill_cfs_bandwidth_runtime(cfs_b);
4650 
4651         if (!throttled) {
4652                 /* mark as potentially idle for the upcoming period */
4653                 cfs_b->idle = 1;
4654                 return 0;
4655         }
4656 
4657         /* account preceding periods in which throttling occurred */
4658         cfs_b->nr_throttled += overrun;
4659 
4660         runtime_expires = cfs_b->runtime_expires;
4661 
4662         /*
4663          * This check is repeated as we are holding onto the new bandwidth while
4664          * we unthrottle. This can potentially race with an unthrottled group
4665          * trying to acquire new bandwidth from the global pool. This can result
4666          * in us over-using our runtime if it is all used during this loop, but
4667          * only by limited amounts in that extreme case.
4668          */
4669         while (throttled && cfs_b->runtime > 0 && !cfs_b->distribute_running) {
4670                 runtime = cfs_b->runtime;
4671                 cfs_b->distribute_running = 1;
4672                 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4673                 /* we can't nest cfs_b->lock while distributing bandwidth */
4674                 runtime = distribute_cfs_runtime(cfs_b, runtime,
4675                                                  runtime_expires);
4676                 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4677 
4678                 cfs_b->distribute_running = 0;
4679                 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4680 
4681                 lsub_positive(&cfs_b->runtime, runtime);
4682         }
4683 
4684         /*
4685          * While we are ensured activity in the period following an
4686          * unthrottle, this also covers the case in which the new bandwidth is
4687          * insufficient to cover the existing bandwidth deficit.  (Forcing the
4688          * timer to remain active while there are any throttled entities.)
4689          */
4690         cfs_b->idle = 0;
4691 
4692         return 0;
4693 
4694 out_deactivate:
4695         return 1;
4696 }
4697 
4698 /* a cfs_rq won't donate quota below this amount */
4699 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4700 /* minimum remaining period time to redistribute slack quota */
4701 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4702 /* how long we wait to gather additional slack before distributing */
4703 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4704 
4705 /*
4706  * Are we near the end of the current quota period?
4707  *
4708  * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4709  * hrtimer base being cleared by hrtimer_start. In the case of
4710  * migrate_hrtimers, base is never cleared, so we are fine.
4711  */
4712 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4713 {
4714         struct hrtimer *refresh_timer = &cfs_b->period_timer;
4715         u64 remaining;
4716 
4717         /* if the call-back is running a quota refresh is already occurring */
4718         if (hrtimer_callback_running(refresh_timer))
4719                 return 1;
4720 
4721         /* is a quota refresh about to occur? */
4722         remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4723         if (remaining < min_expire)
4724                 return 1;
4725 
4726         return 0;
4727 }
4728 
4729 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4730 {
4731         u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4732 
4733         /* if there's a quota refresh soon don't bother with slack */
4734         if (runtime_refresh_within(cfs_b, min_left))
4735                 return;
4736 
4737         hrtimer_start(&cfs_b->slack_timer,
4738                         ns_to_ktime(cfs_bandwidth_slack_period),
4739                         HRTIMER_MODE_REL);
4740 }
4741 
4742 /* we know any runtime found here is valid as update_curr() precedes return */
4743 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4744 {
4745         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4746         s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4747 
4748         if (slack_runtime <= 0)
4749                 return;
4750 
4751         raw_spin_lock(&cfs_b->lock);
4752         if (cfs_b->quota != RUNTIME_INF &&
4753             cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4754                 cfs_b->runtime += slack_runtime;
4755 
4756                 /* we are under rq->lock, defer unthrottling using a timer */
4757                 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4758                     !list_empty(&cfs_b->throttled_cfs_rq))
4759                         start_cfs_slack_bandwidth(cfs_b);
4760         }
4761         raw_spin_unlock(&cfs_b->lock);
4762 
4763         /* even if it's not valid for return we don't want to try again */
4764         cfs_rq->runtime_remaining -= slack_runtime;
4765 }
4766 
4767 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4768 {
4769         if (!cfs_bandwidth_used())
4770                 return;
4771 
4772         if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4773                 return;
4774 
4775         __return_cfs_rq_runtime(cfs_rq);
4776 }
4777 
4778 /*
4779  * This is done with a timer (instead of inline with bandwidth return) since
4780  * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4781  */
4782 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4783 {
4784         u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4785         unsigned long flags;
4786         u64 expires;
4787 
4788         /* confirm we're still not at a refresh boundary */
4789         raw_spin_lock_irqsave(&cfs_b->lock, flags);
4790         if (cfs_b->distribute_running) {
4791                 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4792                 return;
4793         }
4794 
4795         if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4796                 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4797                 return;
4798         }
4799 
4800         if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4801                 runtime = cfs_b->runtime;
4802 
4803         expires = cfs_b->runtime_expires;
4804         if (runtime)
4805                 cfs_b->distribute_running = 1;
4806 
4807         raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4808 
4809         if (!runtime)
4810                 return;
4811 
4812         runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4813 
4814         raw_spin_lock_irqsave(&cfs_b->lock, flags);
4815         if (expires == cfs_b->runtime_expires)
4816                 lsub_positive(&cfs_b->runtime, runtime);
4817         cfs_b->distribute_running = 0;
4818         raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4819 }
4820 
4821 /*
4822  * When a group wakes up we want to make sure that its quota is not already
4823  * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4824  * runtime as update_curr() throttling can not not trigger until it's on-rq.
4825  */
4826 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4827 {
4828         if (!cfs_bandwidth_used())
4829                 return;
4830 
4831         /* an active group must be handled by the update_curr()->put() path */
4832         if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4833                 return;
4834 
4835         /* ensure the group is not already throttled */
4836         if (cfs_rq_throttled(cfs_rq))
4837                 return;
4838 
4839         /* update runtime allocation */
4840         account_cfs_rq_runtime(cfs_rq, 0);
4841         if (cfs_rq->runtime_remaining <= 0)
4842                 throttle_cfs_rq(cfs_rq);
4843 }
4844 
4845 static void sync_throttle(struct task_group *tg, int cpu)
4846 {
4847         struct cfs_rq *pcfs_rq, *cfs_rq;
4848 
4849         if (!cfs_bandwidth_used())
4850                 return;
4851 
4852         if (!tg->parent)
4853                 return;
4854 
4855         cfs_rq = tg->cfs_rq[cpu];
4856         pcfs_rq = tg->parent->cfs_rq[cpu];
4857 
4858         cfs_rq->throttle_count = pcfs_rq->throttle_count;
4859         cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4860 }
4861 
4862 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4863 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4864 {
4865         if (!cfs_bandwidth_used())
4866                 return false;
4867 
4868         if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4869                 return false;
4870 
4871         /*
4872          * it's possible for a throttled entity to be forced into a running
4873          * state (e.g. set_curr_task), in this case we're finished.
