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

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

Version: ~ [ linux-5.11-rc3 ] ~ [ linux-5.10.7 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.89 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.167 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.215 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.251 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.251 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.140 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.85 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 // SPDX-License-Identifier: GPL-2.0
  2 /*
  3  * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
  4  *
  5  *  Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
  6  *
  7  *  Interactivity improvements by Mike Galbraith
  8  *  (C) 2007 Mike Galbraith <efault@gmx.de>
  9  *
 10  *  Various enhancements by Dmitry Adamushko.
 11  *  (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
 12  *
 13  *  Group scheduling enhancements by Srivatsa Vaddagiri
 14  *  Copyright IBM Corporation, 2007
 15  *  Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
 16  *
 17  *  Scaled math optimizations by Thomas Gleixner
 18  *  Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
 19  *
 20  *  Adaptive scheduling granularity, math enhancements by Peter Zijlstra
 21  *  Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
 22  */
 23 #include "sched.h"
 24 
 25 #include <trace/events/sched.h>
 26 
 27 /*
 28  * Targeted preemption latency for CPU-bound tasks:
 29  *
 30  * NOTE: this latency value is not the same as the concept of
 31  * 'timeslice length' - timeslices in CFS are of variable length
 32  * and have no persistent notion like in traditional, time-slice
 33  * based scheduling concepts.
 34  *
 35  * (to see the precise effective timeslice length of your workload,
 36  *  run vmstat and monitor the context-switches (cs) field)
 37  *
 38  * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
 39  */
 40 unsigned int sysctl_sched_latency                       = 6000000ULL;
 41 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 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 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 #endif
 98 
 99 #ifdef CONFIG_CFS_BANDWIDTH
100 /*
101  * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
102  * each time a cfs_rq requests quota.
103  *
104  * Note: in the case that the slice exceeds the runtime remaining (either due
105  * to consumption or the quota being specified to be smaller than the slice)
106  * we will always only issue the remaining available time.
107  *
108  * (default: 5 msec, units: microseconds)
109  */
110 unsigned int sysctl_sched_cfs_bandwidth_slice           = 5000UL;
111 #endif
112 
113 /*
114  * The margin used when comparing utilization with CPU capacity:
115  * util * margin < capacity * 1024
116  *
117  * (default: ~20%)
118  */
119 unsigned int capacity_margin                            = 1280;
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 
252 /* cpu runqueue to which this cfs_rq is attached */
253 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
254 {
255         return cfs_rq->rq;
256 }
257 
258 /* An entity is a task if it doesn't "own" a runqueue */
259 #define entity_is_task(se)      (!se->my_q)
260 
261 static inline struct task_struct *task_of(struct sched_entity *se)
262 {
263         SCHED_WARN_ON(!entity_is_task(se));
264         return container_of(se, struct task_struct, se);
265 }
266 
267 /* Walk up scheduling entities hierarchy */
268 #define for_each_sched_entity(se) \
269                 for (; se; se = se->parent)
270 
271 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
272 {
273         return p->se.cfs_rq;
274 }
275 
276 /* runqueue on which this entity is (to be) queued */
277 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
278 {
279         return se->cfs_rq;
280 }
281 
282 /* runqueue "owned" by this group */
283 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
284 {
285         return grp->my_q;
286 }
287 
288 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
289 {
290         if (!cfs_rq->on_list) {
291                 struct rq *rq = rq_of(cfs_rq);
292                 int cpu = cpu_of(rq);
293                 /*
294                  * Ensure we either appear before our parent (if already
295                  * enqueued) or force our parent to appear after us when it is
296                  * enqueued. The fact that we always enqueue bottom-up
297                  * reduces this to two cases and a special case for the root
298                  * cfs_rq. Furthermore, it also means that we will always reset
299                  * tmp_alone_branch either when the branch is connected
300                  * to a tree or when we reach the beg of the tree
301                  */
302                 if (cfs_rq->tg->parent &&
303                     cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
304                         /*
305                          * If parent is already on the list, we add the child
306                          * just before. Thanks to circular linked property of
307                          * the list, this means to put the child at the tail
308                          * of the list that starts by parent.
309                          */
310                         list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
311                                 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
312                         /*
313                          * The branch is now connected to its tree so we can
314                          * reset tmp_alone_branch to the beginning of the
315                          * list.
316                          */
317                         rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
318                 } else if (!cfs_rq->tg->parent) {
319                         /*
320                          * cfs rq without parent should be put
321                          * at the tail of the list.
322                          */
323                         list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
324                                 &rq->leaf_cfs_rq_list);
325                         /*
326                          * We have reach the beg of a tree so we can reset
327                          * tmp_alone_branch to the beginning of the list.
328                          */
329                         rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
330                 } else {
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 beg of the branch
335                          * where we will add parent.
336                          */
337                         list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
338                                 rq->tmp_alone_branch);
339                         /*
340                          * update tmp_alone_branch to points to the new beg
341                          * of the branch
342                          */
343                         rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
344                 }
345 
346                 cfs_rq->on_list = 1;
347         }
348 }
349 
350 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
351 {
352         if (cfs_rq->on_list) {
353                 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
354                 cfs_rq->on_list = 0;
355         }
356 }
357 
358 /* Iterate thr' all leaf cfs_rq's on a runqueue */
359 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos)                      \
360         list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list,    \
361                                  leaf_cfs_rq_list)
362 
363 /* Do the two (enqueued) entities belong to the same group ? */
364 static inline struct cfs_rq *
365 is_same_group(struct sched_entity *se, struct sched_entity *pse)
366 {
367         if (se->cfs_rq == pse->cfs_rq)
368                 return se->cfs_rq;
369 
370         return NULL;
371 }
372 
373 static inline struct sched_entity *parent_entity(struct sched_entity *se)
374 {
375         return se->parent;
376 }
377 
378 static void
379 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
380 {
381         int se_depth, pse_depth;
382 
383         /*
384          * preemption test can be made between sibling entities who are in the
385          * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
386          * both tasks until we find their ancestors who are siblings of common
387          * parent.
388          */
389 
390         /* First walk up until both entities are at same depth */
391         se_depth = (*se)->depth;
392         pse_depth = (*pse)->depth;
393 
394         while (se_depth > pse_depth) {
395                 se_depth--;
396                 *se = parent_entity(*se);
397         }
398 
399         while (pse_depth > se_depth) {
400                 pse_depth--;
401                 *pse = parent_entity(*pse);
402         }
403 
404         while (!is_same_group(*se, *pse)) {
405                 *se = parent_entity(*se);
406                 *pse = parent_entity(*pse);
407         }
408 }
409 
410 #else   /* !CONFIG_FAIR_GROUP_SCHED */
411 
412 static inline struct task_struct *task_of(struct sched_entity *se)
413 {
414         return container_of(se, struct task_struct, se);
415 }
416 
417 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
418 {
419         return container_of(cfs_rq, struct rq, cfs);
420 }
421 
422 #define entity_is_task(se)      1
423 
424 #define for_each_sched_entity(se) \
425                 for (; se; se = NULL)
426 
427 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
428 {
429         return &task_rq(p)->cfs;
430 }
431 
432 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
433 {
434         struct task_struct *p = task_of(se);
435         struct rq *rq = task_rq(p);
436 
437         return &rq->cfs;
438 }
439 
440 /* runqueue "owned" by this group */
441 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
442 {
443         return NULL;
444 }
445 
446 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
447 {
448 }
449 
450 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
451 {
452 }
453 
454 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos)      \
455                 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
456 
457 static inline struct sched_entity *parent_entity(struct sched_entity *se)
458 {
459         return NULL;
460 }
461 
462 static inline void
463 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
464 {
465 }
466 
467 #endif  /* CONFIG_FAIR_GROUP_SCHED */
468 
469 static __always_inline
470 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
471 
472 /**************************************************************
473  * Scheduling class tree data structure manipulation methods:
474  */
475 
476 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
477 {
478         s64 delta = (s64)(vruntime - max_vruntime);
479         if (delta > 0)
480                 max_vruntime = vruntime;
481 
482         return max_vruntime;
483 }
484 
485 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
486 {
487         s64 delta = (s64)(vruntime - min_vruntime);
488         if (delta < 0)
489                 min_vruntime = vruntime;
490 
491         return min_vruntime;
492 }
493 
494 static inline int entity_before(struct sched_entity *a,
495                                 struct sched_entity *b)
496 {
497         return (s64)(a->vruntime - b->vruntime) < 0;
498 }
499 
500 static void update_min_vruntime(struct cfs_rq *cfs_rq)
501 {
502         struct sched_entity *curr = cfs_rq->curr;
503         struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
504 
505         u64 vruntime = cfs_rq->min_vruntime;
506 
507         if (curr) {
508                 if (curr->on_rq)
509                         vruntime = curr->vruntime;
510                 else
511                         curr = NULL;
512         }
513 
514         if (leftmost) { /* non-empty tree */
515                 struct sched_entity *se;
516                 se = rb_entry(leftmost, struct sched_entity, run_node);
517 
518                 if (!curr)
519                         vruntime = se->vruntime;
520                 else
521                         vruntime = min_vruntime(vruntime, se->vruntime);
522         }
523 
524         /* ensure we never gain time by being placed backwards. */
525         cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
526 #ifndef CONFIG_64BIT
527         smp_wmb();
528         cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
529 #endif
530 }
531 
532 /*
533  * Enqueue an entity into the rb-tree:
534  */
535 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
536 {
537         struct rb_node **link = &cfs_rq->tasks_timeline.rb_root.rb_node;
538         struct rb_node *parent = NULL;
539         struct sched_entity *entry;
540         bool leftmost = true;
541 
542         /*
543          * Find the right place in the rbtree:
544          */
545         while (*link) {
546                 parent = *link;
547                 entry = rb_entry(parent, struct sched_entity, run_node);
548                 /*
549                  * We dont care about collisions. Nodes with
550                  * the same key stay together.
551                  */
552                 if (entity_before(se, entry)) {
553                         link = &parent->rb_left;
554                 } else {
555                         link = &parent->rb_right;
556                         leftmost = false;
557                 }
558         }
559 
560         rb_link_node(&se->run_node, parent, link);
561         rb_insert_color_cached(&se->run_node,
562                                &cfs_rq->tasks_timeline, leftmost);
563 }
564 
565 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
566 {
567         rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
568 }
569 
570 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
571 {
572         struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
573 
574         if (!left)
575                 return NULL;
576 
577         return rb_entry(left, struct sched_entity, run_node);
578 }
579 
580 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
581 {
582         struct rb_node *next = rb_next(&se->run_node);
583 
584         if (!next)
585                 return NULL;
586 
587         return rb_entry(next, struct sched_entity, run_node);
588 }
589 
590 #ifdef CONFIG_SCHED_DEBUG
591 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
592 {
593         struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
594 
595         if (!last)
596                 return NULL;
597 
598         return rb_entry(last, struct sched_entity, run_node);
599 }
600 
601 /**************************************************************
602  * Scheduling class statistics methods:
603  */
604 
605 int sched_proc_update_handler(struct ctl_table *table, int write,
606                 void __user *buffer, size_t *lenp,
607                 loff_t *ppos)
608 {
609         int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
610         unsigned int factor = get_update_sysctl_factor();
611 
612         if (ret || !write)
613                 return ret;
614 
615         sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
616                                         sysctl_sched_min_granularity);
617 
618 #define WRT_SYSCTL(name) \
619         (normalized_sysctl_##name = sysctl_##name / (factor))
620         WRT_SYSCTL(sched_min_granularity);
621         WRT_SYSCTL(sched_latency);
622         WRT_SYSCTL(sched_wakeup_granularity);
623 #undef WRT_SYSCTL
624 
625         return 0;
626 }
627 #endif
628 
629 /*
630  * delta /= w
631  */
632 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
633 {
634         if (unlikely(se->load.weight != NICE_0_LOAD))
635                 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
636 
637         return delta;
638 }
639 
640 /*
641  * The idea is to set a period in which each task runs once.
642  *
643  * When there are too many tasks (sched_nr_latency) we have to stretch
644  * this period because otherwise the slices get too small.
645  *
646  * p = (nr <= nl) ? l : l*nr/nl
647  */
648 static u64 __sched_period(unsigned long nr_running)
649 {
650         if (unlikely(nr_running > sched_nr_latency))
651                 return nr_running * sysctl_sched_min_granularity;
652         else
653                 return sysctl_sched_latency;
654 }
655 
656 /*
657  * We calculate the wall-time slice from the period by taking a part
658  * proportional to the weight.
659  *
660  * s = p*P[w/rw]
661  */
662 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
663 {
664         u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
665 
666         for_each_sched_entity(se) {
667                 struct load_weight *load;
668                 struct load_weight lw;
669 
670                 cfs_rq = cfs_rq_of(se);
671                 load = &cfs_rq->load;
672 
673                 if (unlikely(!se->on_rq)) {
674                         lw = cfs_rq->load;
675 
676                         update_load_add(&lw, se->load.weight);
677                         load = &lw;
678                 }
679                 slice = __calc_delta(slice, se->load.weight, load);
680         }
681         return slice;
682 }
683 
684 /*
685  * We calculate the vruntime slice of a to-be-inserted task.
686  *
687  * vs = s/w
688  */
689 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
690 {
691         return calc_delta_fair(sched_slice(cfs_rq, se), se);
692 }
693 
694 #ifdef CONFIG_SMP
695 
696 #include "sched-pelt.h"
697 
698 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
699 static unsigned long task_h_load(struct task_struct *p);
700 
701 /* Give new sched_entity start runnable values to heavy its load in infant time */
702 void init_entity_runnable_average(struct sched_entity *se)
703 {
704         struct sched_avg *sa = &se->avg;
705 
706         memset(sa, 0, sizeof(*sa));
707 
708         /*
709          * Tasks are intialized with full load to be seen as heavy tasks until
710          * they get a chance to stabilize to their real load level.
711          * Group entities are intialized with zero load to reflect the fact that
712          * nothing has been attached to the task group yet.
713          */
714         if (entity_is_task(se))
715                 sa->runnable_load_avg = sa->load_avg = scale_load_down(se->load.weight);
716 
717         se->runnable_weight = se->load.weight;
718 
719         /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
720 }
721 
722 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
723 static void attach_entity_cfs_rq(struct sched_entity *se);
724 
725 /*
726  * With new tasks being created, their initial util_avgs are extrapolated
727  * based on the cfs_rq's current util_avg:
728  *
729  *   util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
730  *
731  * However, in many cases, the above util_avg does not give a desired
732  * value. Moreover, the sum of the util_avgs may be divergent, such
733  * as when the series is a harmonic series.
734  *
735  * To solve this problem, we also cap the util_avg of successive tasks to
736  * only 1/2 of the left utilization budget:
737  *
738  *   util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
739  *
740  * where n denotes the nth task.
741  *
742  * For example, a simplest series from the beginning would be like:
743  *
744  *  task  util_avg: 512, 256, 128,  64,  32,   16,    8, ...
745  * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
746  *
747  * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
748  * if util_avg > util_avg_cap.
749  */
750 void post_init_entity_util_avg(struct sched_entity *se)
751 {
752         struct cfs_rq *cfs_rq = cfs_rq_of(se);
753         struct sched_avg *sa = &se->avg;
754         long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
755 
756         if (cap > 0) {
757                 if (cfs_rq->avg.util_avg != 0) {
758                         sa->util_avg  = cfs_rq->avg.util_avg * se->load.weight;
759                         sa->util_avg /= (cfs_rq->avg.load_avg + 1);
760 
761                         if (sa->util_avg > cap)
762                                 sa->util_avg = cap;
763                 } else {
764                         sa->util_avg = cap;
765                 }
766         }
767 
768         if (entity_is_task(se)) {
769                 struct task_struct *p = task_of(se);
770                 if (p->sched_class != &fair_sched_class) {
771                         /*
772                          * For !fair tasks do:
773                          *
774                         update_cfs_rq_load_avg(now, cfs_rq);
775                         attach_entity_load_avg(cfs_rq, se, 0);
776                         switched_from_fair(rq, p);
777                          *
778                          * such that the next switched_to_fair() has the
779                          * expected state.
780                          */
781                         se->avg.last_update_time = cfs_rq_clock_task(cfs_rq);
782                         return;
783                 }
784         }
785 
786         attach_entity_cfs_rq(se);
787 }
788 
789 #else /* !CONFIG_SMP */
790 void init_entity_runnable_average(struct sched_entity *se)
791 {
792 }
793 void post_init_entity_util_avg(struct sched_entity *se)
794 {
795 }
796 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
797 {
798 }
799 #endif /* CONFIG_SMP */
800 
801 /*
802  * Update the current task's runtime statistics.
803  */
804 static void update_curr(struct cfs_rq *cfs_rq)
805 {
806         struct sched_entity *curr = cfs_rq->curr;
807         u64 now = rq_clock_task(rq_of(cfs_rq));
808         u64 delta_exec;
809 
810         if (unlikely(!curr))
811                 return;
812 
813         delta_exec = now - curr->exec_start;
814         if (unlikely((s64)delta_exec <= 0))
815                 return;
816 
817         curr->exec_start = now;
818 
819         schedstat_set(curr->statistics.exec_max,
820                       max(delta_exec, curr->statistics.exec_max));
821 
822         curr->sum_exec_runtime += delta_exec;
823         schedstat_add(cfs_rq->exec_clock, delta_exec);
824 
825         curr->vruntime += calc_delta_fair(delta_exec, curr);
826         update_min_vruntime(cfs_rq);
827 
828         if (entity_is_task(curr)) {
829                 struct task_struct *curtask = task_of(curr);
830 
831                 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
832                 cgroup_account_cputime(curtask, delta_exec);
833                 account_group_exec_runtime(curtask, delta_exec);
834         }
835 
836         account_cfs_rq_runtime(cfs_rq, delta_exec);
837 }
838 
839 static void update_curr_fair(struct rq *rq)
840 {
841         update_curr(cfs_rq_of(&rq->curr->se));
842 }
843 
844 static inline void
845 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
846 {
847         u64 wait_start, prev_wait_start;
848 
849         if (!schedstat_enabled())
850                 return;
851 
852         wait_start = rq_clock(rq_of(cfs_rq));
853         prev_wait_start = schedstat_val(se->statistics.wait_start);
854 
855         if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
856             likely(wait_start > prev_wait_start))
857                 wait_start -= prev_wait_start;
858 
859         __schedstat_set(se->statistics.wait_start, wait_start);
860 }
861 
862 static inline void
863 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
864 {
865         struct task_struct *p;
866         u64 delta;
867 
868         if (!schedstat_enabled())
869                 return;
870 
871         delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
872 
873         if (entity_is_task(se)) {
874                 p = task_of(se);
875                 if (task_on_rq_migrating(p)) {
876                         /*
877                          * Preserve migrating task's wait time so wait_start
878                          * time stamp can be adjusted to accumulate wait time
879                          * prior to migration.
880                          */
881                         __schedstat_set(se->statistics.wait_start, delta);
882                         return;
883                 }
884                 trace_sched_stat_wait(p, delta);
885         }
886 
887         __schedstat_set(se->statistics.wait_max,
888                       max(schedstat_val(se->statistics.wait_max), delta));
889         __schedstat_inc(se->statistics.wait_count);
890         __schedstat_add(se->statistics.wait_sum, delta);
891         __schedstat_set(se->statistics.wait_start, 0);
892 }
893 
894 static inline void
895 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
896 {
897         struct task_struct *tsk = NULL;
898         u64 sleep_start, block_start;
899 
900         if (!schedstat_enabled())
901                 return;
902 
903         sleep_start = schedstat_val(se->statistics.sleep_start);
904         block_start = schedstat_val(se->statistics.block_start);
905 
906         if (entity_is_task(se))
907                 tsk = task_of(se);
908 
909         if (sleep_start) {
910                 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
911 
912                 if ((s64)delta < 0)
913                         delta = 0;
914 
915                 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
916                         __schedstat_set(se->statistics.sleep_max, delta);
917 
918                 __schedstat_set(se->statistics.sleep_start, 0);
919                 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
920 
921                 if (tsk) {
922                         account_scheduler_latency(tsk, delta >> 10, 1);
923                         trace_sched_stat_sleep(tsk, delta);
924                 }
925         }
926         if (block_start) {
927                 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
928 
929                 if ((s64)delta < 0)
930                         delta = 0;
931 
932                 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
933                         __schedstat_set(se->statistics.block_max, delta);
934 
935                 __schedstat_set(se->statistics.block_start, 0);
936                 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
937 
938                 if (tsk) {
939                         if (tsk->in_iowait) {
940                                 __schedstat_add(se->statistics.iowait_sum, delta);
941                                 __schedstat_inc(se->statistics.iowait_count);
942                                 trace_sched_stat_iowait(tsk, delta);
943                         }
944 
945                         trace_sched_stat_blocked(tsk, delta);
946 
947                         /*
948                          * Blocking time is in units of nanosecs, so shift by
949                          * 20 to get a milliseconds-range estimation of the
950                          * amount of time that the task spent sleeping:
951                          */
952                         if (unlikely(prof_on == SLEEP_PROFILING)) {
953                                 profile_hits(SLEEP_PROFILING,
954                                                 (void *)get_wchan(tsk),
955                                                 delta >> 20);
956                         }
957                         account_scheduler_latency(tsk, delta >> 10, 0);
958                 }
959         }
960 }
961 
962 /*
963  * Task is being enqueued - update stats:
964  */
965 static inline void
966 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
967 {
968         if (!schedstat_enabled())
969                 return;
970 
971         /*
972          * Are we enqueueing a waiting task? (for current tasks
973          * a dequeue/enqueue event is a NOP)
974          */
975         if (se != cfs_rq->curr)
976                 update_stats_wait_start(cfs_rq, se);
977 
978         if (flags & ENQUEUE_WAKEUP)
979                 update_stats_enqueue_sleeper(cfs_rq, se);
980 }
981 
982 static inline void
983 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
984 {
985 
986         if (!schedstat_enabled())
987                 return;
988 
989         /*
990          * Mark the end of the wait period if dequeueing a
991          * waiting task:
992          */
993         if (se != cfs_rq->curr)
994                 update_stats_wait_end(cfs_rq, se);
995 
996         if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
997                 struct task_struct *tsk = task_of(se);
998 
999                 if (tsk->state & TASK_INTERRUPTIBLE)
1000                         __schedstat_set(se->statistics.sleep_start,
1001                                       rq_clock(rq_of(cfs_rq)));
1002                 if (tsk->state & TASK_UNINTERRUPTIBLE)
1003                         __schedstat_set(se->statistics.block_start,
1004                                       rq_clock(rq_of(cfs_rq)));
1005         }
1006 }
1007 
1008 /*
1009  * We are picking a new current task - update its stats:
1010  */
1011 static inline void
1012 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1013 {
1014         /*
1015          * We are starting a new run period:
1016          */
1017         se->exec_start = rq_clock_task(rq_of(cfs_rq));
1018 }
1019 
1020 /**************************************************
1021  * Scheduling class queueing methods:
1022  */
1023 
1024 #ifdef CONFIG_NUMA_BALANCING
1025 /*
1026  * Approximate time to scan a full NUMA task in ms. The task scan period is
1027  * calculated based on the tasks virtual memory size and
1028  * numa_balancing_scan_size.
1029  */
1030 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1031 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1032 
1033 /* Portion of address space to scan in MB */
1034 unsigned int sysctl_numa_balancing_scan_size = 256;
1035 
1036 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1037 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1038 
1039 struct numa_group {
1040         atomic_t refcount;
1041 
1042         spinlock_t lock; /* nr_tasks, tasks */
1043         int nr_tasks;
1044         pid_t gid;
1045         int active_nodes;
1046 
1047         struct rcu_head rcu;
1048         unsigned long total_faults;
1049         unsigned long max_faults_cpu;
1050         /*
1051          * Faults_cpu is used to decide whether memory should move
1052          * towards the CPU. As a consequence, these stats are weighted
1053          * more by CPU use than by memory faults.
