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 */
40unsigned int sysctl_sched_latency = 6000000ULL;
41static 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 */
54enum 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 */
61unsigned int sysctl_sched_min_granularity = 750000ULL;
62static unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
63
64/*
65 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
66 */
67static 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 */
73unsigned 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 */
84unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
85static unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
86
87const_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 */
93int __weak arch_asym_cpu_priority(int cpu)
94{
95 return -cpu;
96}
97
98/*
99 * The margin used when comparing utilization with CPU capacity:
100 * util * margin < capacity * 1024
101 *
102 * (default: ~20%)
103 */
104static unsigned int capacity_margin = 1280;
105#endif
106
107#ifdef CONFIG_CFS_BANDWIDTH
108/*
109 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
110 * each time a cfs_rq requests quota.
111 *
112 * Note: in the case that the slice exceeds the runtime remaining (either due
113 * to consumption or the quota being specified to be smaller than the slice)
114 * we will always only issue the remaining available time.
115 *
116 * (default: 5 msec, units: microseconds)
117 */
118unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
119#endif
120
121static inline void update_load_add(struct load_weight *lw, unsigned long inc)
122{
123 lw->weight += inc;
124 lw->inv_weight = 0;
125}
126
127static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
128{
129 lw->weight -= dec;
130 lw->inv_weight = 0;
131}
132
133static 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 */
148static 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
169static 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
181void sched_init_granularity(void)
182{
183 update_sysctl();
184}
185
186#define WMULT_CONST (~0U)
187#define WMULT_SHIFT 32
188
189static 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 */
218static 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
244const struct sched_class fair_sched_class;
245
246/**************************************************************
247 * CFS operations on generic schedulable entities:
248 */
249
250#ifdef CONFIG_FAIR_GROUP_SCHED
251static inline struct task_struct *task_of(struct sched_entity *se)
252{
253 SCHED_WARN_ON(!entity_is_task(se));
254 return container_of(se, struct task_struct, se);
255}
256
257/* Walk up scheduling entities hierarchy */
258#define for_each_sched_entity(se) \
259 for (; se; se = se->parent)
260
261static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
262{
263 return p->se.cfs_rq;
264}
265
266/* runqueue on which this entity is (to be) queued */
267static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
268{
269 return se->cfs_rq;
270}
271
272/* runqueue "owned" by this group */
273static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
274{
275 return grp->my_q;
276}
277
278static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
279{
280 struct rq *rq = rq_of(cfs_rq);
281 int cpu = cpu_of(rq);
282
283 if (cfs_rq->on_list)
284 return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
285
286 cfs_rq->on_list = 1;
287
288 /*
289 * Ensure we either appear before our parent (if already
290 * enqueued) or force our parent to appear after us when it is
291 * enqueued. The fact that we always enqueue bottom-up
292 * reduces this to two cases and a special case for the root
293 * cfs_rq. Furthermore, it also means that we will always reset
294 * tmp_alone_branch either when the branch is connected
295 * to a tree or when we reach the top of the tree
296 */
297 if (cfs_rq->tg->parent &&
298 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
299 /*
300 * If parent is already on the list, we add the child
301 * just before. Thanks to circular linked property of
302 * the list, this means to put the child at the tail
303 * of the list that starts by parent.
304 */
305 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
306 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
307 /*
308 * The branch is now connected to its tree so we can
309 * reset tmp_alone_branch to the beginning of the
310 * list.
311 */
312 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
313 return true;
314 }
315
316 if (!cfs_rq->tg->parent) {
317 /*
318 * cfs rq without parent should be put
319 * at the tail of the list.
320 */
321 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
322 &rq->leaf_cfs_rq_list);
323 /*
324 * We have reach the top of a tree so we can reset
325 * tmp_alone_branch to the beginning of the list.
326 */
327 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
328 return true;
329 }
330
331 /*
332 * The parent has not already been added so we want to
333 * make sure that it will be put after us.
334 * tmp_alone_branch points to the begin of the branch
335 * where we will add parent.
336 */
337 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
338 /*
339 * update tmp_alone_branch to points to the new begin
340 * of the branch
341 */
342 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
343 return false;
344}
345
346static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
347{
348 if (cfs_rq->on_list) {
349 struct rq *rq = rq_of(cfs_rq);
350
351 /*
352 * With cfs_rq being unthrottled/throttled during an enqueue,
353 * it can happen the tmp_alone_branch points the a leaf that
354 * we finally want to del. In this case, tmp_alone_branch moves
355 * to the prev element but it will point to rq->leaf_cfs_rq_list
356 * at the end of the enqueue.
357 */
358 if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
359 rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
360
361 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
362 cfs_rq->on_list = 0;
363 }
364}
365
366static inline void assert_list_leaf_cfs_rq(struct rq *rq)
367{
368 SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
369}
370
371/* Iterate thr' all leaf cfs_rq's on a runqueue */
372#define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
373 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
374 leaf_cfs_rq_list)
375
376/* Do the two (enqueued) entities belong to the same group ? */
377static inline struct cfs_rq *
378is_same_group(struct sched_entity *se, struct sched_entity *pse)
379{
380 if (se->cfs_rq == pse->cfs_rq)
381 return se->cfs_rq;
382
383 return NULL;
384}
385
386static inline struct sched_entity *parent_entity(struct sched_entity *se)
387{
388 return se->parent;
389}
390
391static void
392find_matching_se(struct sched_entity **se, struct sched_entity **pse)
393{
394 int se_depth, pse_depth;
395
396 /*
397 * preemption test can be made between sibling entities who are in the
398 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
399 * both tasks until we find their ancestors who are siblings of common
400 * parent.
401 */
402
403 /* First walk up until both entities are at same depth */
404 se_depth = (*se)->depth;
405 pse_depth = (*pse)->depth;
406
407 while (se_depth > pse_depth) {
408 se_depth--;
409 *se = parent_entity(*se);
410 }
411
412 while (pse_depth > se_depth) {
413 pse_depth--;
414 *pse = parent_entity(*pse);
415 }
416
417 while (!is_same_group(*se, *pse)) {
418 *se = parent_entity(*se);
419 *pse = parent_entity(*pse);
420 }
421}
422
423#else /* !CONFIG_FAIR_GROUP_SCHED */
424
425static inline struct task_struct *task_of(struct sched_entity *se)
426{
427 return container_of(se, struct task_struct, se);
428}
429
430#define for_each_sched_entity(se) \
431 for (; se; se = NULL)
432
433static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
434{
435 return &task_rq(p)->cfs;
436}
437
438static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
439{
440 struct task_struct *p = task_of(se);
441 struct rq *rq = task_rq(p);
442
443 return &rq->cfs;
444}
445
446/* runqueue "owned" by this group */
447static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
448{
449 return NULL;
450}
451
452static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
453{
454 return true;
455}
456
457static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
458{
459}
460
461static inline void assert_list_leaf_cfs_rq(struct rq *rq)
462{
463}
464
465#define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
466 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
467
468static inline struct sched_entity *parent_entity(struct sched_entity *se)
469{
470 return NULL;
471}
472
473static inline void
474find_matching_se(struct sched_entity **se, struct sched_entity **pse)
475{
476}
477
478#endif /* CONFIG_FAIR_GROUP_SCHED */
479
480static __always_inline
481void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
482
483/**************************************************************
484 * Scheduling class tree data structure manipulation methods:
485 */
486
487static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
488{
489 s64 delta = (s64)(vruntime - max_vruntime);
490 if (delta > 0)
491 max_vruntime = vruntime;
492
493 return max_vruntime;
494}
495
496static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
497{
498 s64 delta = (s64)(vruntime - min_vruntime);
499 if (delta < 0)
500 min_vruntime = vruntime;
501
502 return min_vruntime;
503}
504
505static inline int entity_before(struct sched_entity *a,
506 struct sched_entity *b)
507{
508 return (s64)(a->vruntime - b->vruntime) < 0;
509}
510
511static void update_min_vruntime(struct cfs_rq *cfs_rq)
512{
513 struct sched_entity *curr = cfs_rq->curr;
514 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
515
516 u64 vruntime = cfs_rq->min_vruntime;
517
518 if (curr) {
519 if (curr->on_rq)
520 vruntime = curr->vruntime;
521 else
522 curr = NULL;
523 }
524
525 if (leftmost) { /* non-empty tree */
526 struct sched_entity *se;
527 se = rb_entry(leftmost, struct sched_entity, run_node);
528
529 if (!curr)
530 vruntime = se->vruntime;
531 else
532 vruntime = min_vruntime(vruntime, se->vruntime);
533 }
534
535 /* ensure we never gain time by being placed backwards. */
536 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
537#ifndef CONFIG_64BIT
538 smp_wmb();
539 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
540#endif
541}
542
543/*
544 * Enqueue an entity into the rb-tree:
545 */
546static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
547{
548 struct rb_node **link = &cfs_rq->tasks_timeline.rb_root.rb_node;
549 struct rb_node *parent = NULL;
550 struct sched_entity *entry;
551 bool leftmost = true;
552
553 /*
554 * Find the right place in the rbtree:
555 */
556 while (*link) {
557 parent = *link;
558 entry = rb_entry(parent, struct sched_entity, run_node);
559 /*
560 * We dont care about collisions. Nodes with
561 * the same key stay together.
562 */
563 if (entity_before(se, entry)) {
564 link = &parent->rb_left;
565 } else {
566 link = &parent->rb_right;
567 leftmost = false;
568 }
569 }
570
571 rb_link_node(&se->run_node, parent, link);
572 rb_insert_color_cached(&se->run_node,
573 &cfs_rq->tasks_timeline, leftmost);
574}
575
576static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
577{
578 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
579}
580
581struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
582{
583 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
584
585 if (!left)
586 return NULL;
587
588 return rb_entry(left, struct sched_entity, run_node);
589}
590
591static struct sched_entity *__pick_next_entity(struct sched_entity *se)
592{
593 struct rb_node *next = rb_next(&se->run_node);
594
595 if (!next)
596 return NULL;
597
598 return rb_entry(next, struct sched_entity, run_node);
599}
600
601#ifdef CONFIG_SCHED_DEBUG
602struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
603{
604 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
605
606 if (!last)
607 return NULL;
608
609 return rb_entry(last, struct sched_entity, run_node);
610}
611
612/**************************************************************
613 * Scheduling class statistics methods:
614 */
615
616int sched_proc_update_handler(struct ctl_table *table, int write,
617 void __user *buffer, size_t *lenp,
618 loff_t *ppos)
619{
620 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
621 unsigned int factor = get_update_sysctl_factor();
622
623 if (ret || !write)
624 return ret;
625
626 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
627 sysctl_sched_min_granularity);
628
629#define WRT_SYSCTL(name) \
630 (normalized_sysctl_##name = sysctl_##name / (factor))
631 WRT_SYSCTL(sched_min_granularity);
632 WRT_SYSCTL(sched_latency);
633 WRT_SYSCTL(sched_wakeup_granularity);
634#undef WRT_SYSCTL
635
636 return 0;
637}
638#endif
639
640/*
641 * delta /= w
642 */
643static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
644{
645 if (unlikely(se->load.weight != NICE_0_LOAD))
646 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
647
648 return delta;
649}
650
651/*
652 * The idea is to set a period in which each task runs once.
653 *
654 * When there are too many tasks (sched_nr_latency) we have to stretch
655 * this period because otherwise the slices get too small.
656 *
657 * p = (nr <= nl) ? l : l*nr/nl
658 */
659static u64 __sched_period(unsigned long nr_running)
660{
661 if (unlikely(nr_running > sched_nr_latency))
662 return nr_running * sysctl_sched_min_granularity;
663 else
664 return sysctl_sched_latency;
665}
666
667/*
668 * We calculate the wall-time slice from the period by taking a part
669 * proportional to the weight.
670 *
671 * s = p*P[w/rw]
672 */
673static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
674{
675 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
676
677 for_each_sched_entity(se) {
678 struct load_weight *load;
679 struct load_weight lw;
680
681 cfs_rq = cfs_rq_of(se);
682 load = &cfs_rq->load;
683
684 if (unlikely(!se->on_rq)) {
685 lw = cfs_rq->load;
686
687 update_load_add(&lw, se->load.weight);
688 load = &lw;
689 }
690 slice = __calc_delta(slice, se->load.weight, load);
691 }
692 return slice;
693}
694
695/*
696 * We calculate the vruntime slice of a to-be-inserted task.
697 *
698 * vs = s/w
699 */
700static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
701{
702 return calc_delta_fair(sched_slice(cfs_rq, se), se);
703}
704
705#include "pelt.h"
706#ifdef CONFIG_SMP
707
708static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
709static unsigned long task_h_load(struct task_struct *p);
710static unsigned long capacity_of(int cpu);
711
712/* Give new sched_entity start runnable values to heavy its load in infant time */
713void init_entity_runnable_average(struct sched_entity *se)
714{
715 struct sched_avg *sa = &se->avg;
716
717 memset(sa, 0, sizeof(*sa));
718
719 /*
720 * Tasks are initialized with full load to be seen as heavy tasks until
721 * they get a chance to stabilize to their real load level.