4874          */
4875         if (cfs_rq_throttled(cfs_rq))
4876                 return true;
4877 
4878         throttle_cfs_rq(cfs_rq);
4879         return true;
4880 }
4881 
4882 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4883 {
4884         struct cfs_bandwidth *cfs_b =
4885                 container_of(timer, struct cfs_bandwidth, slack_timer);
4886 
4887         do_sched_cfs_slack_timer(cfs_b);
4888 
4889         return HRTIMER_NORESTART;
4890 }
4891 
4892 extern const u64 max_cfs_quota_period;
4893 
4894 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4895 {
4896         struct cfs_bandwidth *cfs_b =
4897                 container_of(timer, struct cfs_bandwidth, period_timer);
4898         unsigned long flags;
4899         int overrun;
4900         int idle = 0;
4901         int count = 0;
4902 
4903         raw_spin_lock_irqsave(&cfs_b->lock, flags);
4904         for (;;) {
4905                 overrun = hrtimer_forward_now(timer, cfs_b->period);
4906                 if (!overrun)
4907                         break;
4908 
4909                 if (++count > 3) {
4910                         u64 new, old = ktime_to_ns(cfs_b->period);
4911 
4912                         new = (old * 147) / 128; /* ~115% */
4913                         new = min(new, max_cfs_quota_period);
4914 
4915                         cfs_b->period = ns_to_ktime(new);
4916 
4917                         /* since max is 1s, this is limited to 1e9^2, which fits in u64 */
4918                         cfs_b->quota *= new;
4919                         cfs_b->quota = div64_u64(cfs_b->quota, old);
4920 
4921                         pr_warn_ratelimited(
4922         "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us %lld, cfs_quota_us = %lld)\n",
4923                                 smp_processor_id(),
4924                                 div_u64(new, NSEC_PER_USEC),
4925                                 div_u64(cfs_b->quota, NSEC_PER_USEC));
4926 
4927                         /* reset count so we don't come right back in here */
4928                         count = 0;
4929                 }
4930 
4931                 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
4932         }
4933         if (idle)
4934                 cfs_b->period_active = 0;
4935         raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4936 
4937         return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4938 }
4939 
4940 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4941 {
4942         raw_spin_lock_init(&cfs_b->lock);
4943         cfs_b->runtime = 0;
4944         cfs_b->quota = RUNTIME_INF;
4945         cfs_b->period = ns_to_ktime(default_cfs_period());
4946 
4947         INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4948         hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4949         cfs_b->period_timer.function = sched_cfs_period_timer;
4950         hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4951         cfs_b->slack_timer.function = sched_cfs_slack_timer;
4952         cfs_b->distribute_running = 0;
4953 }
4954 
4955 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4956 {
4957         cfs_rq->runtime_enabled = 0;
4958         INIT_LIST_HEAD(&cfs_rq->throttled_list);
4959 }
4960 
4961 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4962 {
4963         u64 overrun;
4964 
4965         lockdep_assert_held(&cfs_b->lock);
4966 
4967         if (cfs_b->period_active)
4968                 return;
4969 
4970         cfs_b->period_active = 1;
4971         overrun = hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4972         cfs_b->runtime_expires += (overrun + 1) * ktime_to_ns(cfs_b->period);
4973         cfs_b->expires_seq++;
4974         hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4975 }
4976 
4977 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4978 {
4979         /* init_cfs_bandwidth() was not called */
4980         if (!cfs_b->throttled_cfs_rq.next)
4981                 return;
4982 
4983         hrtimer_cancel(&cfs_b->period_timer);
4984         hrtimer_cancel(&cfs_b->slack_timer);
4985 }
4986 
4987 /*
4988  * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
4989  *
4990  * The race is harmless, since modifying bandwidth settings of unhooked group
4991  * bits doesn't do much.
4992  */
4993 
4994 /* cpu online calback */
4995 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4996 {
4997         struct task_group *tg;
4998 
4999         lockdep_assert_held(&rq->lock);
5000 
5001         rcu_read_lock();
5002         list_for_each_entry_rcu(tg, &task_groups, list) {
5003                 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5004                 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5005 
5006                 raw_spin_lock(&cfs_b->lock);
5007                 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5008                 raw_spin_unlock(&cfs_b->lock);
5009         }
5010         rcu_read_unlock();
5011 }
5012 
5013 /* cpu offline callback */
5014 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5015 {
5016         struct task_group *tg;
5017 
5018         lockdep_assert_held(&rq->lock);
5019 
5020         rcu_read_lock();
5021         list_for_each_entry_rcu(tg, &task_groups, list) {
5022                 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5023 
5024                 if (!cfs_rq->runtime_enabled)
5025                         continue;
5026 
5027                 /*
5028                  * clock_task is not advancing so we just need to make sure
5029                  * there's some valid quota amount
5030                  */
5031                 cfs_rq->runtime_remaining = 1;
5032                 /*
5033                  * Offline rq is schedulable till CPU is completely disabled
5034                  * in take_cpu_down(), so we prevent new cfs throttling here.
5035                  */
5036                 cfs_rq->runtime_enabled = 0;
5037 
5038                 if (cfs_rq_throttled(cfs_rq))
5039                         unthrottle_cfs_rq(cfs_rq);
5040         }
5041         rcu_read_unlock();
5042 }
5043 
5044 #else /* CONFIG_CFS_BANDWIDTH */
5045 
5046 static inline bool cfs_bandwidth_used(void)
5047 {
5048         return false;
5049 }
5050 
5051 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
5052 {
5053         return rq_clock_task(rq_of(cfs_rq));
5054 }
5055 
5056 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5057 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5058 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5059 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5060 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5061 
5062 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5063 {
5064         return 0;
5065 }
5066 
5067 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5068 {
5069         return 0;
5070 }
5071 
5072 static inline int throttled_lb_pair(struct task_group *tg,
5073                                     int src_cpu, int dest_cpu)
5074 {
5075         return 0;
5076 }
5077 
5078 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5079 
5080 #ifdef CONFIG_FAIR_GROUP_SCHED
5081 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5082 #endif
5083 
5084 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5085 {
5086         return NULL;
5087 }
5088 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5089 static inline void update_runtime_enabled(struct rq *rq) {}
5090 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5091 
5092 #endif /* CONFIG_CFS_BANDWIDTH */
5093 
5094 /**************************************************
5095  * CFS operations on tasks:
5096  */
5097 
5098 #ifdef CONFIG_SCHED_HRTICK
5099 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5100 {
5101         struct sched_entity *se = &p->se;
5102         struct cfs_rq *cfs_rq = cfs_rq_of(se);
5103 
5104         SCHED_WARN_ON(task_rq(p) != rq);
5105 
5106         if (rq->cfs.h_nr_running > 1) {
5107                 u64 slice = sched_slice(cfs_rq, se);
5108                 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5109                 s64 delta = slice - ran;
5110 
5111                 if (delta < 0) {
5112                         if (rq->curr == p)
5113                                 resched_curr(rq);
5114                         return;
5115                 }
5116                 hrtick_start(rq, delta);
5117         }
5118 }
5119 
5120 /*
5121  * called from enqueue/dequeue and updates the hrtick when the
5122  * current task is from our class and nr_running is low enough
5123  * to matter.
5124  */
5125 static void hrtick_update(struct rq *rq)
5126 {
5127         struct task_struct *curr = rq->curr;
5128 
5129         if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
5130                 return;
5131 
5132         if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5133                 hrtick_start_fair(rq, curr);
5134 }
5135 #else /* !CONFIG_SCHED_HRTICK */
5136 static inline void
5137 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5138 {
5139 }
5140 
5141 static inline void hrtick_update(struct rq *rq)
5142 {
5143 }
5144 #endif
5145 
5146 #ifdef CONFIG_SMP
5147 static inline unsigned long cpu_util(int cpu);
5148 static unsigned long capacity_of(int cpu);
5149 
5150 static inline bool cpu_overutilized(int cpu)
5151 {
5152         return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5153 }
5154 
5155 static inline void update_overutilized_status(struct rq *rq)
5156 {
5157         if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu))
5158                 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
5159 }
5160 #else
5161 static inline void update_overutilized_status(struct rq *rq) { }
5162 #endif
5163 
5164 /*
5165  * The enqueue_task method is called before nr_running is
5166  * increased. Here we update the fair scheduling stats and
5167  * then put the task into the rbtree:
5168  */
5169 static void
5170 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5171 {
5172         struct cfs_rq *cfs_rq;
5173         struct sched_entity *se = &p->se;
5174 
5175         /*
5176          * The code below (indirectly) updates schedutil which looks at
5177          * the cfs_rq utilization to select a frequency.