1054          */
1055         unsigned long *faults_cpu;
1056         unsigned long faults[0];
1057 };
1058 
1059 static inline unsigned long group_faults_priv(struct numa_group *ng);
1060 static inline unsigned long group_faults_shared(struct numa_group *ng);
1061 
1062 static unsigned int task_nr_scan_windows(struct task_struct *p)
1063 {
1064         unsigned long rss = 0;
1065         unsigned long nr_scan_pages;
1066 
1067         /*
1068          * Calculations based on RSS as non-present and empty pages are skipped
1069          * by the PTE scanner and NUMA hinting faults should be trapped based
1070          * on resident pages
1071          */
1072         nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1073         rss = get_mm_rss(p->mm);
1074         if (!rss)
1075                 rss = nr_scan_pages;
1076 
1077         rss = round_up(rss, nr_scan_pages);
1078         return rss / nr_scan_pages;
1079 }
1080 
1081 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1082 #define MAX_SCAN_WINDOW 2560
1083 
1084 static unsigned int task_scan_min(struct task_struct *p)
1085 {
1086         unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1087         unsigned int scan, floor;
1088         unsigned int windows = 1;
1089 
1090         if (scan_size < MAX_SCAN_WINDOW)
1091                 windows = MAX_SCAN_WINDOW / scan_size;
1092         floor = 1000 / windows;
1093 
1094         scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1095         return max_t(unsigned int, floor, scan);
1096 }
1097 
1098 static unsigned int task_scan_start(struct task_struct *p)
1099 {
1100         unsigned long smin = task_scan_min(p);
1101         unsigned long period = smin;
1102 
1103         /* Scale the maximum scan period with the amount of shared memory. */
1104         if (p->numa_group) {
1105                 struct numa_group *ng = p->numa_group;
1106                 unsigned long shared = group_faults_shared(ng);
1107                 unsigned long private = group_faults_priv(ng);
1108 
1109                 period *= atomic_read(&ng->refcount);
1110                 period *= shared + 1;
1111                 period /= private + shared + 1;
1112         }
1113 
1114         return max(smin, period);
1115 }
1116 
1117 static unsigned int task_scan_max(struct task_struct *p)
1118 {
1119         unsigned long smin = task_scan_min(p);
1120         unsigned long smax;
1121 
1122         /* Watch for min being lower than max due to floor calculations */
1123         smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1124 
1125         /* Scale the maximum scan period with the amount of shared memory. */
1126         if (p->numa_group) {
1127                 struct numa_group *ng = p->numa_group;
1128                 unsigned long shared = group_faults_shared(ng);
1129                 unsigned long private = group_faults_priv(ng);
1130                 unsigned long period = smax;
1131 
1132                 period *= atomic_read(&ng->refcount);
1133                 period *= shared + 1;
1134                 period /= private + shared + 1;
1135 
1136                 smax = max(smax, period);
1137         }
1138 
1139         return max(smin, smax);
1140 }
1141 
1142 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1143 {
1144         rq->nr_numa_running += (p->numa_preferred_nid != -1);
1145         rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1146 }
1147 
1148 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1149 {
1150         rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1151         rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1152 }
1153 
1154 /* Shared or private faults. */
1155 #define NR_NUMA_HINT_FAULT_TYPES 2
1156 
1157 /* Memory and CPU locality */
1158 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1159 
1160 /* Averaged statistics, and temporary buffers. */
1161 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1162 
1163 pid_t task_numa_group_id(struct task_struct *p)
1164 {
1165         return p->numa_group ? p->numa_group->gid : 0;
1166 }
1167 
1168 /*
1169  * The averaged statistics, shared & private, memory & CPU,
1170  * occupy the first half of the array. The second half of the
1171  * array is for current counters, which are averaged into the
1172  * first set by task_numa_placement.
1173  */
1174 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1175 {
1176         return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1177 }
1178 
1179 static inline unsigned long task_faults(struct task_struct *p, int nid)
1180 {
1181         if (!p->numa_faults)
1182                 return 0;
1183 
1184         return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1185                 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1186 }
1187 
1188 static inline unsigned long group_faults(struct task_struct *p, int nid)
1189 {
1190         if (!p->numa_group)
1191                 return 0;
1192 
1193         return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1194                 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1195 }
1196 
1197 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1198 {
1199         return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1200                 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1201 }
1202 
1203 static inline unsigned long group_faults_priv(struct numa_group *ng)
1204 {
1205         unsigned long faults = 0;
1206         int node;
1207 
1208         for_each_online_node(node) {
1209                 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1210         }
1211 
1212         return faults;
1213 }
1214 
1215 static inline unsigned long group_faults_shared(struct numa_group *ng)
1216 {
1217         unsigned long faults = 0;
1218         int node;
1219 
1220         for_each_online_node(node) {
1221                 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1222         }
1223 
1224         return faults;
1225 }
1226 
1227 /*
1228  * A node triggering more than 1/3 as many NUMA faults as the maximum is
1229  * considered part of a numa group's pseudo-interleaving set. Migrations
1230  * between these nodes are slowed down, to allow things to settle down.
1231  */
1232 #define ACTIVE_NODE_FRACTION 3
1233 
1234 static bool numa_is_active_node(int nid, struct numa_group *ng)
1235 {
1236         return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1237 }
1238 
1239 /* Handle placement on systems where not all nodes are directly connected. */
1240 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1241                                         int maxdist, bool task)
1242 {
1243         unsigned long score = 0;
1244         int node;
1245 
1246         /*
1247          * All nodes are directly connected, and the same distance
1248          * from each other. No need for fancy placement algorithms.
1249          */
1250         if (sched_numa_topology_type == NUMA_DIRECT)
1251                 return 0;
1252 
1253         /*
1254          * This code is called for each node, introducing N^2 complexity,
1255          * which should be ok given the number of nodes rarely exceeds 8.
1256          */
1257         for_each_online_node(node) {
1258                 unsigned long faults;
1259                 int dist = node_distance(nid, node);
1260 
1261                 /*
1262                  * The furthest away nodes in the system are not interesting
1263                  * for placement; nid was already counted.
1264                  */
1265                 if (dist == sched_max_numa_distance || node == nid)
1266                         continue;
1267 
1268                 /*
1269                  * On systems with a backplane NUMA topology, compare groups
1270                  * of nodes, and move tasks towards the group with the most
1271                  * memory accesses. When comparing two nodes at distance
1272                  * "hoplimit", only nodes closer by than "hoplimit" are part
1273                  * of each group. Skip other nodes.
1274                  */
1275                 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1276                                         dist > maxdist)
1277                         continue;
1278 
1279                 /* Add up the faults from nearby nodes. */
1280                 if (task)
1281                         faults = task_faults(p, node);
1282                 else
1283                         faults = group_faults(p, node);
1284 
1285                 /*
1286                  * On systems with a glueless mesh NUMA topology, there are
1287                  * no fixed "groups of nodes". Instead, nodes that are not
1288                  * directly connected bounce traffic through intermediate
1289                  * nodes; a numa_group can occupy any set of nodes.
1290                  * The further away a node is, the less the faults count.
1291                  * This seems to result in good task placement.
1292                  */
1293                 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1294                         faults *= (sched_max_numa_distance - dist);
1295                         faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1296                 }
1297 
1298                 score += faults;
1299         }
1300 
1301         return score;
1302 }
1303 
1304 /*
1305  * These return the fraction of accesses done by a particular task, or
1306  * task group, on a particular numa node.  The group weight is given a
1307  * larger multiplier, in order to group tasks together that are almost
1308  * evenly spread out between numa nodes.
1309  */
1310 static inline unsigned long task_weight(struct task_struct *p, int nid,
1311                                         int dist)
1312 {
1313         unsigned long faults, total_faults;
1314 
1315         if (!p->numa_faults)
1316                 return 0;
1317 
1318         total_faults = p->total_numa_faults;
1319 
1320         if (!total_faults)
1321                 return 0;
1322 
1323         faults = task_faults(p, nid);
1324         faults += score_nearby_nodes(p, nid, dist, true);
1325 
1326         return 1000 * faults / total_faults;
1327 }
1328 
1329 static inline unsigned long group_weight(struct task_struct *p, int nid,
1330                                          int dist)
1331 {
1332         unsigned long faults, total_faults;
1333 
1334         if (!p->numa_group)
1335                 return 0;
1336 
1337         total_faults = p->numa_group->total_faults;
1338 
1339         if (!total_faults)
1340                 return 0;
1341 
1342         faults = group_faults(p, nid);
1343         faults += score_nearby_nodes(p, nid, dist, false);
1344 
1345         return 1000 * faults / total_faults;
1346 }
1347 
1348 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1349                                 int src_nid, int dst_cpu)
1350 {
1351         struct numa_group *ng = p->numa_group;
1352         int dst_nid = cpu_to_node(dst_cpu);
1353         int last_cpupid, this_cpupid;
1354 
1355         this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1356 
1357         /*
1358          * Multi-stage node selection is used in conjunction with a periodic
1359          * migration fault to build a temporal task<->page relation. By using
1360          * a two-stage filter we remove short/unlikely relations.
1361          *
1362          * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1363          * a task's usage of a particular page (n_p) per total usage of this
1364          * page (n_t) (in a given time-span) to a probability.
1365          *
1366          * Our periodic faults will sample this probability and getting the
1367          * same result twice in a row, given these samples are fully
1368          * independent, is then given by P(n)^2, provided our sample period
1369          * is sufficiently short compared to the usage pattern.
1370          *
1371          * This quadric squishes small probabilities, making it less likely we
1372          * act on an unlikely task<->page relation.
1373          */
1374         last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1375         if (!cpupid_pid_unset(last_cpupid) &&
1376                                 cpupid_to_nid(last_cpupid) != dst_nid)
1377                 return false;
1378 
1379         /* Always allow migrate on private faults */
1380         if (cpupid_match_pid(p, last_cpupid))
1381                 return true;
1382 
1383         /* A shared fault, but p->numa_group has not been set up yet. */
1384         if (!ng)
1385                 return true;
1386 
1387         /*
1388          * Destination node is much more heavily used than the source
1389          * node? Allow migration.
1390          */
1391         if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1392                                         ACTIVE_NODE_FRACTION)
1393                 return true;
1394 
1395         /*
1396          * Distribute memory according to CPU & memory use on each node,
1397          * with 3/4 hysteresis to avoid unnecessary memory migrations:
1398          *
1399          * faults_cpu(dst)   3   faults_cpu(src)
1400          * --------------- * - > ---------------
1401          * faults_mem(dst)   4   faults_mem(src)
1402          */
1403         return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1404                group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1405 }
1406 
1407 static unsigned long weighted_cpuload(struct rq *rq);
1408 static unsigned long source_load(int cpu, int type);
1409 static unsigned long target_load(int cpu, int type);
1410 static unsigned long capacity_of(int cpu);
1411 
1412 /* Cached statistics for all CPUs within a node */
1413 struct numa_stats {
1414         unsigned long nr_running;
1415         unsigned long load;
1416 
1417         /* Total compute capacity of CPUs on a node */
1418         unsigned long compute_capacity;
1419 
1420         /* Approximate capacity in terms of runnable tasks on a node */
1421         unsigned long task_capacity;
1422         int has_free_capacity;
1423 };
1424 
1425 /*
1426  * XXX borrowed from update_sg_lb_stats
1427  */
1428 static void update_numa_stats(struct numa_stats *ns, int nid)
1429 {
1430         int smt, cpu, cpus = 0;
1431         unsigned long capacity;
1432 
1433         memset(ns, 0, sizeof(*ns));
1434         for_each_cpu(cpu, cpumask_of_node(nid)) {
1435                 struct rq *rq = cpu_rq(cpu);
1436 
1437                 ns->nr_running += rq->nr_running;
1438                 ns->load += weighted_cpuload(rq);
1439                 ns->compute_capacity += capacity_of(cpu);
1440 
1441                 cpus++;
1442         }
1443 
1444         /*
1445          * If we raced with hotplug and there are no CPUs left in our mask
1446          * the @ns structure is NULL'ed and task_numa_compare() will
1447          * not find this node attractive.
1448          *
1449          * We'll either bail at !has_free_capacity, or we'll detect a huge
1450          * imbalance and bail there.
1451          */
1452         if (!cpus)
1453                 return;
1454 
1455         /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1456         smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1457         capacity = cpus / smt; /* cores */
1458 
1459         ns->task_capacity = min_t(unsigned, capacity,
1460                 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1461         ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1462 }
1463 
1464 struct task_numa_env {
1465         struct task_struct *p;
1466 
1467         int src_cpu, src_nid;
1468         int dst_cpu, dst_nid;
1469 
1470         struct numa_stats src_stats, dst_stats;
1471 
1472         int imbalance_pct;
1473         int dist;
1474 
1475         struct task_struct *best_task;
1476         long best_imp;
1477         int best_cpu;
1478 };
1479 
1480 static void task_numa_assign(struct task_numa_env *env,
1481                              struct task_struct *p, long imp)
1482 {
1483         if (env->best_task)
1484                 put_task_struct(env->best_task);
1485         if (p)
1486                 get_task_struct(p);
1487 
1488         env->best_task = p;
1489         env->best_imp = imp;
1490         env->best_cpu = env->dst_cpu;
1491 }
1492 
1493 static bool load_too_imbalanced(long src_load, long dst_load,
1494                                 struct task_numa_env *env)
1495 {
1496         long imb, old_imb;
1497         long orig_src_load, orig_dst_load;
1498         long src_capacity, dst_capacity;
1499 
1500         /*
1501          * The load is corrected for the CPU capacity available on each node.
1502          *
1503          * src_load        dst_load
1504          * ------------ vs ---------
1505          * src_capacity    dst_capacity
1506          */
1507         src_capacity = env->src_stats.compute_capacity;
1508         dst_capacity = env->dst_stats.compute_capacity;
1509 
1510         /* We care about the slope of the imbalance, not the direction. */
1511         if (dst_load < src_load)
1512                 swap(dst_load, src_load);
1513 
1514         /* Is the difference below the threshold? */
1515         imb = dst_load * src_capacity * 100 -
1516               src_load * dst_capacity * env->imbalance_pct;
1517         if (imb <= 0)
1518                 return false;
1519 
1520         /*
1521          * The imbalance is above the allowed threshold.
1522          * Compare it with the old imbalance.
1523          */
1524         orig_src_load = env->src_stats.load;
1525         orig_dst_load = env->dst_stats.load;
1526 
1527         if (orig_dst_load < orig_src_load)
1528                 swap(orig_dst_load, orig_src_load);
1529 
1530         old_imb = orig_dst_load * src_capacity * 100 -
1531                   orig_src_load * dst_capacity * env->imbalance_pct;
1532 
1533         /* Would this change make things worse? */
1534         return (imb > old_imb);
1535 }
1536 
1537 /*
1538  * This checks if the overall compute and NUMA accesses of the system would
1539  * be improved if the source tasks was migrated to the target dst_cpu taking
1540  * into account that it might be best if task running on the dst_cpu should
1541  * be exchanged with the source task
1542  */
1543 static void task_numa_compare(struct task_numa_env *env,
1544                               long taskimp, long groupimp)
1545 {
1546         struct rq *src_rq = cpu_rq(env->src_cpu);
1547         struct rq *dst_rq = cpu_rq(env->dst_cpu);
1548         struct task_struct *cur;
1549         long src_load, dst_load;
1550         long load;
1551         long imp = env->p->numa_group ? groupimp : taskimp;
1552         long moveimp = imp;
1553         int dist = env->dist;
1554 
1555         rcu_read_lock();
1556         cur = task_rcu_dereference(&dst_rq->curr);
1557         if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1558                 cur = NULL;
1559 
1560         /*
1561          * Because we have preemption enabled we can get migrated around and
1562          * end try selecting ourselves (current == env->p) as a swap candidate.
1563          */
1564         if (cur == env->p)
1565                 goto unlock;
1566 
1567         /*
1568          * "imp" is the fault differential for the source task between the
1569          * source and destination node. Calculate the total differential for
1570          * the source task and potential destination task. The more negative
1571          * the value is, the more rmeote accesses that would be expected to
1572          * be incurred if the tasks were swapped.
1573          */
1574         if (cur) {
1575                 /* Skip this swap candidate if cannot move to the source CPU: */
1576                 if (!cpumask_test_cpu(env->src_cpu, &cur->cpus_allowed))
1577                         goto unlock;
1578 
1579                 /*
1580                  * If dst and source tasks are in the same NUMA group, or not
1581                  * in any group then look only at task weights.
1582                  */
1583                 if (cur->numa_group == env->p->numa_group) {
1584                         imp = taskimp + task_weight(cur, env->src_nid, dist) -
1585                               task_weight(cur, env->dst_nid, dist);
1586                         /*
1587                          * Add some hysteresis to prevent swapping the
1588                          * tasks within a group over tiny differences.
1589                          */
1590                         if (cur->numa_group)
1591                                 imp -= imp/16;
1592                 } else {
1593                         /*
1594                          * Compare the group weights. If a task is all by
1595                          * itself (not part of a group), use the task weight
1596                          * instead.
1597                          */
1598                         if (cur->numa_group)
1599                                 imp += group_weight(cur, env->src_nid, dist) -
1600                                        group_weight(cur, env->dst_nid, dist);
1601                         else
1602                                 imp += task_weight(cur, env->src_nid, dist) -
1603                                        task_weight(cur, env->dst_nid, dist);
1604                 }
1605         }
1606 
1607         if (imp <= env->best_imp && moveimp <= env->best_imp)
1608                 goto unlock;
1609 
1610         if (!cur) {
1611                 /* Is there capacity at our destination? */
1612                 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1613                     !env->dst_stats.has_free_capacity)
1614                         goto unlock;
1615 
1616                 goto balance;
1617         }
1618 
1619         /* Balance doesn't matter much if we're running a task per CPU: */
1620         if (imp > env->best_imp && src_rq->nr_running == 1 &&
1621                         dst_rq->nr_running == 1)
1622                 goto assign;
1623 
1624         /*
1625          * In the overloaded case, try and keep the load balanced.
1626          */
1627 balance:
1628         load = task_h_load(env->p);
1629         dst_load = env->dst_stats.load + load;
1630         src_load = env->src_stats.load - load;
1631 
1632         if (moveimp > imp && moveimp > env->best_imp) {
1633                 /*
1634                  * If the improvement from just moving env->p direction is
1635                  * better than swapping tasks around, check if a move is
1636                  * possible. Store a slightly smaller score than moveimp,
1637                  * so an actually idle CPU will win.
1638                  */
1639                 if (!load_too_imbalanced(src_load, dst_load, env)) {
1640                         imp = moveimp - 1;
1641                         cur = NULL;
1642                         goto assign;
1643                 }
1644         }
1645 
1646         if (imp <= env->best_imp)
1647                 goto unlock;
1648 
1649         if (cur) {
1650                 load = task_h_load(cur);
1651                 dst_load -= load;
1652                 src_load += load;
1653         }
1654 
1655         if (load_too_imbalanced(src_load, dst_load, env))
1656                 goto unlock;
1657 
1658         /*
1659          * One idle CPU per node is evaluated for a task numa move.
1660          * Call select_idle_sibling to maybe find a better one.
1661          */
1662         if (!cur) {
1663                 /*
1664                  * select_idle_siblings() uses an per-CPU cpumask that
1665                  * can be used from IRQ context.
1666                  */
1667                 local_irq_disable();
1668                 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1669                                                    env->dst_cpu);
1670                 local_irq_enable();
1671         }
1672 
1673 assign:
1674         task_numa_assign(env, cur, imp);
1675 unlock:
1676         rcu_read_unlock();
1677 }
1678 
1679 static void task_numa_find_cpu(struct task_numa_env *env,
1680                                 long taskimp, long groupimp)
1681 {
1682         int cpu;
1683 
1684         for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1685                 /* Skip this CPU if the source task cannot migrate */
1686                 if (!cpumask_test_cpu(cpu, &env->p->cpus_allowed))
1687                         continue;
1688 
1689                 env->dst_cpu = cpu;
1690                 task_numa_compare(env, taskimp, groupimp);
1691         }
1692 }
1693 
1694 /* Only move tasks to a NUMA node less busy than the current node. */
1695 static bool numa_has_capacity(struct task_numa_env *env)
1696 {
1697         struct numa_stats *src = &env->src_stats;
1698         struct numa_stats *dst = &env->dst_stats;
1699 
1700         if (src->has_free_capacity && !dst->has_free_capacity)
1701                 return false;
1702 
1703         /*
1704          * Only consider a task move if the source has a higher load
1705          * than the destination, corrected for CPU capacity on each node.
1706          *
1707          *      src->load                dst->load
1708          * --------------------- vs ---------------------
1709          * src->compute_capacity    dst->compute_capacity
1710          */
1711         if (src->load * dst->compute_capacity * env->imbalance_pct >
1712 
1713             dst->load * src->compute_capacity * 100)
1714                 return true;
1715 
1716         return false;
1717 }
1718 
1719 static int task_numa_migrate(struct task_struct *p)
1720 {
1721         struct task_numa_env env = {
1722                 .p = p,
1723 
1724                 .src_cpu = task_cpu(p),
1725                 .src_nid = task_node(p),
1726 
1727                 .imbalance_pct = 112,
1728 
1729                 .best_task = NULL,
1730                 .best_imp = 0,
1731                 .best_cpu = -1,
1732         };
1733         struct sched_domain *sd;
1734         unsigned long taskweight, groupweight;
1735         int nid, ret, dist;
1736         long taskimp, groupimp;
1737 
1738         /*
1739          * Pick the lowest SD_NUMA domain, as that would have the smallest
1740          * imbalance and would be the first to start moving tasks about.
1741          *
1742          * And we want to avoid any moving of tasks about, as that would create
1743          * random movement of tasks -- counter the numa conditions we're trying
1744          * to satisfy here.
1745          */
1746         rcu_read_lock();
1747         sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1748         if (sd)
1749                 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1750         rcu_read_unlock();
1751 
1752         /*
1753          * Cpusets can break the scheduler domain tree into smaller
1754          * balance domains, some of which do not cross NUMA boundaries.
1755          * Tasks that are "trapped" in such domains cannot be migrated
1756          * elsewhere, so there is no point in (re)trying.
1757          */
1758         if (unlikely(!sd)) {
1759                 p->numa_preferred_nid = task_node(p);
1760                 return -EINVAL;
1761         }
1762 
1763         env.dst_nid = p->numa_preferred_nid;
1764         dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1765         taskweight = task_weight(p, env.src_nid, dist);
1766         groupweight = group_weight(p, env.src_nid, dist);
1767         update_numa_stats(&env.src_stats, env.src_nid);
1768         taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1769         groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1770         update_numa_stats(&env.dst_stats, env.dst_nid);
1771 
1772         /* Try to find a spot on the preferred nid. */
1773         if (numa_has_capacity(&env))
1774                 task_numa_find_cpu(&env, taskimp, groupimp);
1775 
1776         /*
1777          * Look at other nodes in these cases:
1778          * - there is no space available on the preferred_nid
1779          * - the task is part of a numa_group that is interleaved across
1780          *   multiple NUMA nodes; in order to better consolidate the group,
1781          *   we need to check other locations.
1782          */
1783         if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1784                 for_each_online_node(nid) {
1785                         if (nid == env.src_nid || nid == p->numa_preferred_nid)
1786                                 continue;
1787 
1788                         dist = node_distance(env.src_nid, env.dst_nid);
1789                         if (sched_numa_topology_type == NUMA_BACKPLANE &&
1790                                                 dist != env.dist) {
1791                                 taskweight = task_weight(p, env.src_nid, dist);
1792                                 groupweight = group_weight(p, env.src_nid, dist);
1793                         }
1794 
1795                         /* Only consider nodes where both task and groups benefit */
1796                         taskimp = task_weight(p, nid, dist) - taskweight;
1797                         groupimp = group_weight(p, nid, dist) - groupweight;
1798                         if (taskimp < 0 && groupimp < 0)
1799                                 continue;
1800 
1801                         env.dist = dist;
1802                         env.dst_nid = nid;
1803                         update_numa_stats(&env.dst_stats, env.dst_nid);
1804                         if (numa_has_capacity(&env))
1805                                 task_numa_find_cpu(&env, taskimp, groupimp);
1806                 }
1807         }
1808 
1809         /*
1810          * If the task is part of a workload that spans multiple NUMA nodes,
1811          * and is migrating into one of the workload's active nodes, remember
1812          * this node as the task's preferred numa node, so the workload can
1813          * settle down.
1814          * A task that migrated to a second choice node will be better off
1815          * trying for a better one later. Do not set the preferred node here.