722 * Group entities are initialized with zero load to reflect the fact that
723 * nothing has been attached to the task group yet.
724 */
725 if (entity_is_task(se))
726 sa->runnable_load_avg = sa->load_avg = scale_load_down(se->load.weight);
727
728 se->runnable_weight = se->load.weight;
729
730 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
731}
732
733static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
734static void attach_entity_cfs_rq(struct sched_entity *se);
735
736/*
737 * With new tasks being created, their initial util_avgs are extrapolated
738 * based on the cfs_rq's current util_avg:
739 *
740 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
741 *
742 * However, in many cases, the above util_avg does not give a desired
743 * value. Moreover, the sum of the util_avgs may be divergent, such
744 * as when the series is a harmonic series.
745 *
746 * To solve this problem, we also cap the util_avg of successive tasks to
747 * only 1/2 of the left utilization budget:
748 *
749 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
750 *
751 * where n denotes the nth task and cpu_scale the CPU capacity.
752 *
753 * For example, for a CPU with 1024 of capacity, a simplest series from
754 * the beginning would be like:
755 *
756 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
757 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
758 *
759 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
760 * if util_avg > util_avg_cap.
761 */
762void post_init_entity_util_avg(struct task_struct *p)
763{
764 struct sched_entity *se = &p->se;
765 struct cfs_rq *cfs_rq = cfs_rq_of(se);
766 struct sched_avg *sa = &se->avg;
767 long cpu_scale = arch_scale_cpu_capacity(NULL, cpu_of(rq_of(cfs_rq)));
768 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
769
770 if (cap > 0) {
771 if (cfs_rq->avg.util_avg != 0) {
772 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
773 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
774
775 if (sa->util_avg > cap)
776 sa->util_avg = cap;
777 } else {
778 sa->util_avg = cap;
779 }
780 }
781
782 if (p->sched_class != &fair_sched_class) {
783 /*
784 * For !fair tasks do:
785 *
786 update_cfs_rq_load_avg(now, cfs_rq);
787 attach_entity_load_avg(cfs_rq, se, 0);
788 switched_from_fair(rq, p);
789 *
790 * such that the next switched_to_fair() has the
791 * expected state.
792 */
793 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
794 return;
795 }
796
797 attach_entity_cfs_rq(se);
798}
799
800#else /* !CONFIG_SMP */
801void init_entity_runnable_average(struct sched_entity *se)
802{
803}
804void post_init_entity_util_avg(struct task_struct *p)
805{
806}
807static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
808{
809}
810#endif /* CONFIG_SMP */
811
812/*
813 * Update the current task's runtime statistics.
814 */
815static void update_curr(struct cfs_rq *cfs_rq)
816{
817 struct sched_entity *curr = cfs_rq->curr;
818 u64 now = rq_clock_task(rq_of(cfs_rq));
819 u64 delta_exec;
820
821 if (unlikely(!curr))
822 return;
823
824 delta_exec = now - curr->exec_start;
825 if (unlikely((s64)delta_exec <= 0))
826 return;
827
828 curr->exec_start = now;
829
830 schedstat_set(curr->statistics.exec_max,
831 max(delta_exec, curr->statistics.exec_max));
832
833 curr->sum_exec_runtime += delta_exec;
834 schedstat_add(cfs_rq->exec_clock, delta_exec);
835
836 curr->vruntime += calc_delta_fair(delta_exec, curr);
837 update_min_vruntime(cfs_rq);
838
839 if (entity_is_task(curr)) {
840 struct task_struct *curtask = task_of(curr);
841
842 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
843 cgroup_account_cputime(curtask, delta_exec);
844 account_group_exec_runtime(curtask, delta_exec);
845 }
846
847 account_cfs_rq_runtime(cfs_rq, delta_exec);
848}
849
850static void update_curr_fair(struct rq *rq)
851{
852 update_curr(cfs_rq_of(&rq->curr->se));
853}
854
855static inline void
856update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
857{
858 u64 wait_start, prev_wait_start;
859
860 if (!schedstat_enabled())
861 return;
862
863 wait_start = rq_clock(rq_of(cfs_rq));
864 prev_wait_start = schedstat_val(se->statistics.wait_start);
865
866 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
867 likely(wait_start > prev_wait_start))
868 wait_start -= prev_wait_start;
869
870 __schedstat_set(se->statistics.wait_start, wait_start);
871}
872
873static inline void
874update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
875{
876 struct task_struct *p;
877 u64 delta;
878
879 if (!schedstat_enabled())
880 return;
881
882 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
883
884 if (entity_is_task(se)) {
885 p = task_of(se);
886 if (task_on_rq_migrating(p)) {
887 /*
888 * Preserve migrating task's wait time so wait_start
889 * time stamp can be adjusted to accumulate wait time
890 * prior to migration.
891 */
892 __schedstat_set(se->statistics.wait_start, delta);
893 return;
894 }
895 trace_sched_stat_wait(p, delta);
896 }
897
898 __schedstat_set(se->statistics.wait_max,
899 max(schedstat_val(se->statistics.wait_max), delta));
900 __schedstat_inc(se->statistics.wait_count);
901 __schedstat_add(se->statistics.wait_sum, delta);
902 __schedstat_set(se->statistics.wait_start, 0);
903}
904
905static inline void
906update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
907{
908 struct task_struct *tsk = NULL;
909 u64 sleep_start, block_start;
910
911 if (!schedstat_enabled())
912 return;
913
914 sleep_start = schedstat_val(se->statistics.sleep_start);
915 block_start = schedstat_val(se->statistics.block_start);
916
917 if (entity_is_task(se))
918 tsk = task_of(se);
919
920 if (sleep_start) {
921 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
922
923 if ((s64)delta < 0)
924 delta = 0;
925
926 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
927 __schedstat_set(se->statistics.sleep_max, delta);
928
929 __schedstat_set(se->statistics.sleep_start, 0);
930 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
931
932 if (tsk) {
933 account_scheduler_latency(tsk, delta >> 10, 1);
934 trace_sched_stat_sleep(tsk, delta);
935 }
936 }
937 if (block_start) {
938 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
939
940 if ((s64)delta < 0)
941 delta = 0;
942
943 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
944 __schedstat_set(se->statistics.block_max, delta);
945
946 __schedstat_set(se->statistics.block_start, 0);
947 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
948
949 if (tsk) {
950 if (tsk->in_iowait) {
951 __schedstat_add(se->statistics.iowait_sum, delta);
952 __schedstat_inc(se->statistics.iowait_count);
953 trace_sched_stat_iowait(tsk, delta);
954 }
955
956 trace_sched_stat_blocked(tsk, delta);
957
958 /*
959 * Blocking time is in units of nanosecs, so shift by
960 * 20 to get a milliseconds-range estimation of the
961 * amount of time that the task spent sleeping:
962 */
963 if (unlikely(prof_on == SLEEP_PROFILING)) {
964 profile_hits(SLEEP_PROFILING,
965 (void *)get_wchan(tsk),
966 delta >> 20);
967 }
968 account_scheduler_latency(tsk, delta >> 10, 0);
969 }
970 }
971}
972
973/*
974 * Task is being enqueued - update stats:
975 */
976static inline void
977update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
978{
979 if (!schedstat_enabled())
980 return;
981
982 /*
983 * Are we enqueueing a waiting task? (for current tasks
984 * a dequeue/enqueue event is a NOP)
985 */
986 if (se != cfs_rq->curr)
987 update_stats_wait_start(cfs_rq, se);
988
989 if (flags & ENQUEUE_WAKEUP)
990 update_stats_enqueue_sleeper(cfs_rq, se);
991}
992
993static inline void
994update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
995{
996
997 if (!schedstat_enabled())
998 return;
999
1000 /*
1001 * Mark the end of the wait period if dequeueing a
1002 * waiting task:
1003 */
1004 if (se != cfs_rq->curr)
1005 update_stats_wait_end(cfs_rq, se);
1006
1007 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1008 struct task_struct *tsk = task_of(se);
1009
1010 if (tsk->state & TASK_INTERRUPTIBLE)
1011 __schedstat_set(se->statistics.sleep_start,
1012 rq_clock(rq_of(cfs_rq)));
1013 if (tsk->state & TASK_UNINTERRUPTIBLE)
1014 __schedstat_set(se->statistics.block_start,
1015 rq_clock(rq_of(cfs_rq)));
1016 }
1017}
1018
1019/*
1020 * We are picking a new current task - update its stats:
1021 */
1022static inline void
1023update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1024{
1025 /*
1026 * We are starting a new run period:
1027 */
1028 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1029}
1030
1031/**************************************************
1032 * Scheduling class queueing methods:
1033 */
1034
1035#ifdef CONFIG_NUMA_BALANCING
1036/*
1037 * Approximate time to scan a full NUMA task in ms. The task scan period is
1038 * calculated based on the tasks virtual memory size and
1039 * numa_balancing_scan_size.
1040 */
1041unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1042unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1043
1044/* Portion of address space to scan in MB */
1045unsigned int sysctl_numa_balancing_scan_size = 256;
1046
1047/* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1048unsigned int sysctl_numa_balancing_scan_delay = 1000;
1049
1050struct numa_group {
1051 refcount_t refcount;
1052
1053 spinlock_t lock; /* nr_tasks, tasks */
1054 int nr_tasks;
1055 pid_t gid;
1056 int active_nodes;
1057
1058 struct rcu_head rcu;
1059 unsigned long total_faults;
1060 unsigned long max_faults_cpu;
1061 /*
1062 * Faults_cpu is used to decide whether memory should move
1063 * towards the CPU. As a consequence, these stats are weighted
1064 * more by CPU use than by memory faults.
1065 */
1066 unsigned long *faults_cpu;
1067 unsigned long faults[0];
1068};
1069
1070static inline unsigned long group_faults_priv(struct numa_group *ng);
1071static inline unsigned long group_faults_shared(struct numa_group *ng);
1072
1073static unsigned int task_nr_scan_windows(struct task_struct *p)
1074{
1075 unsigned long rss = 0;
1076 unsigned long nr_scan_pages;
1077
1078 /*
1079 * Calculations based on RSS as non-present and empty pages are skipped
1080 * by the PTE scanner and NUMA hinting faults should be trapped based
1081 * on resident pages
1082 */
1083 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1084 rss = get_mm_rss(p->mm);
1085 if (!rss)
1086 rss = nr_scan_pages;
1087
1088 rss = round_up(rss, nr_scan_pages);
1089 return rss / nr_scan_pages;
1090}
1091
1092/* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1093#define MAX_SCAN_WINDOW 2560
1094
1095static unsigned int task_scan_min(struct task_struct *p)
1096{
1097 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1098 unsigned int scan, floor;
1099 unsigned int windows = 1;
1100
1101 if (scan_size < MAX_SCAN_WINDOW)
1102 windows = MAX_SCAN_WINDOW / scan_size;
1103 floor = 1000 / windows;
1104
1105 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1106 return max_t(unsigned int, floor, scan);
1107}
1108
1109static unsigned int task_scan_start(struct task_struct *p)
1110{
1111 unsigned long smin = task_scan_min(p);
1112 unsigned long period = smin;
1113
1114 /* Scale the maximum scan period with the amount of shared memory. */
1115 if (p->numa_group) {
1116 struct numa_group *ng = p->numa_group;
1117 unsigned long shared = group_faults_shared(ng);
1118 unsigned long private = group_faults_priv(ng);
1119
1120 period *= refcount_read(&ng->refcount);
1121 period *= shared + 1;
1122 period /= private + shared + 1;
1123 }
1124
1125 return max(smin, period);
1126}
1127
1128static unsigned int task_scan_max(struct task_struct *p)
1129{
1130 unsigned long smin = task_scan_min(p);
1131 unsigned long smax;
1132
1133 /* Watch for min being lower than max due to floor calculations */
1134 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1135
1136 /* Scale the maximum scan period with the amount of shared memory. */
1137 if (p->numa_group) {
1138 struct numa_group *ng = p->numa_group;
1139 unsigned long shared = group_faults_shared(ng);
1140 unsigned long private = group_faults_priv(ng);
1141 unsigned long period = smax;
1142
1143 period *= refcount_read(&ng->refcount);
1144 period *= shared + 1;
1145 period /= private + shared + 1;
1146
1147 smax = max(smax, period);
1148 }
1149
1150 return max(smin, smax);
1151}
1152
1153void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
1154{
1155 int mm_users = 0;
1156 struct mm_struct *mm = p->mm;
1157
1158 if (mm) {
1159 mm_users = atomic_read(&mm->mm_users);
1160 if (mm_users == 1) {
1161 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1162 mm->numa_scan_seq = 0;
1163 }
1164 }
1165 p->node_stamp = 0;
1166 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
1167 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1168 p->numa_work.next = &p->numa_work;
1169 p->numa_faults = NULL;
1170 p->numa_group = NULL;
1171 p->last_task_numa_placement = 0;
1172 p->last_sum_exec_runtime = 0;
1173
1174 /* New address space, reset the preferred nid */
1175 if (!(clone_flags & CLONE_VM)) {
1176 p->numa_preferred_nid = NUMA_NO_NODE;
1177 return;
1178 }
1179
1180 /*
1181 * New thread, keep existing numa_preferred_nid which should be copied
1182 * already by arch_dup_task_struct but stagger when scans start.