5178          * Let's add the task's estimated utilization to the cfs_rq's
5179          * estimated utilization, before we update schedutil.
5180          */
5181         util_est_enqueue(&rq->cfs, p);
5182 
5183         /*
5184          * If in_iowait is set, the code below may not trigger any cpufreq
5185          * utilization updates, so do it here explicitly with the IOWAIT flag
5186          * passed.
5187          */
5188         if (p->in_iowait)
5189                 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5190 
5191         for_each_sched_entity(se) {
5192                 if (se->on_rq)
5193                         break;
5194                 cfs_rq = cfs_rq_of(se);
5195                 enqueue_entity(cfs_rq, se, flags);
5196 
5197                 /*
5198                  * end evaluation on encountering a throttled cfs_rq
5199                  *
5200                  * note: in the case of encountering a throttled cfs_rq we will
5201                  * post the final h_nr_running increment below.
5202                  */
5203                 if (cfs_rq_throttled(cfs_rq))
5204                         break;
5205                 cfs_rq->h_nr_running++;
5206 
5207                 flags = ENQUEUE_WAKEUP;
5208         }
5209 
5210         for_each_sched_entity(se) {
5211                 cfs_rq = cfs_rq_of(se);
5212                 cfs_rq->h_nr_running++;
5213 
5214                 if (cfs_rq_throttled(cfs_rq))
5215                         break;
5216 
5217                 update_load_avg(cfs_rq, se, UPDATE_TG);
5218                 update_cfs_group(se);
5219         }
5220 
5221         if (!se) {
5222                 add_nr_running(rq, 1);
5223                 /*
5224                  * Since new tasks are assigned an initial util_avg equal to
5225                  * half of the spare capacity of their CPU, tiny tasks have the
5226                  * ability to cross the overutilized threshold, which will
5227                  * result in the load balancer ruining all the task placement
5228                  * done by EAS. As a way to mitigate that effect, do not account
5229                  * for the first enqueue operation of new tasks during the
5230                  * overutilized flag detection.
5231                  *
5232                  * A better way of solving this problem would be to wait for
5233                  * the PELT signals of tasks to converge before taking them
5234                  * into account, but that is not straightforward to implement,
5235                  * and the following generally works well enough in practice.
5236                  */
5237                 if (flags & ENQUEUE_WAKEUP)
5238                         update_overutilized_status(rq);
5239 
5240         }
5241 
5242         if (cfs_bandwidth_used()) {
5243                 /*
5244                  * When bandwidth control is enabled; the cfs_rq_throttled()
5245                  * breaks in the above iteration can result in incomplete
5246                  * leaf list maintenance, resulting in triggering the assertion
5247                  * below.
5248                  */
5249                 for_each_sched_entity(se) {
5250                         cfs_rq = cfs_rq_of(se);
5251 
5252                         if (list_add_leaf_cfs_rq(cfs_rq))
5253                                 break;
5254                 }
5255         }
5256 
5257         assert_list_leaf_cfs_rq(rq);
5258 
5259         hrtick_update(rq);
5260 }
5261 
5262 static void set_next_buddy(struct sched_entity *se);
5263 
5264 /*
5265  * The dequeue_task method is called before nr_running is
5266  * decreased. We remove the task from the rbtree and
5267  * update the fair scheduling stats:
5268  */
5269 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5270 {
5271         struct cfs_rq *cfs_rq;
5272         struct sched_entity *se = &p->se;
5273         int task_sleep = flags & DEQUEUE_SLEEP;
5274 
5275         for_each_sched_entity(se) {
5276                 cfs_rq = cfs_rq_of(se);
5277                 dequeue_entity(cfs_rq, se, flags);
5278 
5279                 /*
5280                  * end evaluation on encountering a throttled cfs_rq
5281                  *
5282                  * note: in the case of encountering a throttled cfs_rq we will
5283                  * post the final h_nr_running decrement below.
5284                 */
5285                 if (cfs_rq_throttled(cfs_rq))
5286                         break;
5287                 cfs_rq->h_nr_running--;
5288 
5289                 /* Don't dequeue parent if it has other entities besides us */
5290                 if (cfs_rq->load.weight) {
5291                         /* Avoid re-evaluating load for this entity: */
5292                         se = parent_entity(se);
5293                         /*
5294                          * Bias pick_next to pick a task from this cfs_rq, as
5295                          * p is sleeping when it is within its sched_slice.
5296                          */
5297                         if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5298                                 set_next_buddy(se);
5299                         break;
5300                 }
5301                 flags |= DEQUEUE_SLEEP;
5302         }
5303 
5304         for_each_sched_entity(se) {
5305                 cfs_rq = cfs_rq_of(se);
5306                 cfs_rq->h_nr_running--;
5307 
5308                 if (cfs_rq_throttled(cfs_rq))
5309                         break;
5310 
5311                 update_load_avg(cfs_rq, se, UPDATE_TG);
5312                 update_cfs_group(se);
5313         }
5314 
5315         if (!se)
5316                 sub_nr_running(rq, 1);
5317 
5318         util_est_dequeue(&rq->cfs, p, task_sleep);
5319         hrtick_update(rq);
5320 }
5321 
5322 #ifdef CONFIG_SMP
5323 
5324 /* Working cpumask for: load_balance, load_balance_newidle. */
5325 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5326 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5327 
5328 #ifdef CONFIG_NO_HZ_COMMON
5329 /*
5330  * per rq 'load' arrray crap; XXX kill this.
5331  */
5332 
5333 /*
5334  * The exact cpuload calculated at every tick would be:
5335  *
5336  *   load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
5337  *
5338  * If a CPU misses updates for n ticks (as it was idle) and update gets
5339  * called on the n+1-th tick when CPU may be busy, then we have:
5340  *
5341  *   load_n   = (1 - 1/2^i)^n * load_0
5342  *   load_n+1 = (1 - 1/2^i)   * load_n + (1/2^i) * cur_load
5343  *
5344  * decay_load_missed() below does efficient calculation of
5345  *
5346  *   load' = (1 - 1/2^i)^n * load
5347  *
5348  * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
5349  * This allows us to precompute the above in said factors, thereby allowing the
5350  * reduction of an arbitrary n in O(log_2 n) steps. (See also
5351  * fixed_power_int())
5352  *
5353  * The calculation is approximated on a 128 point scale.
5354  */
5355 #define DEGRADE_SHIFT           7
5356 
5357 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
5358 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
5359         {   0,   0,  0,  0,  0,  0, 0, 0 },
5360         {  64,  32,  8,  0,  0,  0, 0, 0 },
5361         {  96,  72, 40, 12,  1,  0, 0, 0 },
5362         { 112,  98, 75, 43, 15,  1, 0, 0 },
5363         { 120, 112, 98, 76, 45, 16, 2, 0 }
5364 };
5365 
5366 /*
5367  * Update cpu_load for any missed ticks, due to tickless idle. The backlog
5368  * would be when CPU is idle and so we just decay the old load without
5369  * adding any new load.