1816          */
1817         if (p->numa_group) {
1818                 struct numa_group *ng = p->numa_group;
1819 
1820                 if (env.best_cpu == -1)
1821                         nid = env.src_nid;
1822                 else
1823                         nid = env.dst_nid;
1824 
1825                 if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng))
1826                         sched_setnuma(p, env.dst_nid);
1827         }
1828 
1829         /* No better CPU than the current one was found. */
1830         if (env.best_cpu == -1)
1831                 return -EAGAIN;
1832 
1833         /*
1834          * Reset the scan period if the task is being rescheduled on an
1835          * alternative node to recheck if the tasks is now properly placed.
1836          */
1837         p->numa_scan_period = task_scan_start(p);
1838 
1839         if (env.best_task == NULL) {
1840                 ret = migrate_task_to(p, env.best_cpu);
1841                 if (ret != 0)
1842                         trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1843                 return ret;
1844         }
1845 
1846         ret = migrate_swap(p, env.best_task);
1847         if (ret != 0)
1848                 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1849         put_task_struct(env.best_task);
1850         return ret;
1851 }
1852 
1853 /* Attempt to migrate a task to a CPU on the preferred node. */
1854 static void numa_migrate_preferred(struct task_struct *p)
1855 {
1856         unsigned long interval = HZ;
1857 
1858         /* This task has no NUMA fault statistics yet */
1859         if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1860                 return;
1861 
1862         /* Periodically retry migrating the task to the preferred node */
1863         interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1864         p->numa_migrate_retry = jiffies + interval;
1865 
1866         /* Success if task is already running on preferred CPU */
1867         if (task_node(p) == p->numa_preferred_nid)
1868                 return;
1869 
1870         /* Otherwise, try migrate to a CPU on the preferred node */
1871         task_numa_migrate(p);
1872 }
1873 
1874 /*
1875  * Find out how many nodes on the workload is actively running on. Do this by
1876  * tracking the nodes from which NUMA hinting faults are triggered. This can
1877  * be different from the set of nodes where the workload's memory is currently
1878  * located.
1879  */
1880 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1881 {
1882         unsigned long faults, max_faults = 0;
1883         int nid, active_nodes = 0;
1884 
1885         for_each_online_node(nid) {
1886                 faults = group_faults_cpu(numa_group, nid);
1887                 if (faults > max_faults)
1888                         max_faults = faults;
1889         }
1890 
1891         for_each_online_node(nid) {
1892                 faults = group_faults_cpu(numa_group, nid);
1893                 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1894                         active_nodes++;
1895         }
1896 
1897         numa_group->max_faults_cpu = max_faults;
1898         numa_group->active_nodes = active_nodes;
1899 }
1900 
1901 /*
1902  * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1903  * increments. The more local the fault statistics are, the higher the scan
1904  * period will be for the next scan window. If local/(local+remote) ratio is
1905  * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1906  * the scan period will decrease. Aim for 70% local accesses.
1907  */
1908 #define NUMA_PERIOD_SLOTS 10
1909 #define NUMA_PERIOD_THRESHOLD 7
1910 
1911 /*
1912  * Increase the scan period (slow down scanning) if the majority of
1913  * our memory is already on our local node, or if the majority of
1914  * the page accesses are shared with other processes.
1915  * Otherwise, decrease the scan period.
1916  */
1917 static void update_task_scan_period(struct task_struct *p,
1918                         unsigned long shared, unsigned long private)
1919 {
1920         unsigned int period_slot;
1921         int lr_ratio, ps_ratio;
1922         int diff;
1923 
1924         unsigned long remote = p->numa_faults_locality[0];
1925         unsigned long local = p->numa_faults_locality[1];
1926 
1927         /*
1928          * If there were no record hinting faults then either the task is
1929          * completely idle or all activity is areas that are not of interest
1930          * to automatic numa balancing. Related to that, if there were failed
1931          * migration then it implies we are migrating too quickly or the local
1932          * node is overloaded. In either case, scan slower
1933          */
1934         if (local + shared == 0 || p->numa_faults_locality[2]) {
1935                 p->numa_scan_period = min(p->numa_scan_period_max,
1936                         p->numa_scan_period << 1);
1937 
1938                 p->mm->numa_next_scan = jiffies +
1939                         msecs_to_jiffies(p->numa_scan_period);
1940 
1941                 return;
1942         }
1943 
1944         /*
1945          * Prepare to scale scan period relative to the current period.
1946          *       == NUMA_PERIOD_THRESHOLD scan period stays the same
1947          *       <  NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1948          *       >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1949          */
1950         period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1951         lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1952         ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
1953 
1954         if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
1955                 /*
1956                  * Most memory accesses are local. There is no need to
1957                  * do fast NUMA scanning, since memory is already local.
1958                  */
1959                 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
1960                 if (!slot)
1961                         slot = 1;
1962                 diff = slot * period_slot;
1963         } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
1964                 /*
1965                  * Most memory accesses are shared with other tasks.
1966                  * There is no point in continuing fast NUMA scanning,
1967                  * since other tasks may just move the memory elsewhere.
1968                  */
1969                 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
1970                 if (!slot)
1971                         slot = 1;
1972                 diff = slot * period_slot;
1973         } else {
1974                 /*
1975                  * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
1976                  * yet they are not on the local NUMA node. Speed up
1977                  * NUMA scanning to get the memory moved over.
1978                  */
1979                 int ratio = max(lr_ratio, ps_ratio);
1980                 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1981         }
1982 
1983         p->numa_scan_period = clamp(p->numa_scan_period + diff,
1984                         task_scan_min(p), task_scan_max(p));
1985         memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1986 }
1987 
1988 /*
1989  * Get the fraction of time the task has been running since the last
1990  * NUMA placement cycle. The scheduler keeps similar statistics, but
1991  * decays those on a 32ms period, which is orders of magnitude off
1992  * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1993  * stats only if the task is so new there are no NUMA statistics yet.
1994  */
1995 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1996 {
1997         u64 runtime, delta, now;
1998         /* Use the start of this time slice to avoid calculations. */
1999         now = p->se.exec_start;
2000         runtime = p->se.sum_exec_runtime;
2001 
2002         if (p->last_task_numa_placement) {
2003                 delta = runtime - p->last_sum_exec_runtime;
2004                 *period = now - p->last_task_numa_placement;
2005         } else {
2006                 delta = p->se.avg.load_sum;
2007                 *period = LOAD_AVG_MAX;
2008         }
2009 
2010         p->last_sum_exec_runtime = runtime;
2011         p->last_task_numa_placement = now;
2012 
2013         return delta;
2014 }
2015 
2016 /*
2017  * Determine the preferred nid for a task in a numa_group. This needs to
2018  * be done in a way that produces consistent results with group_weight,
2019  * otherwise workloads might not converge.
2020  */
2021 static int preferred_group_nid(struct task_struct *p, int nid)
2022 {
2023         nodemask_t nodes;
2024         int dist;
2025 
2026         /* Direct connections between all NUMA nodes. */
2027         if (sched_numa_topology_type == NUMA_DIRECT)
2028                 return nid;
2029 
2030         /*
2031          * On a system with glueless mesh NUMA topology, group_weight
2032          * scores nodes according to the number of NUMA hinting faults on
2033          * both the node itself, and on nearby nodes.
2034          */
2035         if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2036                 unsigned long score, max_score = 0;
2037                 int node, max_node = nid;
2038 
2039                 dist = sched_max_numa_distance;
2040 
2041                 for_each_online_node(node) {
2042                         score = group_weight(p, node, dist);
2043                         if (score > max_score) {
2044                                 max_score = score;
2045                                 max_node = node;
2046                         }
2047                 }
2048                 return max_node;
2049         }
2050 
2051         /*
2052          * Finding the preferred nid in a system with NUMA backplane
2053          * interconnect topology is more involved. The goal is to locate
2054          * tasks from numa_groups near each other in the system, and
2055          * untangle workloads from different sides of the system. This requires
2056          * searching down the hierarchy of node groups, recursively searching
2057          * inside the highest scoring group of nodes. The nodemask tricks
2058          * keep the complexity of the search down.
2059          */
2060         nodes = node_online_map;
2061         for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2062                 unsigned long max_faults = 0;
2063                 nodemask_t max_group = NODE_MASK_NONE;
2064                 int a, b;
2065 
2066                 /* Are there nodes at this distance from each other? */
2067                 if (!find_numa_distance(dist))
2068                         continue;
2069 
2070                 for_each_node_mask(a, nodes) {
2071                         unsigned long faults = 0;
2072                         nodemask_t this_group;
2073                         nodes_clear(this_group);
2074 
2075                         /* Sum group's NUMA faults; includes a==b case. */
2076                         for_each_node_mask(b, nodes) {
2077                                 if (node_distance(a, b) < dist) {
2078                                         faults += group_faults(p, b);
2079                                         node_set(b, this_group);
2080                                         node_clear(b, nodes);
2081                                 }
2082                         }
2083 
2084                         /* Remember the top group. */
2085                         if (faults > max_faults) {
2086                                 max_faults = faults;
2087                                 max_group = this_group;
2088                                 /*
2089                                  * subtle: at the smallest distance there is
2090                                  * just one node left in each "group", the
2091                                  * winner is the preferred nid.
2092                                  */
2093                                 nid = a;
2094                         }
2095                 }
2096                 /* Next round, evaluate the nodes within max_group. */
2097                 if (!max_faults)
2098                         break;
2099                 nodes = max_group;
2100         }
2101         return nid;
2102 }
2103 
2104 static void task_numa_placement(struct task_struct *p)
2105 {
2106         int seq, nid, max_nid = -1, max_group_nid = -1;
2107         unsigned long max_faults = 0, max_group_faults = 0;
2108         unsigned long fault_types[2] = { 0, 0 };
2109         unsigned long total_faults;
2110         u64 runtime, period;
2111         spinlock_t *group_lock = NULL;
2112 
2113         /*
2114          * The p->mm->numa_scan_seq field gets updated without
2115          * exclusive access. Use READ_ONCE() here to ensure
2116          * that the field is read in a single access:
2117          */
2118         seq = READ_ONCE(p->mm->numa_scan_seq);
2119         if (p->numa_scan_seq == seq)
2120                 return;
2121         p->numa_scan_seq = seq;
2122         p->numa_scan_period_max = task_scan_max(p);
2123 
2124         total_faults = p->numa_faults_locality[0] +
2125                        p->numa_faults_locality[1];
2126         runtime = numa_get_avg_runtime(p, &period);
2127 
2128         /* If the task is part of a group prevent parallel updates to group stats */
2129         if (p->numa_group) {
2130                 group_lock = &p->numa_group->lock;
2131                 spin_lock_irq(group_lock);
2132         }
2133 
2134         /* Find the node with the highest number of faults */
2135         for_each_online_node(nid) {
2136                 /* Keep track of the offsets in numa_faults array */
2137                 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2138                 unsigned long faults = 0, group_faults = 0;
2139                 int priv;
2140 
2141                 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2142                         long diff, f_diff, f_weight;
2143 
2144                         mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2145                         membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2146                         cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2147                         cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2148 
2149                         /* Decay existing window, copy faults since last scan */
2150                         diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2151                         fault_types[priv] += p->numa_faults[membuf_idx];
2152                         p->numa_faults[membuf_idx] = 0;
2153 
2154                         /*
2155                          * Normalize the faults_from, so all tasks in a group
2156                          * count according to CPU use, instead of by the raw
2157                          * number of faults. Tasks with little runtime have
2158                          * little over-all impact on throughput, and thus their
2159                          * faults are less important.
2160                          */
2161                         f_weight = div64_u64(runtime << 16, period + 1);
2162                         f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2163                                    (total_faults + 1);
2164                         f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2165                         p->numa_faults[cpubuf_idx] = 0;
2166 
2167                         p->numa_faults[mem_idx] += diff;
2168                         p->numa_faults[cpu_idx] += f_diff;
2169                         faults += p->numa_faults[mem_idx];
2170                         p->total_numa_faults += diff;
2171                         if (p->numa_group) {
2172                                 /*
2173                                  * safe because we can only change our own group
2174                                  *
2175                                  * mem_idx represents the offset for a given
2176                                  * nid and priv in a specific region because it
2177                                  * is at the beginning of the numa_faults array.
2178                                  */
2179                                 p->numa_group->faults[mem_idx] += diff;
2180                                 p->numa_group->faults_cpu[mem_idx] += f_diff;
2181                                 p->numa_group->total_faults += diff;
2182                                 group_faults += p->numa_group->faults[mem_idx];
2183                         }
2184                 }
2185 
2186                 if (faults > max_faults) {
2187                         max_faults = faults;
2188                         max_nid = nid;
2189                 }
2190 
2191                 if (group_faults > max_group_faults) {
2192                         max_group_faults = group_faults;
2193                         max_group_nid = nid;
2194                 }
2195         }
2196 
2197         update_task_scan_period(p, fault_types[0], fault_types[1]);
2198 
2199         if (p->numa_group) {
2200                 numa_group_count_active_nodes(p->numa_group);
2201                 spin_unlock_irq(group_lock);
2202                 max_nid = preferred_group_nid(p, max_group_nid);
2203         }
2204 
2205         if (max_faults) {
2206                 /* Set the new preferred node */
2207                 if (max_nid != p->numa_preferred_nid)
2208                         sched_setnuma(p, max_nid);
2209 
2210                 if (task_node(p) != p->numa_preferred_nid)
2211                         numa_migrate_preferred(p);
2212         }
2213 }
2214 
2215 static inline int get_numa_group(struct numa_group *grp)
2216 {
2217         return atomic_inc_not_zero(&grp->refcount);
2218 }
2219 
2220 static inline void put_numa_group(struct numa_group *grp)
2221 {
2222         if (atomic_dec_and_test(&grp->refcount))
2223                 kfree_rcu(grp, rcu);
2224 }
2225 
2226 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2227                         int *priv)
2228 {
2229         struct numa_group *grp, *my_grp;
2230         struct task_struct *tsk;
2231         bool join = false;
2232         int cpu = cpupid_to_cpu(cpupid);
2233         int i;
2234 
2235         if (unlikely(!p->numa_group)) {
2236                 unsigned int size = sizeof(struct numa_group) +
2237                                     4*nr_node_ids*sizeof(unsigned long);
2238 
2239                 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2240                 if (!grp)
2241                         return;
2242 
2243                 atomic_set(&grp->refcount, 1);
2244                 grp->active_nodes = 1;
2245                 grp->max_faults_cpu = 0;
2246                 spin_lock_init(&grp->lock);
2247                 grp->gid = p->pid;
2248                 /* Second half of the array tracks nids where faults happen */
2249                 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2250                                                 nr_node_ids;
2251 
2252                 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2253                         grp->faults[i] = p->numa_faults[i];
2254 
2255                 grp->total_faults = p->total_numa_faults;
2256 
2257                 grp->nr_tasks++;
2258                 rcu_assign_pointer(p->numa_group, grp);
2259         }
2260 
2261         rcu_read_lock();
2262         tsk = READ_ONCE(cpu_rq(cpu)->curr);
2263 
2264         if (!cpupid_match_pid(tsk, cpupid))
2265                 goto no_join;
2266 
2267         grp = rcu_dereference(tsk->numa_group);
2268         if (!grp)
2269                 goto no_join;
2270 
2271         my_grp = p->numa_group;
2272         if (grp == my_grp)
2273                 goto no_join;
2274 
2275         /*
2276          * Only join the other group if its bigger; if we're the bigger group,
2277          * the other task will join us.
2278          */
2279         if (my_grp->nr_tasks > grp->nr_tasks)
2280                 goto no_join;
2281 
2282         /*
2283          * Tie-break on the grp address.
2284          */
2285         if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2286                 goto no_join;
2287 
2288         /* Always join threads in the same process. */
2289         if (tsk->mm == current->mm)
2290                 join = true;
2291 
2292         /* Simple filter to avoid false positives due to PID collisions */
2293         if (flags & TNF_SHARED)
2294                 join = true;
2295 
2296         /* Update priv based on whether false sharing was detected */
2297         *priv = !join;
2298 
2299         if (join && !get_numa_group(grp))
2300                 goto no_join;
2301 
2302         rcu_read_unlock();
2303 
2304         if (!join)
2305                 return;
2306 
2307         BUG_ON(irqs_disabled());
2308         double_lock_irq(&my_grp->lock, &grp->lock);
2309 
2310         for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2311                 my_grp->faults[i] -= p->numa_faults[i];
2312                 grp->faults[i] += p->numa_faults[i];
2313         }
2314         my_grp->total_faults -= p->total_numa_faults;
2315         grp->total_faults += p->total_numa_faults;
2316 
2317         my_grp->nr_tasks--;
2318         grp->nr_tasks++;
2319 
2320         spin_unlock(&my_grp->lock);
2321         spin_unlock_irq(&grp->lock);
2322 
2323         rcu_assign_pointer(p->numa_group, grp);
2324 
2325         put_numa_group(my_grp);
2326         return;
2327 
2328 no_join:
2329         rcu_read_unlock();
2330         return;
2331 }
2332 
2333 void task_numa_free(struct task_struct *p)
2334 {
2335         struct numa_group *grp = p->numa_group;
2336         void *numa_faults = p->numa_faults;
2337         unsigned long flags;
2338         int i;
2339 
2340         if (grp) {
2341                 spin_lock_irqsave(&grp->lock, flags);
2342                 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2343                         grp->faults[i] -= p->numa_faults[i];
2344                 grp->total_faults -= p->total_numa_faults;
2345 
2346                 grp->nr_tasks--;
2347                 spin_unlock_irqrestore(&grp->lock, flags);
2348                 RCU_INIT_POINTER(p->numa_group, NULL);
2349                 put_numa_group(grp);
2350         }
2351 
2352         p->numa_faults = NULL;
2353         kfree(numa_faults);
2354 }
2355 
2356 /*
2357  * Got a PROT_NONE fault for a page on @node.
2358  */
2359 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2360 {
2361         struct task_struct *p = current;
2362         bool migrated = flags & TNF_MIGRATED;
2363         int cpu_node = task_node(current);
2364         int local = !!(flags & TNF_FAULT_LOCAL);
2365         struct numa_group *ng;
2366         int priv;
2367 
2368         if (!static_branch_likely(&sched_numa_balancing))
2369                 return;
2370 
2371         /* for example, ksmd faulting in a user's mm */
2372         if (!p->mm)
2373                 return;
2374 
2375         /* Allocate buffer to track faults on a per-node basis */
2376         if (unlikely(!p->numa_faults)) {
2377                 int size = sizeof(*p->numa_faults) *
2378                            NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2379 
2380                 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2381                 if (!p->numa_faults)
2382                         return;
2383 
2384                 p->total_numa_faults = 0;
2385                 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2386         }
2387 
2388         /*
2389          * First accesses are treated as private, otherwise consider accesses
2390          * to be private if the accessing pid has not changed
2391          */
2392         if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2393                 priv = 1;
2394         } else {
2395                 priv = cpupid_match_pid(p, last_cpupid);
2396                 if (!priv && !(flags & TNF_NO_GROUP))
2397                         task_numa_group(p, last_cpupid, flags, &priv);
2398         }
2399 
2400         /*
2401          * If a workload spans multiple NUMA nodes, a shared fault that
2402          * occurs wholly within the set of nodes that the workload is
2403          * actively using should be counted as local. This allows the
2404          * scan rate to slow down when a workload has settled down.
2405          */
2406         ng = p->numa_group;
2407         if (!priv && !local && ng && ng->active_nodes > 1 &&
2408                                 numa_is_active_node(cpu_node, ng) &&
2409                                 numa_is_active_node(mem_node, ng))
2410                 local = 1;
2411 
2412         task_numa_placement(p);
2413 
2414         /*
2415          * Retry task to preferred node migration periodically, in case it
2416          * case it previously failed, or the scheduler moved us.
2417          */
2418         if (time_after(jiffies, p->numa_migrate_retry))
2419                 numa_migrate_preferred(p);
2420 
2421         if (migrated)
2422                 p->numa_pages_migrated += pages;
2423         if (flags & TNF_MIGRATE_FAIL)
2424                 p->numa_faults_locality[2] += pages;
2425 
2426         p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2427         p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2428         p->numa_faults_locality[local] += pages;
2429 }
2430 
2431 static void reset_ptenuma_scan(struct task_struct *p)
2432 {
2433         /*
2434          * We only did a read acquisition of the mmap sem, so
2435          * p->mm->numa_scan_seq is written to without exclusive access
2436          * and the update is not guaranteed to be atomic. That's not
2437          * much of an issue though, since this is just used for
2438          * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2439          * expensive, to avoid any form of compiler optimizations:
2440          */
2441         WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2442         p->mm->numa_scan_offset = 0;
2443 }
2444 
2445 /*
2446  * The expensive part of numa migration is done from task_work context.
2447  * Triggered from task_tick_numa().
2448  */
2449 void task_numa_work(struct callback_head *work)
2450 {
2451         unsigned long migrate, next_scan, now = jiffies;
2452         struct task_struct *p = current;
2453         struct mm_struct *mm = p->mm;
2454         u64 runtime = p->se.sum_exec_runtime;
2455         struct vm_area_struct *vma;
2456         unsigned long start, end;
2457         unsigned long nr_pte_updates = 0;
2458         long pages, virtpages;
2459 
2460         SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2461 
2462         work->next = work; /* protect against double add */
2463         /*
2464          * Who cares about NUMA placement when they're dying.
2465          *
2466          * NOTE: make sure not to dereference p->mm before this check,
2467          * exit_task_work() happens _after_ exit_mm() so we could be called
2468          * without p->mm even though we still had it when we enqueued this
2469          * work.
2470          */
2471         if (p->flags & PF_EXITING)
2472                 return;
2473 
2474         if (!mm->numa_next_scan) {
2475                 mm->numa_next_scan = now +
2476                         msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2477         }
2478 
2479         /*
2480          * Enforce maximal scan/migration frequency..
2481          */
2482         migrate = mm->numa_next_scan;
2483         if (time_before(now, migrate))
2484                 return;
2485 
2486         if (p->numa_scan_period == 0) {
2487                 p->numa_scan_period_max = task_scan_max(p);
2488                 p->numa_scan_period = task_scan_start(p);
2489         }
2490 
2491         next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2492         if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2493                 return;
2494 
2495         /*
2496          * Delay this task enough that another task of this mm will likely win
2497          * the next time around.
2498          */
2499         p->node_stamp += 2 * TICK_NSEC;
2500 
2501         start = mm->numa_scan_offset;
2502         pages = sysctl_numa_balancing_scan_size;
2503         pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2504         virtpages = pages * 8;     /* Scan up to this much virtual space */
2505         if (!pages)
2506                 return;
2507 
2508 
2509         if (!down_read_trylock(&mm->mmap_sem))
2510                 return;
2511         vma = find_vma(mm, start);
2512         if (!vma) {
2513                 reset_ptenuma_scan(p);
2514                 start = 0;
2515                 vma = mm->mmap;
2516         }
2517         for (; vma; vma = vma->vm_next) {
2518                 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2519                         is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2520                         continue;
2521                 }
2522 
2523                 /*
2524                  * Shared library pages mapped by multiple processes are not
2525                  * migrated as it is expected they are cache replicated. Avoid
2526                  * hinting faults in read-only file-backed mappings or the vdso
2527                  * as migrating the pages will be of marginal benefit.
2528                  */
2529                 if (!vma->vm_mm ||
2530                     (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2531                         continue;
2532 
2533                 /*
2534                  * Skip inaccessible VMAs to avoid any confusion between
2535                  * PROT_NONE and NUMA hinting ptes
2536                  */
2537                 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2538                         continue;
2539 
2540                 do {
2541                         start = max(start, vma->vm_start);
2542                         end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2543                         end = min(end, vma->vm_end);
2544                         nr_pte_updates = change_prot_numa(vma, start, end);
2545 
2546                         /*
2547                          * Try to scan sysctl_numa_balancing_size worth of
2548                          * hpages that have at least one present PTE that
2549                          * is not already pte-numa. If the VMA contains
2550                          * areas that are unused or already full of prot_numa
2551                          * PTEs, scan up to virtpages, to skip through those
2552                          * areas faster.
2553                          */
2554                         if (nr_pte_updates)
2555                                 pages -= (end - start) >> PAGE_SHIFT;
2556                         virtpages -= (end - start) >> PAGE_SHIFT;
2557 
2558                         start = end;
2559                         if (pages <= 0 || virtpages <= 0)
2560                                 goto out;
2561 
2562                         cond_resched();
2563                 } while (end != vma->vm_end);
2564         }
2565 
2566 out:
2567         /*
2568          * It is possible to reach the end of the VMA list but the last few
2569          * VMAs are not guaranteed to the vma_migratable. If they are not, we
2570          * would find the !migratable VMA on the next scan but not reset the
2571          * scanner to the start so check it now.