1183 */
1184 if (mm) {
1185 unsigned int delay;
1186
1187 delay = min_t(unsigned int, task_scan_max(current),
1188 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
1189 delay += 2 * TICK_NSEC;
1190 p->node_stamp = delay;
1191 }
1192}
1193
1194static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1195{
1196 rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
1197 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1198}
1199
1200static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1201{
1202 rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1203 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1204}
1205
1206/* Shared or private faults. */
1207#define NR_NUMA_HINT_FAULT_TYPES 2
1208
1209/* Memory and CPU locality */
1210#define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1211
1212/* Averaged statistics, and temporary buffers. */
1213#define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1214
1215pid_t task_numa_group_id(struct task_struct *p)
1216{
1217 return p->numa_group ? p->numa_group->gid : 0;
1218}
1219
1220/*
1221 * The averaged statistics, shared & private, memory & CPU,
1222 * occupy the first half of the array. The second half of the
1223 * array is for current counters, which are averaged into the
1224 * first set by task_numa_placement.
1225 */
1226static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1227{
1228 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1229}
1230
1231static inline unsigned long task_faults(struct task_struct *p, int nid)
1232{
1233 if (!p->numa_faults)
1234 return 0;
1235
1236 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1237 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1238}
1239
1240static inline unsigned long group_faults(struct task_struct *p, int nid)
1241{
1242 if (!p->numa_group)
1243 return 0;
1244
1245 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1246 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1247}
1248
1249static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1250{
1251 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1252 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1253}
1254
1255static inline unsigned long group_faults_priv(struct numa_group *ng)
1256{
1257 unsigned long faults = 0;
1258 int node;
1259
1260 for_each_online_node(node) {
1261 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1262 }
1263
1264 return faults;
1265}
1266
1267static inline unsigned long group_faults_shared(struct numa_group *ng)
1268{
1269 unsigned long faults = 0;
1270 int node;
1271
1272 for_each_online_node(node) {
1273 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1274 }
1275
1276 return faults;
1277}
1278
1279/*
1280 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1281 * considered part of a numa group's pseudo-interleaving set. Migrations
1282 * between these nodes are slowed down, to allow things to settle down.
1283 */
1284#define ACTIVE_NODE_FRACTION 3
1285
1286static bool numa_is_active_node(int nid, struct numa_group *ng)
1287{
1288 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1289}
1290
1291/* Handle placement on systems where not all nodes are directly connected. */
1292static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1293 int maxdist, bool task)
1294{
1295 unsigned long score = 0;
1296 int node;
1297
1298 /*
1299 * All nodes are directly connected, and the same distance
1300 * from each other. No need for fancy placement algorithms.
1301 */
1302 if (sched_numa_topology_type == NUMA_DIRECT)
1303 return 0;
1304
1305 /*
1306 * This code is called for each node, introducing N^2 complexity,
1307 * which should be ok given the number of nodes rarely exceeds 8.
1308 */
1309 for_each_online_node(node) {
1310 unsigned long faults;
1311 int dist = node_distance(nid, node);
1312
1313 /*
1314 * The furthest away nodes in the system are not interesting
1315 * for placement; nid was already counted.
1316 */
1317 if (dist == sched_max_numa_distance || node == nid)
1318 continue;
1319
1320 /*
1321 * On systems with a backplane NUMA topology, compare groups
1322 * of nodes, and move tasks towards the group with the most
1323 * memory accesses. When comparing two nodes at distance
1324 * "hoplimit", only nodes closer by than "hoplimit" are part
1325 * of each group. Skip other nodes.
1326 */
1327 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1328 dist >= maxdist)
1329 continue;
1330
1331 /* Add up the faults from nearby nodes. */
1332 if (task)
1333 faults = task_faults(p, node);
1334 else
1335 faults = group_faults(p, node);
1336
1337 /*
1338 * On systems with a glueless mesh NUMA topology, there are
1339 * no fixed "groups of nodes". Instead, nodes that are not
1340 * directly connected bounce traffic through intermediate
1341 * nodes; a numa_group can occupy any set of nodes.
1342 * The further away a node is, the less the faults count.
1343 * This seems to result in good task placement.
1344 */
1345 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1346 faults *= (sched_max_numa_distance - dist);
1347 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1348 }
1349
1350 score += faults;
1351 }
1352
1353 return score;
1354}
1355
1356/*
1357 * These return the fraction of accesses done by a particular task, or
1358 * task group, on a particular numa node. The group weight is given a
1359 * larger multiplier, in order to group tasks together that are almost
1360 * evenly spread out between numa nodes.
1361 */
1362static inline unsigned long task_weight(struct task_struct *p, int nid,
1363 int dist)
1364{
1365 unsigned long faults, total_faults;
1366
1367 if (!p->numa_faults)
1368 return 0;
1369
1370 total_faults = p->total_numa_faults;
1371
1372 if (!total_faults)
1373 return 0;
1374
1375 faults = task_faults(p, nid);
1376 faults += score_nearby_nodes(p, nid, dist, true);
1377
1378 return 1000 * faults / total_faults;
1379}
1380
1381static inline unsigned long group_weight(struct task_struct *p, int nid,
1382 int dist)
1383{
1384 unsigned long faults, total_faults;
1385
1386 if (!p->numa_group)
1387 return 0;
1388
1389 total_faults = p->numa_group->total_faults;
1390
1391 if (!total_faults)
1392 return 0;
1393
1394 faults = group_faults(p, nid);
1395 faults += score_nearby_nodes(p, nid, dist, false);
1396
1397 return 1000 * faults / total_faults;
1398}
1399
1400bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1401 int src_nid, int dst_cpu)
1402{
1403 struct numa_group *ng = p->numa_group;
1404 int dst_nid = cpu_to_node(dst_cpu);
1405 int last_cpupid, this_cpupid;
1406
1407 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1408 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1409
1410 /*
1411 * Allow first faults or private faults to migrate immediately early in
1412 * the lifetime of a task. The magic number 4 is based on waiting for
1413 * two full passes of the "multi-stage node selection" test that is
1414 * executed below.
1415 */
1416 if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1417 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1418 return true;
1419
1420 /*
1421 * Multi-stage node selection is used in conjunction with a periodic
1422 * migration fault to build a temporal task<->page relation. By using
1423 * a two-stage filter we remove short/unlikely relations.
1424 *
1425 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1426 * a task's usage of a particular page (n_p) per total usage of this
1427 * page (n_t) (in a given time-span) to a probability.
1428 *
1429 * Our periodic faults will sample this probability and getting the
1430 * same result twice in a row, given these samples are fully
1431 * independent, is then given by P(n)^2, provided our sample period
1432 * is sufficiently short compared to the usage pattern.
1433 *
1434 * This quadric squishes small probabilities, making it less likely we
1435 * act on an unlikely task<->page relation.
1436 */
1437 if (!cpupid_pid_unset(last_cpupid) &&
1438 cpupid_to_nid(last_cpupid) != dst_nid)
1439 return false;
1440
1441 /* Always allow migrate on private faults */
1442 if (cpupid_match_pid(p, last_cpupid))
1443 return true;
1444
1445 /* A shared fault, but p->numa_group has not been set up yet. */
1446 if (!ng)
1447 return true;
1448
1449 /*
1450 * Destination node is much more heavily used than the source
1451 * node? Allow migration.
1452 */
1453 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1454 ACTIVE_NODE_FRACTION)
1455 return true;
1456
1457 /*
1458 * Distribute memory according to CPU & memory use on each node,
1459 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1460 *
1461 * faults_cpu(dst) 3 faults_cpu(src)
1462 * --------------- * - > ---------------
1463 * faults_mem(dst) 4 faults_mem(src)
1464 */
1465 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1466 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1467}
1468
1469static unsigned long weighted_cpuload(struct rq *rq);
1470static unsigned long source_load(int cpu, int type);
1471static unsigned long target_load(int cpu, int type);
1472
1473/* Cached statistics for all CPUs within a node */
1474struct numa_stats {
1475 unsigned long load;
1476
1477 /* Total compute capacity of CPUs on a node */
1478 unsigned long compute_capacity;
1479};
1480
1481/*
1482 * XXX borrowed from update_sg_lb_stats
1483 */
1484static void update_numa_stats(struct numa_stats *ns, int nid)
1485{
1486 int cpu;
1487
1488 memset(ns, 0, sizeof(*ns));
1489 for_each_cpu(cpu, cpumask_of_node(nid)) {
1490 struct rq *rq = cpu_rq(cpu);
1491
1492 ns->load += weighted_cpuload(rq);
1493 ns->compute_capacity += capacity_of(cpu);
1494 }
1495
1496}
1497
1498struct task_numa_env {
1499 struct task_struct *p;
1500
1501 int src_cpu, src_nid;
1502 int dst_cpu, dst_nid;
1503
1504 struct numa_stats src_stats, dst_stats;
1505
1506 int imbalance_pct;
1507 int dist;
1508
1509 struct task_struct *best_task;
1510 long best_imp;
1511 int best_cpu;
1512};
1513
1514static void task_numa_assign(struct task_numa_env *env,
1515 struct task_struct *p, long imp)
1516{
1517 struct rq *rq = cpu_rq(env->dst_cpu);
1518
1519 /* Bail out if run-queue part of active NUMA balance. */
1520 if (xchg(&rq->numa_migrate_on, 1))
1521 return;
1522
1523 /*
1524 * Clear previous best_cpu/rq numa-migrate flag, since task now
1525 * found a better CPU to move/swap.
1526 */
1527 if (env->best_cpu != -1) {
1528 rq = cpu_rq(env->best_cpu);
1529 WRITE_ONCE(rq->numa_migrate_on, 0);
1530 }
1531
1532 if (env->best_task)
1533 put_task_struct(env->best_task);
1534 if (p)
1535 get_task_struct(p);
1536
1537 env->best_task = p;
1538 env->best_imp = imp;
1539 env->best_cpu = env->dst_cpu;
1540}
1541
1542static bool load_too_imbalanced(long src_load, long dst_load,
1543 struct task_numa_env *env)
1544{
1545 long imb, old_imb;
1546 long orig_src_load, orig_dst_load;
1547 long src_capacity, dst_capacity;
1548
1549 /*
1550 * The load is corrected for the CPU capacity available on each node.
1551 *
1552 * src_load dst_load
1553 * ------------ vs ---------
1554 * src_capacity dst_capacity
1555 */
1556 src_capacity = env->src_stats.compute_capacity;
1557 dst_capacity = env->dst_stats.compute_capacity;
1558
1559 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1560
1561 orig_src_load = env->src_stats.load;
1562 orig_dst_load = env->dst_stats.load;
1563
1564 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1565
1566 /* Would this change make things worse? */
1567 return (imb > old_imb);
1568}
1569
1570/*
1571 * Maximum NUMA importance can be 1998 (2*999);
1572 * SMALLIMP @ 30 would be close to 1998/64.
1573 * Used to deter task migration.
1574 */
1575#define SMALLIMP 30
1576
1577/*
1578 * This checks if the overall compute and NUMA accesses of the system would
1579 * be improved if the source tasks was migrated to the target dst_cpu taking
1580 * into account that it might be best if task running on the dst_cpu should
1581 * be exchanged with the source task
1582 */
1583static void task_numa_compare(struct task_numa_env *env,
1584 long taskimp, long groupimp, bool maymove)
1585{
1586 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1587 struct task_struct *cur;
1588 long src_load, dst_load;
1589 long load;
1590 long imp = env->p->numa_group ? groupimp : taskimp;
1591 long moveimp = imp;
1592 int dist = env->dist;
1593
1594 if (READ_ONCE(dst_rq->numa_migrate_on))
1595 return;
1596
1597 rcu_read_lock();
1598 cur = task_rcu_dereference(&dst_rq->curr);
1599 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1600 cur = NULL;
1601
1602 /*
1603 * Because we have preemption enabled we can get migrated around and
1604 * end try selecting ourselves (current == env->p) as a swap candidate.
1605 */
1606 if (cur == env->p)
1607 goto unlock;
1608
1609 if (!cur) {
1610 if (maymove && moveimp >= env->best_imp)
1611 goto assign;
1612 else
1613 goto unlock;
1614 }
1615
1616 /*
1617 * "imp" is the fault differential for the source task between the
1618 * source and destination node. Calculate the total differential for
1619 * the source task and potential destination task. The more negative
1620 * the value is, the more remote accesses that would be expected to
1621 * be incurred if the tasks were swapped.
1622 */
1623 /* Skip this swap candidate if cannot move to the source cpu */
1624 if (!cpumask_test_cpu(env->src_cpu, &cur->cpus_allowed))
1625 goto unlock;
1626
1627 /*
1628 * If dst and source tasks are in the same NUMA group, or not
1629 * in any group then look only at task weights.