5370  */
5371 static unsigned long
5372 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
5373 {
5374         int j = 0;
5375 
5376         if (!missed_updates)
5377                 return load;
5378 
5379         if (missed_updates >= degrade_zero_ticks[idx])
5380                 return 0;
5381 
5382         if (idx == 1)
5383                 return load >> missed_updates;
5384 
5385         while (missed_updates) {
5386                 if (missed_updates % 2)
5387                         load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
5388 
5389                 missed_updates >>= 1;
5390                 j++;
5391         }
5392         return load;
5393 }
5394 
5395 static struct {
5396         cpumask_var_t idle_cpus_mask;
5397         atomic_t nr_cpus;
5398         int has_blocked;                /* Idle CPUS has blocked load */
5399         unsigned long next_balance;     /* in jiffy units */
5400         unsigned long next_blocked;     /* Next update of blocked load in jiffies */
5401 } nohz ____cacheline_aligned;
5402 
5403 #endif /* CONFIG_NO_HZ_COMMON */
5404 
5405 /**
5406  * __cpu_load_update - update the rq->cpu_load[] statistics
5407  * @this_rq: The rq to update statistics for
5408  * @this_load: The current load
5409  * @pending_updates: The number of missed updates
5410  *
5411  * Update rq->cpu_load[] statistics. This function is usually called every
5412  * scheduler tick (TICK_NSEC).
5413  *
5414  * This function computes a decaying average:
5415  *
5416  *   load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
5417  *
5418  * Because of NOHZ it might not get called on every tick which gives need for
5419  * the @pending_updates argument.
5420  *
5421  *   load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
5422  *             = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
5423  *             = A * (A * load[i]_n-2 + B) + B
5424  *             = A * (A * (A * load[i]_n-3 + B) + B) + B
5425  *             = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5426  *             = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5427  *             = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5428  *             = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5429  *
5430  * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5431  * any change in load would have resulted in the tick being turned back on.
5432  *
5433  * For regular NOHZ, this reduces to:
5434  *
5435  *   load[i]_n = (1 - 1/2^i)^n * load[i]_0
5436  *
5437  * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5438  * term.
5439  */
5440 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
5441                             unsigned long pending_updates)
5442 {
5443         unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
5444         int i, scale;
5445 
5446         this_rq->nr_load_updates++;
5447 
5448         /* Update our load: */
5449         this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
5450         for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
5451                 unsigned long old_load, new_load;
5452 
5453                 /* scale is effectively 1 << i now, and >> i divides by scale */
5454 
5455                 old_load = this_rq->cpu_load[i];
5456 #ifdef CONFIG_NO_HZ_COMMON
5457                 old_load = decay_load_missed(old_load, pending_updates - 1, i);
5458                 if (tickless_load) {
5459                         old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
5460                         /*
5461                          * old_load can never be a negative value because a
5462                          * decayed tickless_load cannot be greater than the
5463                          * original tickless_load.
5464                          */
5465                         old_load += tickless_load;
5466                 }
5467 #endif
5468                 new_load = this_load;
5469                 /*
5470                  * Round up the averaging division if load is increasing. This
5471                  * prevents us from getting stuck on 9 if the load is 10, for
5472                  * example.
5473                  */
5474                 if (new_load > old_load)
5475                         new_load += scale - 1;
5476 
5477                 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
5478         }
5479 }
5480 
5481 /* Used instead of source_load when we know the type == 0 */
5482 static unsigned long weighted_cpuload(struct rq *rq)
5483 {
5484         return cfs_rq_runnable_load_avg(&rq->cfs);
5485 }
5486 
5487 #ifdef CONFIG_NO_HZ_COMMON
5488 /*
5489  * There is no sane way to deal with nohz on smp when using jiffies because the
5490  * CPU doing the jiffies update might drift wrt the CPU doing the jiffy reading
5491  * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5492  *
5493  * Therefore we need to avoid the delta approach from the regular tick when
5494  * possible since that would seriously skew the load calculation. This is why we
5495  * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5496  * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5497  * loop exit, nohz_idle_balance, nohz full exit...)
5498  *
5499  * This means we might still be one tick off for nohz periods.
5500  */
5501 
5502 static void cpu_load_update_nohz(struct rq *this_rq,
5503                                  unsigned long curr_jiffies,
5504                                  unsigned long load)
5505 {
5506         unsigned long pending_updates;
5507 
5508         pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5509         if (pending_updates) {
5510                 this_rq->last_load_update_tick = curr_jiffies;
5511                 /*
5512                  * In the regular NOHZ case, we were idle, this means load 0.
5513                  * In the NOHZ_FULL case, we were non-idle, we should consider
5514                  * its weighted load.
5515                  */
5516                 cpu_load_update(this_rq, load, pending_updates);
5517         }
5518 }
5519 
5520 /*
5521  * Called from nohz_idle_balance() to update the load ratings before doing the
5522  * idle balance.
5523  */
5524 static void cpu_load_update_idle(struct rq *this_rq)
5525 {
5526         /*
5527          * bail if there's load or we're actually up-to-date.
5528          */
5529         if (weighted_cpuload(this_rq))
5530                 return;
5531 
5532         cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
5533 }
5534 
5535 /*
5536  * Record CPU load on nohz entry so we know the tickless load to account
5537  * on nohz exit. cpu_load[0] happens then to be updated more frequently
5538  * than other cpu_load[idx] but it should be fine as cpu_load readers
5539  * shouldn't rely into synchronized cpu_load[*] updates.
5540  */
5541 void cpu_load_update_nohz_start(void)
5542 {
5543         struct rq *this_rq = this_rq();
5544 
5545         /*
5546          * This is all lockless but should be fine. If weighted_cpuload changes
5547          * concurrently we'll exit nohz. And cpu_load write can race with
5548          * cpu_load_update_idle() but both updater would be writing the same.
5549          */
5550         this_rq->cpu_load[0] = weighted_cpuload(this_rq);
5551 }
5552 
5553 /*
5554  * Account the tickless load in the end of a nohz frame.
5555  */
5556 void cpu_load_update_nohz_stop(void)
5557 {
5558         unsigned long curr_jiffies = READ_ONCE(jiffies);
5559         struct rq *this_rq = this_rq();
5560         unsigned long load;
5561         struct rq_flags rf;
5562 
5563         if (curr_jiffies == this_rq->last_load_update_tick)
5564                 return;
5565 
5566         load = weighted_cpuload(this_rq);
5567         rq_lock(this_rq, &rf);
5568         update_rq_clock(this_rq);
5569         cpu_load_update_nohz(this_rq, curr_jiffies, load);
5570         rq_unlock(this_rq, &rf);
5571 }
5572 #else /* !CONFIG_NO_HZ_COMMON */
5573 static inline void cpu_load_update_nohz(struct rq *this_rq,
5574                                         unsigned long curr_jiffies,
5575                                         unsigned long load) { }
5576 #endif /* CONFIG_NO_HZ_COMMON */
5577 
5578 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
5579 {
5580 #ifdef CONFIG_NO_HZ_COMMON
5581         /* See the mess around cpu_load_update_nohz(). */
5582         this_rq->last_load_update_tick = READ_ONCE(jiffies);
5583 #endif
5584         cpu_load_update(this_rq, load, 1);
5585 }
5586 
5587 /*
5588  * Called from scheduler_tick()
5589  */
5590 void cpu_load_update_active(struct rq *this_rq)
5591 {
5592         unsigned long load = weighted_cpuload(this_rq);
5593 
5594         if (tick_nohz_tick_stopped())
5595                 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
5596         else
5597                 cpu_load_update_periodic(this_rq, load);
5598 }
5599 
5600 /*
5601  * Return a low guess at the load of a migration-source CPU weighted
5602  * according to the scheduling class and "nice" value.
5603  *
5604  * We want to under-estimate the load of migration sources, to
5605  * balance conservatively.
5606  */
5607 static unsigned long source_load(int cpu, int type)
5608 {
5609         struct rq *rq = cpu_rq(cpu);
5610         unsigned long total = weighted_cpuload(rq);
5611 
5612         if (type == 0 || !sched_feat(LB_BIAS))
5613                 return total;
5614 
5615         return min(rq->cpu_load[type-1], total);
5616 }
5617 
5618 /*
5619  * Return a high guess at the load of a migration-target CPU weighted
5620  * according to the scheduling class and "nice" value.