2572          */
2573         if (vma)
2574                 mm->numa_scan_offset = start;
2575         else
2576                 reset_ptenuma_scan(p);
2577         up_read(&mm->mmap_sem);
2578 
2579         /*
2580          * Make sure tasks use at least 32x as much time to run other code
2581          * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2582          * Usually update_task_scan_period slows down scanning enough; on an
2583          * overloaded system we need to limit overhead on a per task basis.
2584          */
2585         if (unlikely(p->se.sum_exec_runtime != runtime)) {
2586                 u64 diff = p->se.sum_exec_runtime - runtime;
2587                 p->node_stamp += 32 * diff;
2588         }
2589 }
2590 
2591 /*
2592  * Drive the periodic memory faults..
2593  */
2594 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2595 {
2596         struct callback_head *work = &curr->numa_work;
2597         u64 period, now;
2598 
2599         /*
2600          * We don't care about NUMA placement if we don't have memory.
2601          */
2602         if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2603                 return;
2604 
2605         /*
2606          * Using runtime rather than walltime has the dual advantage that
2607          * we (mostly) drive the selection from busy threads and that the
2608          * task needs to have done some actual work before we bother with
2609          * NUMA placement.
2610          */
2611         now = curr->se.sum_exec_runtime;
2612         period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2613 
2614         if (now > curr->node_stamp + period) {
2615                 if (!curr->node_stamp)
2616                         curr->numa_scan_period = task_scan_start(curr);
2617                 curr->node_stamp += period;
2618 
2619                 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2620                         init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2621                         task_work_add(curr, work, true);
2622                 }
2623         }
2624 }
2625 
2626 #else
2627 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2628 {
2629 }
2630 
2631 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2632 {
2633 }
2634 
2635 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2636 {
2637 }
2638 
2639 #endif /* CONFIG_NUMA_BALANCING */
2640 
2641 static void
2642 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2643 {
2644         update_load_add(&cfs_rq->load, se->load.weight);
2645         if (!parent_entity(se))
2646                 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2647 #ifdef CONFIG_SMP
2648         if (entity_is_task(se)) {
2649                 struct rq *rq = rq_of(cfs_rq);
2650 
2651                 account_numa_enqueue(rq, task_of(se));
2652                 list_add(&se->group_node, &rq->cfs_tasks);
2653         }
2654 #endif
2655         cfs_rq->nr_running++;
2656 }
2657 
2658 static void
2659 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2660 {
2661         update_load_sub(&cfs_rq->load, se->load.weight);
2662         if (!parent_entity(se))
2663                 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2664 #ifdef CONFIG_SMP
2665         if (entity_is_task(se)) {
2666                 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2667                 list_del_init(&se->group_node);
2668         }
2669 #endif
2670         cfs_rq->nr_running--;
2671 }
2672 
2673 /*
2674  * Signed add and clamp on underflow.
2675  *
2676  * Explicitly do a load-store to ensure the intermediate value never hits
2677  * memory. This allows lockless observations without ever seeing the negative
2678  * values.
2679  */
2680 #define add_positive(_ptr, _val) do {                           \
2681         typeof(_ptr) ptr = (_ptr);                              \
2682         typeof(_val) val = (_val);                              \
2683         typeof(*ptr) res, var = READ_ONCE(*ptr);                \
2684                                                                 \
2685         res = var + val;                                        \
2686                                                                 \
2687         if (val < 0 && res > var)                               \
2688                 res = 0;                                        \
2689                                                                 \
2690         WRITE_ONCE(*ptr, res);                                  \
2691 } while (0)
2692 
2693 /*
2694  * Unsigned subtract and clamp on underflow.
2695  *
2696  * Explicitly do a load-store to ensure the intermediate value never hits
2697  * memory. This allows lockless observations without ever seeing the negative
2698  * values.
2699  */
2700 #define sub_positive(_ptr, _val) do {                           \
2701         typeof(_ptr) ptr = (_ptr);                              \
2702         typeof(*ptr) val = (_val);                              \
2703         typeof(*ptr) res, var = READ_ONCE(*ptr);                \
2704         res = var - val;                                        \
2705         if (res > var)                                          \
2706                 res = 0;                                        \
2707         WRITE_ONCE(*ptr, res);                                  \
2708 } while (0)
2709 
2710 #ifdef CONFIG_SMP
2711 /*
2712  * XXX we want to get rid of these helpers and use the full load resolution.
2713  */
2714 static inline long se_weight(struct sched_entity *se)
2715 {
2716         return scale_load_down(se->load.weight);
2717 }
2718 
2719 static inline long se_runnable(struct sched_entity *se)
2720 {
2721         return scale_load_down(se->runnable_weight);
2722 }
2723 
2724 static inline void
2725 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2726 {
2727         cfs_rq->runnable_weight += se->runnable_weight;
2728 
2729         cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg;
2730         cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum;
2731 }
2732 
2733 static inline void
2734 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2735 {
2736         cfs_rq->runnable_weight -= se->runnable_weight;
2737 
2738         sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg);
2739         sub_positive(&cfs_rq->avg.runnable_load_sum,
2740                      se_runnable(se) * se->avg.runnable_load_sum);
2741 }
2742 
2743 static inline void
2744 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2745 {
2746         cfs_rq->avg.load_avg += se->avg.load_avg;
2747         cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
2748 }
2749 
2750 static inline void
2751 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2752 {
2753         sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2754         sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
2755 }
2756 #else
2757 static inline void
2758 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2759 static inline void
2760 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2761 static inline void
2762 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2763 static inline void
2764 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2765 #endif
2766 
2767 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2768                             unsigned long weight, unsigned long runnable)
2769 {
2770         if (se->on_rq) {
2771                 /* commit outstanding execution time */
2772                 if (cfs_rq->curr == se)
2773                         update_curr(cfs_rq);
2774                 account_entity_dequeue(cfs_rq, se);
2775                 dequeue_runnable_load_avg(cfs_rq, se);
2776         }
2777         dequeue_load_avg(cfs_rq, se);
2778 
2779         se->runnable_weight = runnable;
2780         update_load_set(&se->load, weight);
2781 
2782 #ifdef CONFIG_SMP
2783         do {
2784                 u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
2785 
2786                 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
2787                 se->avg.runnable_load_avg =
2788                         div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider);
2789         } while (0);
2790 #endif
2791 
2792         enqueue_load_avg(cfs_rq, se);
2793         if (se->on_rq) {
2794                 account_entity_enqueue(cfs_rq, se);
2795                 enqueue_runnable_load_avg(cfs_rq, se);
2796         }
2797 }
2798 
2799 void reweight_task(struct task_struct *p, int prio)
2800 {
2801         struct sched_entity *se = &p->se;
2802         struct cfs_rq *cfs_rq = cfs_rq_of(se);
2803         struct load_weight *load = &se->load;
2804         unsigned long weight = scale_load(sched_prio_to_weight[prio]);
2805 
2806         reweight_entity(cfs_rq, se, weight, weight);
2807         load->inv_weight = sched_prio_to_wmult[prio];
2808 }
2809 
2810 #ifdef CONFIG_FAIR_GROUP_SCHED
2811 #ifdef CONFIG_SMP
2812 /*
2813  * All this does is approximate the hierarchical proportion which includes that
2814  * global sum we all love to hate.
2815  *
2816  * That is, the weight of a group entity, is the proportional share of the
2817  * group weight based on the group runqueue weights. That is:
2818  *
2819  *                     tg->weight * grq->load.weight
2820  *   ge->load.weight = -----------------------------               (1)
2821  *                        \Sum grq->load.weight
2822  *
2823  * Now, because computing that sum is prohibitively expensive to compute (been
2824  * there, done that) we approximate it with this average stuff. The average
2825  * moves slower and therefore the approximation is cheaper and more stable.
2826  *
2827  * So instead of the above, we substitute:
2828  *
2829  *   grq->load.weight -> grq->avg.load_avg                         (2)
2830  *
2831  * which yields the following:
2832  *
2833  *                     tg->weight * grq->avg.load_avg
2834  *   ge->load.weight = ------------------------------              (3)
2835  *                              tg->load_avg
2836  *
2837  * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2838  *
2839  * That is shares_avg, and it is right (given the approximation (2)).
2840  *
2841  * The problem with it is that because the average is slow -- it was designed
2842  * to be exactly that of course -- this leads to transients in boundary
2843  * conditions. In specific, the case where the group was idle and we start the
2844  * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2845  * yielding bad latency etc..
2846  *
2847  * Now, in that special case (1) reduces to:
2848  *
2849  *                     tg->weight * grq->load.weight
2850  *   ge->load.weight = ----------------------------- = tg->weight   (4)
2851  *                          grp->load.weight
2852  *
2853  * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2854  *
2855  * So what we do is modify our approximation (3) to approach (4) in the (near)
2856  * UP case, like:
2857  *
2858  *   ge->load.weight =
2859  *
2860  *              tg->weight * grq->load.weight
2861  *     ---------------------------------------------------         (5)
2862  *     tg->load_avg - grq->avg.load_avg + grq->load.weight
2863  *
2864  * But because grq->load.weight can drop to 0, resulting in a divide by zero,
2865  * we need to use grq->avg.load_avg as its lower bound, which then gives:
2866  *
2867  *
2868  *                     tg->weight * grq->load.weight
2869  *   ge->load.weight = -----------------------------               (6)
2870  *                              tg_load_avg'
2871  *
2872  * Where:
2873  *
2874  *   tg_load_avg' = tg->load_avg - grq->avg.load_avg +
2875  *                  max(grq->load.weight, grq->avg.load_avg)
2876  *
2877  * And that is shares_weight and is icky. In the (near) UP case it approaches
2878  * (4) while in the normal case it approaches (3). It consistently
2879  * overestimates the ge->load.weight and therefore:
2880  *
2881  *   \Sum ge->load.weight >= tg->weight
2882  *
2883  * hence icky!
2884  */
2885 static long calc_group_shares(struct cfs_rq *cfs_rq)
2886 {
2887         long tg_weight, tg_shares, load, shares;
2888         struct task_group *tg = cfs_rq->tg;
2889 
2890         tg_shares = READ_ONCE(tg->shares);
2891 
2892         load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
2893 
2894         tg_weight = atomic_long_read(&tg->load_avg);
2895 
2896         /* Ensure tg_weight >= load */
2897         tg_weight -= cfs_rq->tg_load_avg_contrib;
2898         tg_weight += load;
2899 
2900         shares = (tg_shares * load);
2901         if (tg_weight)
2902                 shares /= tg_weight;
2903 
2904         /*
2905          * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2906          * of a group with small tg->shares value. It is a floor value which is
2907          * assigned as a minimum load.weight to the sched_entity representing
2908          * the group on a CPU.
2909          *
2910          * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2911          * on an 8-core system with 8 tasks each runnable on one CPU shares has
2912          * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2913          * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2914          * instead of 0.
2915          */
2916         return clamp_t(long, shares, MIN_SHARES, tg_shares);
2917 }
2918 
2919 /*
2920  * This calculates the effective runnable weight for a group entity based on
2921  * the group entity weight calculated above.
2922  *
2923  * Because of the above approximation (2), our group entity weight is
2924  * an load_avg based ratio (3). This means that it includes blocked load and
2925  * does not represent the runnable weight.
2926  *
2927  * Approximate the group entity's runnable weight per ratio from the group
2928  * runqueue:
2929  *
2930  *                                           grq->avg.runnable_load_avg
2931  *   ge->runnable_weight = ge->load.weight * -------------------------- (7)
2932  *                                               grq->avg.load_avg
2933  *
2934  * However, analogous to above, since the avg numbers are slow, this leads to
2935  * transients in the from-idle case. Instead we use:
2936  *
2937  *   ge->runnable_weight = ge->load.weight *
2938  *
2939  *              max(grq->avg.runnable_load_avg, grq->runnable_weight)
2940  *              -----------------------------------------------------   (8)
2941  *                    max(grq->avg.load_avg, grq->load.weight)
2942  *
2943  * Where these max() serve both to use the 'instant' values to fix the slow
2944  * from-idle and avoid the /0 on to-idle, similar to (6).
2945  */
2946 static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares)
2947 {
2948         long runnable, load_avg;
2949 
2950         load_avg = max(cfs_rq->avg.load_avg,
2951                        scale_load_down(cfs_rq->load.weight));
2952 
2953         runnable = max(cfs_rq->avg.runnable_load_avg,
2954                        scale_load_down(cfs_rq->runnable_weight));
2955 
2956         runnable *= shares;
2957         if (load_avg)
2958                 runnable /= load_avg;
2959 
2960         return clamp_t(long, runnable, MIN_SHARES, shares);
2961 }
2962 #endif /* CONFIG_SMP */
2963 
2964 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2965 
2966 /*
2967  * Recomputes the group entity based on the current state of its group
2968  * runqueue.
2969  */
2970 static void update_cfs_group(struct sched_entity *se)
2971 {
2972         struct cfs_rq *gcfs_rq = group_cfs_rq(se);
2973         long shares, runnable;
2974 
2975         if (!gcfs_rq)
2976                 return;
2977 
2978         if (throttled_hierarchy(gcfs_rq))
2979                 return;
2980 
2981 #ifndef CONFIG_SMP
2982         runnable = shares = READ_ONCE(gcfs_rq->tg->shares);
2983 
2984         if (likely(se->load.weight == shares))
2985                 return;
2986 #else
2987         shares   = calc_group_shares(gcfs_rq);
2988         runnable = calc_group_runnable(gcfs_rq, shares);
2989 #endif
2990 
2991         reweight_entity(cfs_rq_of(se), se, shares, runnable);
2992 }
2993 
2994 #else /* CONFIG_FAIR_GROUP_SCHED */
2995 static inline void update_cfs_group(struct sched_entity *se)
2996 {
2997 }
2998 #endif /* CONFIG_FAIR_GROUP_SCHED */
2999 
3000 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3001 {
3002         struct rq *rq = rq_of(cfs_rq);
3003 
3004         if (&rq->cfs == cfs_rq || (flags & SCHED_CPUFREQ_MIGRATION)) {
3005                 /*
3006                  * There are a few boundary cases this might miss but it should
3007                  * get called often enough that that should (hopefully) not be
3008                  * a real problem.
3009                  *
3010                  * It will not get called when we go idle, because the idle
3011                  * thread is a different class (!fair), nor will the utilization
3012                  * number include things like RT tasks.
3013                  *
3014                  * As is, the util number is not freq-invariant (we'd have to
3015                  * implement arch_scale_freq_capacity() for that).
3016                  *
3017                  * See cpu_util().
3018                  */
3019                 cpufreq_update_util(rq, flags);
3020         }
3021 }
3022 
3023 #ifdef CONFIG_SMP
3024 /*
3025  * Approximate:
3026  *   val * y^n,    where y^32 ~= 0.5 (~1 scheduling period)
3027  */
3028 static u64 decay_load(u64 val, u64 n)
3029 {
3030         unsigned int local_n;
3031 
3032         if (unlikely(n > LOAD_AVG_PERIOD * 63))
3033                 return 0;
3034 
3035         /* after bounds checking we can collapse to 32-bit */
3036         local_n = n;
3037 
3038         /*
3039          * As y^PERIOD = 1/2, we can combine
3040          *    y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
3041          * With a look-up table which covers y^n (n<PERIOD)
3042          *
3043          * To achieve constant time decay_load.
3044          */
3045         if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
3046                 val >>= local_n / LOAD_AVG_PERIOD;
3047                 local_n %= LOAD_AVG_PERIOD;
3048         }
3049 
3050         val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
3051         return val;
3052 }
3053 
3054 static u32 __accumulate_pelt_segments(u64 periods, u32 d1, u32 d3)
3055 {
3056         u32 c1, c2, c3 = d3; /* y^0 == 1 */
3057 
3058         /*
3059          * c1 = d1 y^p
3060          */
3061         c1 = decay_load((u64)d1, periods);
3062 
3063         /*
3064          *            p-1
3065          * c2 = 1024 \Sum y^n
3066          *            n=1
3067          *
3068          *              inf        inf
3069          *    = 1024 ( \Sum y^n - \Sum y^n - y^0 )
3070          *              n=0        n=p
3071          */
3072         c2 = LOAD_AVG_MAX - decay_load(LOAD_AVG_MAX, periods) - 1024;
3073 
3074         return c1 + c2 + c3;
3075 }
3076 
3077 /*
3078  * Accumulate the three separate parts of the sum; d1 the remainder
3079  * of the last (incomplete) period, d2 the span of full periods and d3
3080  * the remainder of the (incomplete) current period.
3081  *
3082  *           d1          d2           d3
3083  *           ^           ^            ^
3084  *           |           |            |
3085  *         |<->|<----------------->|<--->|
3086  * ... |---x---|------| ... |------|-----x (now)
3087  *
3088  *                           p-1
3089  * u' = (u + d1) y^p + 1024 \Sum y^n + d3 y^0
3090  *                           n=1
3091  *
3092  *    = u y^p +                                 (Step 1)
3093  *
3094  *                     p-1
3095  *      d1 y^p + 1024 \Sum y^n + d3 y^0         (Step 2)
3096  *                     n=1
3097  */
3098 static __always_inline u32
3099 accumulate_sum(u64 delta, int cpu, struct sched_avg *sa,
3100                unsigned long load, unsigned long runnable, int running)
3101 {
3102         unsigned long scale_freq, scale_cpu;
3103         u32 contrib = (u32)delta; /* p == 0 -> delta < 1024 */
3104         u64 periods;
3105 
3106         scale_freq = arch_scale_freq_capacity(cpu);
3107         scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
3108 
3109         delta += sa->period_contrib;
3110         periods = delta / 1024; /* A period is 1024us (~1ms) */
3111 
3112         /*
3113          * Step 1: decay old *_sum if we crossed period boundaries.
3114          */
3115         if (periods) {
3116                 sa->load_sum = decay_load(sa->load_sum, periods);
3117                 sa->runnable_load_sum =
3118                         decay_load(sa->runnable_load_sum, periods);
3119                 sa->util_sum = decay_load((u64)(sa->util_sum), periods);
3120 
3121                 /*
3122                  * Step 2
3123                  */
3124                 delta %= 1024;
3125                 contrib = __accumulate_pelt_segments(periods,
3126                                 1024 - sa->period_contrib, delta);
3127         }
3128         sa->period_contrib = delta;
3129 
3130         contrib = cap_scale(contrib, scale_freq);
3131         if (load)
3132                 sa->load_sum += load * contrib;
3133         if (runnable)
3134                 sa->runnable_load_sum += runnable * contrib;
3135         if (running)
3136                 sa->util_sum += contrib * scale_cpu;
3137 
3138         return periods;
3139 }
3140 
3141 /*
3142  * We can represent the historical contribution to runnable average as the
3143  * coefficients of a geometric series.  To do this we sub-divide our runnable
3144  * history into segments of approximately 1ms (1024us); label the segment that
3145  * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
3146  *
3147  * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
3148  *      p0            p1           p2
3149  *     (now)       (~1ms ago)  (~2ms ago)
3150  *
3151  * Let u_i denote the fraction of p_i that the entity was runnable.
3152  *
3153  * We then designate the fractions u_i as our co-efficients, yielding the
3154  * following representation of historical load:
3155  *   u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
3156  *
3157  * We choose y based on the with of a reasonably scheduling period, fixing:
3158  *   y^32 = 0.5
3159  *
3160  * This means that the contribution to load ~32ms ago (u_32) will be weighted
3161  * approximately half as much as the contribution to load within the last ms
3162  * (u_0).
3163  *
3164  * When a period "rolls over" and we have new u_0`, multiplying the previous
3165  * sum again by y is sufficient to update:
3166  *   load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
3167  *            = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
3168  */
3169 static __always_inline int
3170 ___update_load_sum(u64 now, int cpu, struct sched_avg *sa,
3171                   unsigned long load, unsigned long runnable, int running)
3172 {
3173         u64 delta;
3174 
3175         delta = now - sa->last_update_time;
3176         /*
3177          * This should only happen when time goes backwards, which it
3178          * unfortunately does during sched clock init when we swap over to TSC.
3179          */
3180         if ((s64)delta < 0) {
3181                 sa->last_update_time = now;
3182                 return 0;
3183         }
3184 
3185         /*
3186          * Use 1024ns as the unit of measurement since it's a reasonable
3187          * approximation of 1us and fast to compute.
3188          */
3189         delta >>= 10;
3190         if (!delta)
3191                 return 0;
3192 
3193         sa->last_update_time += delta << 10;
3194 
3195         /*
3196          * running is a subset of runnable (weight) so running can't be set if
3197          * runnable is clear. But there are some corner cases where the current
3198          * se has been already dequeued but cfs_rq->curr still points to it.
3199          * This means that weight will be 0 but not running for a sched_entity
3200          * but also for a cfs_rq if the latter becomes idle. As an example,
3201          * this happens during idle_balance() which calls
3202          * update_blocked_averages()
3203          */
3204         if (!load)
3205                 runnable = running = 0;
3206 
3207         /*
3208          * Now we know we crossed measurement unit boundaries. The *_avg
3209          * accrues by two steps:
3210          *
3211          * Step 1: accumulate *_sum since last_update_time. If we haven't
3212          * crossed period boundaries, finish.
3213          */
3214         if (!accumulate_sum(delta, cpu, sa, load, runnable, running))
3215                 return 0;
3216 
3217         return 1;
3218 }
3219 
3220 static __always_inline void
3221 ___update_load_avg(struct sched_avg *sa, unsigned long load, unsigned long runnable)
3222 {
3223         u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3224 
3225         /*
3226          * Step 2: update *_avg.
3227          */
3228         sa->load_avg = div_u64(load * sa->load_sum, divider);
3229         sa->runnable_load_avg = div_u64(runnable * sa->runnable_load_sum, divider);
3230         sa->util_avg = sa->util_sum / divider;
3231 }
3232 
3233 /*
3234  * When a task is dequeued, its estimated utilization should not be update if
3235  * its util_avg has not been updated at least once.
3236  * This flag is used to synchronize util_avg updates with util_est updates.
3237  * We map this information into the LSB bit of the utilization saved at
3238  * dequeue time (i.e. util_est.dequeued).
3239  */
3240 #define UTIL_AVG_UNCHANGED 0x1
3241 
3242 static inline void cfs_se_util_change(struct sched_avg *avg)
3243 {
3244         unsigned int enqueued;
3245 
3246         if (!sched_feat(UTIL_EST))
3247                 return;
3248 
3249         /* Avoid store if the flag has been already set */
3250         enqueued = avg->util_est.enqueued;
3251         if (!(enqueued & UTIL_AVG_UNCHANGED))
3252                 return;
3253 
3254         /* Reset flag to report util_avg has been updated */
3255         enqueued &= ~UTIL_AVG_UNCHANGED;
3256         WRITE_ONCE(avg->util_est.enqueued, enqueued);
3257 }
3258 
3259 /*
3260  * sched_entity:
3261  *
3262  *   task:
3263  *     se_runnable() == se_weight()
3264  *
3265  *   group: [ see update_cfs_group() ]
3266  *     se_weight()   = tg->weight * grq->load_avg / tg->load_avg
3267  *     se_runnable() = se_weight(se) * grq->runnable_load_avg / grq->load_avg
3268  *
3269  *   load_sum := runnable_sum
3270  *   load_avg = se_weight(se) * runnable_avg
3271  *
3272  *   runnable_load_sum := runnable_sum
3273  *   runnable_load_avg = se_runnable(se) * runnable_avg
3274  *
3275  * XXX collapse load_sum and runnable_load_sum
3276  *
3277  * cfq_rs:
3278  *
3279  *   load_sum = \Sum se_weight(se) * se->avg.load_sum
3280  *   load_avg = \Sum se->avg.load_avg
3281  *
3282  *   runnable_load_sum = \Sum se_runnable(se) * se->avg.runnable_load_sum
3283  *   runnable_load_avg = \Sum se->avg.runable_load_avg
3284  */
3285 
3286 static int
3287 __update_load_avg_blocked_se(u64 now, int cpu, struct sched_entity *se)
3288 {
3289         if (entity_is_task(se))
3290                 se->runnable_weight = se->load.weight;
3291 
3292         if (___update_load_sum(now, cpu, &se->avg, 0, 0, 0)) {
3293                 ___update_load_avg(&se->avg, se_weight(se), se_runnable(se));
3294                 return 1;
3295         }
3296 
3297         return 0;
3298 }
3299 
3300 static int
3301 __update_load_avg_se(u64 now, int cpu, struct cfs_rq *cfs_rq, struct sched_entity *se)
3302 {
3303         if (entity_is_task(se))
3304                 se->runnable_weight = se->load.weight;
3305 
3306         if (___update_load_sum(now, cpu, &se->avg, !!se->on_rq, !!se->on_rq,
3307                                 cfs_rq->curr == se)) {
3308 
3309                 ___update_load_avg(&se->avg, se_weight(se), se_runnable(se));
3310                 cfs_se_util_change(&se->avg);
3311                 return 1;
3312         }
3313 
3314         return 0;
3315 }
3316 
3317 static int
3318 __update_load_avg_cfs_rq(u64 now, int cpu, struct cfs_rq *cfs_rq)
3319 {
3320         if (___update_load_sum(now, cpu, &cfs_rq->avg,
3321                                 scale_load_down(cfs_rq->load.weight),
3322                                 scale_load_down(cfs_rq->runnable_weight),
3323                                 cfs_rq->curr != NULL)) {
3324 
3325                 ___update_load_avg(&cfs_rq->avg, 1, 1);
3326                 return 1;
3327         }
3328 
3329         return 0;
3330 }
3331 
3332 #ifdef CONFIG_FAIR_GROUP_SCHED
3333 /**
3334  * update_tg_load_avg - update the tg's load avg
3335  * @cfs_rq: the cfs_rq whose avg changed
3336  * @force: update regardless of how small the difference
3337  *
3338  * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3339  * However, because tg->load_avg is a global value there are performance
3340  * considerations.