1630 */
1631 if (cur->numa_group == env->p->numa_group) {
1632 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1633 task_weight(cur, env->dst_nid, dist);
1634 /*
1635 * Add some hysteresis to prevent swapping the
1636 * tasks within a group over tiny differences.
1637 */
1638 if (cur->numa_group)
1639 imp -= imp / 16;
1640 } else {
1641 /*
1642 * Compare the group weights. If a task is all by itself
1643 * (not part of a group), use the task weight instead.
1644 */
1645 if (cur->numa_group && env->p->numa_group)
1646 imp += group_weight(cur, env->src_nid, dist) -
1647 group_weight(cur, env->dst_nid, dist);
1648 else
1649 imp += task_weight(cur, env->src_nid, dist) -
1650 task_weight(cur, env->dst_nid, dist);
1651 }
1652
1653 if (maymove && moveimp > imp && moveimp > env->best_imp) {
1654 imp = moveimp;
1655 cur = NULL;
1656 goto assign;
1657 }
1658
1659 /*
1660 * If the NUMA importance is less than SMALLIMP,
1661 * task migration might only result in ping pong
1662 * of tasks and also hurt performance due to cache
1663 * misses.
1664 */
1665 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
1666 goto unlock;
1667
1668 /*
1669 * In the overloaded case, try and keep the load balanced.
1670 */
1671 load = task_h_load(env->p) - task_h_load(cur);
1672 if (!load)
1673 goto assign;
1674
1675 dst_load = env->dst_stats.load + load;
1676 src_load = env->src_stats.load - load;
1677
1678 if (load_too_imbalanced(src_load, dst_load, env))
1679 goto unlock;
1680
1681assign:
1682 /*
1683 * One idle CPU per node is evaluated for a task numa move.
1684 * Call select_idle_sibling to maybe find a better one.
1685 */
1686 if (!cur) {
1687 /*
1688 * select_idle_siblings() uses an per-CPU cpumask that
1689 * can be used from IRQ context.
1690 */
1691 local_irq_disable();
1692 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1693 env->dst_cpu);
1694 local_irq_enable();
1695 }
1696
1697 task_numa_assign(env, cur, imp);
1698unlock:
1699 rcu_read_unlock();
1700}
1701
1702static void task_numa_find_cpu(struct task_numa_env *env,
1703 long taskimp, long groupimp)
1704{
1705 long src_load, dst_load, load;
1706 bool maymove = false;
1707 int cpu;
1708
1709 load = task_h_load(env->p);
1710 dst_load = env->dst_stats.load + load;
1711 src_load = env->src_stats.load - load;
1712
1713 /*
1714 * If the improvement from just moving env->p direction is better
1715 * than swapping tasks around, check if a move is possible.
1716 */
1717 maymove = !load_too_imbalanced(src_load, dst_load, env);
1718
1719 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1720 /* Skip this CPU if the source task cannot migrate */
1721 if (!cpumask_test_cpu(cpu, &env->p->cpus_allowed))
1722 continue;
1723
1724 env->dst_cpu = cpu;
1725 task_numa_compare(env, taskimp, groupimp, maymove);
1726 }
1727}
1728
1729static int task_numa_migrate(struct task_struct *p)
1730{
1731 struct task_numa_env env = {
1732 .p = p,
1733
1734 .src_cpu = task_cpu(p),
1735 .src_nid = task_node(p),
1736
1737 .imbalance_pct = 112,
1738
1739 .best_task = NULL,
1740 .best_imp = 0,
1741 .best_cpu = -1,
1742 };
1743 struct sched_domain *sd;
1744 struct rq *best_rq;
1745 unsigned long taskweight, groupweight;
1746 int nid, ret, dist;
1747 long taskimp, groupimp;
1748
1749 /*
1750 * Pick the lowest SD_NUMA domain, as that would have the smallest
1751 * imbalance and would be the first to start moving tasks about.
1752 *
1753 * And we want to avoid any moving of tasks about, as that would create
1754 * random movement of tasks -- counter the numa conditions we're trying
1755 * to satisfy here.
1756 */
1757 rcu_read_lock();
1758 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1759 if (sd)
1760 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1761 rcu_read_unlock();
1762
1763 /*
1764 * Cpusets can break the scheduler domain tree into smaller
1765 * balance domains, some of which do not cross NUMA boundaries.
1766 * Tasks that are "trapped" in such domains cannot be migrated
1767 * elsewhere, so there is no point in (re)trying.
1768 */
1769 if (unlikely(!sd)) {
1770 sched_setnuma(p, task_node(p));
1771 return -EINVAL;
1772 }
1773
1774 env.dst_nid = p->numa_preferred_nid;
1775 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1776 taskweight = task_weight(p, env.src_nid, dist);
1777 groupweight = group_weight(p, env.src_nid, dist);
1778 update_numa_stats(&env.src_stats, env.src_nid);
1779 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1780 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1781 update_numa_stats(&env.dst_stats, env.dst_nid);
1782
1783 /* Try to find a spot on the preferred nid. */
1784 task_numa_find_cpu(&env, taskimp, groupimp);
1785
1786 /*
1787 * Look at other nodes in these cases:
1788 * - there is no space available on the preferred_nid
1789 * - the task is part of a numa_group that is interleaved across
1790 * multiple NUMA nodes; in order to better consolidate the group,
1791 * we need to check other locations.
1792 */
1793 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1794 for_each_online_node(nid) {
1795 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1796 continue;
1797
1798 dist = node_distance(env.src_nid, env.dst_nid);
1799 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1800 dist != env.dist) {
1801 taskweight = task_weight(p, env.src_nid, dist);
1802 groupweight = group_weight(p, env.src_nid, dist);
1803 }
1804
1805 /* Only consider nodes where both task and groups benefit */
1806 taskimp = task_weight(p, nid, dist) - taskweight;
1807 groupimp = group_weight(p, nid, dist) - groupweight;
1808 if (taskimp < 0 && groupimp < 0)
1809 continue;
1810
1811 env.dist = dist;
1812 env.dst_nid = nid;
1813 update_numa_stats(&env.dst_stats, env.dst_nid);
1814 task_numa_find_cpu(&env, taskimp, groupimp);
1815 }
1816 }
1817
1818 /*
1819 * If the task is part of a workload that spans multiple NUMA nodes,
1820 * and is migrating into one of the workload's active nodes, remember
1821 * this node as the task's preferred numa node, so the workload can
1822 * settle down.
1823 * A task that migrated to a second choice node will be better off
1824 * trying for a better one later. Do not set the preferred node here.
1825 */
1826 if (p->numa_group) {
1827 if (env.best_cpu == -1)
1828 nid = env.src_nid;
1829 else
1830 nid = cpu_to_node(env.best_cpu);
1831
1832 if (nid != p->numa_preferred_nid)
1833 sched_setnuma(p, nid);
1834 }
1835
1836 /* No better CPU than the current one was found. */
1837 if (env.best_cpu == -1)
1838 return -EAGAIN;
1839
1840 best_rq = cpu_rq(env.best_cpu);
1841 if (env.best_task == NULL) {
1842 ret = migrate_task_to(p, env.best_cpu);
1843 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1844 if (ret != 0)
1845 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1846 return ret;
1847 }
1848
1849 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
1850 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1851
1852 if (ret != 0)
1853 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1854 put_task_struct(env.best_task);
1855 return ret;
1856}
1857
1858/* Attempt to migrate a task to a CPU on the preferred node. */
1859static void numa_migrate_preferred(struct task_struct *p)
1860{
1861 unsigned long interval = HZ;
1862
1863 /* This task has no NUMA fault statistics yet */
1864 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
1865 return;
1866
1867 /* Periodically retry migrating the task to the preferred node */
1868 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1869 p->numa_migrate_retry = jiffies + interval;
1870
1871 /* Success if task is already running on preferred CPU */
1872 if (task_node(p) == p->numa_preferred_nid)
1873 return;
1874
1875 /* Otherwise, try migrate to a CPU on the preferred node */
1876 task_numa_migrate(p);
1877}
1878
1879/*
1880 * Find out how many nodes on the workload is actively running on. Do this by
1881 * tracking the nodes from which NUMA hinting faults are triggered. This can
1882 * be different from the set of nodes where the workload's memory is currently
1883 * located.
1884 */
1885static void numa_group_count_active_nodes(struct numa_group *numa_group)
1886{
1887 unsigned long faults, max_faults = 0;
1888 int nid, active_nodes = 0;
1889
1890 for_each_online_node(nid) {
1891 faults = group_faults_cpu(numa_group, nid);
1892 if (faults > max_faults)
1893 max_faults = faults;
1894 }
1895
1896 for_each_online_node(nid) {
1897 faults = group_faults_cpu(numa_group, nid);
1898 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1899 active_nodes++;
1900 }
1901
1902 numa_group->max_faults_cpu = max_faults;
1903 numa_group->active_nodes = active_nodes;
1904}
1905
1906/*
1907 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1908 * increments. The more local the fault statistics are, the higher the scan
1909 * period will be for the next scan window. If local/(local+remote) ratio is
1910 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1911 * the scan period will decrease. Aim for 70% local accesses.
1912 */
1913#define NUMA_PERIOD_SLOTS 10
1914#define NUMA_PERIOD_THRESHOLD 7
1915
1916/*
1917 * Increase the scan period (slow down scanning) if the majority of
1918 * our memory is already on our local node, or if the majority of
1919 * the page accesses are shared with other processes.
1920 * Otherwise, decrease the scan period.
1921 */
1922static void update_task_scan_period(struct task_struct *p,
1923 unsigned long shared, unsigned long private)
1924{
1925 unsigned int period_slot;
1926 int lr_ratio, ps_ratio;
1927 int diff;
1928
1929 unsigned long remote = p->numa_faults_locality[0];
1930 unsigned long local = p->numa_faults_locality[1];
1931
1932 /*
1933 * If there were no record hinting faults then either the task is
1934 * completely idle or all activity is areas that are not of interest
1935 * to automatic numa balancing. Related to that, if there were failed
1936 * migration then it implies we are migrating too quickly or the local
1937 * node is overloaded. In either case, scan slower
1938 */
1939 if (local + shared == 0 || p->numa_faults_locality[2]) {
1940 p->numa_scan_period = min(p->numa_scan_period_max,
1941 p->numa_scan_period << 1);
1942
1943 p->mm->numa_next_scan = jiffies +
1944 msecs_to_jiffies(p->numa_scan_period);
1945
1946 return;
1947 }
1948
1949 /*
1950 * Prepare to scale scan period relative to the current period.
1951 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1952 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1953 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1954 */
1955 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1956 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1957 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
1958
1959 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
1960 /*
1961 * Most memory accesses are local. There is no need to
1962 * do fast NUMA scanning, since memory is already local.
1963 */
1964 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
1965 if (!slot)
1966 slot = 1;
1967 diff = slot * period_slot;
1968 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
1969 /*
1970 * Most memory accesses are shared with other tasks.
1971 * There is no point in continuing fast NUMA scanning,
1972 * since other tasks may just move the memory elsewhere.
1973 */
1974 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
1975 if (!slot)
1976 slot = 1;
1977 diff = slot * period_slot;
1978 } else {
1979 /*
1980 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
1981 * yet they are not on the local NUMA node. Speed up
1982 * NUMA scanning to get the memory moved over.
1983 */
1984 int ratio = max(lr_ratio, ps_ratio);
1985 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1986 }
1987
1988 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1989 task_scan_min(p), task_scan_max(p));
1990 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1991}
1992
1993/*
1994 * Get the fraction of time the task has been running since the last
1995 * NUMA placement cycle. The scheduler keeps similar statistics, but
1996 * decays those on a 32ms period, which is orders of magnitude off
1997 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1998 * stats only if the task is so new there are no NUMA statistics yet.
1999 */
2000static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2001{
2002 u64 runtime, delta, now;
2003 /* Use the start of this time slice to avoid calculations. */
2004 now = p->se.exec_start;
2005 runtime = p->se.sum_exec_runtime;
2006
2007 if (p->last_task_numa_placement) {
2008 delta = runtime - p->last_sum_exec_runtime;
2009 *period = now - p->last_task_numa_placement;
2010 } else {
2011 delta = p->se.avg.load_sum;
2012 *period = LOAD_AVG_MAX;
2013 }
2014
2015 p->last_sum_exec_runtime = runtime;
2016 p->last_task_numa_placement = now;
2017
2018 return delta;
2019}
2020
2021/*
2022 * Determine the preferred nid for a task in a numa_group. This needs to
2023 * be done in a way that produces consistent results with group_weight,
2024 * otherwise workloads might not converge.
2025 */
2026static int preferred_group_nid(struct task_struct *p, int nid)
2027{
2028 nodemask_t nodes;
2029 int dist;
2030
2031 /* Direct connections between all NUMA nodes. */
2032 if (sched_numa_topology_type == NUMA_DIRECT)
2033 return nid;
2034
2035 /*
2036 * On a system with glueless mesh NUMA topology, group_weight
2037 * scores nodes according to the number of NUMA hinting faults on
2038 * both the node itself, and on nearby nodes.