5621  */
5622 static unsigned long target_load(int cpu, int type)
5623 {
5624         struct rq *rq = cpu_rq(cpu);
5625         unsigned long total = weighted_cpuload(rq);
5626 
5627         if (type == 0 || !sched_feat(LB_BIAS))
5628                 return total;
5629 
5630         return max(rq->cpu_load[type-1], total);
5631 }
5632 
5633 static unsigned long capacity_of(int cpu)
5634 {
5635         return cpu_rq(cpu)->cpu_capacity;
5636 }
5637 
5638 static unsigned long cpu_avg_load_per_task(int cpu)
5639 {
5640         struct rq *rq = cpu_rq(cpu);
5641         unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5642         unsigned long load_avg = weighted_cpuload(rq);
5643 
5644         if (nr_running)
5645                 return load_avg / nr_running;
5646 
5647         return 0;
5648 }
5649 
5650 static void record_wakee(struct task_struct *p)
5651 {
5652         /*
5653          * Only decay a single time; tasks that have less then 1 wakeup per
5654          * jiffy will not have built up many flips.
5655          */
5656         if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5657                 current->wakee_flips >>= 1;
5658                 current->wakee_flip_decay_ts = jiffies;
5659         }
5660 
5661         if (current->last_wakee != p) {
5662                 current->last_wakee = p;
5663                 current->wakee_flips++;
5664         }
5665 }
5666 
5667 /*
5668  * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5669  *
5670  * A waker of many should wake a different task than the one last awakened
5671  * at a frequency roughly N times higher than one of its wakees.
5672  *
5673  * In order to determine whether we should let the load spread vs consolidating
5674  * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5675  * partner, and a factor of lls_size higher frequency in the other.
5676  *
5677  * With both conditions met, we can be relatively sure that the relationship is
5678  * non-monogamous, with partner count exceeding socket size.
5679  *
5680  * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5681  * whatever is irrelevant, spread criteria is apparent partner count exceeds
5682  * socket size.
5683  */
5684 static int wake_wide(struct task_struct *p)
5685 {
5686         unsigned int master = current->wakee_flips;
5687         unsigned int slave = p->wakee_flips;
5688         int factor = this_cpu_read(sd_llc_size);
5689 
5690         if (master < slave)
5691                 swap(master, slave);
5692         if (slave < factor || master < slave * factor)
5693                 return 0;
5694         return 1;
5695 }
5696 
5697 /*
5698  * The purpose of wake_affine() is to quickly determine on which CPU we can run
5699  * soonest. For the purpose of speed we only consider the waking and previous
5700  * CPU.
5701  *
5702  * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5703  *                      cache-affine and is (or will be) idle.
5704  *
5705  * wake_affine_weight() - considers the weight to reflect the average
5706  *                        scheduling latency of the CPUs. This seems to work
5707  *                        for the overloaded case.
5708  */
5709 static int
5710 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5711 {
5712         /*
5713          * If this_cpu is idle, it implies the wakeup is from interrupt
5714          * context. Only allow the move if cache is shared. Otherwise an
5715          * interrupt intensive workload could force all tasks onto one
5716          * node depending on the IO topology or IRQ affinity settings.
5717          *
5718          * If the prev_cpu is idle and cache affine then avoid a migration.
5719          * There is no guarantee that the cache hot data from an interrupt
5720          * is more important than cache hot data on the prev_cpu and from
5721          * a cpufreq perspective, it's better to have higher utilisation
5722          * on one CPU.
5723          */
5724         if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5725                 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5726 
5727         if (sync && cpu_rq(this_cpu)->nr_running == 1)
5728                 return this_cpu;
5729 
5730         return nr_cpumask_bits;
5731 }
5732 
5733 static int
5734 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5735                    int this_cpu, int prev_cpu, int sync)
5736 {
5737         s64 this_eff_load, prev_eff_load;
5738         unsigned long task_load;
5739 
5740         this_eff_load = target_load(this_cpu, sd->wake_idx);
5741 
5742         if (sync) {
5743                 unsigned long current_load = task_h_load(current);
5744 
5745                 if (current_load > this_eff_load)
5746                         return this_cpu;
5747 
5748                 this_eff_load -= current_load;
5749         }
5750 
5751         task_load = task_h_load(p);
5752 
5753         this_eff_load += task_load;
5754         if (sched_feat(WA_BIAS))
5755                 this_eff_load *= 100;
5756         this_eff_load *= capacity_of(prev_cpu);
5757 
5758         prev_eff_load = source_load(prev_cpu, sd->wake_idx);
5759         prev_eff_load -= task_load;
5760         if (sched_feat(WA_BIAS))
5761                 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5762         prev_eff_load *= capacity_of(this_cpu);
5763 
5764         /*
5765          * If sync, adjust the weight of prev_eff_load such that if
5766          * prev_eff == this_eff that select_idle_sibling() will consider
5767          * stacking the wakee on top of the waker if no other CPU is
5768          * idle.
5769          */
5770         if (sync)
5771                 prev_eff_load += 1;
5772 
5773         return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5774 }
5775 
5776 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5777                        int this_cpu, int prev_cpu, int sync)
5778 {
5779         int target = nr_cpumask_bits;
5780 
5781         if (sched_feat(WA_IDLE))
5782                 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5783 
5784         if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5785                 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5786 
5787         schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5788         if (target == nr_cpumask_bits)
5789                 return prev_cpu;
5790 
5791         schedstat_inc(sd->ttwu_move_affine);
5792         schedstat_inc(p->se.statistics.nr_wakeups_affine);
5793         return target;
5794 }
5795 
5796 static unsigned long cpu_util_without(int cpu, struct task_struct *p);
5797 
5798 static unsigned long capacity_spare_without(int cpu, struct task_struct *p)
5799 {
5800         return max_t(long, capacity_of(cpu) - cpu_util_without(cpu, p), 0);
5801 }
5802 
5803 /*
5804  * find_idlest_group finds and returns the least busy CPU group within the
5805  * domain.
5806  *
5807  * Assumes p is allowed on at least one CPU in sd.
5808  */
5809 static struct sched_group *
5810 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5811                   int this_cpu, int sd_flag)
5812 {
5813         struct sched_group *idlest = NULL, *group = sd->groups;
5814         struct sched_group *most_spare_sg = NULL;
5815         unsigned long min_runnable_load = ULONG_MAX;
5816         unsigned long this_runnable_load = ULONG_MAX;
5817         unsigned long min_avg_load = ULONG_MAX, this_avg_load = ULONG_MAX;
5818         unsigned long most_spare = 0, this_spare = 0;
5819         int load_idx = sd->forkexec_idx;
5820         int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5821         unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5822                                 (sd->imbalance_pct-100) / 100;
5823 
5824         if (sd_flag & SD_BALANCE_WAKE)
5825                 load_idx = sd->wake_idx;
5826 
5827         do {
5828                 unsigned long load, avg_load, runnable_load;
5829                 unsigned long spare_cap, max_spare_cap;
5830                 int local_group;
5831                 int i;
5832 
5833                 /* Skip over this group if it has no CPUs allowed */
5834                 if (!cpumask_intersects(sched_group_span(group),
5835                                         &p->cpus_allowed))
5836                         continue;
5837 
5838                 local_group = cpumask_test_cpu(this_cpu,
5839                                                sched_group_span(group));
5840 
5841                 /*
5842                  * Tally up the load of all CPUs in the group and find
5843                  * the group containing the CPU with most spare capacity.