3341  *
3342  * In order to avoid having to look at the other cfs_rq's, we use a
3343  * differential update where we store the last value we propagated. This in
3344  * turn allows skipping updates if the differential is 'small'.
3345  *
3346  * Updating tg's load_avg is necessary before update_cfs_share().
3347  */
3348 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3349 {
3350         long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3351 
3352         /*
3353          * No need to update load_avg for root_task_group as it is not used.
3354          */
3355         if (cfs_rq->tg == &root_task_group)
3356                 return;
3357 
3358         if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3359                 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3360                 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3361         }
3362 }
3363 
3364 /*
3365  * Called within set_task_rq() right before setting a task's CPU. The
3366  * caller only guarantees p->pi_lock is held; no other assumptions,
3367  * including the state of rq->lock, should be made.
3368  */
3369 void set_task_rq_fair(struct sched_entity *se,
3370                       struct cfs_rq *prev, struct cfs_rq *next)
3371 {
3372         u64 p_last_update_time;
3373         u64 n_last_update_time;
3374 
3375         if (!sched_feat(ATTACH_AGE_LOAD))
3376                 return;
3377 
3378         /*
3379          * We are supposed to update the task to "current" time, then its up to
3380          * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3381          * getting what current time is, so simply throw away the out-of-date
3382          * time. This will result in the wakee task is less decayed, but giving
3383          * the wakee more load sounds not bad.
3384          */
3385         if (!(se->avg.last_update_time && prev))
3386                 return;
3387 
3388 #ifndef CONFIG_64BIT
3389         {
3390                 u64 p_last_update_time_copy;
3391                 u64 n_last_update_time_copy;
3392 
3393                 do {
3394                         p_last_update_time_copy = prev->load_last_update_time_copy;
3395                         n_last_update_time_copy = next->load_last_update_time_copy;
3396 
3397                         smp_rmb();
3398 
3399                         p_last_update_time = prev->avg.last_update_time;
3400                         n_last_update_time = next->avg.last_update_time;
3401 
3402                 } while (p_last_update_time != p_last_update_time_copy ||
3403                          n_last_update_time != n_last_update_time_copy);
3404         }
3405 #else
3406         p_last_update_time = prev->avg.last_update_time;
3407         n_last_update_time = next->avg.last_update_time;
3408 #endif
3409         __update_load_avg_blocked_se(p_last_update_time, cpu_of(rq_of(prev)), se);
3410         se->avg.last_update_time = n_last_update_time;
3411 }
3412 
3413 
3414 /*
3415  * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3416  * propagate its contribution. The key to this propagation is the invariant
3417  * that for each group:
3418  *
3419  *   ge->avg == grq->avg                                                (1)
3420  *
3421  * _IFF_ we look at the pure running and runnable sums. Because they
3422  * represent the very same entity, just at different points in the hierarchy.
3423  *
3424  * Per the above update_tg_cfs_util() is trivial and simply copies the running
3425  * sum over (but still wrong, because the group entity and group rq do not have
3426  * their PELT windows aligned).
3427  *
3428  * However, update_tg_cfs_runnable() is more complex. So we have:
3429  *
3430  *   ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg          (2)
3431  *
3432  * And since, like util, the runnable part should be directly transferable,
3433  * the following would _appear_ to be the straight forward approach:
3434  *
3435  *   grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg       (3)
3436  *
3437  * And per (1) we have:
3438  *
3439  *   ge->avg.runnable_avg == grq->avg.runnable_avg
3440  *
3441  * Which gives:
3442  *
3443  *                      ge->load.weight * grq->avg.load_avg
3444  *   ge->avg.load_avg = -----------------------------------             (4)
3445  *                               grq->load.weight
3446  *
3447  * Except that is wrong!
3448  *
3449  * Because while for entities historical weight is not important and we
3450  * really only care about our future and therefore can consider a pure
3451  * runnable sum, runqueues can NOT do this.
3452  *
3453  * We specifically want runqueues to have a load_avg that includes
3454  * historical weights. Those represent the blocked load, the load we expect
3455  * to (shortly) return to us. This only works by keeping the weights as
3456  * integral part of the sum. We therefore cannot decompose as per (3).
3457  *
3458  * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3459  * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3460  * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3461  * runnable section of these tasks overlap (or not). If they were to perfectly
3462  * align the rq as a whole would be runnable 2/3 of the time. If however we
3463  * always have at least 1 runnable task, the rq as a whole is always runnable.
3464  *
3465  * So we'll have to approximate.. :/
3466  *
3467  * Given the constraint:
3468  *
3469  *   ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3470  *
3471  * We can construct a rule that adds runnable to a rq by assuming minimal
3472  * overlap.
3473  *
3474  * On removal, we'll assume each task is equally runnable; which yields:
3475  *
3476  *   grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3477  *
3478  * XXX: only do this for the part of runnable > running ?
3479  *
3480  */
3481 
3482 static inline void
3483 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3484 {
3485         long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3486 
3487         /* Nothing to update */
3488         if (!delta)
3489                 return;
3490 
3491         /*
3492          * The relation between sum and avg is:
3493          *
3494          *   LOAD_AVG_MAX - 1024 + sa->period_contrib
3495          *
3496          * however, the PELT windows are not aligned between grq and gse.
3497          */
3498 
3499         /* Set new sched_entity's utilization */
3500         se->avg.util_avg = gcfs_rq->avg.util_avg;
3501         se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3502 
3503         /* Update parent cfs_rq utilization */
3504         add_positive(&cfs_rq->avg.util_avg, delta);
3505         cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3506 }
3507 
3508 static inline void
3509 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3510 {
3511         long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3512         unsigned long runnable_load_avg, load_avg;
3513         u64 runnable_load_sum, load_sum = 0;
3514         s64 delta_sum;
3515 
3516         if (!runnable_sum)
3517                 return;
3518 
3519         gcfs_rq->prop_runnable_sum = 0;
3520 
3521         if (runnable_sum >= 0) {
3522                 /*
3523                  * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3524                  * the CPU is saturated running == runnable.
3525                  */
3526                 runnable_sum += se->avg.load_sum;
3527                 runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
3528         } else {
3529                 /*
3530                  * Estimate the new unweighted runnable_sum of the gcfs_rq by
3531                  * assuming all tasks are equally runnable.
3532                  */
3533                 if (scale_load_down(gcfs_rq->load.weight)) {
3534                         load_sum = div_s64(gcfs_rq->avg.load_sum,
3535                                 scale_load_down(gcfs_rq->load.weight));
3536                 }
3537 
3538                 /* But make sure to not inflate se's runnable */
3539                 runnable_sum = min(se->avg.load_sum, load_sum);
3540         }
3541 
3542         /*
3543          * runnable_sum can't be lower than running_sum
3544          * As running sum is scale with CPU capacity wehreas the runnable sum
3545          * is not we rescale running_sum 1st
3546          */
3547         running_sum = se->avg.util_sum /
3548                 arch_scale_cpu_capacity(NULL, cpu_of(rq_of(cfs_rq)));
3549         runnable_sum = max(runnable_sum, running_sum);
3550 
3551         load_sum = (s64)se_weight(se) * runnable_sum;
3552         load_avg = div_s64(load_sum, LOAD_AVG_MAX);
3553 
3554         delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3555         delta_avg = load_avg - se->avg.load_avg;
3556 
3557         se->avg.load_sum = runnable_sum;
3558         se->avg.load_avg = load_avg;
3559         add_positive(&cfs_rq->avg.load_avg, delta_avg);
3560         add_positive(&cfs_rq->avg.load_sum, delta_sum);
3561 
3562         runnable_load_sum = (s64)se_runnable(se) * runnable_sum;
3563         runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX);
3564         delta_sum = runnable_load_sum - se_weight(se) * se->avg.runnable_load_sum;
3565         delta_avg = runnable_load_avg - se->avg.runnable_load_avg;
3566 
3567         se->avg.runnable_load_sum = runnable_sum;
3568         se->avg.runnable_load_avg = runnable_load_avg;
3569 
3570         if (se->on_rq) {
3571                 add_positive(&cfs_rq->avg.runnable_load_avg, delta_avg);
3572                 add_positive(&cfs_rq->avg.runnable_load_sum, delta_sum);
3573         }
3574 }
3575 
3576 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3577 {
3578         cfs_rq->propagate = 1;
3579         cfs_rq->prop_runnable_sum += runnable_sum;
3580 }
3581 
3582 /* Update task and its cfs_rq load average */
3583 static inline int propagate_entity_load_avg(struct sched_entity *se)
3584 {
3585         struct cfs_rq *cfs_rq, *gcfs_rq;
3586 
3587         if (entity_is_task(se))
3588                 return 0;
3589 
3590         gcfs_rq = group_cfs_rq(se);
3591         if (!gcfs_rq->propagate)
3592                 return 0;
3593 
3594         gcfs_rq->propagate = 0;
3595 
3596         cfs_rq = cfs_rq_of(se);
3597 
3598         add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3599 
3600         update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3601         update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3602 
3603         return 1;
3604 }
3605 
3606 /*
3607  * Check if we need to update the load and the utilization of a blocked
3608  * group_entity:
3609  */
3610 static inline bool skip_blocked_update(struct sched_entity *se)
3611 {
3612         struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3613 
3614         /*
3615          * If sched_entity still have not zero load or utilization, we have to
3616          * decay it:
3617          */
3618         if (se->avg.load_avg || se->avg.util_avg)
3619                 return false;
3620 
3621         /*
3622          * If there is a pending propagation, we have to update the load and
3623          * the utilization of the sched_entity:
3624          */
3625         if (gcfs_rq->propagate)
3626                 return false;
3627 
3628         /*
3629          * Otherwise, the load and the utilization of the sched_entity is
3630          * already zero and there is no pending propagation, so it will be a
3631          * waste of time to try to decay it:
3632          */
3633         return true;
3634 }
3635 
3636 #else /* CONFIG_FAIR_GROUP_SCHED */
3637 
3638 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3639 
3640 static inline int propagate_entity_load_avg(struct sched_entity *se)
3641 {
3642         return 0;
3643 }
3644 
3645 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3646 
3647 #endif /* CONFIG_FAIR_GROUP_SCHED */
3648 
3649 /**
3650  * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3651  * @now: current time, as per cfs_rq_clock_task()
3652  * @cfs_rq: cfs_rq to update
3653  *
3654  * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3655  * avg. The immediate corollary is that all (fair) tasks must be attached, see
3656  * post_init_entity_util_avg().
3657  *
3658  * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3659  *
3660  * Returns true if the load decayed or we removed load.
3661  *
3662  * Since both these conditions indicate a changed cfs_rq->avg.load we should
3663  * call update_tg_load_avg() when this function returns true.
3664  */
3665 static inline int
3666 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3667 {
3668         unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0;
3669         struct sched_avg *sa = &cfs_rq->avg;
3670         int decayed = 0;
3671 
3672         if (cfs_rq->removed.nr) {
3673                 unsigned long r;
3674                 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3675 
3676                 raw_spin_lock(&cfs_rq->removed.lock);
3677                 swap(cfs_rq->removed.util_avg, removed_util);
3678                 swap(cfs_rq->removed.load_avg, removed_load);
3679                 swap(cfs_rq->removed.runnable_sum, removed_runnable_sum);
3680                 cfs_rq->removed.nr = 0;
3681                 raw_spin_unlock(&cfs_rq->removed.lock);
3682 
3683                 r = removed_load;
3684                 sub_positive(&sa->load_avg, r);
3685                 sub_positive(&sa->load_sum, r * divider);
3686 
3687                 r = removed_util;
3688                 sub_positive(&sa->util_avg, r);
3689                 sub_positive(&sa->util_sum, r * divider);
3690 
3691                 add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum);
3692 
3693                 decayed = 1;
3694         }
3695 
3696         decayed |= __update_load_avg_cfs_rq(now, cpu_of(rq_of(cfs_rq)), cfs_rq);
3697 
3698 #ifndef CONFIG_64BIT
3699         smp_wmb();
3700         cfs_rq->load_last_update_time_copy = sa->last_update_time;
3701 #endif
3702 
3703         if (decayed)
3704                 cfs_rq_util_change(cfs_rq, 0);
3705 
3706         return decayed;
3707 }
3708 
3709 /**
3710  * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3711  * @cfs_rq: cfs_rq to attach to
3712  * @se: sched_entity to attach
3713  *
3714  * Must call update_cfs_rq_load_avg() before this, since we rely on
3715  * cfs_rq->avg.last_update_time being current.
3716  */
3717 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3718 {
3719         u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3720 
3721         /*
3722          * When we attach the @se to the @cfs_rq, we must align the decay
3723          * window because without that, really weird and wonderful things can
3724          * happen.
3725          *
3726          * XXX illustrate
3727          */
3728         se->avg.last_update_time = cfs_rq->avg.last_update_time;
3729         se->avg.period_contrib = cfs_rq->avg.period_contrib;
3730 
3731         /*
3732          * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3733          * period_contrib. This isn't strictly correct, but since we're
3734          * entirely outside of the PELT hierarchy, nobody cares if we truncate
3735          * _sum a little.
3736          */
3737         se->avg.util_sum = se->avg.util_avg * divider;
3738 
3739         se->avg.load_sum = divider;
3740         if (se_weight(se)) {
3741                 se->avg.load_sum =
3742                         div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3743         }
3744 
3745         se->avg.runnable_load_sum = se->avg.load_sum;
3746 
3747         enqueue_load_avg(cfs_rq, se);
3748         cfs_rq->avg.util_avg += se->avg.util_avg;
3749         cfs_rq->avg.util_sum += se->avg.util_sum;
3750 
3751         add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3752 
3753         cfs_rq_util_change(cfs_rq, flags);
3754 }
3755 
3756 /**
3757  * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3758  * @cfs_rq: cfs_rq to detach from
3759  * @se: sched_entity to detach
3760  *
3761  * Must call update_cfs_rq_load_avg() before this, since we rely on
3762  * cfs_rq->avg.last_update_time being current.
3763  */
3764 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3765 {
3766         dequeue_load_avg(cfs_rq, se);
3767         sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3768         sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3769 
3770         add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3771 
3772         cfs_rq_util_change(cfs_rq, 0);
3773 }
3774 
3775 /*
3776  * Optional action to be done while updating the load average
3777  */
3778 #define UPDATE_TG       0x1
3779 #define SKIP_AGE_LOAD   0x2
3780 #define DO_ATTACH       0x4
3781 
3782 /* Update task and its cfs_rq load average */
3783 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3784 {
3785         u64 now = cfs_rq_clock_task(cfs_rq);
3786         struct rq *rq = rq_of(cfs_rq);
3787         int cpu = cpu_of(rq);
3788         int decayed;
3789 
3790         /*
3791          * Track task load average for carrying it to new CPU after migrated, and
3792          * track group sched_entity load average for task_h_load calc in migration
3793          */
3794         if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3795                 __update_load_avg_se(now, cpu, cfs_rq, se);
3796 
3797         decayed  = update_cfs_rq_load_avg(now, cfs_rq);
3798         decayed |= propagate_entity_load_avg(se);
3799 
3800         if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3801 
3802                 /*
3803                  * DO_ATTACH means we're here from enqueue_entity().
3804                  * !last_update_time means we've passed through
3805                  * migrate_task_rq_fair() indicating we migrated.
3806                  *
3807                  * IOW we're enqueueing a task on a new CPU.
3808                  */
3809                 attach_entity_load_avg(cfs_rq, se, SCHED_CPUFREQ_MIGRATION);
3810                 update_tg_load_avg(cfs_rq, 0);
3811 
3812         } else if (decayed && (flags & UPDATE_TG))
3813                 update_tg_load_avg(cfs_rq, 0);
3814 }
3815 
3816 #ifndef CONFIG_64BIT
3817 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3818 {
3819         u64 last_update_time_copy;
3820         u64 last_update_time;
3821 
3822         do {
3823                 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3824                 smp_rmb();
3825                 last_update_time = cfs_rq->avg.last_update_time;
3826         } while (last_update_time != last_update_time_copy);
3827 
3828         return last_update_time;
3829 }
3830 #else
3831 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3832 {
3833         return cfs_rq->avg.last_update_time;
3834 }
3835 #endif
3836 
3837 /*
3838  * Synchronize entity load avg of dequeued entity without locking
3839  * the previous rq.
3840  */
3841 void sync_entity_load_avg(struct sched_entity *se)
3842 {
3843         struct cfs_rq *cfs_rq = cfs_rq_of(se);
3844         u64 last_update_time;
3845 
3846         last_update_time = cfs_rq_last_update_time(cfs_rq);
3847         __update_load_avg_blocked_se(last_update_time, cpu_of(rq_of(cfs_rq)), se);
3848 }
3849 
3850 /*
3851  * Task first catches up with cfs_rq, and then subtract
3852  * itself from the cfs_rq (task must be off the queue now).
3853  */
3854 void remove_entity_load_avg(struct sched_entity *se)
3855 {
3856         struct cfs_rq *cfs_rq = cfs_rq_of(se);
3857         unsigned long flags;
3858 
3859         /*
3860          * tasks cannot exit without having gone through wake_up_new_task() ->
3861          * post_init_entity_util_avg() which will have added things to the
3862          * cfs_rq, so we can remove unconditionally.
3863          *
3864          * Similarly for groups, they will have passed through
3865          * post_init_entity_util_avg() before unregister_sched_fair_group()
3866          * calls this.
3867          */
3868 
3869         sync_entity_load_avg(se);
3870 
3871         raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3872         ++cfs_rq->removed.nr;
3873         cfs_rq->removed.util_avg        += se->avg.util_avg;
3874         cfs_rq->removed.load_avg        += se->avg.load_avg;
3875         cfs_rq->removed.runnable_sum    += se->avg.load_sum; /* == runnable_sum */
3876         raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3877 }
3878 
3879 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3880 {
3881         return cfs_rq->avg.runnable_load_avg;
3882 }
3883 
3884 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3885 {
3886         return cfs_rq->avg.load_avg;
3887 }
3888 
3889 static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3890 
3891 static inline unsigned long task_util(struct task_struct *p)
3892 {
3893         return READ_ONCE(p->se.avg.util_avg);
3894 }
3895 
3896 static inline unsigned long _task_util_est(struct task_struct *p)
3897 {
3898         struct util_est ue = READ_ONCE(p->se.avg.util_est);
3899 
3900         return max(ue.ewma, ue.enqueued);
3901 }
3902 
3903 static inline unsigned long task_util_est(struct task_struct *p)
3904 {
3905         return max(task_util(p), _task_util_est(p));
3906 }
3907 
3908 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3909                                     struct task_struct *p)
3910 {
3911         unsigned int enqueued;
3912 
3913         if (!sched_feat(UTIL_EST))
3914                 return;
3915 
3916         /* Update root cfs_rq's estimated utilization */
3917         enqueued  = cfs_rq->avg.util_est.enqueued;
3918         enqueued += (_task_util_est(p) | UTIL_AVG_UNCHANGED);
3919         WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3920 }
3921 
3922 /*
3923  * Check if a (signed) value is within a specified (unsigned) margin,
3924  * based on the observation that:
3925  *
3926  *     abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3927  *
3928  * NOTE: this only works when value + maring < INT_MAX.
3929  */
3930 static inline bool within_margin(int value, int margin)
3931 {
3932         return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3933 }
3934 
3935 static void
3936 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep)
3937 {
3938         long last_ewma_diff;
3939         struct util_est ue;
3940 
3941         if (!sched_feat(UTIL_EST))
3942                 return;
3943 
3944         /* Update root cfs_rq's estimated utilization */
3945         ue.enqueued  = cfs_rq->avg.util_est.enqueued;
3946         ue.enqueued -= min_t(unsigned int, ue.enqueued,
3947                              (_task_util_est(p) | UTIL_AVG_UNCHANGED));
3948         WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued);
3949 
3950         /*
3951          * Skip update of task's estimated utilization when the task has not
3952          * yet completed an activation, e.g. being migrated.
3953          */
3954         if (!task_sleep)
3955                 return;
3956 
3957         /*
3958          * If the PELT values haven't changed since enqueue time,
3959          * skip the util_est update.
3960          */
3961         ue = p->se.avg.util_est;
3962         if (ue.enqueued & UTIL_AVG_UNCHANGED)
3963                 return;
3964 
3965         /*
3966          * Skip update of task's estimated utilization when its EWMA is
3967          * already ~1% close to its last activation value.
3968          */
3969         ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
3970         last_ewma_diff = ue.enqueued - ue.ewma;
3971         if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
3972                 return;
3973 
3974         /*
3975          * Update Task's estimated utilization
3976          *
3977          * When *p completes an activation we can consolidate another sample
3978          * of the task size. This is done by storing the current PELT value
3979          * as ue.enqueued and by using this value to update the Exponential
3980          * Weighted Moving Average (EWMA):
3981          *
3982          *  ewma(t) = w *  task_util(p) + (1-w) * ewma(t-1)
3983          *          = w *  task_util(p) +         ewma(t-1)  - w * ewma(t-1)
3984          *          = w * (task_util(p) -         ewma(t-1)) +     ewma(t-1)
3985          *          = w * (      last_ewma_diff            ) +     ewma(t-1)
3986          *          = w * (last_ewma_diff  +  ewma(t-1) / w)
3987          *
3988          * Where 'w' is the weight of new samples, which is configured to be
3989          * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
3990          */
3991         ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
3992         ue.ewma  += last_ewma_diff;
3993         ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
3994         WRITE_ONCE(p->se.avg.util_est, ue);
3995 }
3996 
3997 #else /* CONFIG_SMP */
3998 
3999 static inline int
4000 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
4001 {
4002         return 0;
4003 }
4004 
4005 #define UPDATE_TG       0x0
4006 #define SKIP_AGE_LOAD   0x0
4007 #define DO_ATTACH       0x0
4008 
4009 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
4010 {
4011         cfs_rq_util_change(cfs_rq, 0);
4012 }
4013 
4014 static inline void remove_entity_load_avg(struct sched_entity *se) {}
4015 
4016 static inline void
4017 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) {}
4018 static inline void
4019 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4020 
4021 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
4022 {
4023         return 0;
4024 }
4025 
4026 static inline void
4027 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4028 
4029 static inline void
4030 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p,
4031                  bool task_sleep) {}
4032 
4033 #endif /* CONFIG_SMP */
4034 
4035 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
4036 {
4037 #ifdef CONFIG_SCHED_DEBUG
4038         s64 d = se->vruntime - cfs_rq->min_vruntime;
4039 
4040         if (d < 0)
4041                 d = -d;
4042 
4043         if (d > 3*sysctl_sched_latency)
4044                 schedstat_inc(cfs_rq->nr_spread_over);
4045 #endif
4046 }
4047 
4048 static void
4049 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
4050 {
4051         u64 vruntime = cfs_rq->min_vruntime;
4052 
4053         /*
4054          * The 'current' period is already promised to the current tasks,
4055          * however the extra weight of the new task will slow them down a
4056          * little, place the new task so that it fits in the slot that
4057          * stays open at the end.