2039 */
2040 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2041 unsigned long score, max_score = 0;
2042 int node, max_node = nid;
2043
2044 dist = sched_max_numa_distance;
2045
2046 for_each_online_node(node) {
2047 score = group_weight(p, node, dist);
2048 if (score > max_score) {
2049 max_score = score;
2050 max_node = node;
2051 }
2052 }
2053 return max_node;
2054 }
2055
2056 /*
2057 * Finding the preferred nid in a system with NUMA backplane
2058 * interconnect topology is more involved. The goal is to locate
2059 * tasks from numa_groups near each other in the system, and
2060 * untangle workloads from different sides of the system. This requires
2061 * searching down the hierarchy of node groups, recursively searching
2062 * inside the highest scoring group of nodes. The nodemask tricks
2063 * keep the complexity of the search down.
2064 */
2065 nodes = node_online_map;
2066 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2067 unsigned long max_faults = 0;
2068 nodemask_t max_group = NODE_MASK_NONE;
2069 int a, b;
2070
2071 /* Are there nodes at this distance from each other? */
2072 if (!find_numa_distance(dist))
2073 continue;
2074
2075 for_each_node_mask(a, nodes) {
2076 unsigned long faults = 0;
2077 nodemask_t this_group;
2078 nodes_clear(this_group);
2079
2080 /* Sum group's NUMA faults; includes a==b case. */
2081 for_each_node_mask(b, nodes) {
2082 if (node_distance(a, b) < dist) {
2083 faults += group_faults(p, b);
2084 node_set(b, this_group);
2085 node_clear(b, nodes);
2086 }
2087 }
2088
2089 /* Remember the top group. */
2090 if (faults > max_faults) {
2091 max_faults = faults;
2092 max_group = this_group;
2093 /*
2094 * subtle: at the smallest distance there is
2095 * just one node left in each "group", the
2096 * winner is the preferred nid.
2097 */
2098 nid = a;
2099 }
2100 }
2101 /* Next round, evaluate the nodes within max_group. */
2102 if (!max_faults)
2103 break;
2104 nodes = max_group;
2105 }
2106 return nid;
2107}
2108
2109static void task_numa_placement(struct task_struct *p)
2110{
2111 int seq, nid, max_nid = NUMA_NO_NODE;
2112 unsigned long max_faults = 0;
2113 unsigned long fault_types[2] = { 0, 0 };
2114 unsigned long total_faults;
2115 u64 runtime, period;
2116 spinlock_t *group_lock = NULL;
2117
2118 /*
2119 * The p->mm->numa_scan_seq field gets updated without
2120 * exclusive access. Use READ_ONCE() here to ensure
2121 * that the field is read in a single access:
2122 */
2123 seq = READ_ONCE(p->mm->numa_scan_seq);
2124 if (p->numa_scan_seq == seq)
2125 return;
2126 p->numa_scan_seq = seq;
2127 p->numa_scan_period_max = task_scan_max(p);
2128
2129 total_faults = p->numa_faults_locality[0] +
2130 p->numa_faults_locality[1];
2131 runtime = numa_get_avg_runtime(p, &period);
2132
2133 /* If the task is part of a group prevent parallel updates to group stats */
2134 if (p->numa_group) {
2135 group_lock = &p->numa_group->lock;
2136 spin_lock_irq(group_lock);
2137 }
2138
2139 /* Find the node with the highest number of faults */
2140 for_each_online_node(nid) {
2141 /* Keep track of the offsets in numa_faults array */
2142 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2143 unsigned long faults = 0, group_faults = 0;
2144 int priv;
2145
2146 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2147 long diff, f_diff, f_weight;
2148
2149 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2150 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2151 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2152 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2153
2154 /* Decay existing window, copy faults since last scan */
2155 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2156 fault_types[priv] += p->numa_faults[membuf_idx];
2157 p->numa_faults[membuf_idx] = 0;
2158
2159 /*
2160 * Normalize the faults_from, so all tasks in a group
2161 * count according to CPU use, instead of by the raw
2162 * number of faults. Tasks with little runtime have
2163 * little over-all impact on throughput, and thus their
2164 * faults are less important.
2165 */
2166 f_weight = div64_u64(runtime << 16, period + 1);
2167 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2168 (total_faults + 1);
2169 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2170 p->numa_faults[cpubuf_idx] = 0;
2171
2172 p->numa_faults[mem_idx] += diff;
2173 p->numa_faults[cpu_idx] += f_diff;
2174 faults += p->numa_faults[mem_idx];
2175 p->total_numa_faults += diff;
2176 if (p->numa_group) {
2177 /*
2178 * safe because we can only change our own group
2179 *
2180 * mem_idx represents the offset for a given
2181 * nid and priv in a specific region because it
2182 * is at the beginning of the numa_faults array.
2183 */
2184 p->numa_group->faults[mem_idx] += diff;
2185 p->numa_group->faults_cpu[mem_idx] += f_diff;
2186 p->numa_group->total_faults += diff;
2187 group_faults += p->numa_group->faults[mem_idx];
2188 }
2189 }
2190
2191 if (!p->numa_group) {
2192 if (faults > max_faults) {
2193 max_faults = faults;
2194 max_nid = nid;
2195 }
2196 } else if (group_faults > max_faults) {
2197 max_faults = group_faults;
2198 max_nid = nid;
2199 }
2200 }
2201
2202 if (p->numa_group) {
2203 numa_group_count_active_nodes(p->numa_group);
2204 spin_unlock_irq(group_lock);
2205 max_nid = preferred_group_nid(p, max_nid);
2206 }
2207
2208 if (max_faults) {
2209 /* Set the new preferred node */
2210 if (max_nid != p->numa_preferred_nid)
2211 sched_setnuma(p, max_nid);
2212 }
2213
2214 update_task_scan_period(p, fault_types[0], fault_types[1]);
2215}
2216
2217static inline int get_numa_group(struct numa_group *grp)
2218{
2219 return refcount_inc_not_zero(&grp->refcount);
2220}
2221
2222static inline void put_numa_group(struct numa_group *grp)
2223{
2224 if (refcount_dec_and_test(&grp->refcount))
2225 kfree_rcu(grp, rcu);
2226}
2227
2228static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2229 int *priv)
2230{
2231 struct numa_group *grp, *my_grp;
2232 struct task_struct *tsk;
2233 bool join = false;
2234 int cpu = cpupid_to_cpu(cpupid);
2235 int i;
2236
2237 if (unlikely(!p->numa_group)) {
2238 unsigned int size = sizeof(struct numa_group) +
2239 4*nr_node_ids*sizeof(unsigned long);
2240
2241 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2242 if (!grp)
2243 return;
2244
2245 refcount_set(&grp->refcount, 1);
2246 grp->active_nodes = 1;
2247 grp->max_faults_cpu = 0;
2248 spin_lock_init(&grp->lock);
2249 grp->gid = p->pid;
2250 /* Second half of the array tracks nids where faults happen */
2251 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2252 nr_node_ids;
2253
2254 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2255 grp->faults[i] = p->numa_faults[i];
2256
2257 grp->total_faults = p->total_numa_faults;
2258
2259 grp->nr_tasks++;
2260 rcu_assign_pointer(p->numa_group, grp);
2261 }
2262
2263 rcu_read_lock();
2264 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2265
2266 if (!cpupid_match_pid(tsk, cpupid))
2267 goto no_join;
2268
2269 grp = rcu_dereference(tsk->numa_group);
2270 if (!grp)
2271 goto no_join;
2272
2273 my_grp = p->numa_group;
2274 if (grp == my_grp)
2275 goto no_join;
2276
2277 /*
2278 * Only join the other group if its bigger; if we're the bigger group,
2279 * the other task will join us.
2280 */
2281 if (my_grp->nr_tasks > grp->nr_tasks)
2282 goto no_join;
2283
2284 /*
2285 * Tie-break on the grp address.
2286 */
2287 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2288 goto no_join;
2289
2290 /* Always join threads in the same process. */
2291 if (tsk->mm == current->mm)
2292 join = true;
2293
2294 /* Simple filter to avoid false positives due to PID collisions */
2295 if (flags & TNF_SHARED)
2296 join = true;
2297
2298 /* Update priv based on whether false sharing was detected */
2299 *priv = !join;
2300
2301 if (join && !get_numa_group(grp))
2302 goto no_join;
2303
2304 rcu_read_unlock();
2305
2306 if (!join)
2307 return;
2308
2309 BUG_ON(irqs_disabled());
2310 double_lock_irq(&my_grp->lock, &grp->lock);
2311
2312 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2313 my_grp->faults[i] -= p->numa_faults[i];
2314 grp->faults[i] += p->numa_faults[i];
2315 }
2316 my_grp->total_faults -= p->total_numa_faults;
2317 grp->total_faults += p->total_numa_faults;
2318
2319 my_grp->nr_tasks--;
2320 grp->nr_tasks++;
2321
2322 spin_unlock(&my_grp->lock);
2323 spin_unlock_irq(&grp->lock);
2324
2325 rcu_assign_pointer(p->numa_group, grp);
2326
2327 put_numa_group(my_grp);
2328 return;
2329
2330no_join:
2331 rcu_read_unlock();
2332 return;
2333}
2334
2335void task_numa_free(struct task_struct *p)
2336{
2337 struct numa_group *grp = p->numa_group;
2338 void *numa_faults = p->numa_faults;
2339 unsigned long flags;
2340 int i;
2341
2342 if (grp) {
2343 spin_lock_irqsave(&grp->lock, flags);
2344 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2345 grp->faults[i] -= p->numa_faults[i];
2346 grp->total_faults -= p->total_numa_faults;
2347
2348 grp->nr_tasks--;
2349 spin_unlock_irqrestore(&grp->lock, flags);
2350 RCU_INIT_POINTER(p->numa_group, NULL);
2351 put_numa_group(grp);
2352 }
2353
2354 p->numa_faults = NULL;
2355 kfree(numa_faults);
2356}
2357
2358/*
2359 * Got a PROT_NONE fault for a page on @node.
2360 */
2361void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2362{
2363 struct task_struct *p = current;
2364 bool migrated = flags & TNF_MIGRATED;
2365 int cpu_node = task_node(current);
2366 int local = !!(flags & TNF_FAULT_LOCAL);
2367 struct numa_group *ng;
2368 int priv;
2369
2370 if (!static_branch_likely(&sched_numa_balancing))
2371 return;
2372
2373 /* for example, ksmd faulting in a user's mm */
2374 if (!p->mm)
2375 return;
2376
2377 /* Allocate buffer to track faults on a per-node basis */
2378 if (unlikely(!p->numa_faults)) {
2379 int size = sizeof(*p->numa_faults) *
2380 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2381
2382 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2383 if (!p->numa_faults)
2384 return;
2385
2386 p->total_numa_faults = 0;
2387 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2388 }
2389
2390 /*
2391 * First accesses are treated as private, otherwise consider accesses
2392 * to be private if the accessing pid has not changed
2393 */
2394 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2395 priv = 1;
2396 } else {
2397 priv = cpupid_match_pid(p, last_cpupid);
2398 if (!priv && !(flags & TNF_NO_GROUP))
2399 task_numa_group(p, last_cpupid, flags, &priv);
2400 }
2401
2402 /*
2403 * If a workload spans multiple NUMA nodes, a shared fault that
2404 * occurs wholly within the set of nodes that the workload is
2405 * actively using should be counted as local. This allows the
2406 * scan rate to slow down when a workload has settled down.
2407 */
2408 ng = p->numa_group;
2409 if (!priv && !local && ng && ng->active_nodes > 1 &&
2410 numa_is_active_node(cpu_node, ng) &&
2411 numa_is_active_node(mem_node, ng))
2412 local = 1;
2413
2414 /*
2415 * Retry to migrate task to preferred node periodically, in case it
2416 * previously failed, or the scheduler moved us.
2417 */
2418 if (time_after(jiffies, p->numa_migrate_retry)) {
2419 task_numa_placement(p);
2420 numa_migrate_preferred(p);
2421 }
2422
2423 if (migrated)
2424 p->numa_pages_migrated += pages;
2425 if (flags & TNF_MIGRATE_FAIL)
2426 p->numa_faults_locality[2] += pages;
2427
2428 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2429 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2430 p->numa_faults_locality[local] += pages;
2431}
2432
2433static void reset_ptenuma_scan(struct task_struct *p)
2434{
2435 /*
2436 * We only did a read acquisition of the mmap sem, so
2437 * p->mm->numa_scan_seq is written to without exclusive access
2438 * and the update is not guaranteed to be atomic. That's not
2439 * much of an issue though, since this is just used for
2440 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2441 * expensive, to avoid any form of compiler optimizations:
2442 */
2443 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2444 p->mm->numa_scan_offset = 0;
2445}
2446
2447/*
2448 * The expensive part of numa migration is done from task_work context.
2449 * Triggered from task_tick_numa().