5844                  */
5845                 avg_load = 0;
5846                 runnable_load = 0;
5847                 max_spare_cap = 0;
5848 
5849                 for_each_cpu(i, sched_group_span(group)) {
5850                         /* Bias balancing toward CPUs of our domain */
5851                         if (local_group)
5852                                 load = source_load(i, load_idx);
5853                         else
5854                                 load = target_load(i, load_idx);
5855 
5856                         runnable_load += load;
5857 
5858                         avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5859 
5860                         spare_cap = capacity_spare_without(i, p);
5861 
5862                         if (spare_cap > max_spare_cap)
5863                                 max_spare_cap = spare_cap;
5864                 }
5865 
5866                 /* Adjust by relative CPU capacity of the group */
5867                 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5868                                         group->sgc->capacity;
5869                 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5870                                         group->sgc->capacity;
5871 
5872                 if (local_group) {
5873                         this_runnable_load = runnable_load;
5874                         this_avg_load = avg_load;
5875                         this_spare = max_spare_cap;
5876                 } else {
5877                         if (min_runnable_load > (runnable_load + imbalance)) {
5878                                 /*
5879                                  * The runnable load is significantly smaller
5880                                  * so we can pick this new CPU:
5881                                  */
5882                                 min_runnable_load = runnable_load;
5883                                 min_avg_load = avg_load;
5884                                 idlest = group;
5885                         } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5886                                    (100*min_avg_load > imbalance_scale*avg_load)) {
5887                                 /*
5888                                  * The runnable loads are close so take the
5889                                  * blocked load into account through avg_load:
5890                                  */
5891                                 min_avg_load = avg_load;
5892                                 idlest = group;
5893                         }
5894 
5895                         if (most_spare < max_spare_cap) {
5896                                 most_spare = max_spare_cap;
5897                                 most_spare_sg = group;
5898                         }
5899                 }
5900         } while (group = group->next, group != sd->groups);
5901 
5902         /*
5903          * The cross-over point between using spare capacity or least load
5904          * is too conservative for high utilization tasks on partially
5905          * utilized systems if we require spare_capacity > task_util(p),
5906          * so we allow for some task stuffing by using
5907          * spare_capacity > task_util(p)/2.
5908          *
5909          * Spare capacity can't be used for fork because the utilization has
5910          * not been set yet, we must first select a rq to compute the initial
5911          * utilization.
5912          */
5913         if (sd_flag & SD_BALANCE_FORK)
5914                 goto skip_spare;
5915 
5916         if (this_spare > task_util(p) / 2 &&
5917             imbalance_scale*this_spare > 100*most_spare)
5918                 return NULL;
5919 
5920         if (most_spare > task_util(p) / 2)
5921                 return most_spare_sg;
5922 
5923 skip_spare:
5924         if (!idlest)
5925                 return NULL;
5926 
5927         /*
5928          * When comparing groups across NUMA domains, it's possible for the
5929          * local domain to be very lightly loaded relative to the remote
5930          * domains but "imbalance" skews the comparison making remote CPUs
5931          * look much more favourable. When considering cross-domain, add
5932          * imbalance to the runnable load on the remote node and consider
5933          * staying local.
5934          */
5935         if ((sd->flags & SD_NUMA) &&
5936             min_runnable_load + imbalance >= this_runnable_load)
5937                 return NULL;
5938 
5939         if (min_runnable_load > (this_runnable_load + imbalance))
5940                 return NULL;
5941 
5942         if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5943              (100*this_avg_load < imbalance_scale*min_avg_load))
5944                 return NULL;
5945 
5946         return idlest;
5947 }
5948 
5949 /*
5950  * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5951  */
5952 static int
5953 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5954 {
5955         unsigned long load, min_load = ULONG_MAX;
5956         unsigned int min_exit_latency = UINT_MAX;
5957         u64 latest_idle_timestamp = 0;
5958         int least_loaded_cpu = this_cpu;
5959         int shallowest_idle_cpu = -1;
5960         int i;
5961 
5962         /* Check if we have any choice: */
5963         if (group->group_weight == 1)
5964                 return cpumask_first(sched_group_span(group));
5965 
5966         /* Traverse only the allowed CPUs */
5967         for_each_cpu_and(i, sched_group_span(group), &p->cpus_allowed) {
5968                 if (available_idle_cpu(i)) {
5969                         struct rq *rq = cpu_rq(i);
5970                         struct cpuidle_state *idle = idle_get_state(rq);
5971                         if (idle && idle->exit_latency < min_exit_latency) {
5972                                 /*
5973                                  * We give priority to a CPU whose idle state
5974                                  * has the smallest exit latency irrespective
5975                                  * of any idle timestamp.
5976                                  */
5977                                 min_exit_latency = idle->exit_latency;
5978                                 latest_idle_timestamp = rq->idle_stamp;
5979                                 shallowest_idle_cpu = i;
5980                         } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5981                                    rq->idle_stamp > latest_idle_timestamp) {
5982                                 /*
5983                                  * If equal or no active idle state, then
5984                                  * the most recently idled CPU might have
5985                                  * a warmer cache.
5986                                  */
5987                                 latest_idle_timestamp = rq->idle_stamp;
5988                                 shallowest_idle_cpu = i;
5989                         }
5990                 } else if (shallowest_idle_cpu == -1) {
5991                         load = weighted_cpuload(cpu_rq(i));
5992                         if (load < min_load) {
5993                                 min_load = load;
5994                                 least_loaded_cpu = i;
5995                         }
5996                 }
5997         }
5998 
5999         return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
6000 }
6001 
6002 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
6003                                   int cpu, int prev_cpu, int sd_flag)
6004 {
6005         int new_cpu = cpu;
6006 
6007         if (!cpumask_intersects(sched_domain_span(sd), &p->cpus_allowed))
6008                 return prev_cpu;
6009 
6010         /*
6011          * We need task's util for capacity_spare_without, sync it up to
6012          * prev_cpu's last_update_time.
6013          */
6014         if (!(sd_flag & SD_BALANCE_FORK))
6015                 sync_entity_load_avg(&p->se);
6016 
6017         while (sd) {
6018                 struct sched_group *group;
6019                 struct sched_domain *tmp;
6020                 int weight;
6021 
6022                 if (!(sd->flags & sd_flag)) {
6023                         sd = sd->child;
6024                         continue;
6025                 }
6026 
6027                 group = find_idlest_group(sd, p, cpu, sd_flag);
6028                 if (!group) {
6029                         sd = sd->child;
6030                         continue;
6031                 }
6032 
6033                 new_cpu = find_idlest_group_cpu(group, p, cpu);
6034                 if (new_cpu == cpu) {
6035                         /* Now try balancing at a lower domain level of 'cpu': */
6036                         sd = sd->child;
6037                         continue;
6038                 }
6039 
6040                 /* Now try balancing at a lower domain level of 'new_cpu': */
6041                 cpu = new_cpu;
6042                 weight = sd->span_weight;
6043                 sd = NULL;
6044                 for_each_domain(cpu, tmp) {
6045                         if (weight <= tmp->span_weight)
6046                                 break;
6047                         if (tmp->flags & sd_flag)
6048                                 sd = tmp;
6049                 }
6050         }
6051 
6052         return new_cpu;
6053 }
6054 
6055 #ifdef CONFIG_SCHED_SMT
6056 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
6057 EXPORT_SYMBOL_GPL(sched_smt_present);
6058 
6059 static inline void set_idle_cores(int cpu, int val)
6060 {
6061         struct sched_domain_shared *sds;
6062 
6063         sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6064         if (sds)
6065                 WRITE_ONCE(sds->has_idle_cores, val);
6066 }
6067 
6068 static inline bool test_idle_cores(int cpu, bool def)
6069 {
6070         struct sched_domain_shared *sds;
6071 
6072         sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6073         if (sds)
6074                 return READ_ONCE(sds->has_idle_cores);
6075 
6076         return def;
6077 }
6078 
6079 /*
6080  * Scans the local SMT mask to see if the entire core is idle, and records this
6081  * information in sd_llc_shared->has_idle_cores.