4058          */
4059         if (initial && sched_feat(START_DEBIT))
4060                 vruntime += sched_vslice(cfs_rq, se);
4061 
4062         /* sleeps up to a single latency don't count. */
4063         if (!initial) {
4064                 unsigned long thresh = sysctl_sched_latency;
4065 
4066                 /*
4067                  * Halve their sleep time's effect, to allow
4068                  * for a gentler effect of sleepers:
4069                  */
4070                 if (sched_feat(GENTLE_FAIR_SLEEPERS))
4071                         thresh >>= 1;
4072 
4073                 vruntime -= thresh;
4074         }
4075 
4076         /* ensure we never gain time by being placed backwards. */
4077         se->vruntime = max_vruntime(se->vruntime, vruntime);
4078 }
4079 
4080 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
4081 
4082 static inline void check_schedstat_required(void)
4083 {
4084 #ifdef CONFIG_SCHEDSTATS
4085         if (schedstat_enabled())
4086                 return;
4087 
4088         /* Force schedstat enabled if a dependent tracepoint is active */
4089         if (trace_sched_stat_wait_enabled()    ||
4090                         trace_sched_stat_sleep_enabled()   ||
4091                         trace_sched_stat_iowait_enabled()  ||
4092                         trace_sched_stat_blocked_enabled() ||
4093                         trace_sched_stat_runtime_enabled())  {
4094                 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
4095                              "stat_blocked and stat_runtime require the "
4096                              "kernel parameter schedstats=enable or "
4097                              "kernel.sched_schedstats=1\n");
4098         }
4099 #endif
4100 }
4101 
4102 
4103 /*
4104  * MIGRATION
4105  *
4106  *      dequeue
4107  *        update_curr()
4108  *          update_min_vruntime()
4109  *        vruntime -= min_vruntime
4110  *
4111  *      enqueue
4112  *        update_curr()
4113  *          update_min_vruntime()
4114  *        vruntime += min_vruntime
4115  *
4116  * this way the vruntime transition between RQs is done when both
4117  * min_vruntime are up-to-date.
4118  *
4119  * WAKEUP (remote)
4120  *
4121  *      ->migrate_task_rq_fair() (p->state == TASK_WAKING)
4122  *        vruntime -= min_vruntime
4123  *
4124  *      enqueue
4125  *        update_curr()
4126  *          update_min_vruntime()
4127  *        vruntime += min_vruntime
4128  *
4129  * this way we don't have the most up-to-date min_vruntime on the originating
4130  * CPU and an up-to-date min_vruntime on the destination CPU.
4131  */
4132 
4133 static void
4134 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4135 {
4136         bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
4137         bool curr = cfs_rq->curr == se;
4138 
4139         /*
4140          * If we're the current task, we must renormalise before calling
4141          * update_curr().
4142          */
4143         if (renorm && curr)
4144                 se->vruntime += cfs_rq->min_vruntime;
4145 
4146         update_curr(cfs_rq);
4147 
4148         /*
4149          * Otherwise, renormalise after, such that we're placed at the current
4150          * moment in time, instead of some random moment in the past. Being
4151          * placed in the past could significantly boost this task to the
4152          * fairness detriment of existing tasks.
4153          */
4154         if (renorm && !curr)
4155                 se->vruntime += cfs_rq->min_vruntime;
4156 
4157         /*
4158          * When enqueuing a sched_entity, we must:
4159          *   - Update loads to have both entity and cfs_rq synced with now.
4160          *   - Add its load to cfs_rq->runnable_avg
4161          *   - For group_entity, update its weight to reflect the new share of
4162          *     its group cfs_rq
4163          *   - Add its new weight to cfs_rq->load.weight
4164          */
4165         update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
4166         update_cfs_group(se);
4167         enqueue_runnable_load_avg(cfs_rq, se);
4168         account_entity_enqueue(cfs_rq, se);
4169 
4170         if (flags & ENQUEUE_WAKEUP)
4171                 place_entity(cfs_rq, se, 0);
4172 
4173         check_schedstat_required();
4174         update_stats_enqueue(cfs_rq, se, flags);
4175         check_spread(cfs_rq, se);
4176         if (!curr)
4177                 __enqueue_entity(cfs_rq, se);
4178         se->on_rq = 1;
4179 
4180         if (cfs_rq->nr_running == 1) {
4181                 list_add_leaf_cfs_rq(cfs_rq);
4182                 check_enqueue_throttle(cfs_rq);
4183         }
4184 }
4185 
4186 static void __clear_buddies_last(struct sched_entity *se)
4187 {
4188         for_each_sched_entity(se) {
4189                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4190                 if (cfs_rq->last != se)
4191                         break;
4192 
4193                 cfs_rq->last = NULL;
4194         }
4195 }
4196 
4197 static void __clear_buddies_next(struct sched_entity *se)
4198 {
4199         for_each_sched_entity(se) {
4200                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4201                 if (cfs_rq->next != se)
4202                         break;
4203 
4204                 cfs_rq->next = NULL;
4205         }
4206 }
4207 
4208 static void __clear_buddies_skip(struct sched_entity *se)
4209 {
4210         for_each_sched_entity(se) {
4211                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4212                 if (cfs_rq->skip != se)
4213                         break;
4214 
4215                 cfs_rq->skip = NULL;
4216         }
4217 }
4218 
4219 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4220 {
4221         if (cfs_rq->last == se)
4222                 __clear_buddies_last(se);
4223 
4224         if (cfs_rq->next == se)
4225                 __clear_buddies_next(se);
4226 
4227         if (cfs_rq->skip == se)
4228                 __clear_buddies_skip(se);
4229 }
4230 
4231 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4232 
4233 static void
4234 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4235 {
4236         /*
4237          * Update run-time statistics of the 'current'.
4238          */
4239         update_curr(cfs_rq);
4240 
4241         /*
4242          * When dequeuing a sched_entity, we must:
4243          *   - Update loads to have both entity and cfs_rq synced with now.
4244          *   - Substract its load from the cfs_rq->runnable_avg.
4245          *   - Substract its previous weight from cfs_rq->load.weight.
4246          *   - For group entity, update its weight to reflect the new share
4247          *     of its group cfs_rq.
4248          */
4249         update_load_avg(cfs_rq, se, UPDATE_TG);
4250         dequeue_runnable_load_avg(cfs_rq, se);
4251 
4252         update_stats_dequeue(cfs_rq, se, flags);
4253 
4254         clear_buddies(cfs_rq, se);
4255 
4256         if (se != cfs_rq->curr)
4257                 __dequeue_entity(cfs_rq, se);
4258         se->on_rq = 0;
4259         account_entity_dequeue(cfs_rq, se);
4260 
4261         /*
4262          * Normalize after update_curr(); which will also have moved
4263          * min_vruntime if @se is the one holding it back. But before doing
4264          * update_min_vruntime() again, which will discount @se's position and
4265          * can move min_vruntime forward still more.
4266          */
4267         if (!(flags & DEQUEUE_SLEEP))
4268                 se->vruntime -= cfs_rq->min_vruntime;
4269 
4270         /* return excess runtime on last dequeue */
4271         return_cfs_rq_runtime(cfs_rq);
4272 
4273         update_cfs_group(se);
4274 
4275         /*
4276          * Now advance min_vruntime if @se was the entity holding it back,
4277          * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4278          * put back on, and if we advance min_vruntime, we'll be placed back
4279          * further than we started -- ie. we'll be penalized.
4280          */
4281         if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
4282                 update_min_vruntime(cfs_rq);
4283 }
4284 
4285 /*
4286  * Preempt the current task with a newly woken task if needed:
4287  */
4288 static void
4289 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4290 {
4291         unsigned long ideal_runtime, delta_exec;
4292         struct sched_entity *se;
4293         s64 delta;
4294 
4295         ideal_runtime = sched_slice(cfs_rq, curr);
4296         delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4297         if (delta_exec > ideal_runtime) {
4298                 resched_curr(rq_of(cfs_rq));
4299                 /*
4300                  * The current task ran long enough, ensure it doesn't get
4301                  * re-elected due to buddy favours.
4302                  */
4303                 clear_buddies(cfs_rq, curr);
4304                 return;
4305         }
4306 
4307         /*
4308          * Ensure that a task that missed wakeup preemption by a
4309          * narrow margin doesn't have to wait for a full slice.
4310          * This also mitigates buddy induced latencies under load.
4311          */
4312         if (delta_exec < sysctl_sched_min_granularity)
4313                 return;
4314 
4315         se = __pick_first_entity(cfs_rq);
4316         delta = curr->vruntime - se->vruntime;
4317 
4318         if (delta < 0)
4319                 return;
4320 
4321         if (delta > ideal_runtime)
4322                 resched_curr(rq_of(cfs_rq));
4323 }
4324 
4325 static void
4326 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4327 {
4328         /* 'current' is not kept within the tree. */
4329         if (se->on_rq) {
4330                 /*
4331                  * Any task has to be enqueued before it get to execute on
4332                  * a CPU. So account for the time it spent waiting on the
4333                  * runqueue.
4334                  */
4335                 update_stats_wait_end(cfs_rq, se);
4336                 __dequeue_entity(cfs_rq, se);
4337                 update_load_avg(cfs_rq, se, UPDATE_TG);
4338         }
4339 
4340         update_stats_curr_start(cfs_rq, se);
4341         cfs_rq->curr = se;
4342 
4343         /*
4344          * Track our maximum slice length, if the CPU's load is at
4345          * least twice that of our own weight (i.e. dont track it
4346          * when there are only lesser-weight tasks around):
4347          */
4348         if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
4349                 schedstat_set(se->statistics.slice_max,
4350                         max((u64)schedstat_val(se->statistics.slice_max),
4351                             se->sum_exec_runtime - se->prev_sum_exec_runtime));
4352         }
4353 
4354         se->prev_sum_exec_runtime = se->sum_exec_runtime;
4355 }
4356 
4357 static int
4358 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4359 
4360 /*
4361  * Pick the next process, keeping these things in mind, in this order:
4362  * 1) keep things fair between processes/task groups
4363  * 2) pick the "next" process, since someone really wants that to run
4364  * 3) pick the "last" process, for cache locality
4365  * 4) do not run the "skip" process, if something else is available
4366  */
4367 static struct sched_entity *
4368 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4369 {
4370         struct sched_entity *left = __pick_first_entity(cfs_rq);
4371         struct sched_entity *se;
4372 
4373         /*
4374          * If curr is set we have to see if its left of the leftmost entity
4375          * still in the tree, provided there was anything in the tree at all.
4376          */
4377         if (!left || (curr && entity_before(curr, left)))
4378                 left = curr;
4379 
4380         se = left; /* ideally we run the leftmost entity */
4381 
4382         /*
4383          * Avoid running the skip buddy, if running something else can
4384          * be done without getting too unfair.
4385          */
4386         if (cfs_rq->skip == se) {
4387                 struct sched_entity *second;
4388 
4389                 if (se == curr) {
4390                         second = __pick_first_entity(cfs_rq);
4391                 } else {
4392                         second = __pick_next_entity(se);
4393                         if (!second || (curr && entity_before(curr, second)))
4394                                 second = curr;
4395                 }
4396 
4397                 if (second && wakeup_preempt_entity(second, left) < 1)
4398                         se = second;
4399         }
4400 
4401         /*
4402          * Prefer last buddy, try to return the CPU to a preempted task.
4403          */
4404         if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4405                 se = cfs_rq->last;
4406 
4407         /*
4408          * Someone really wants this to run. If it's not unfair, run it.
4409          */
4410         if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4411                 se = cfs_rq->next;
4412 
4413         clear_buddies(cfs_rq, se);
4414 
4415         return se;
4416 }
4417 
4418 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4419 
4420 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4421 {
4422         /*
4423          * If still on the runqueue then deactivate_task()
4424          * was not called and update_curr() has to be done:
4425          */
4426         if (prev->on_rq)
4427                 update_curr(cfs_rq);
4428 
4429         /* throttle cfs_rqs exceeding runtime */
4430         check_cfs_rq_runtime(cfs_rq);
4431 
4432         check_spread(cfs_rq, prev);
4433 
4434         if (prev->on_rq) {
4435                 update_stats_wait_start(cfs_rq, prev);
4436                 /* Put 'current' back into the tree. */
4437                 __enqueue_entity(cfs_rq, prev);
4438                 /* in !on_rq case, update occurred at dequeue */
4439                 update_load_avg(cfs_rq, prev, 0);
4440         }
4441         cfs_rq->curr = NULL;
4442 }
4443 
4444 static void
4445 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4446 {
4447         /*
4448          * Update run-time statistics of the 'current'.
4449          */
4450         update_curr(cfs_rq);
4451 
4452         /*
4453          * Ensure that runnable average is periodically updated.
4454          */
4455         update_load_avg(cfs_rq, curr, UPDATE_TG);
4456         update_cfs_group(curr);
4457 
4458 #ifdef CONFIG_SCHED_HRTICK
4459         /*
4460          * queued ticks are scheduled to match the slice, so don't bother
4461          * validating it and just reschedule.
4462          */
4463         if (queued) {
4464                 resched_curr(rq_of(cfs_rq));
4465                 return;
4466         }
4467         /*
4468          * don't let the period tick interfere with the hrtick preemption
4469          */
4470         if (!sched_feat(DOUBLE_TICK) &&
4471                         hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4472                 return;
4473 #endif
4474 
4475         if (cfs_rq->nr_running > 1)
4476                 check_preempt_tick(cfs_rq, curr);
4477 }
4478 
4479 
4480 /**************************************************
4481  * CFS bandwidth control machinery
4482  */
4483 
4484 #ifdef CONFIG_CFS_BANDWIDTH
4485 
4486 #ifdef HAVE_JUMP_LABEL
4487 static struct static_key __cfs_bandwidth_used;
4488 
4489 static inline bool cfs_bandwidth_used(void)
4490 {
4491         return static_key_false(&__cfs_bandwidth_used);
4492 }
4493 
4494 void cfs_bandwidth_usage_inc(void)
4495 {
4496         static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4497 }
4498 
4499 void cfs_bandwidth_usage_dec(void)
4500 {
4501         static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4502 }
4503 #else /* HAVE_JUMP_LABEL */
4504 static bool cfs_bandwidth_used(void)
4505 {
4506         return true;
4507 }
4508 
4509 void cfs_bandwidth_usage_inc(void) {}
4510 void cfs_bandwidth_usage_dec(void) {}
4511 #endif /* HAVE_JUMP_LABEL */
4512 
4513 /*
4514  * default period for cfs group bandwidth.
4515  * default: 0.1s, units: nanoseconds
4516  */
4517 static inline u64 default_cfs_period(void)
4518 {
4519         return 100000000ULL;
4520 }
4521 
4522 static inline u64 sched_cfs_bandwidth_slice(void)
4523 {
4524         return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4525 }
4526 
4527 /*
4528  * Replenish runtime according to assigned quota and update expiration time.
4529  * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4530  * additional synchronization around rq->lock.
4531  *
4532  * requires cfs_b->lock
4533  */
4534 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4535 {
4536         u64 now;
4537 
4538         if (cfs_b->quota == RUNTIME_INF)
4539                 return;
4540 
4541         now = sched_clock_cpu(smp_processor_id());
4542         cfs_b->runtime = cfs_b->quota;
4543         cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
4544         cfs_b->expires_seq++;
4545 }
4546 
4547 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4548 {
4549         return &tg->cfs_bandwidth;
4550 }
4551 
4552 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4553 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4554 {
4555         if (unlikely(cfs_rq->throttle_count))
4556                 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4557 
4558         return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4559 }
4560 
4561 /* returns 0 on failure to allocate runtime */
4562 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4563 {
4564         struct task_group *tg = cfs_rq->tg;
4565         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4566         u64 amount = 0, min_amount, expires;
4567         int expires_seq;
4568 
4569         /* note: this is a positive sum as runtime_remaining <= 0 */
4570         min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4571 
4572         raw_spin_lock(&cfs_b->lock);
4573         if (cfs_b->quota == RUNTIME_INF)
4574                 amount = min_amount;
4575         else {
4576                 start_cfs_bandwidth(cfs_b);
4577 
4578                 if (cfs_b->runtime > 0) {
4579                         amount = min(cfs_b->runtime, min_amount);
4580                         cfs_b->runtime -= amount;
4581                         cfs_b->idle = 0;
4582                 }
4583         }
4584         expires_seq = cfs_b->expires_seq;
4585         expires = cfs_b->runtime_expires;
4586         raw_spin_unlock(&cfs_b->lock);
4587 
4588         cfs_rq->runtime_remaining += amount;
4589         /*
4590          * we may have advanced our local expiration to account for allowed
4591          * spread between our sched_clock and the one on which runtime was
4592          * issued.
4593          */
4594         if (cfs_rq->expires_seq != expires_seq) {
4595                 cfs_rq->expires_seq = expires_seq;
4596                 cfs_rq->runtime_expires = expires;
4597         }
4598 
4599         return cfs_rq->runtime_remaining > 0;
4600 }
4601 
4602 /*
4603  * Note: This depends on the synchronization provided by sched_clock and the
4604  * fact that rq->clock snapshots this value.
4605  */
4606 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4607 {
4608         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4609 
4610         /* if the deadline is ahead of our clock, nothing to do */
4611         if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
4612                 return;
4613 
4614         if (cfs_rq->runtime_remaining < 0)
4615                 return;
4616 
4617         /*
4618          * If the local deadline has passed we have to consider the
4619          * possibility that our sched_clock is 'fast' and the global deadline
4620          * has not truly expired.
4621          *
4622          * Fortunately we can check determine whether this the case by checking
4623          * whether the global deadline(cfs_b->expires_seq) has advanced.
4624          */
4625         if (cfs_rq->expires_seq == cfs_b->expires_seq) {
4626                 /* extend local deadline, drift is bounded above by 2 ticks */
4627                 cfs_rq->runtime_expires += TICK_NSEC;
4628         } else {
4629                 /* global deadline is ahead, expiration has passed */
4630                 cfs_rq->runtime_remaining = 0;
4631         }
4632 }
4633 
4634 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4635 {
4636         /* dock delta_exec before expiring quota (as it could span periods) */
4637         cfs_rq->runtime_remaining -= delta_exec;
4638         expire_cfs_rq_runtime(cfs_rq);
4639 
4640         if (likely(cfs_rq->runtime_remaining > 0))
4641                 return;
4642 
4643         /*
4644          * if we're unable to extend our runtime we resched so that the active
4645          * hierarchy can be throttled
4646          */
4647         if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4648                 resched_curr(rq_of(cfs_rq));
4649 }
4650 
4651 static __always_inline
4652 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4653 {
4654         if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4655                 return;
4656 
4657         __account_cfs_rq_runtime(cfs_rq, delta_exec);
4658 }
4659 
4660 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4661 {
4662         return cfs_bandwidth_used() && cfs_rq->throttled;
4663 }
4664 
4665 /* check whether cfs_rq, or any parent, is throttled */
4666 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4667 {
4668         return cfs_bandwidth_used() && cfs_rq->throttle_count;
4669 }
4670 
4671 /*
4672  * Ensure that neither of the group entities corresponding to src_cpu or
4673  * dest_cpu are members of a throttled hierarchy when performing group
4674  * load-balance operations.
4675  */
4676 static inline int throttled_lb_pair(struct task_group *tg,
4677                                     int src_cpu, int dest_cpu)
4678 {
4679         struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4680 
4681         src_cfs_rq = tg->cfs_rq[src_cpu];
4682         dest_cfs_rq = tg->cfs_rq[dest_cpu];
4683 
4684         return throttled_hierarchy(src_cfs_rq) ||
4685                throttled_hierarchy(dest_cfs_rq);
4686 }
4687 
4688 /* updated child weight may affect parent so we have to do this bottom up */
4689 static int tg_unthrottle_up(struct task_group *tg, void *data)
4690 {
4691         struct rq *rq = data;
4692         struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4693 
4694         cfs_rq->throttle_count--;
4695         if (!cfs_rq->throttle_count) {
4696                 /* adjust cfs_rq_clock_task() */
4697                 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4698                                              cfs_rq->throttled_clock_task;
4699         }
4700 
4701         return 0;
4702 }
4703 
4704 static int tg_throttle_down(struct task_group *tg, void *data)
4705 {
4706         struct rq *rq = data;
4707         struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4708 
4709         /* group is entering throttled state, stop time */
4710         if (!cfs_rq->throttle_count)
4711                 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4712         cfs_rq->throttle_count++;
4713 
4714         return 0;
4715 }
4716 
4717 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4718 {
4719         struct rq *rq = rq_of(cfs_rq);
4720         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4721         struct sched_entity *se;
4722         long task_delta, dequeue = 1;
4723         bool empty;
4724 
4725         se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4726 
4727         /* freeze hierarchy runnable averages while throttled */
4728         rcu_read_lock();
4729         walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4730         rcu_read_unlock();
4731 
4732         task_delta = cfs_rq->h_nr_running;
4733         for_each_sched_entity(se) {
4734                 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4735                 /* throttled entity or throttle-on-deactivate */
4736                 if (!se->on_rq)
4737                         break;
4738 
4739                 if (dequeue)
4740                         dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4741                 qcfs_rq->h_nr_running -= task_delta;
4742 
4743                 if (qcfs_rq->load.weight)
4744                         dequeue = 0;
4745         }
4746 
4747         if (!se)
4748                 sub_nr_running(rq, task_delta);
4749 
4750         cfs_rq->throttled = 1;
4751         cfs_rq->throttled_clock = rq_clock(rq);
4752         raw_spin_lock(&cfs_b->lock);
4753         empty = list_empty(&cfs_b->throttled_cfs_rq);
4754 
4755         /*
4756          * Add to the _head_ of the list, so that an already-started
4757          * distribute_cfs_runtime will not see us
4758          */
4759         list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4760 
4761         /*
4762          * If we're the first throttled task, make sure the bandwidth
4763          * timer is running.
4764          */
4765         if (empty)
4766                 start_cfs_bandwidth(cfs_b);
4767 
4768         raw_spin_unlock(&cfs_b->lock);
4769 }
4770 
4771 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4772 {
4773         struct rq *rq = rq_of(cfs_rq);
4774         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4775         struct sched_entity *se;
4776         int enqueue = 1;
4777         long task_delta;
4778 
4779         se = cfs_rq->tg->se[cpu_of(rq)];
4780 
4781         cfs_rq->throttled = 0;
4782 
4783         update_rq_clock(rq);
4784 
4785         raw_spin_lock(&cfs_b->lock);
4786         cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4787         list_del_rcu(&cfs_rq->throttled_list);
4788         raw_spin_unlock(&cfs_b->lock);
4789 
4790         /* update hierarchical throttle state */
4791         walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4792 
4793         if (!cfs_rq->load.weight)
4794                 return;
4795 
4796         task_delta = cfs_rq->h_nr_running;
4797         for_each_sched_entity(se) {
4798                 if (se->on_rq)
4799                         enqueue = 0;
4800 
4801                 cfs_rq = cfs_rq_of(se);
4802                 if (enqueue)
4803                         enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4804                 cfs_rq->h_nr_running += task_delta;
4805 
4806                 if (cfs_rq_throttled(cfs_rq))
4807                         break;
4808         }
4809 
4810         if (!se)
4811                 add_nr_running(rq, task_delta);
4812 
4813         /* Determine whether we need to wake up potentially idle CPU: */
4814         if (rq->curr == rq->idle && rq->cfs.nr_running)
4815                 resched_curr(rq);
4816 }
4817 
4818 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4819                 u64 remaining, u64 expires)
4820 {
4821         struct cfs_rq *cfs_rq;
4822         u64 runtime;
4823         u64 starting_runtime = remaining;
4824 
4825         rcu_read_lock();
4826         list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4827                                 throttled_list) {
4828                 struct rq *rq = rq_of(cfs_rq);
4829                 struct rq_flags rf;
4830 
4831                 rq_lock(rq, &rf);
4832                 if (!cfs_rq_throttled(cfs_rq))
4833                         goto next;
4834 
4835                 runtime = -cfs_rq->runtime_remaining + 1;
4836                 if (runtime > remaining)
4837                         runtime = remaining;
4838                 remaining -= runtime;
4839 
4840                 cfs_rq->runtime_remaining += runtime;
4841                 cfs_rq->runtime_expires = expires;
4842 
4843                 /* we check whether we're throttled above */
4844                 if (cfs_rq->runtime_remaining > 0)
4845                         unthrottle_cfs_rq(cfs_rq);
4846 
4847 next:
4848                 rq_unlock(rq, &rf);
4849 
4850                 if (!remaining)
4851                         break;
4852         }
4853         rcu_read_unlock();
4854 
4855         return starting_runtime - remaining;
4856 }
4857 
4858 /*
4859  * Responsible for refilling a task_group's bandwidth and unthrottling its
4860  * cfs_rqs as appropriate. If there has been no activity within the last
4861  * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4862  * used to track this state.