2450 */
2451void task_numa_work(struct callback_head *work)
2452{
2453 unsigned long migrate, next_scan, now = jiffies;
2454 struct task_struct *p = current;
2455 struct mm_struct *mm = p->mm;
2456 u64 runtime = p->se.sum_exec_runtime;
2457 struct vm_area_struct *vma;
2458 unsigned long start, end;
2459 unsigned long nr_pte_updates = 0;
2460 long pages, virtpages;
2461
2462 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2463
2464 work->next = work; /* protect against double add */
2465 /*
2466 * Who cares about NUMA placement when they're dying.
2467 *
2468 * NOTE: make sure not to dereference p->mm before this check,
2469 * exit_task_work() happens _after_ exit_mm() so we could be called
2470 * without p->mm even though we still had it when we enqueued this
2471 * work.
2472 */
2473 if (p->flags & PF_EXITING)
2474 return;
2475
2476 if (!mm->numa_next_scan) {
2477 mm->numa_next_scan = now +
2478 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2479 }
2480
2481 /*
2482 * Enforce maximal scan/migration frequency..
2483 */
2484 migrate = mm->numa_next_scan;
2485 if (time_before(now, migrate))
2486 return;
2487
2488 if (p->numa_scan_period == 0) {
2489 p->numa_scan_period_max = task_scan_max(p);
2490 p->numa_scan_period = task_scan_start(p);
2491 }
2492
2493 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2494 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2495 return;
2496
2497 /*
2498 * Delay this task enough that another task of this mm will likely win
2499 * the next time around.
2500 */
2501 p->node_stamp += 2 * TICK_NSEC;
2502
2503 start = mm->numa_scan_offset;
2504 pages = sysctl_numa_balancing_scan_size;
2505 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2506 virtpages = pages * 8; /* Scan up to this much virtual space */
2507 if (!pages)
2508 return;
2509
2510
2511 if (!down_read_trylock(&mm->mmap_sem))
2512 return;
2513 vma = find_vma(mm, start);
2514 if (!vma) {
2515 reset_ptenuma_scan(p);
2516 start = 0;
2517 vma = mm->mmap;
2518 }
2519 for (; vma; vma = vma->vm_next) {
2520 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2521 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2522 continue;
2523 }
2524
2525 /*
2526 * Shared library pages mapped by multiple processes are not
2527 * migrated as it is expected they are cache replicated. Avoid
2528 * hinting faults in read-only file-backed mappings or the vdso
2529 * as migrating the pages will be of marginal benefit.
2530 */
2531 if (!vma->vm_mm ||
2532 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2533 continue;
2534
2535 /*
2536 * Skip inaccessible VMAs to avoid any confusion between
2537 * PROT_NONE and NUMA hinting ptes
2538 */
2539 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2540 continue;
2541
2542 do {
2543 start = max(start, vma->vm_start);
2544 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2545 end = min(end, vma->vm_end);
2546 nr_pte_updates = change_prot_numa(vma, start, end);
2547
2548 /*
2549 * Try to scan sysctl_numa_balancing_size worth of
2550 * hpages that have at least one present PTE that
2551 * is not already pte-numa. If the VMA contains
2552 * areas that are unused or already full of prot_numa
2553 * PTEs, scan up to virtpages, to skip through those
2554 * areas faster.
2555 */
2556 if (nr_pte_updates)
2557 pages -= (end - start) >> PAGE_SHIFT;
2558 virtpages -= (end - start) >> PAGE_SHIFT;
2559
2560 start = end;
2561 if (pages <= 0 || virtpages <= 0)
2562 goto out;
2563
2564 cond_resched();
2565 } while (end != vma->vm_end);
2566 }
2567
2568out:
2569 /*
2570 * It is possible to reach the end of the VMA list but the last few
2571 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2572 * would find the !migratable VMA on the next scan but not reset the
2573 * scanner to the start so check it now.
2574 */
2575 if (vma)
2576 mm->numa_scan_offset = start;
2577 else
2578 reset_ptenuma_scan(p);
2579 up_read(&mm->mmap_sem);
2580
2581 /*
2582 * Make sure tasks use at least 32x as much time to run other code
2583 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2584 * Usually update_task_scan_period slows down scanning enough; on an
2585 * overloaded system we need to limit overhead on a per task basis.
2586 */
2587 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2588 u64 diff = p->se.sum_exec_runtime - runtime;
2589 p->node_stamp += 32 * diff;
2590 }
2591}
2592
2593/*
2594 * Drive the periodic memory faults..
2595 */
2596void task_tick_numa(struct rq *rq, struct task_struct *curr)
2597{
2598 struct callback_head *work = &curr->numa_work;
2599 u64 period, now;
2600
2601 /*
2602 * We don't care about NUMA placement if we don't have memory.
2603 */
2604 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2605 return;
2606
2607 /*
2608 * Using runtime rather than walltime has the dual advantage that
2609 * we (mostly) drive the selection from busy threads and that the
2610 * task needs to have done some actual work before we bother with
2611 * NUMA placement.
2612 */
2613 now = curr->se.sum_exec_runtime;
2614 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2615
2616 if (now > curr->node_stamp + period) {
2617 if (!curr->node_stamp)
2618 curr->numa_scan_period = task_scan_start(curr);
2619 curr->node_stamp += period;
2620
2621 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2622 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2623 task_work_add(curr, work, true);
2624 }
2625 }
2626}
2627
2628static void update_scan_period(struct task_struct *p, int new_cpu)
2629{
2630 int src_nid = cpu_to_node(task_cpu(p));
2631 int dst_nid = cpu_to_node(new_cpu);
2632
2633 if (!static_branch_likely(&sched_numa_balancing))
2634 return;
2635
2636 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
2637 return;
2638
2639 if (src_nid == dst_nid)
2640 return;
2641
2642 /*
2643 * Allow resets if faults have been trapped before one scan
2644 * has completed. This is most likely due to a new task that
2645 * is pulled cross-node due to wakeups or load balancing.
2646 */
2647 if (p->numa_scan_seq) {
2648 /*
2649 * Avoid scan adjustments if moving to the preferred
2650 * node or if the task was not previously running on
2651 * the preferred node.
2652 */
2653 if (dst_nid == p->numa_preferred_nid ||
2654 (p->numa_preferred_nid != NUMA_NO_NODE &&
2655 src_nid != p->numa_preferred_nid))
2656 return;
2657 }
2658
2659 p->numa_scan_period = task_scan_start(p);
2660}
2661
2662#else
2663static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2664{
2665}
2666
2667static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2668{
2669}
2670
2671static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2672{
2673}
2674
2675static inline void update_scan_period(struct task_struct *p, int new_cpu)
2676{
2677}
2678
2679#endif /* CONFIG_NUMA_BALANCING */
2680
2681static void
2682account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2683{
2684 update_load_add(&cfs_rq->load, se->load.weight);
2685 if (!parent_entity(se))
2686 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2687#ifdef CONFIG_SMP
2688 if (entity_is_task(se)) {
2689 struct rq *rq = rq_of(cfs_rq);
2690
2691 account_numa_enqueue(rq, task_of(se));
2692 list_add(&se->group_node, &rq->cfs_tasks);
2693 }
2694#endif
2695 cfs_rq->nr_running++;
2696}
2697
2698static void
2699account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2700{
2701 update_load_sub(&cfs_rq->load, se->load.weight);
2702 if (!parent_entity(se))
2703 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2704#ifdef CONFIG_SMP
2705 if (entity_is_task(se)) {
2706 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2707 list_del_init(&se->group_node);
2708 }
2709#endif
2710 cfs_rq->nr_running--;
2711}
2712
2713/*
2714 * Signed add and clamp on underflow.
2715 *
2716 * Explicitly do a load-store to ensure the intermediate value never hits
2717 * memory. This allows lockless observations without ever seeing the negative
2718 * values.
2719 */
2720#define add_positive(_ptr, _val) do { \
2721 typeof(_ptr) ptr = (_ptr); \
2722 typeof(_val) val = (_val); \
2723 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2724 \
2725 res = var + val; \
2726 \
2727 if (val < 0 && res > var) \
2728 res = 0; \
2729 \
2730 WRITE_ONCE(*ptr, res); \
2731} while (0)
2732
2733/*
2734 * Unsigned subtract and clamp on underflow.
2735 *
2736 * Explicitly do a load-store to ensure the intermediate value never hits
2737 * memory. This allows lockless observations without ever seeing the negative
2738 * values.
2739 */
2740#define sub_positive(_ptr, _val) do { \
2741 typeof(_ptr) ptr = (_ptr); \
2742 typeof(*ptr) val = (_val); \
2743 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2744 res = var - val; \
2745 if (res > var) \
2746 res = 0; \
2747 WRITE_ONCE(*ptr, res); \
2748} while (0)
2749
2750/*
2751 * Remove and clamp on negative, from a local variable.
2752 *
2753 * A variant of sub_positive(), which does not use explicit load-store
2754 * and is thus optimized for local variable updates.
2755 */
2756#define lsub_positive(_ptr, _val) do { \
2757 typeof(_ptr) ptr = (_ptr); \
2758 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
2759} while (0)
2760
2761#ifdef CONFIG_SMP
2762static inline void
2763enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2764{
2765 cfs_rq->runnable_weight += se->runnable_weight;
2766
2767 cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg;
2768 cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum;
2769}
2770
2771static inline void
2772dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2773{
2774 cfs_rq->runnable_weight -= se->runnable_weight;
2775
2776 sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg);
2777 sub_positive(&cfs_rq->avg.runnable_load_sum,
2778 se_runnable(se) * se->avg.runnable_load_sum);
2779}
2780
2781static inline void
2782enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2783{
2784 cfs_rq->avg.load_avg += se->avg.load_avg;
2785 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
2786}
2787
2788static inline void
2789dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2790{
2791 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2792 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
2793}
2794#else
2795static inline void
2796enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2797static inline void
2798dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2799static inline void
2800enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2801static inline void
2802dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2803#endif
2804
2805static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2806 unsigned long weight, unsigned long runnable)
2807{
2808 if (se->on_rq) {
2809 /* commit outstanding execution time */
2810 if (cfs_rq->curr == se)
2811 update_curr(cfs_rq);
2812 account_entity_dequeue(cfs_rq, se);
2813 dequeue_runnable_load_avg(cfs_rq, se);
2814 }
2815 dequeue_load_avg(cfs_rq, se);
2816
2817 se->runnable_weight = runnable;
2818 update_load_set(&se->load, weight);
2819
2820#ifdef CONFIG_SMP
2821 do {
2822 u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
2823
2824 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
2825 se->avg.runnable_load_avg =
2826 div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider);
2827 } while (0);
2828#endif
2829
2830 enqueue_load_avg(cfs_rq, se);
2831 if (se->on_rq) {
2832 account_entity_enqueue(cfs_rq, se);
2833 enqueue_runnable_load_avg(cfs_rq, se);
2834 }
2835}
2836
2837void reweight_task(struct task_struct *p, int prio)
2838{
2839 struct sched_entity *se = &p->se;
2840 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2841 struct load_weight *load = &se->load;
2842 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
2843
2844 reweight_entity(cfs_rq, se, weight, weight);
2845 load->inv_weight = sched_prio_to_wmult[prio];
2846}
2847
2848#ifdef CONFIG_FAIR_GROUP_SCHED
2849#ifdef CONFIG_SMP
2850/*
2851 * All this does is approximate the hierarchical proportion which includes that
2852 * global sum we all love to hate.
2853 *
2854 * That is, the weight of a group entity, is the proportional share of the
2855 * group weight based on the group runqueue weights. That is:
2856 *
2857 * tg->weight * grq->load.weight
2858 * ge->load.weight = ----------------------------- (1)
2859 * \Sum grq->load.weight
2860 *
2861 * Now, because computing that sum is prohibitively expensive to compute (been
2862 * there, done that) we approximate it with this average stuff. The average
2863 * moves slower and therefore the approximation is cheaper and more stable.
2864 *
2865 * So instead of the above, we substitute:
2866 *
2867 * grq->load.weight -> grq->avg.load_avg (2)
2868 *
2869 * which yields the following:
2870 *
2871 * tg->weight * grq->avg.load_avg
2872 * ge->load.weight = ------------------------------ (3)
2873 * tg->load_avg
2874 *
2875 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2876 *
2877 * That is shares_avg, and it is right (given the approximation (2)).
2878 *
2879 * The problem with it is that because the average is slow -- it was designed
2880 * to be exactly that of course -- this leads to transients in boundary
2881 * conditions. In specific, the case where the group was idle and we start the
2882 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2883 * yielding bad latency etc..