6082  *
6083  * Since SMT siblings share all cache levels, inspecting this limited remote
6084  * state should be fairly cheap.
6085  */
6086 void __update_idle_core(struct rq *rq)
6087 {
6088         int core = cpu_of(rq);
6089         int cpu;
6090 
6091         rcu_read_lock();
6092         if (test_idle_cores(core, true))
6093                 goto unlock;
6094 
6095         for_each_cpu(cpu, cpu_smt_mask(core)) {
6096                 if (cpu == core)
6097                         continue;
6098 
6099                 if (!available_idle_cpu(cpu))
6100                         goto unlock;
6101         }
6102 
6103         set_idle_cores(core, 1);
6104 unlock:
6105         rcu_read_unlock();
6106 }
6107 
6108 /*
6109  * Scan the entire LLC domain for idle cores; this dynamically switches off if
6110  * there are no idle cores left in the system; tracked through
6111  * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6112  */
6113 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6114 {
6115         struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6116         int core, cpu;
6117 
6118         if (!static_branch_likely(&sched_smt_present))
6119                 return -1;
6120 
6121         if (!test_idle_cores(target, false))
6122                 return -1;
6123 
6124         cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
6125 
6126         for_each_cpu_wrap(core, cpus, target) {
6127                 bool idle = true;
6128 
6129                 for_each_cpu(cpu, cpu_smt_mask(core)) {
6130                         __cpumask_clear_cpu(cpu, cpus);
6131                         if (!available_idle_cpu(cpu))
6132                                 idle = false;
6133                 }
6134 
6135                 if (idle)
6136                         return core;
6137         }
6138 
6139         /*
6140          * Failed to find an idle core; stop looking for one.
6141          */
6142         set_idle_cores(target, 0);
6143 
6144         return -1;
6145 }
6146 
6147 /*
6148  * Scan the local SMT mask for idle CPUs.
6149  */
6150 static int select_idle_smt(struct task_struct *p, int target)
6151 {
6152         int cpu;
6153 
6154         if (!static_branch_likely(&sched_smt_present))
6155                 return -1;
6156 
6157         for_each_cpu(cpu, cpu_smt_mask(target)) {
6158                 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6159                         continue;
6160                 if (available_idle_cpu(cpu))
6161                         return cpu;
6162         }
6163 
6164         return -1;
6165 }
6166 
6167 #else /* CONFIG_SCHED_SMT */
6168 
6169 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6170 {
6171         return -1;
6172 }
6173 
6174 static inline int select_idle_smt(struct task_struct *p, int target)
6175 {
6176         return -1;
6177 }
6178 
6179 #endif /* CONFIG_SCHED_SMT */
6180 
6181 /*
6182  * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6183  * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6184  * average idle time for this rq (as found in rq->avg_idle).
6185  */
6186 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
6187 {
6188         struct sched_domain *this_sd;
6189         u64 avg_cost, avg_idle;
6190         u64 time, cost;
6191         s64 delta;
6192         int cpu, nr = INT_MAX;
6193 
6194         this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6195         if (!this_sd)
6196                 return -1;
6197 
6198         /*
6199          * Due to large variance we need a large fuzz factor; hackbench in
6200          * particularly is sensitive here.
6201          */
6202         avg_idle = this_rq()->avg_idle / 512;
6203         avg_cost = this_sd->avg_scan_cost + 1;
6204 
6205         if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
6206                 return -1;
6207 
6208         if (sched_feat(SIS_PROP)) {
6209                 u64 span_avg = sd->span_weight * avg_idle;
6210                 if (span_avg > 4*avg_cost)
6211                         nr = div_u64(span_avg, avg_cost);
6212                 else
6213                         nr = 4;
6214         }
6215 
6216         time = local_clock();
6217 
6218         for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
6219                 if (!--nr)
6220                         return -1;
6221                 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6222                         continue;
6223                 if (available_idle_cpu(cpu))
6224                         break;
6225         }
6226 
6227         time = local_clock() - time;
6228         cost = this_sd->avg_scan_cost;
6229         delta = (s64)(time - cost) / 8;
6230         this_sd->avg_scan_cost += delta;
6231 
6232         return cpu;
6233 }
6234 
6235 /*
6236  * Try and locate an idle core/thread in the LLC cache domain.
6237  */
6238 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6239 {
6240         struct sched_domain *sd;
6241         int i, recent_used_cpu;
6242 
6243         if (available_idle_cpu(target))
6244                 return target;
6245 
6246         /*
6247          * If the previous CPU is cache affine and idle, don't be stupid:
6248          */
6249         if (prev != target && cpus_share_cache(prev, target) && available_idle_cpu(prev))
6250                 return prev;
6251 
6252         /* Check a recently used CPU as a potential idle candidate: */
6253         recent_used_cpu = p->recent_used_cpu;
6254         if (recent_used_cpu != prev &&
6255             recent_used_cpu != target &&
6256             cpus_share_cache(recent_used_cpu, target) &&
6257             available_idle_cpu(recent_used_cpu) &&
6258             cpumask_test_cpu(p->recent_used_cpu, &p->cpus_allowed)) {
6259                 /*
6260                  * Replace recent_used_cpu with prev as it is a potential
6261                  * candidate for the next wake:
6262                  */
6263                 p->recent_used_cpu = prev;
6264                 return recent_used_cpu;
6265         }
6266 
6267         sd = rcu_dereference(per_cpu(sd_llc, target));
6268         if (!sd)
6269                 return target;
6270 
6271         i = select_idle_core(p, sd, target);
6272         if ((unsigned)i < nr_cpumask_bits)
6273                 return i;
6274 
6275         i = select_idle_cpu(p, sd, target);
6276         if ((unsigned)i < nr_cpumask_bits)
6277                 return i;
6278 
6279         i = select_idle_smt(p, target);
6280         if ((unsigned)i < nr_cpumask_bits)
6281                 return i;
6282 
6283         return target;
6284 }
6285 
6286 /**
6287  * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
6288  * @cpu: the CPU to get the utilization of
6289  *
6290  * The unit of the return value must be the one of capacity so we can compare
6291  * the utilization with the capacity of the CPU that is available for CFS task
6292  * (ie cpu_capacity).
6293  *
6294  * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6295  * recent utilization of currently non-runnable tasks on a CPU. It represents
6296  * the amount of utilization of a CPU in the range [0..capacity_orig] where
6297  * capacity_orig is the cpu_capacity available at the highest frequency
6298  * (arch_scale_freq_capacity()).
6299  * The utilization of a CPU converges towards a sum equal to or less than the
6300  * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6301  * the running time on this CPU scaled by capacity_curr.
6302  *
6303  * The estimated utilization of a CPU is defined to be the maximum between its
6304  * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6305  * currently RUNNABLE on that CPU.
6306  * This allows to properly represent the expected utilization of a CPU which
6307  * has just got a big task running since a long sleep period. At the same time
6308  * however it preserves the benefits of the "blocked utilization" in
6309  * describing the potential for other tasks waking up on the same CPU.
6310  *
6311  * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6312  * higher than capacity_orig because of unfortunate rounding in
6313  * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6314  * the average stabilizes with the new running time. We need to check that the
6315  * utilization stays within the range of [0..capacity_orig] and cap it if
6316  * necessary. Without utilization capping, a group could be seen as overloaded
6317  * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6318  * available capacity. We allow utilization to overshoot capacity_curr (but not
6319  * capacity_orig) as it useful for predicting the capacity required after task
6320  * migrations (scheduler-driven DVFS).