4863  */
4864 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4865 {
4866         u64 runtime, runtime_expires;
4867         int throttled;
4868 
4869         /* no need to continue the timer with no bandwidth constraint */
4870         if (cfs_b->quota == RUNTIME_INF)
4871                 goto out_deactivate;
4872 
4873         throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4874         cfs_b->nr_periods += overrun;
4875 
4876         /*
4877          * idle depends on !throttled (for the case of a large deficit), and if
4878          * we're going inactive then everything else can be deferred
4879          */
4880         if (cfs_b->idle && !throttled)
4881                 goto out_deactivate;
4882 
4883         __refill_cfs_bandwidth_runtime(cfs_b);
4884 
4885         if (!throttled) {
4886                 /* mark as potentially idle for the upcoming period */
4887                 cfs_b->idle = 1;
4888                 return 0;
4889         }
4890 
4891         /* account preceding periods in which throttling occurred */
4892         cfs_b->nr_throttled += overrun;
4893 
4894         runtime_expires = cfs_b->runtime_expires;
4895 
4896         /*
4897          * This check is repeated as we are holding onto the new bandwidth while
4898          * we unthrottle. This can potentially race with an unthrottled group
4899          * trying to acquire new bandwidth from the global pool. This can result
4900          * in us over-using our runtime if it is all used during this loop, but
4901          * only by limited amounts in that extreme case.
4902          */
4903         while (throttled && cfs_b->runtime > 0) {
4904                 runtime = cfs_b->runtime;
4905                 raw_spin_unlock(&cfs_b->lock);
4906                 /* we can't nest cfs_b->lock while distributing bandwidth */
4907                 runtime = distribute_cfs_runtime(cfs_b, runtime,
4908                                                  runtime_expires);
4909                 raw_spin_lock(&cfs_b->lock);
4910 
4911                 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4912 
4913                 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4914         }
4915 
4916         /*
4917          * While we are ensured activity in the period following an
4918          * unthrottle, this also covers the case in which the new bandwidth is
4919          * insufficient to cover the existing bandwidth deficit.  (Forcing the
4920          * timer to remain active while there are any throttled entities.)
4921          */
4922         cfs_b->idle = 0;
4923 
4924         return 0;
4925 
4926 out_deactivate:
4927         return 1;
4928 }
4929 
4930 /* a cfs_rq won't donate quota below this amount */
4931 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4932 /* minimum remaining period time to redistribute slack quota */
4933 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4934 /* how long we wait to gather additional slack before distributing */
4935 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4936 
4937 /*
4938  * Are we near the end of the current quota period?
4939  *
4940  * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4941  * hrtimer base being cleared by hrtimer_start. In the case of
4942  * migrate_hrtimers, base is never cleared, so we are fine.
4943  */
4944 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4945 {
4946         struct hrtimer *refresh_timer = &cfs_b->period_timer;
4947         u64 remaining;
4948 
4949         /* if the call-back is running a quota refresh is already occurring */
4950         if (hrtimer_callback_running(refresh_timer))
4951                 return 1;
4952 
4953         /* is a quota refresh about to occur? */
4954         remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4955         if (remaining < min_expire)
4956                 return 1;
4957 
4958         return 0;
4959 }
4960 
4961 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4962 {
4963         u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4964 
4965         /* if there's a quota refresh soon don't bother with slack */
4966         if (runtime_refresh_within(cfs_b, min_left))
4967                 return;
4968 
4969         hrtimer_start(&cfs_b->slack_timer,
4970                         ns_to_ktime(cfs_bandwidth_slack_period),
4971                         HRTIMER_MODE_REL);
4972 }
4973 
4974 /* we know any runtime found here is valid as update_curr() precedes return */
4975 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4976 {
4977         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4978         s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4979 
4980         if (slack_runtime <= 0)
4981                 return;
4982 
4983         raw_spin_lock(&cfs_b->lock);
4984         if (cfs_b->quota != RUNTIME_INF &&
4985             cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4986                 cfs_b->runtime += slack_runtime;
4987 
4988                 /* we are under rq->lock, defer unthrottling using a timer */
4989                 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4990                     !list_empty(&cfs_b->throttled_cfs_rq))
4991                         start_cfs_slack_bandwidth(cfs_b);
4992         }
4993         raw_spin_unlock(&cfs_b->lock);
4994 
4995         /* even if it's not valid for return we don't want to try again */
4996         cfs_rq->runtime_remaining -= slack_runtime;
4997 }
4998 
4999 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5000 {
5001         if (!cfs_bandwidth_used())
5002                 return;
5003 
5004         if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
5005                 return;
5006 
5007         __return_cfs_rq_runtime(cfs_rq);
5008 }
5009 
5010 /*
5011  * This is done with a timer (instead of inline with bandwidth return) since
5012  * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
5013  */
5014 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
5015 {
5016         u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
5017         u64 expires;
5018 
5019         /* confirm we're still not at a refresh boundary */
5020         raw_spin_lock(&cfs_b->lock);
5021         if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
5022                 raw_spin_unlock(&cfs_b->lock);
5023                 return;
5024         }
5025 
5026         if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
5027                 runtime = cfs_b->runtime;
5028 
5029         expires = cfs_b->runtime_expires;
5030         raw_spin_unlock(&cfs_b->lock);
5031 
5032         if (!runtime)
5033                 return;
5034 
5035         runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
5036 
5037         raw_spin_lock(&cfs_b->lock);
5038         if (expires == cfs_b->runtime_expires)
5039                 cfs_b->runtime -= min(runtime, cfs_b->runtime);
5040         raw_spin_unlock(&cfs_b->lock);
5041 }
5042 
5043 /*
5044  * When a group wakes up we want to make sure that its quota is not already
5045  * expired/exceeded, otherwise it may be allowed to steal additional ticks of
5046  * runtime as update_curr() throttling can not not trigger until it's on-rq.
5047  */
5048 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
5049 {
5050         if (!cfs_bandwidth_used())
5051                 return;
5052 
5053         /* an active group must be handled by the update_curr()->put() path */
5054         if (!cfs_rq->runtime_enabled || cfs_rq->curr)
5055                 return;
5056 
5057         /* ensure the group is not already throttled */
5058         if (cfs_rq_throttled(cfs_rq))
5059                 return;
5060 
5061         /* update runtime allocation */
5062         account_cfs_rq_runtime(cfs_rq, 0);
5063         if (cfs_rq->runtime_remaining <= 0)
5064                 throttle_cfs_rq(cfs_rq);
5065 }
5066 
5067 static void sync_throttle(struct task_group *tg, int cpu)
5068 {
5069         struct cfs_rq *pcfs_rq, *cfs_rq;
5070 
5071         if (!cfs_bandwidth_used())
5072                 return;
5073 
5074         if (!tg->parent)
5075                 return;
5076 
5077         cfs_rq = tg->cfs_rq[cpu];
5078         pcfs_rq = tg->parent->cfs_rq[cpu];
5079 
5080         cfs_rq->throttle_count = pcfs_rq->throttle_count;
5081         cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
5082 }
5083 
5084 /* conditionally throttle active cfs_rq's from put_prev_entity() */
5085 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5086 {
5087         if (!cfs_bandwidth_used())
5088                 return false;
5089 
5090         if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
5091                 return false;
5092 
5093         /*
5094          * it's possible for a throttled entity to be forced into a running
5095          * state (e.g. set_curr_task), in this case we're finished.
5096          */
5097         if (cfs_rq_throttled(cfs_rq))
5098                 return true;
5099 
5100         throttle_cfs_rq(cfs_rq);
5101         return true;
5102 }
5103 
5104 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
5105 {
5106         struct cfs_bandwidth *cfs_b =
5107                 container_of(timer, struct cfs_bandwidth, slack_timer);
5108 
5109         do_sched_cfs_slack_timer(cfs_b);
5110 
5111         return HRTIMER_NORESTART;
5112 }
5113 
5114 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
5115 {
5116         struct cfs_bandwidth *cfs_b =
5117                 container_of(timer, struct cfs_bandwidth, period_timer);
5118         int overrun;
5119         int idle = 0;
5120 
5121         raw_spin_lock(&cfs_b->lock);
5122         for (;;) {
5123                 overrun = hrtimer_forward_now(timer, cfs_b->period);
5124                 if (!overrun)
5125                         break;
5126 
5127                 idle = do_sched_cfs_period_timer(cfs_b, overrun);
5128         }
5129         if (idle)
5130                 cfs_b->period_active = 0;
5131         raw_spin_unlock(&cfs_b->lock);
5132 
5133         return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
5134 }
5135 
5136 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5137 {
5138         raw_spin_lock_init(&cfs_b->lock);
5139         cfs_b->runtime = 0;
5140         cfs_b->quota = RUNTIME_INF;
5141         cfs_b->period = ns_to_ktime(default_cfs_period());
5142 
5143         INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
5144         hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
5145         cfs_b->period_timer.function = sched_cfs_period_timer;
5146         hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5147         cfs_b->slack_timer.function = sched_cfs_slack_timer;
5148 }
5149 
5150 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5151 {
5152         cfs_rq->runtime_enabled = 0;
5153         INIT_LIST_HEAD(&cfs_rq->throttled_list);
5154 }
5155 
5156 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5157 {
5158         lockdep_assert_held(&cfs_b->lock);
5159 
5160         if (!cfs_b->period_active) {
5161                 cfs_b->period_active = 1;
5162                 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
5163                 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
5164         }
5165 }
5166 
5167 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5168 {
5169         /* init_cfs_bandwidth() was not called */
5170         if (!cfs_b->throttled_cfs_rq.next)
5171                 return;
5172 
5173         hrtimer_cancel(&cfs_b->period_timer);
5174         hrtimer_cancel(&cfs_b->slack_timer);
5175 }
5176 
5177 /*
5178  * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5179  *
5180  * The race is harmless, since modifying bandwidth settings of unhooked group
5181  * bits doesn't do much.
5182  */
5183 
5184 /* cpu online calback */
5185 static void __maybe_unused update_runtime_enabled(struct rq *rq)
5186 {
5187         struct task_group *tg;
5188 
5189         lockdep_assert_held(&rq->lock);
5190 
5191         rcu_read_lock();
5192         list_for_each_entry_rcu(tg, &task_groups, list) {
5193                 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5194                 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5195 
5196                 raw_spin_lock(&cfs_b->lock);
5197                 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5198                 raw_spin_unlock(&cfs_b->lock);
5199         }
5200         rcu_read_unlock();
5201 }
5202 
5203 /* cpu offline callback */
5204 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5205 {
5206         struct task_group *tg;
5207 
5208         lockdep_assert_held(&rq->lock);
5209 
5210         rcu_read_lock();
5211         list_for_each_entry_rcu(tg, &task_groups, list) {
5212                 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5213 
5214                 if (!cfs_rq->runtime_enabled)
5215                         continue;
5216 
5217                 /*
5218                  * clock_task is not advancing so we just need to make sure
5219                  * there's some valid quota amount
5220                  */
5221                 cfs_rq->runtime_remaining = 1;
5222                 /*
5223                  * Offline rq is schedulable till CPU is completely disabled
5224                  * in take_cpu_down(), so we prevent new cfs throttling here.
5225                  */
5226                 cfs_rq->runtime_enabled = 0;
5227 
5228                 if (cfs_rq_throttled(cfs_rq))
5229                         unthrottle_cfs_rq(cfs_rq);
5230         }
5231         rcu_read_unlock();
5232 }
5233 
5234 #else /* CONFIG_CFS_BANDWIDTH */
5235 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
5236 {
5237         return rq_clock_task(rq_of(cfs_rq));
5238 }
5239 
5240 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5241 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5242 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5243 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5244 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5245 
5246 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5247 {
5248         return 0;
5249 }
5250 
5251 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5252 {
5253         return 0;
5254 }
5255 
5256 static inline int throttled_lb_pair(struct task_group *tg,
5257                                     int src_cpu, int dest_cpu)
5258 {
5259         return 0;
5260 }
5261 
5262 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5263 
5264 #ifdef CONFIG_FAIR_GROUP_SCHED
5265 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5266 #endif
5267 
5268 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5269 {
5270         return NULL;
5271 }
5272 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5273 static inline void update_runtime_enabled(struct rq *rq) {}
5274 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5275 
5276 #endif /* CONFIG_CFS_BANDWIDTH */
5277 
5278 /**************************************************
5279  * CFS operations on tasks:
5280  */
5281 
5282 #ifdef CONFIG_SCHED_HRTICK
5283 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5284 {
5285         struct sched_entity *se = &p->se;
5286         struct cfs_rq *cfs_rq = cfs_rq_of(se);
5287 
5288         SCHED_WARN_ON(task_rq(p) != rq);
5289 
5290         if (rq->cfs.h_nr_running > 1) {
5291                 u64 slice = sched_slice(cfs_rq, se);
5292                 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5293                 s64 delta = slice - ran;
5294 
5295                 if (delta < 0) {
5296                         if (rq->curr == p)
5297                                 resched_curr(rq);
5298                         return;
5299                 }
5300                 hrtick_start(rq, delta);
5301         }
5302 }
5303 
5304 /*
5305  * called from enqueue/dequeue and updates the hrtick when the
5306  * current task is from our class and nr_running is low enough
5307  * to matter.
5308  */
5309 static void hrtick_update(struct rq *rq)
5310 {
5311         struct task_struct *curr = rq->curr;
5312 
5313         if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
5314                 return;
5315 
5316         if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5317                 hrtick_start_fair(rq, curr);
5318 }
5319 #else /* !CONFIG_SCHED_HRTICK */
5320 static inline void
5321 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5322 {
5323 }
5324 
5325 static inline void hrtick_update(struct rq *rq)
5326 {
5327 }
5328 #endif
5329 
5330 /*
5331  * The enqueue_task method is called before nr_running is
5332  * increased. Here we update the fair scheduling stats and
5333  * then put the task into the rbtree:
5334  */
5335 static void
5336 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5337 {
5338         struct cfs_rq *cfs_rq;
5339         struct sched_entity *se = &p->se;
5340 
5341         /*
5342          * If in_iowait is set, the code below may not trigger any cpufreq
5343          * utilization updates, so do it here explicitly with the IOWAIT flag
5344          * passed.
5345          */
5346         if (p->in_iowait)
5347                 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5348 
5349         for_each_sched_entity(se) {
5350                 if (se->on_rq)
5351                         break;
5352                 cfs_rq = cfs_rq_of(se);
5353                 enqueue_entity(cfs_rq, se, flags);
5354 
5355                 /*
5356                  * end evaluation on encountering a throttled cfs_rq
5357                  *
5358                  * note: in the case of encountering a throttled cfs_rq we will
5359                  * post the final h_nr_running increment below.
5360                  */
5361                 if (cfs_rq_throttled(cfs_rq))
5362                         break;
5363                 cfs_rq->h_nr_running++;
5364 
5365                 flags = ENQUEUE_WAKEUP;
5366         }
5367 
5368         for_each_sched_entity(se) {
5369                 cfs_rq = cfs_rq_of(se);
5370                 cfs_rq->h_nr_running++;
5371 
5372                 if (cfs_rq_throttled(cfs_rq))
5373                         break;
5374 
5375                 update_load_avg(cfs_rq, se, UPDATE_TG);
5376                 update_cfs_group(se);
5377         }
5378 
5379         if (!se)
5380                 add_nr_running(rq, 1);
5381 
5382         util_est_enqueue(&rq->cfs, p);
5383         hrtick_update(rq);
5384 }
5385 
5386 static void set_next_buddy(struct sched_entity *se);
5387 
5388 /*
5389  * The dequeue_task method is called before nr_running is
5390  * decreased. We remove the task from the rbtree and
5391  * update the fair scheduling stats:
5392  */
5393 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5394 {
5395         struct cfs_rq *cfs_rq;
5396         struct sched_entity *se = &p->se;
5397         int task_sleep = flags & DEQUEUE_SLEEP;
5398 
5399         for_each_sched_entity(se) {
5400                 cfs_rq = cfs_rq_of(se);
5401                 dequeue_entity(cfs_rq, se, flags);
5402 
5403                 /*
5404                  * end evaluation on encountering a throttled cfs_rq
5405                  *
5406                  * note: in the case of encountering a throttled cfs_rq we will
5407                  * post the final h_nr_running decrement below.
5408                 */
5409                 if (cfs_rq_throttled(cfs_rq))
5410                         break;
5411                 cfs_rq->h_nr_running--;
5412 
5413                 /* Don't dequeue parent if it has other entities besides us */
5414                 if (cfs_rq->load.weight) {
5415                         /* Avoid re-evaluating load for this entity: */
5416                         se = parent_entity(se);
5417                         /*
5418                          * Bias pick_next to pick a task from this cfs_rq, as
5419                          * p is sleeping when it is within its sched_slice.
5420                          */
5421                         if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5422                                 set_next_buddy(se);
5423                         break;
5424                 }
5425                 flags |= DEQUEUE_SLEEP;
5426         }
5427 
5428         for_each_sched_entity(se) {
5429                 cfs_rq = cfs_rq_of(se);
5430                 cfs_rq->h_nr_running--;
5431 
5432                 if (cfs_rq_throttled(cfs_rq))
5433                         break;
5434 
5435                 update_load_avg(cfs_rq, se, UPDATE_TG);
5436                 update_cfs_group(se);
5437         }
5438 
5439         if (!se)
5440                 sub_nr_running(rq, 1);
5441 
5442         util_est_dequeue(&rq->cfs, p, task_sleep);
5443         hrtick_update(rq);
5444 }
5445 
5446 #ifdef CONFIG_SMP
5447 
5448 /* Working cpumask for: load_balance, load_balance_newidle. */
5449 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5450 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5451 
5452 #ifdef CONFIG_NO_HZ_COMMON
5453 /*
5454  * per rq 'load' arrray crap; XXX kill this.
5455  */
5456 
5457 /*
5458  * The exact cpuload calculated at every tick would be:
5459  *
5460  *   load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
5461  *
5462  * If a CPU misses updates for n ticks (as it was idle) and update gets
5463  * called on the n+1-th tick when CPU may be busy, then we have:
5464  *
5465  *   load_n   = (1 - 1/2^i)^n * load_0
5466  *   load_n+1 = (1 - 1/2^i)   * load_n + (1/2^i) * cur_load
5467  *
5468  * decay_load_missed() below does efficient calculation of
5469  *
5470  *   load' = (1 - 1/2^i)^n * load
5471  *
5472  * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
5473  * This allows us to precompute the above in said factors, thereby allowing the
5474  * reduction of an arbitrary n in O(log_2 n) steps. (See also
5475  * fixed_power_int())
5476  *
5477  * The calculation is approximated on a 128 point scale.
5478  */
5479 #define DEGRADE_SHIFT           7
5480 
5481 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
5482 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
5483         {   0,   0,  0,  0,  0,  0, 0, 0 },
5484         {  64,  32,  8,  0,  0,  0, 0, 0 },
5485         {  96,  72, 40, 12,  1,  0, 0, 0 },
5486         { 112,  98, 75, 43, 15,  1, 0, 0 },
5487         { 120, 112, 98, 76, 45, 16, 2, 0 }
5488 };
5489 
5490 /*
5491  * Update cpu_load for any missed ticks, due to tickless idle. The backlog
5492  * would be when CPU is idle and so we just decay the old load without
5493  * adding any new load.
5494  */
5495 static unsigned long
5496 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
5497 {
5498         int j = 0;
5499 
5500         if (!missed_updates)
5501                 return load;
5502 
5503         if (missed_updates >= degrade_zero_ticks[idx])
5504                 return 0;
5505 
5506         if (idx == 1)
5507                 return load >> missed_updates;
5508 
5509         while (missed_updates) {
5510                 if (missed_updates % 2)
5511                         load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
5512 
5513                 missed_updates >>= 1;
5514                 j++;
5515         }
5516         return load;
5517 }
5518 
5519 static struct {
5520         cpumask_var_t idle_cpus_mask;
5521         atomic_t nr_cpus;
5522         int has_blocked;                /* Idle CPUS has blocked load */
5523         unsigned long next_balance;     /* in jiffy units */
5524         unsigned long next_blocked;     /* Next update of blocked load in jiffies */
5525 } nohz ____cacheline_aligned;
5526 
5527 #endif /* CONFIG_NO_HZ_COMMON */
5528 
5529 /**
5530  * __cpu_load_update - update the rq->cpu_load[] statistics
5531  * @this_rq: The rq to update statistics for
5532  * @this_load: The current load
5533  * @pending_updates: The number of missed updates
5534  *
5535  * Update rq->cpu_load[] statistics. This function is usually called every
5536  * scheduler tick (TICK_NSEC).
5537  *
5538  * This function computes a decaying average:
5539  *
5540  *   load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
5541  *
5542  * Because of NOHZ it might not get called on every tick which gives need for
5543  * the @pending_updates argument.
5544  *
5545  *   load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
5546  *             = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
5547  *             = A * (A * load[i]_n-2 + B) + B
5548  *             = A * (A * (A * load[i]_n-3 + B) + B) + B
5549  *             = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5550  *             = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5551  *             = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5552  *             = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5553  *
5554  * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5555  * any change in load would have resulted in the tick being turned back on.
5556  *
5557  * For regular NOHZ, this reduces to:
5558  *
5559  *   load[i]_n = (1 - 1/2^i)^n * load[i]_0
5560  *
5561  * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5562  * term.
5563  */
5564 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
5565                             unsigned long pending_updates)
5566 {
5567         unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
5568         int i, scale;
5569 
5570         this_rq->nr_load_updates++;
5571 
5572         /* Update our load: */
5573         this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
5574         for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
5575                 unsigned long old_load, new_load;
5576 
5577                 /* scale is effectively 1 << i now, and >> i divides by scale */
5578 
5579                 old_load = this_rq->cpu_load[i];
5580 #ifdef CONFIG_NO_HZ_COMMON
5581                 old_load = decay_load_missed(old_load, pending_updates - 1, i);
5582                 if (tickless_load) {
5583                         old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
5584                         /*
5585                          * old_load can never be a negative value because a
5586                          * decayed tickless_load cannot be greater than the
5587                          * original tickless_load.
5588                          */
5589                         old_load += tickless_load;
5590                 }
5591 #endif
5592                 new_load = this_load;
5593                 /*
5594                  * Round up the averaging division if load is increasing. This
5595                  * prevents us from getting stuck on 9 if the load is 10, for
5596                  * example.
5597                  */
5598                 if (new_load > old_load)
5599                         new_load += scale - 1;
5600 
5601                 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
5602         }
5603 
5604         sched_avg_update(this_rq);
5605 }
5606 
5607 /* Used instead of source_load when we know the type == 0 */
5608 static unsigned long weighted_cpuload(struct rq *rq)
5609 {
5610         return cfs_rq_runnable_load_avg(&rq->cfs);
5611 }
5612 
5613 #ifdef CONFIG_NO_HZ_COMMON
5614 /*
5615  * There is no sane way to deal with nohz on smp when using jiffies because the
5616  * CPU doing the jiffies update might drift wrt the CPU doing the jiffy reading
5617  * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5618  *
5619  * Therefore we need to avoid the delta approach from the regular tick when
5620  * possible since that would seriously skew the load calculation. This is why we
5621  * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5622  * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5623  * loop exit, nohz_idle_balance, nohz full exit...)
5624  *
5625  * This means we might still be one tick off for nohz periods.
5626  */
5627 
5628 static void cpu_load_update_nohz(struct rq *this_rq,
5629                                  unsigned long curr_jiffies,
5630                                  unsigned long load)
5631 {
5632         unsigned long pending_updates;
5633 
5634         pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5635         if (pending_updates) {
5636                 this_rq->last_load_update_tick = curr_jiffies;
5637                 /*
5638                  * In the regular NOHZ case, we were idle, this means load 0.