2884 *
2885 * Now, in that special case (1) reduces to:
2886 *
2887 * tg->weight * grq->load.weight
2888 * ge->load.weight = ----------------------------- = tg->weight (4)
2889 * grp->load.weight
2890 *
2891 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2892 *
2893 * So what we do is modify our approximation (3) to approach (4) in the (near)
2894 * UP case, like:
2895 *
2896 * ge->load.weight =
2897 *
2898 * tg->weight * grq->load.weight
2899 * --------------------------------------------------- (5)
2900 * tg->load_avg - grq->avg.load_avg + grq->load.weight
2901 *
2902 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
2903 * we need to use grq->avg.load_avg as its lower bound, which then gives:
2904 *
2905 *
2906 * tg->weight * grq->load.weight
2907 * ge->load.weight = ----------------------------- (6)
2908 * tg_load_avg'
2909 *
2910 * Where:
2911 *
2912 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
2913 * max(grq->load.weight, grq->avg.load_avg)
2914 *
2915 * And that is shares_weight and is icky. In the (near) UP case it approaches
2916 * (4) while in the normal case it approaches (3). It consistently
2917 * overestimates the ge->load.weight and therefore:
2918 *
2919 * \Sum ge->load.weight >= tg->weight
2920 *
2921 * hence icky!
2922 */
2923static long calc_group_shares(struct cfs_rq *cfs_rq)
2924{
2925 long tg_weight, tg_shares, load, shares;
2926 struct task_group *tg = cfs_rq->tg;
2927
2928 tg_shares = READ_ONCE(tg->shares);
2929
2930 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
2931
2932 tg_weight = atomic_long_read(&tg->load_avg);
2933
2934 /* Ensure tg_weight >= load */
2935 tg_weight -= cfs_rq->tg_load_avg_contrib;
2936 tg_weight += load;
2937
2938 shares = (tg_shares * load);
2939 if (tg_weight)
2940 shares /= tg_weight;
2941
2942 /*
2943 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2944 * of a group with small tg->shares value. It is a floor value which is
2945 * assigned as a minimum load.weight to the sched_entity representing
2946 * the group on a CPU.
2947 *
2948 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2949 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2950 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2951 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2952 * instead of 0.
2953 */
2954 return clamp_t(long, shares, MIN_SHARES, tg_shares);
2955}
2956
2957/*
2958 * This calculates the effective runnable weight for a group entity based on
2959 * the group entity weight calculated above.
2960 *
2961 * Because of the above approximation (2), our group entity weight is
2962 * an load_avg based ratio (3). This means that it includes blocked load and
2963 * does not represent the runnable weight.
2964 *
2965 * Approximate the group entity's runnable weight per ratio from the group
2966 * runqueue:
2967 *
2968 * grq->avg.runnable_load_avg
2969 * ge->runnable_weight = ge->load.weight * -------------------------- (7)
2970 * grq->avg.load_avg
2971 *
2972 * However, analogous to above, since the avg numbers are slow, this leads to
2973 * transients in the from-idle case. Instead we use:
2974 *
2975 * ge->runnable_weight = ge->load.weight *
2976 *
2977 * max(grq->avg.runnable_load_avg, grq->runnable_weight)
2978 * ----------------------------------------------------- (8)
2979 * max(grq->avg.load_avg, grq->load.weight)
2980 *
2981 * Where these max() serve both to use the 'instant' values to fix the slow
2982 * from-idle and avoid the /0 on to-idle, similar to (6).
2983 */
2984static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares)
2985{
2986 long runnable, load_avg;
2987
2988 load_avg = max(cfs_rq->avg.load_avg,
2989 scale_load_down(cfs_rq->load.weight));
2990
2991 runnable = max(cfs_rq->avg.runnable_load_avg,
2992 scale_load_down(cfs_rq->runnable_weight));
2993
2994 runnable *= shares;
2995 if (load_avg)
2996 runnable /= load_avg;
2997
2998 return clamp_t(long, runnable, MIN_SHARES, shares);
2999}
3000#endif /* CONFIG_SMP */
3001
3002static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3003
3004/*
3005 * Recomputes the group entity based on the current state of its group
3006 * runqueue.
3007 */
3008static void update_cfs_group(struct sched_entity *se)
3009{
3010 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3011 long shares, runnable;
3012
3013 if (!gcfs_rq)
3014 return;
3015
3016 if (throttled_hierarchy(gcfs_rq))
3017 return;
3018
3019#ifndef CONFIG_SMP
3020 runnable = shares = READ_ONCE(gcfs_rq->tg->shares);
3021
3022 if (likely(se->load.weight == shares))
3023 return;
3024#else
3025 shares = calc_group_shares(gcfs_rq);
3026 runnable = calc_group_runnable(gcfs_rq, shares);
3027#endif
3028
3029 reweight_entity(cfs_rq_of(se), se, shares, runnable);
3030}
3031
3032#else /* CONFIG_FAIR_GROUP_SCHED */
3033static inline void update_cfs_group(struct sched_entity *se)
3034{
3035}
3036#endif /* CONFIG_FAIR_GROUP_SCHED */
3037
3038static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3039{
3040 struct rq *rq = rq_of(cfs_rq);
3041
3042 if (&rq->cfs == cfs_rq || (flags & SCHED_CPUFREQ_MIGRATION)) {
3043 /*
3044 * There are a few boundary cases this might miss but it should
3045 * get called often enough that that should (hopefully) not be
3046 * a real problem.
3047 *
3048 * It will not get called when we go idle, because the idle
3049 * thread is a different class (!fair), nor will the utilization
3050 * number include things like RT tasks.
3051 *
3052 * As is, the util number is not freq-invariant (we'd have to
3053 * implement arch_scale_freq_capacity() for that).
3054 *
3055 * See cpu_util().
3056 */
3057 cpufreq_update_util(rq, flags);
3058 }
3059}
3060
3061#ifdef CONFIG_SMP
3062#ifdef CONFIG_FAIR_GROUP_SCHED
3063/**
3064 * update_tg_load_avg - update the tg's load avg
3065 * @cfs_rq: the cfs_rq whose avg changed
3066 * @force: update regardless of how small the difference
3067 *
3068 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3069 * However, because tg->load_avg is a global value there are performance
3070 * considerations.
3071 *
3072 * In order to avoid having to look at the other cfs_rq's, we use a
3073 * differential update where we store the last value we propagated. This in
3074 * turn allows skipping updates if the differential is 'small'.
3075 *
3076 * Updating tg's load_avg is necessary before update_cfs_share().
3077 */
3078static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3079{
3080 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3081
3082 /*
3083 * No need to update load_avg for root_task_group as it is not used.
3084 */
3085 if (cfs_rq->tg == &root_task_group)
3086 return;
3087
3088 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3089 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3090 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3091 }
3092}
3093
3094/*
3095 * Called within set_task_rq() right before setting a task's CPU. The
3096 * caller only guarantees p->pi_lock is held; no other assumptions,
3097 * including the state of rq->lock, should be made.
3098 */
3099void set_task_rq_fair(struct sched_entity *se,
3100 struct cfs_rq *prev, struct cfs_rq *next)
3101{
3102 u64 p_last_update_time;
3103 u64 n_last_update_time;
3104
3105 if (!sched_feat(ATTACH_AGE_LOAD))
3106 return;
3107
3108 /*
3109 * We are supposed to update the task to "current" time, then its up to
3110 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3111 * getting what current time is, so simply throw away the out-of-date
3112 * time. This will result in the wakee task is less decayed, but giving
3113 * the wakee more load sounds not bad.
3114 */
3115 if (!(se->avg.last_update_time && prev))
3116 return;
3117
3118#ifndef CONFIG_64BIT
3119 {
3120 u64 p_last_update_time_copy;
3121 u64 n_last_update_time_copy;
3122
3123 do {
3124 p_last_update_time_copy = prev->load_last_update_time_copy;
3125 n_last_update_time_copy = next->load_last_update_time_copy;
3126
3127 smp_rmb();
3128
3129 p_last_update_time = prev->avg.last_update_time;
3130 n_last_update_time = next->avg.last_update_time;
3131
3132 } while (p_last_update_time != p_last_update_time_copy ||
3133 n_last_update_time != n_last_update_time_copy);
3134 }
3135#else
3136 p_last_update_time = prev->avg.last_update_time;
3137 n_last_update_time = next->avg.last_update_time;
3138#endif
3139 __update_load_avg_blocked_se(p_last_update_time, se);
3140 se->avg.last_update_time = n_last_update_time;
3141}
3142
3143
3144/*
3145 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3146 * propagate its contribution. The key to this propagation is the invariant
3147 * that for each group:
3148 *
3149 * ge->avg == grq->avg (1)
3150 *
3151 * _IFF_ we look at the pure running and runnable sums. Because they
3152 * represent the very same entity, just at different points in the hierarchy.
3153 *
3154 * Per the above update_tg_cfs_util() is trivial and simply copies the running
3155 * sum over (but still wrong, because the group entity and group rq do not have
3156 * their PELT windows aligned).
3157 *
3158 * However, update_tg_cfs_runnable() is more complex. So we have:
3159 *
3160 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3161 *
3162 * And since, like util, the runnable part should be directly transferable,
3163 * the following would _appear_ to be the straight forward approach:
3164 *
3165 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3166 *
3167 * And per (1) we have:
3168 *
3169 * ge->avg.runnable_avg == grq->avg.runnable_avg
3170 *
3171 * Which gives:
3172 *
3173 * ge->load.weight * grq->avg.load_avg
3174 * ge->avg.load_avg = ----------------------------------- (4)
3175 * grq->load.weight
3176 *
3177 * Except that is wrong!
3178 *
3179 * Because while for entities historical weight is not important and we
3180 * really only care about our future and therefore can consider a pure
3181 * runnable sum, runqueues can NOT do this.
3182 *
3183 * We specifically want runqueues to have a load_avg that includes
3184 * historical weights. Those represent the blocked load, the load we expect
3185 * to (shortly) return to us. This only works by keeping the weights as
3186 * integral part of the sum. We therefore cannot decompose as per (3).
3187 *
3188 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3189 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3190 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3191 * runnable section of these tasks overlap (or not). If they were to perfectly
3192 * align the rq as a whole would be runnable 2/3 of the time. If however we
3193 * always have at least 1 runnable task, the rq as a whole is always runnable.
3194 *
3195 * So we'll have to approximate.. :/
3196 *
3197 * Given the constraint:
3198 *
3199 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3200 *
3201 * We can construct a rule that adds runnable to a rq by assuming minimal
3202 * overlap.
3203 *
3204 * On removal, we'll assume each task is equally runnable; which yields:
3205 *
3206 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3207 *
3208 * XXX: only do this for the part of runnable > running ?
3209 *
3210 */
3211
3212static inline void
3213update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3214{
3215 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3216
3217 /* Nothing to update */
3218 if (!delta)
3219 return;
3220
3221 /*
3222 * The relation between sum and avg is:
3223 *
3224 * LOAD_AVG_MAX - 1024 + sa->period_contrib
3225 *
3226 * however, the PELT windows are not aligned between grq and gse.
3227 */
3228
3229 /* Set new sched_entity's utilization */
3230 se->avg.util_avg = gcfs_rq->avg.util_avg;
3231 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3232
3233 /* Update parent cfs_rq utilization */
3234 add_positive(&cfs_rq->avg.util_avg, delta);
3235 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3236}
3237
3238static inline void
3239update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3240{
3241 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3242 unsigned long runnable_load_avg, load_avg;
3243 u64 runnable_load_sum, load_sum = 0;
3244 s64 delta_sum;
3245
3246 if (!runnable_sum)
3247 return;
3248
3249 gcfs_rq->prop_runnable_sum = 0;
3250
3251 if (runnable_sum >= 0) {
3252 /*
3253 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3254 * the CPU is saturated running == runnable.
3255 */
3256 runnable_sum += se->avg.load_sum;
3257 runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
3258 } else {
3259 /*
3260 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3261 * assuming all tasks are equally runnable.