6321  *
6322  * Return: the (estimated) utilization for the specified CPU
6323  */
6324 static inline unsigned long cpu_util(int cpu)
6325 {
6326         struct cfs_rq *cfs_rq;
6327         unsigned int util;
6328 
6329         cfs_rq = &cpu_rq(cpu)->cfs;
6330         util = READ_ONCE(cfs_rq->avg.util_avg);
6331 
6332         if (sched_feat(UTIL_EST))
6333                 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6334 
6335         return min_t(unsigned long, util, capacity_orig_of(cpu));
6336 }
6337 
6338 /*
6339  * cpu_util_without: compute cpu utilization without any contributions from *p
6340  * @cpu: the CPU which utilization is requested
6341  * @p: the task which utilization should be discounted
6342  *
6343  * The utilization of a CPU is defined by the utilization of tasks currently
6344  * enqueued on that CPU as well as tasks which are currently sleeping after an
6345  * execution on that CPU.
6346  *
6347  * This method returns the utilization of the specified CPU by discounting the
6348  * utilization of the specified task, whenever the task is currently
6349  * contributing to the CPU utilization.
6350  */
6351 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
6352 {
6353         struct cfs_rq *cfs_rq;
6354         unsigned int util;
6355 
6356         /* Task has no contribution or is new */
6357         if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6358                 return cpu_util(cpu);
6359 
6360         cfs_rq = &cpu_rq(cpu)->cfs;
6361         util = READ_ONCE(cfs_rq->avg.util_avg);
6362 
6363         /* Discount task's util from CPU's util */
6364         lsub_positive(&util, task_util(p));
6365 
6366         /*
6367          * Covered cases:
6368          *
6369          * a) if *p is the only task sleeping on this CPU, then:
6370          *      cpu_util (== task_util) > util_est (== 0)
6371          *    and thus we return:
6372          *      cpu_util_without = (cpu_util - task_util) = 0
6373          *
6374          * b) if other tasks are SLEEPING on this CPU, which is now exiting
6375          *    IDLE, then:
6376          *      cpu_util >= task_util
6377          *      cpu_util > util_est (== 0)
6378          *    and thus we discount *p's blocked utilization to return:
6379          *      cpu_util_without = (cpu_util - task_util) >= 0
6380          *
6381          * c) if other tasks are RUNNABLE on that CPU and
6382          *      util_est > cpu_util
6383          *    then we use util_est since it returns a more restrictive
6384          *    estimation of the spare capacity on that CPU, by just
6385          *    considering the expected utilization of tasks already
6386          *    runnable on that CPU.
6387          *
6388          * Cases a) and b) are covered by the above code, while case c) is
6389          * covered by the following code when estimated utilization is
6390          * enabled.
6391          */
6392         if (sched_feat(UTIL_EST)) {
6393                 unsigned int estimated =
6394                         READ_ONCE(cfs_rq->avg.util_est.enqueued);
6395 
6396                 /*
6397                  * Despite the following checks we still have a small window
6398                  * for a possible race, when an execl's select_task_rq_fair()
6399                  * races with LB's detach_task():
6400                  *
6401                  *   detach_task()
6402                  *     p->on_rq = TASK_ON_RQ_MIGRATING;
6403                  *     ---------------------------------- A
6404                  *     deactivate_task()                   \
6405                  *       dequeue_task()                     + RaceTime
6406                  *         util_est_dequeue()              /
6407                  *     ---------------------------------- B
6408                  *
6409                  * The additional check on "current == p" it's required to
6410                  * properly fix the execl regression and it helps in further
6411                  * reducing the chances for the above race.
6412                  */
6413                 if (unlikely(task_on_rq_queued(p) || current == p))
6414                         lsub_positive(&estimated, _task_util_est(p));
6415 
6416                 util = max(util, estimated);
6417         }
6418 
6419         /*
6420          * Utilization (estimated) can exceed the CPU capacity, thus let's
6421          * clamp to the maximum CPU capacity to ensure consistency with
6422          * the cpu_util call.
6423          */
6424         return min_t(unsigned long, util, capacity_orig_of(cpu));
6425 }
6426 
6427 /*
6428  * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6429  * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6430  *
6431  * In that case WAKE_AFFINE doesn't make sense and we'll let
6432  * BALANCE_WAKE sort things out.
6433  */
6434 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6435 {
6436         long min_cap, max_cap;
6437 
6438         if (!static_branch_unlikely(&sched_asym_cpucapacity))
6439                 return 0;
6440 
6441         min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6442         max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
6443 
6444         /* Minimum capacity is close to max, no need to abort wake_affine */
6445         if (max_cap - min_cap < max_cap >> 3)
6446                 return 0;
6447 
6448         /* Bring task utilization in sync with prev_cpu */
6449         sync_entity_load_avg(&p->se);
6450 
6451         return !task_fits_capacity(p, min_cap);
6452 }
6453 
6454 /*
6455  * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
6456  * to @dst_cpu.
6457  */
6458 static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
6459 {
6460         struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6461         unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg);
6462 
6463         /*
6464          * If @p migrates from @cpu to another, remove its contribution. Or,
6465          * if @p migrates from another CPU to @cpu, add its contribution. In
6466          * the other cases, @cpu is not impacted by the migration, so the
6467          * util_avg should already be correct.
6468          */
6469         if (task_cpu(p) == cpu && dst_cpu != cpu)
6470                 sub_positive(&util, task_util(p));
6471         else if (task_cpu(p) != cpu && dst_cpu == cpu)
6472                 util += task_util(p);
6473 
6474         if (sched_feat(UTIL_EST)) {
6475                 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
6476 
6477                 /*
6478                  * During wake-up, the task isn't enqueued yet and doesn't
6479                  * appear in the cfs_rq->avg.util_est.enqueued of any rq,
6480                  * so just add it (if needed) to "simulate" what will be
6481                  * cpu_util() after the task has been enqueued.
6482                  */
6483                 if (dst_cpu == cpu)
6484                         util_est += _task_util_est(p);
6485 
6486                 util = max(util, util_est);
6487         }
6488 
6489         return min(util, capacity_orig_of(cpu));
6490 }
6491 
6492 /*
6493  * compute_energy(): Estimates the energy that would be consumed if @p was
6494  * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
6495  * landscape of the * CPUs after the task migration, and uses the Energy Model
6496  * to compute what would be the energy if we decided to actually migrate that
6497  * task.
6498  */
6499 static long
6500 compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd)
6501 {
6502         long util, max_util, sum_util, energy = 0;
6503         int cpu;
6504 
6505         for (; pd; pd = pd->next) {
6506                 max_util = sum_util = 0;
6507                 /*
6508                  * The capacity state of CPUs of the current rd can be driven by
6509                  * CPUs of another rd if they belong to the same performance
6510                  * domain. So, account for the utilization of these CPUs too
6511                  * by masking pd with cpu_online_mask instead of the rd span.
6512                  *
6513                  * If an entire performance domain is outside of the current rd,
6514                  * it will not appear in its pd list and will not be accounted
6515                  * by compute_energy().
6516                  */
6517                 for_each_cpu_and(cpu, perf_domain_span(pd), cpu_online_mask) {
6518                         util = cpu_util_next(cpu, p, dst_cpu);
6519                         util = schedutil_energy_util(cpu, util);
6520                         max_util = max(util, max_util);
6521                         sum_util += util;
6522                 }
6523 
6524                 energy += em_pd_energy(pd->em_pd, max_util, sum_util);