5639                  * In the NOHZ_FULL case, we were non-idle, we should consider
5640                  * its weighted load.
5641                  */
5642                 cpu_load_update(this_rq, load, pending_updates);
5643         }
5644 }
5645 
5646 /*
5647  * Called from nohz_idle_balance() to update the load ratings before doing the
5648  * idle balance.
5649  */
5650 static void cpu_load_update_idle(struct rq *this_rq)
5651 {
5652         /*
5653          * bail if there's load or we're actually up-to-date.
5654          */
5655         if (weighted_cpuload(this_rq))
5656                 return;
5657 
5658         cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
5659 }
5660 
5661 /*
5662  * Record CPU load on nohz entry so we know the tickless load to account
5663  * on nohz exit. cpu_load[0] happens then to be updated more frequently
5664  * than other cpu_load[idx] but it should be fine as cpu_load readers
5665  * shouldn't rely into synchronized cpu_load[*] updates.
5666  */
5667 void cpu_load_update_nohz_start(void)
5668 {
5669         struct rq *this_rq = this_rq();
5670 
5671         /*
5672          * This is all lockless but should be fine. If weighted_cpuload changes
5673          * concurrently we'll exit nohz. And cpu_load write can race with
5674          * cpu_load_update_idle() but both updater would be writing the same.
5675          */
5676         this_rq->cpu_load[0] = weighted_cpuload(this_rq);
5677 }
5678 
5679 /*
5680  * Account the tickless load in the end of a nohz frame.
5681  */
5682 void cpu_load_update_nohz_stop(void)
5683 {
5684         unsigned long curr_jiffies = READ_ONCE(jiffies);
5685         struct rq *this_rq = this_rq();
5686         unsigned long load;
5687         struct rq_flags rf;
5688 
5689         if (curr_jiffies == this_rq->last_load_update_tick)
5690                 return;
5691 
5692         load = weighted_cpuload(this_rq);
5693         rq_lock(this_rq, &rf);
5694         update_rq_clock(this_rq);
5695         cpu_load_update_nohz(this_rq, curr_jiffies, load);
5696         rq_unlock(this_rq, &rf);
5697 }
5698 #else /* !CONFIG_NO_HZ_COMMON */
5699 static inline void cpu_load_update_nohz(struct rq *this_rq,
5700                                         unsigned long curr_jiffies,
5701                                         unsigned long load) { }
5702 #endif /* CONFIG_NO_HZ_COMMON */
5703 
5704 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
5705 {
5706 #ifdef CONFIG_NO_HZ_COMMON
5707         /* See the mess around cpu_load_update_nohz(). */
5708         this_rq->last_load_update_tick = READ_ONCE(jiffies);
5709 #endif
5710         cpu_load_update(this_rq, load, 1);
5711 }
5712 
5713 /*
5714  * Called from scheduler_tick()
5715  */
5716 void cpu_load_update_active(struct rq *this_rq)
5717 {
5718         unsigned long load = weighted_cpuload(this_rq);
5719 
5720         if (tick_nohz_tick_stopped())
5721                 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
5722         else
5723                 cpu_load_update_periodic(this_rq, load);
5724 }
5725 
5726 /*
5727  * Return a low guess at the load of a migration-source CPU weighted
5728  * according to the scheduling class and "nice" value.
5729  *
5730  * We want to under-estimate the load of migration sources, to
5731  * balance conservatively.
5732  */
5733 static unsigned long source_load(int cpu, int type)
5734 {
5735         struct rq *rq = cpu_rq(cpu);
5736         unsigned long total = weighted_cpuload(rq);
5737 
5738         if (type == 0 || !sched_feat(LB_BIAS))
5739                 return total;
5740 
5741         return min(rq->cpu_load[type-1], total);
5742 }
5743 
5744 /*
5745  * Return a high guess at the load of a migration-target CPU weighted
5746  * according to the scheduling class and "nice" value.
5747  */
5748 static unsigned long target_load(int cpu, int type)
5749 {
5750         struct rq *rq = cpu_rq(cpu);
5751         unsigned long total = weighted_cpuload(rq);
5752 
5753         if (type == 0 || !sched_feat(LB_BIAS))
5754                 return total;
5755 
5756         return max(rq->cpu_load[type-1], total);
5757 }
5758 
5759 static unsigned long capacity_of(int cpu)
5760 {
5761         return cpu_rq(cpu)->cpu_capacity;
5762 }
5763 
5764 static unsigned long capacity_orig_of(int cpu)
5765 {
5766         return cpu_rq(cpu)->cpu_capacity_orig;
5767 }
5768 
5769 static unsigned long cpu_avg_load_per_task(int cpu)
5770 {
5771         struct rq *rq = cpu_rq(cpu);
5772         unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5773         unsigned long load_avg = weighted_cpuload(rq);
5774 
5775         if (nr_running)
5776                 return load_avg / nr_running;
5777 
5778         return 0;
5779 }
5780 
5781 static void record_wakee(struct task_struct *p)
5782 {
5783         /*
5784          * Only decay a single time; tasks that have less then 1 wakeup per
5785          * jiffy will not have built up many flips.
5786          */
5787         if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5788                 current->wakee_flips >>= 1;
5789                 current->wakee_flip_decay_ts = jiffies;
5790         }
5791 
5792         if (current->last_wakee != p) {
5793                 current->last_wakee = p;
5794                 current->wakee_flips++;
5795         }
5796 }
5797 
5798 /*
5799  * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5800  *
5801  * A waker of many should wake a different task than the one last awakened
5802  * at a frequency roughly N times higher than one of its wakees.
5803  *
5804  * In order to determine whether we should let the load spread vs consolidating
5805  * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5806  * partner, and a factor of lls_size higher frequency in the other.
5807  *
5808  * With both conditions met, we can be relatively sure that the relationship is
5809  * non-monogamous, with partner count exceeding socket size.
5810  *
5811  * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5812  * whatever is irrelevant, spread criteria is apparent partner count exceeds
5813  * socket size.
5814  */
5815 static int wake_wide(struct task_struct *p)
5816 {
5817         unsigned int master = current->wakee_flips;
5818         unsigned int slave = p->wakee_flips;
5819         int factor = this_cpu_read(sd_llc_size);
5820 
5821         if (master < slave)
5822                 swap(master, slave);
5823         if (slave < factor || master < slave * factor)
5824                 return 0;
5825         return 1;
5826 }
5827 
5828 /*
5829  * The purpose of wake_affine() is to quickly determine on which CPU we can run
5830  * soonest. For the purpose of speed we only consider the waking and previous
5831  * CPU.
5832  *
5833  * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5834  *                      cache-affine and is (or will be) idle.
5835  *
5836  * wake_affine_weight() - considers the weight to reflect the average
5837  *                        scheduling latency of the CPUs. This seems to work
5838  *                        for the overloaded case.
5839  */
5840 static int
5841 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5842 {
5843         /*
5844          * If this_cpu is idle, it implies the wakeup is from interrupt
5845          * context. Only allow the move if cache is shared. Otherwise an
5846          * interrupt intensive workload could force all tasks onto one
5847          * node depending on the IO topology or IRQ affinity settings.
5848          *
5849          * If the prev_cpu is idle and cache affine then avoid a migration.
5850          * There is no guarantee that the cache hot data from an interrupt
5851          * is more important than cache hot data on the prev_cpu and from
5852          * a cpufreq perspective, it's better to have higher utilisation
5853          * on one CPU.
5854          */
5855         if (idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5856                 return idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5857 
5858         if (sync && cpu_rq(this_cpu)->nr_running == 1)
5859                 return this_cpu;
5860 
5861         return nr_cpumask_bits;
5862 }
5863 
5864 static int
5865 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5866                    int this_cpu, int prev_cpu, int sync)
5867 {
5868         s64 this_eff_load, prev_eff_load;
5869         unsigned long task_load;
5870 
5871         this_eff_load = target_load(this_cpu, sd->wake_idx);
5872 
5873         if (sync) {
5874                 unsigned long current_load = task_h_load(current);
5875 
5876                 if (current_load > this_eff_load)
5877                         return this_cpu;
5878 
5879                 this_eff_load -= current_load;
5880         }
5881 
5882         task_load = task_h_load(p);
5883 
5884         this_eff_load += task_load;
5885         if (sched_feat(WA_BIAS))
5886                 this_eff_load *= 100;
5887         this_eff_load *= capacity_of(prev_cpu);
5888 
5889         prev_eff_load = source_load(prev_cpu, sd->wake_idx);
5890         prev_eff_load -= task_load;
5891         if (sched_feat(WA_BIAS))
5892                 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5893         prev_eff_load *= capacity_of(this_cpu);
5894 
5895         /*
5896          * If sync, adjust the weight of prev_eff_load such that if
5897          * prev_eff == this_eff that select_idle_sibling() will consider
5898          * stacking the wakee on top of the waker if no other CPU is
5899          * idle.
5900          */
5901         if (sync)
5902                 prev_eff_load += 1;
5903 
5904         return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5905 }
5906 
5907 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5908                        int this_cpu, int prev_cpu, int sync)
5909 {
5910         int target = nr_cpumask_bits;
5911 
5912         if (sched_feat(WA_IDLE))
5913                 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5914 
5915         if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5916                 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5917 
5918         schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5919         if (target == nr_cpumask_bits)
5920                 return prev_cpu;
5921 
5922         schedstat_inc(sd->ttwu_move_affine);
5923         schedstat_inc(p->se.statistics.nr_wakeups_affine);
5924         return target;
5925 }
5926 
5927 static unsigned long cpu_util_wake(int cpu, struct task_struct *p);
5928 
5929 static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
5930 {
5931         return max_t(long, capacity_of(cpu) - cpu_util_wake(cpu, p), 0);
5932 }
5933 
5934 /*
5935  * find_idlest_group finds and returns the least busy CPU group within the
5936  * domain.
5937  *
5938  * Assumes p is allowed on at least one CPU in sd.
5939  */
5940 static struct sched_group *
5941 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5942                   int this_cpu, int sd_flag)
5943 {
5944         struct sched_group *idlest = NULL, *group = sd->groups;
5945         struct sched_group *most_spare_sg = NULL;
5946         unsigned long min_runnable_load = ULONG_MAX;
5947         unsigned long this_runnable_load = ULONG_MAX;
5948         unsigned long min_avg_load = ULONG_MAX, this_avg_load = ULONG_MAX;
5949         unsigned long most_spare = 0, this_spare = 0;
5950         int load_idx = sd->forkexec_idx;
5951         int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5952         unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5953                                 (sd->imbalance_pct-100) / 100;
5954 
5955         if (sd_flag & SD_BALANCE_WAKE)
5956                 load_idx = sd->wake_idx;
5957 
5958         do {
5959                 unsigned long load, avg_load, runnable_load;
5960                 unsigned long spare_cap, max_spare_cap;
5961                 int local_group;
5962                 int i;
5963 
5964                 /* Skip over this group if it has no CPUs allowed */
5965                 if (!cpumask_intersects(sched_group_span(group),
5966                                         &p->cpus_allowed))
5967                         continue;
5968 
5969                 local_group = cpumask_test_cpu(this_cpu,
5970                                                sched_group_span(group));
5971 
5972                 /*
5973                  * Tally up the load of all CPUs in the group and find
5974                  * the group containing the CPU with most spare capacity.
5975                  */
5976                 avg_load = 0;
5977                 runnable_load = 0;
5978                 max_spare_cap = 0;
5979 
5980                 for_each_cpu(i, sched_group_span(group)) {
5981                         /* Bias balancing toward CPUs of our domain */
5982                         if (local_group)
5983                                 load = source_load(i, load_idx);
5984                         else
5985                                 load = target_load(i, load_idx);
5986 
5987                         runnable_load += load;
5988 
5989                         avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5990 
5991                         spare_cap = capacity_spare_wake(i, p);
5992 
5993                         if (spare_cap > max_spare_cap)
5994                                 max_spare_cap = spare_cap;
5995                 }
5996 
5997                 /* Adjust by relative CPU capacity of the group */
5998                 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5999                                         group->sgc->capacity;
6000                 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
6001                                         group->sgc->capacity;
6002 
6003                 if (local_group) {
6004                         this_runnable_load = runnable_load;
6005                         this_avg_load = avg_load;
6006                         this_spare = max_spare_cap;
6007                 } else {
6008                         if (min_runnable_load > (runnable_load + imbalance)) {
6009                                 /*
6010                                  * The runnable load is significantly smaller
6011                                  * so we can pick this new CPU:
6012                                  */
6013                                 min_runnable_load = runnable_load;
6014                                 min_avg_load = avg_load;
6015                                 idlest = group;
6016                         } else if ((runnable_load < (min_runnable_load + imbalance)) &&
6017                                    (100*min_avg_load > imbalance_scale*avg_load)) {
6018                                 /*
6019                                  * The runnable loads are close so take the
6020                                  * blocked load into account through avg_load:
6021                                  */
6022                                 min_avg_load = avg_load;
6023                                 idlest = group;
6024                         }
6025 
6026                         if (most_spare < max_spare_cap) {
6027                                 most_spare = max_spare_cap;
6028                                 most_spare_sg = group;
6029                         }
6030                 }
6031         } while (group = group->next, group != sd->groups);
6032 
6033         /*
6034          * The cross-over point between using spare capacity or least load
6035          * is too conservative for high utilization tasks on partially
6036          * utilized systems if we require spare_capacity > task_util(p),
6037          * so we allow for some task stuffing by using
6038          * spare_capacity > task_util(p)/2.
6039          *
6040          * Spare capacity can't be used for fork because the utilization has
6041          * not been set yet, we must first select a rq to compute the initial
6042          * utilization.
6043          */
6044         if (sd_flag & SD_BALANCE_FORK)
6045                 goto skip_spare;
6046 
6047         if (this_spare > task_util(p) / 2 &&
6048             imbalance_scale*this_spare > 100*most_spare)
6049                 return NULL;
6050 
6051         if (most_spare > task_util(p) / 2)
6052                 return most_spare_sg;
6053 
6054 skip_spare:
6055         if (!idlest)
6056                 return NULL;
6057 
6058         /*
6059          * When comparing groups across NUMA domains, it's possible for the
6060          * local domain to be very lightly loaded relative to the remote
6061          * domains but "imbalance" skews the comparison making remote CPUs
6062          * look much more favourable. When considering cross-domain, add
6063          * imbalance to the runnable load on the remote node and consider
6064          * staying local.
6065          */
6066         if ((sd->flags & SD_NUMA) &&
6067             min_runnable_load + imbalance >= this_runnable_load)
6068                 return NULL;
6069 
6070         if (min_runnable_load > (this_runnable_load + imbalance))
6071                 return NULL;
6072 
6073         if ((this_runnable_load < (min_runnable_load + imbalance)) &&
6074              (100*this_avg_load < imbalance_scale*min_avg_load))
6075                 return NULL;
6076 
6077         return idlest;
6078 }
6079 
6080 /*
6081  * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
6082  */
6083 static int
6084 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
6085 {
6086         unsigned long load, min_load = ULONG_MAX;
6087         unsigned int min_exit_latency = UINT_MAX;
6088         u64 latest_idle_timestamp = 0;
6089         int least_loaded_cpu = this_cpu;
6090         int shallowest_idle_cpu = -1;
6091         int i;
6092 
6093         /* Check if we have any choice: */
6094         if (group->group_weight == 1)
6095                 return cpumask_first(sched_group_span(group));
6096 
6097         /* Traverse only the allowed CPUs */
6098         for_each_cpu_and(i, sched_group_span(group), &p->cpus_allowed) {
6099                 if (idle_cpu(i)) {
6100                         struct rq *rq = cpu_rq(i);
6101                         struct cpuidle_state *idle = idle_get_state(rq);
6102                         if (idle && idle->exit_latency < min_exit_latency) {
6103                                 /*
6104                                  * We give priority to a CPU whose idle state
6105                                  * has the smallest exit latency irrespective
6106                                  * of any idle timestamp.
6107                                  */
6108                                 min_exit_latency = idle->exit_latency;
6109                                 latest_idle_timestamp = rq->idle_stamp;
6110                                 shallowest_idle_cpu = i;
6111                         } else if ((!idle || idle->exit_latency == min_exit_latency) &&
6112                                    rq->idle_stamp > latest_idle_timestamp) {
6113                                 /*
6114                                  * If equal or no active idle state, then
6115                                  * the most recently idled CPU might have
6116                                  * a warmer cache.
6117                                  */
6118                                 latest_idle_timestamp = rq->idle_stamp;
6119                                 shallowest_idle_cpu = i;
6120                         }
6121                 } else if (shallowest_idle_cpu == -1) {
6122                         load = weighted_cpuload(cpu_rq(i));
6123                         if (load < min_load) {
6124                                 min_load = load;
6125                                 least_loaded_cpu = i;
6126                         }
6127                 }
6128         }
6129 
6130         return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
6131 }
6132 
6133 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
6134                                   int cpu, int prev_cpu, int sd_flag)
6135 {
6136         int new_cpu = cpu;
6137 
6138         if (!cpumask_intersects(sched_domain_span(sd), &p->cpus_allowed))
6139                 return prev_cpu;
6140 
6141         while (sd) {
6142                 struct sched_group *group;
6143                 struct sched_domain *tmp;
6144                 int weight;
6145 
6146                 if (!(sd->flags & sd_flag)) {
6147                         sd = sd->child;
6148                         continue;
6149                 }
6150 
6151                 group = find_idlest_group(sd, p, cpu, sd_flag);
6152                 if (!group) {
6153                         sd = sd->child;
6154                         continue;
6155                 }
6156 
6157                 new_cpu = find_idlest_group_cpu(group, p, cpu);
6158                 if (new_cpu == cpu) {
6159                         /* Now try balancing at a lower domain level of 'cpu': */
6160                         sd = sd->child;
6161                         continue;
6162                 }
6163 
6164                 /* Now try balancing at a lower domain level of 'new_cpu': */
6165                 cpu = new_cpu;
6166                 weight = sd->span_weight;
6167                 sd = NULL;
6168                 for_each_domain(cpu, tmp) {
6169                         if (weight <= tmp->span_weight)
6170                                 break;
6171                         if (tmp->flags & sd_flag)
6172                                 sd = tmp;
6173                 }
6174         }
6175 
6176         return new_cpu;
6177 }
6178 
6179 #ifdef CONFIG_SCHED_SMT
6180 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
6181 
6182 static inline void set_idle_cores(int cpu, int val)
6183 {
6184         struct sched_domain_shared *sds;
6185 
6186         sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6187         if (sds)
6188                 WRITE_ONCE(sds->has_idle_cores, val);
6189 }
6190 
6191 static inline bool test_idle_cores(int cpu, bool def)
6192 {
6193         struct sched_domain_shared *sds;
6194 
6195         sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6196         if (sds)
6197                 return READ_ONCE(sds->has_idle_cores);
6198 
6199         return def;
6200 }
6201 
6202 /*
6203  * Scans the local SMT mask to see if the entire core is idle, and records this
6204  * information in sd_llc_shared->has_idle_cores.
6205  *
6206  * Since SMT siblings share all cache levels, inspecting this limited remote
6207  * state should be fairly cheap.
6208  */
6209 void __update_idle_core(struct rq *rq)
6210 {
6211         int core = cpu_of(rq);
6212         int cpu;
6213 
6214         rcu_read_lock();
6215         if (test_idle_cores(core, true))
6216                 goto unlock;
6217 
6218         for_each_cpu(cpu, cpu_smt_mask(core)) {
6219                 if (cpu == core)
6220                         continue;
6221 
6222                 if (!idle_cpu(cpu))
6223                         goto unlock;
6224         }
6225 
6226         set_idle_cores(core, 1);
6227 unlock:
6228         rcu_read_unlock();
6229 }
6230 
6231 /*
6232  * Scan the entire LLC domain for idle cores; this dynamically switches off if
6233  * there are no idle cores left in the system; tracked through
6234  * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6235  */
6236 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6237 {
6238         struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6239         int core, cpu;
6240 
6241         if (!static_branch_likely(&sched_smt_present))
6242                 return -1;
6243 
6244         if (!test_idle_cores(target, false))
6245                 return -1;
6246 
6247         cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
6248 
6249         for_each_cpu_wrap(core, cpus, target) {
6250                 bool idle = true;
6251 
6252                 for_each_cpu(cpu, cpu_smt_mask(core)) {
6253                         cpumask_clear_cpu(cpu, cpus);
6254                         if (!idle_cpu(cpu))
6255                                 idle = false;
6256                 }
6257 
6258                 if (idle)
6259                         return core;
6260         }
6261 
6262         /*
6263          * Failed to find an idle core; stop looking for one.
6264          */
6265         set_idle_cores(target, 0);
6266 
6267         return -1;
6268 }
6269 
6270 /*
6271  * Scan the local SMT mask for idle CPUs.
6272  */
6273 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6274 {
6275         int cpu;
6276 
6277         if (!static_branch_likely(&sched_smt_present))
6278                 return -1;
6279 
6280         for_each_cpu(cpu, cpu_smt_mask(target)) {
6281                 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6282                         continue;
6283                 if (idle_cpu(cpu))
6284                         return cpu;
6285         }
6286 
6287         return -1;
6288 }
6289 
6290 #else /* CONFIG_SCHED_SMT */
6291 
6292 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6293 {
6294         return -1;
6295 }
6296 
6297 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6298 {
6299         return -1;
6300 }
6301 
6302 #endif /* CONFIG_SCHED_SMT */
6303 
6304 /*
6305  * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6306  * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6307  * average idle time for this rq (as found in rq->avg_idle).
6308  */
6309 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
6310 {
6311         struct sched_domain *this_sd;
6312         u64 avg_cost, avg_idle;
6313         u64 time, cost;
6314         s64 delta;
6315         int cpu, nr = INT_MAX;
6316 
6317         this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6318         if (!this_sd)
6319                 return -1;
6320 
6321         /*
6322          * Due to large variance we need a large fuzz factor; hackbench in
6323          * particularly is sensitive here.
6324          */
6325         avg_idle = this_rq()->avg_idle / 512;
6326         avg_cost = this_sd->avg_scan_cost + 1;
6327 
6328         if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
6329                 return -1;
6330 
6331         if (sched_feat(SIS_PROP)) {
6332                 u64 span_avg = sd->span_weight * avg_idle;
6333                 if (span_avg > 4*avg_cost)
6334                         nr = div_u64(span_avg, avg_cost);
6335                 else
6336                         nr = 4;
6337         }
6338 
6339         time = local_clock();
6340 
6341         for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
6342                 if (!--nr)
6343                         return -1;
6344                 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6345                         continue;
6346                 if (idle_cpu(cpu))
6347                         break;
6348         }
6349 
6350         time = local_clock() - time;
6351         cost = this_sd->avg_scan_cost;
6352         delta = (s64)(time - cost) / 8;
6353         this_sd->avg_scan_cost += delta;
6354 
6355         return cpu;
6356 }
6357 
6358 /*
6359  * Try and locate an idle core/thread in the LLC cache domain.
6360  */
6361 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6362 {
6363         struct sched_domain *sd;
6364         int i, recent_used_cpu;
6365 
6366         if (idle_cpu(target))
6367                 return target;
6368 
6369         /*
6370          * If the previous CPU is cache affine and idle, don't be stupid:
6371          */
6372         if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev))
6373                 return prev;
6374 
6375         /* Check a recently used CPU as a potential idle candidate: */
6376         recent_used_cpu = p->recent_used_cpu;
6377         if (recent_used_cpu != prev &&
6378             recent_used_cpu != target &&
6379             cpus_share_cache(recent_used_cpu, target) &&
6380             idle_cpu(recent_used_cpu) &&
6381             cpumask_test_cpu(p->recent_used_cpu, &p->cpus_allowed)) {
6382                 /*
6383                  * Replace recent_used_cpu with prev as it is a potential
6384                  * candidate for the next wake:
6385                  */
6386                 p->recent_used_cpu = prev;
6387                 return recent_used_cpu;
6388         }
6389 
6390         sd = rcu_dereference(per_cpu(sd_llc, target));
6391         if (!sd)
6392                 return target;
6393 
6394         i = select_idle_core(p, sd, target);
6395         if ((unsigned)i < nr_cpumask_bits)
6396                 return i;
6397 
6398         i = select_idle_cpu(p, sd, target);
6399         if ((unsigned)i < nr_cpumask_bits)
6400                 return i;
6401 
6402         i = select_idle_smt(p, sd, target);