3262 */
3263 if (scale_load_down(gcfs_rq->load.weight)) {
3264 load_sum = div_s64(gcfs_rq->avg.load_sum,
3265 scale_load_down(gcfs_rq->load.weight));
3266 }
3267
3268 /* But make sure to not inflate se's runnable */
3269 runnable_sum = min(se->avg.load_sum, load_sum);
3270 }
3271
3272 /*
3273 * runnable_sum can't be lower than running_sum
3274 * Rescale running sum to be in the same range as runnable sum
3275 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
3276 * runnable_sum is in [0 : LOAD_AVG_MAX]
3277 */
3278 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
3279 runnable_sum = max(runnable_sum, running_sum);
3280
3281 load_sum = (s64)se_weight(se) * runnable_sum;
3282 load_avg = div_s64(load_sum, LOAD_AVG_MAX);
3283
3284 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3285 delta_avg = load_avg - se->avg.load_avg;
3286
3287 se->avg.load_sum = runnable_sum;
3288 se->avg.load_avg = load_avg;
3289 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3290 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3291
3292 runnable_load_sum = (s64)se_runnable(se) * runnable_sum;
3293 runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX);
3294 delta_sum = runnable_load_sum - se_weight(se) * se->avg.runnable_load_sum;
3295 delta_avg = runnable_load_avg - se->avg.runnable_load_avg;
3296
3297 se->avg.runnable_load_sum = runnable_sum;
3298 se->avg.runnable_load_avg = runnable_load_avg;
3299
3300 if (se->on_rq) {
3301 add_positive(&cfs_rq->avg.runnable_load_avg, delta_avg);
3302 add_positive(&cfs_rq->avg.runnable_load_sum, delta_sum);
3303 }
3304}
3305
3306static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3307{
3308 cfs_rq->propagate = 1;
3309 cfs_rq->prop_runnable_sum += runnable_sum;
3310}
3311
3312/* Update task and its cfs_rq load average */
3313static inline int propagate_entity_load_avg(struct sched_entity *se)
3314{
3315 struct cfs_rq *cfs_rq, *gcfs_rq;
3316
3317 if (entity_is_task(se))
3318 return 0;
3319
3320 gcfs_rq = group_cfs_rq(se);
3321 if (!gcfs_rq->propagate)
3322 return 0;
3323
3324 gcfs_rq->propagate = 0;
3325
3326 cfs_rq = cfs_rq_of(se);
3327
3328 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3329
3330 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3331 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3332
3333 return 1;
3334}
3335
3336/*
3337 * Check if we need to update the load and the utilization of a blocked
3338 * group_entity:
3339 */
3340static inline bool skip_blocked_update(struct sched_entity *se)
3341{
3342 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3343
3344 /*
3345 * If sched_entity still have not zero load or utilization, we have to
3346 * decay it:
3347 */
3348 if (se->avg.load_avg || se->avg.util_avg)
3349 return false;
3350
3351 /*
3352 * If there is a pending propagation, we have to update the load and
3353 * the utilization of the sched_entity:
3354 */
3355 if (gcfs_rq->propagate)
3356 return false;
3357
3358 /*
3359 * Otherwise, the load and the utilization of the sched_entity is
3360 * already zero and there is no pending propagation, so it will be a
3361 * waste of time to try to decay it:
3362 */
3363 return true;
3364}
3365
3366#else /* CONFIG_FAIR_GROUP_SCHED */
3367
3368static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3369
3370static inline int propagate_entity_load_avg(struct sched_entity *se)
3371{
3372 return 0;
3373}
3374
3375static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3376
3377#endif /* CONFIG_FAIR_GROUP_SCHED */
3378
3379/**
3380 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3381 * @now: current time, as per cfs_rq_clock_pelt()
3382 * @cfs_rq: cfs_rq to update
3383 *
3384 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3385 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3386 * post_init_entity_util_avg().
3387 *
3388 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3389 *
3390 * Returns true if the load decayed or we removed load.
3391 *
3392 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3393 * call update_tg_load_avg() when this function returns true.
3394 */
3395static inline int
3396update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3397{
3398 unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0;
3399 struct sched_avg *sa = &cfs_rq->avg;
3400 int decayed = 0;
3401
3402 if (cfs_rq->removed.nr) {
3403 unsigned long r;
3404 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3405
3406 raw_spin_lock(&cfs_rq->removed.lock);
3407 swap(cfs_rq->removed.util_avg, removed_util);
3408 swap(cfs_rq->removed.load_avg, removed_load);
3409 swap(cfs_rq->removed.runnable_sum, removed_runnable_sum);
3410 cfs_rq->removed.nr = 0;
3411 raw_spin_unlock(&cfs_rq->removed.lock);
3412
3413 r = removed_load;
3414 sub_positive(&sa->load_avg, r);
3415 sub_positive(&sa->load_sum, r * divider);
3416
3417 r = removed_util;
3418 sub_positive(&sa->util_avg, r);
3419 sub_positive(&sa->util_sum, r * divider);
3420
3421 add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum);
3422
3423 decayed = 1;
3424 }
3425
3426 decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
3427
3428#ifndef CONFIG_64BIT
3429 smp_wmb();
3430 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3431#endif
3432
3433 if (decayed)
3434 cfs_rq_util_change(cfs_rq, 0);
3435
3436 return decayed;
3437}
3438
3439/**
3440 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3441 * @cfs_rq: cfs_rq to attach to
3442 * @se: sched_entity to attach
3443 * @flags: migration hints
3444 *
3445 * Must call update_cfs_rq_load_avg() before this, since we rely on
3446 * cfs_rq->avg.last_update_time being current.
3447 */
3448static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3449{
3450 u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3451
3452 /*
3453 * When we attach the @se to the @cfs_rq, we must align the decay
3454 * window because without that, really weird and wonderful things can
3455 * happen.
3456 *
3457 * XXX illustrate
3458 */
3459 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3460 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3461
3462 /*
3463 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3464 * period_contrib. This isn't strictly correct, but since we're
3465 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3466 * _sum a little.
3467 */
3468 se->avg.util_sum = se->avg.util_avg * divider;
3469
3470 se->avg.load_sum = divider;
3471 if (se_weight(se)) {
3472 se->avg.load_sum =
3473 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3474 }
3475
3476 se->avg.runnable_load_sum = se->avg.load_sum;
3477
3478 enqueue_load_avg(cfs_rq, se);
3479 cfs_rq->avg.util_avg += se->avg.util_avg;
3480 cfs_rq->avg.util_sum += se->avg.util_sum;
3481
3482 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3483
3484 cfs_rq_util_change(cfs_rq, flags);
3485}
3486
3487/**
3488 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3489 * @cfs_rq: cfs_rq to detach from
3490 * @se: sched_entity to detach
3491 *
3492 * Must call update_cfs_rq_load_avg() before this, since we rely on
3493 * cfs_rq->avg.last_update_time being current.
3494 */
3495static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3496{
3497 dequeue_load_avg(cfs_rq, se);
3498 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3499 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3500
3501 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3502
3503 cfs_rq_util_change(cfs_rq, 0);
3504}
3505
3506/*
3507 * Optional action to be done while updating the load average
3508 */
3509#define UPDATE_TG 0x1
3510#define SKIP_AGE_LOAD 0x2
3511#define DO_ATTACH 0x4
3512
3513/* Update task and its cfs_rq load average */
3514static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3515{
3516 u64 now = cfs_rq_clock_pelt(cfs_rq);
3517 int decayed;
3518
3519 /*
3520 * Track task load average for carrying it to new CPU after migrated, and
3521 * track group sched_entity load average for task_h_load calc in migration
3522 */
3523 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3524 __update_load_avg_se(now, cfs_rq, se);
3525
3526 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3527 decayed |= propagate_entity_load_avg(se);
3528
3529 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3530
3531 /*
3532 * DO_ATTACH means we're here from enqueue_entity().
3533 * !last_update_time means we've passed through
3534 * migrate_task_rq_fair() indicating we migrated.
3535 *
3536 * IOW we're enqueueing a task on a new CPU.
3537 */
3538 attach_entity_load_avg(cfs_rq, se, SCHED_CPUFREQ_MIGRATION);
3539 update_tg_load_avg(cfs_rq, 0);
3540
3541 } else if (decayed && (flags & UPDATE_TG))
3542 update_tg_load_avg(cfs_rq, 0);
3543}
3544
3545#ifndef CONFIG_64BIT
3546static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3547{
3548 u64 last_update_time_copy;
3549 u64 last_update_time;
3550
3551 do {
3552 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3553 smp_rmb();
3554 last_update_time = cfs_rq->avg.last_update_time;
3555 } while (last_update_time != last_update_time_copy);
3556
3557 return last_update_time;
3558}
3559#else
3560static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3561{
3562 return cfs_rq->avg.last_update_time;
3563}
3564#endif
3565
3566/*
3567 * Synchronize entity load avg of dequeued entity without locking
3568 * the previous rq.
3569 */
3570void sync_entity_load_avg(struct sched_entity *se)
3571{
3572 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3573 u64 last_update_time;
3574
3575 last_update_time = cfs_rq_last_update_time(cfs_rq);
3576 __update_load_avg_blocked_se(last_update_time, se);
3577}
3578
3579/*
3580 * Task first catches up with cfs_rq, and then subtract
3581 * itself from the cfs_rq (task must be off the queue now).
3582 */
3583void remove_entity_load_avg(struct sched_entity *se)
3584{
3585 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3586 unsigned long flags;
3587
3588 /*
3589 * tasks cannot exit without having gone through wake_up_new_task() ->
3590 * post_init_entity_util_avg() which will have added things to the
3591 * cfs_rq, so we can remove unconditionally.
3592 */
3593
3594 sync_entity_load_avg(se);
3595
3596 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3597 ++cfs_rq->removed.nr;
3598 cfs_rq->removed.util_avg += se->avg.util_avg;
3599 cfs_rq->removed.load_avg += se->avg.load_avg;
3600 cfs_rq->removed.runnable_sum += se->avg.load_sum; /* == runnable_sum */
3601 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3602}
3603
3604static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3605{
3606 return cfs_rq->avg.runnable_load_avg;
3607}
3608
3609static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3610{
3611 return cfs_rq->avg.load_avg;
3612}
3613
3614static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3615
3616static inline unsigned long task_util(struct task_struct *p)
3617{
3618 return READ_ONCE(p->se.avg.util_avg);
3619}
3620
3621static inline unsigned long _task_util_est(struct task_struct *p)
3622{
3623 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3624
3625 return (max(ue.ewma, ue.enqueued) | UTIL_AVG_UNCHANGED);
3626}
3627
3628static inline unsigned long task_util_est(struct task_struct *p)
3629{
3630 return max(task_util(p), _task_util_est(p));
3631}
3632
3633static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3634 struct task_struct *p)
3635{
3636 unsigned int enqueued;
3637
3638 if (!sched_feat(UTIL_EST))
3639 return;
3640
3641 /* Update root cfs_rq's estimated utilization */
3642 enqueued = cfs_rq->avg.util_est.enqueued;
3643 enqueued += _task_util_est(p);
3644 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3645}
3646
3647/*
3648 * Check if a (signed) value is within a specified (unsigned) margin,
3649 * based on the observation that:
3650 *
3651 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3652 *
3653 * NOTE: this only works when value + maring < INT_MAX.
3654 */
3655static inline bool within_margin(int value, int margin)
3656{
3657 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3658}
3659
3660static void
3661util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep)
3662{
3663 long last_ewma_diff;
3664 struct util_est ue;
3665 int cpu;
3666
3667 if (!sched_feat(UTIL_EST))
3668 return;
3669
3670 /* Update root cfs_rq's estimated utilization */
3671 ue.enqueued = cfs_rq->avg.util_est.enqueued;
3672 ue.enqueued -= min_t(unsigned int, ue.enqueued, _task_util_est(p));
3673 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued);
3674
3675 /*
3676 * Skip update of task's estimated utilization when the task has not
3677 * yet completed an activation, e.g. being migrated.
3678 */
3679 if (!task_sleep)
3680 return;
3681
3682 /*
3683 * If the PELT values haven't changed since enqueue time,
3684 * skip the util_est update.
3685 */
3686 ue = p->se.avg.util_est;
3687 if (ue.enqueued & UTIL_AVG_UNCHANGED)
3688 return;
3689
3690 /*
3691 * Skip update of task's estimated utilization when its EWMA is
3692 * already ~1% close to its last activation value.
3693 */
3694 ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
3695 last_ewma_diff = ue.enqueued - ue.ewma;
3696 if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
3697 return;
3698
3699 /*
3700 * To avoid overestimation of actual task utilization, skip updates if
3701 * we cannot grant there is idle time in this CPU.
3702 */
3703 cpu = cpu_of(rq_of(cfs_rq));
3704 if (task_util(p) > capacity_orig_of(cpu))
3705 return;
3706
3707 /*
3708 * Update Task's estimated utilization
3709 *
3710 * When *p completes an activation we can consolidate another sample
3711 * of the task size. This is done by storing the current PELT value
3712 * as ue.enqueued and by using this value to update the Exponential
3713 * Weighted Moving Average (EWMA):
3714 *
3715 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
3716 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
3717 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
3718 * = w * ( last_ewma_diff ) + ewma(t-1)
3719 * = w * (last_ewma_diff + ewma(t-1) / w)
3720 *
3721 * Where 'w' is the weight of new samples, which is configured to be
3722 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
3723 */
3724 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
3725 ue.ewma += last_ewma_diff;
3726 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
3727 WRITE_ONCE(p->se.avg.util_est, ue);
3728}
3729
3730static inline int task_fits_capacity(struct task_struct *p, long capacity)
3731{
3732 return capacity * 1024 > task_util_est(p) * capacity_margin;
3733}
3734
3735static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
3736{
3737 if (!static_branch_unlikely(&sched_asym_cpucapacity))
3738 return;
3739
3740 if (!p) {
3741 rq->misfit_task_load = 0;
3742 return;
3743 }
3744
3745 if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
3746 rq->misfit_task_load = 0;
3747 return;
3748 }
3749
3750 rq->misfit_task_load = task_h_load(p);
3751}
3752
3753#else /* CONFIG_SMP */
3754
3755#define UPDATE_TG 0x0
3756#define SKIP_AGE_LOAD 0x0
3757#define DO_ATTACH 0x0
3758
3759static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
3760{
3761 cfs_rq_util_change(cfs_rq, 0);
3762}
3763
3764static inline void remove_entity_load_avg(struct sched_entity *se) {}
3765
3766static inline void
3767attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) {}
3768static inline void
3769detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3770
3771static inline int idle_balance(struct rq *rq, struct