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"
25 #include <trace/events/sched.h>
27 /*
28 * Targeted preemption latency for CPU-bound tasks:
29 *
30 * NOTE: this latency value is not the same as the concept of
31 * 'timeslice length' - timeslices in CFS are of variable length
32 * and have no persistent notion like in traditional, time-slice
33 * based scheduling concepts.
34 *
35 * (to see the precise effective timeslice length of your workload,
36 * run vmstat and monitor the context-switches (cs) field)
37 *
38 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
39 */
40 unsigned int sysctl_sched_latency = 6000000ULL;
41 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
43 /*
44 * The initial- and re-scaling of tunables is configurable
45 *
46 * Options are:
47 *
48 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
49 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
50 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
51 *
52 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
53 */
54 enum sched_tunable_scaling sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
56 /*
57 * Minimal preemption granularity for CPU-bound tasks:
58 *
59 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
60 */
61 unsigned int sysctl_sched_min_granularity = 750000ULL;
62 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
64 /*
65 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
66 */
67 static unsigned int sched_nr_latency = 8;
69 /*
70 * After fork, child runs first. If set to 0 (default) then
71 * parent will (try to) run first.
72 */
73 unsigned int sysctl_sched_child_runs_first __read_mostly;
75 /*
76 * SCHED_OTHER wake-up granularity.
77 *
78 * This option delays the preemption effects of decoupled workloads
79 * and reduces their over-scheduling. Synchronous workloads will still
80 * have immediate wakeup/sleep latencies.
81 *
82 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
83 */
84 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
85 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
87 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
89 #ifdef CONFIG_SMP
90 /*
91 * For asym packing, by default the lower numbered CPU has higher priority.
92 */
93 int __weak arch_asym_cpu_priority(int cpu)
94 {
95 return -cpu;
96 }
97 #endif
99 #ifdef CONFIG_CFS_BANDWIDTH
100 /*
101 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
102 * each time a cfs_rq requests quota.
103 *
104 * Note: in the case that the slice exceeds the runtime remaining (either due
105 * to consumption or the quota being specified to be smaller than the slice)
106 * we will always only issue the remaining available time.
107 *
108 * (default: 5 msec, units: microseconds)
109 */
110 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
111 #endif
113 /*
114 * The margin used when comparing utilization with CPU capacity:
115 * util * margin < capacity * 1024
116 *
117 * (default: ~20%)
118 */
119 unsigned int capacity_margin = 1280;
121 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
122 {
123 lw->weight += inc;
124 lw->inv_weight = 0;
125 }
127 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
128 {
129 lw->weight -= dec;
130 lw->inv_weight = 0;
131 }
133 static inline void update_load_set(struct load_weight *lw, unsigned long w)
134 {
135 lw->weight = w;
136 lw->inv_weight = 0;
137 }
139 /*
140 * Increase the granularity value when there are more CPUs,
141 * because with more CPUs the 'effective latency' as visible
142 * to users decreases. But the relationship is not linear,
143 * so pick a second-best guess by going with the log2 of the
144 * number of CPUs.
145 *
146 * This idea comes from the SD scheduler of Con Kolivas:
147 */
148 static unsigned int get_update_sysctl_factor(void)
149 {
150 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
151 unsigned int factor;
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 }
166 return factor;
167 }
169 static void update_sysctl(void)
170 {
171 unsigned int factor = get_update_sysctl_factor();
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 }
181 void sched_init_granularity(void)
182 {
183 update_sysctl();
184 }
186 #define WMULT_CONST (~0U)
187 #define WMULT_SHIFT 32
189 static void __update_inv_weight(struct load_weight *lw)
190 {
191 unsigned long w;
193 if (likely(lw->inv_weight))
194 return;
196 w = scale_load_down(lw->weight);
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 }
206 /*
207 * delta_exec * weight / lw.weight
208 * OR
209 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
210 *
211 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
212 * we're guaranteed shift stays positive because inv_weight is guaranteed to
213 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
214 *
215 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
216 * weight/lw.weight <= 1, and therefore our shift will also be positive.
217 */
218 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
219 {
220 u64 fact = scale_load_down(weight);
221 int shift = WMULT_SHIFT;
223 __update_inv_weight(lw);
225 if (unlikely(fact >> 32)) {
226 while (fact >> 32) {
227 fact >>= 1;
228 shift--;
229 }
230 }
232 /* hint to use a 32x32->64 mul */
233 fact = (u64)(u32)fact * lw->inv_weight;
235 while (fact >> 32) {
236 fact >>= 1;
237 shift--;
238 }
240 return mul_u64_u32_shr(delta_exec, fact, shift);
241 }
244 const struct sched_class fair_sched_class;
246 /**************************************************************
247 * CFS operations on generic schedulable entities:
248 */
250 #ifdef CONFIG_FAIR_GROUP_SCHED
252 /* cpu runqueue to which this cfs_rq is attached */
253 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
254 {
255 return cfs_rq->rq;
256 }
258 static inline struct task_struct *task_of(struct sched_entity *se)
259 {
260 SCHED_WARN_ON(!entity_is_task(se));
261 return container_of(se, struct task_struct, se);
262 }
264 /* Walk up scheduling entities hierarchy */
265 #define for_each_sched_entity(se) \
266 for (; se; se = se->parent)
268 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
269 {
270 return p->se.cfs_rq;
271 }
273 /* runqueue on which this entity is (to be) queued */
274 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
275 {
276 return se->cfs_rq;
277 }
279 /* runqueue "owned" by this group */
280 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
281 {
282 return grp->my_q;
283 }
285 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
286 {
287 if (!cfs_rq->on_list) {
288 struct rq *rq = rq_of(cfs_rq);
289 int cpu = cpu_of(rq);
290 /*
291 * Ensure we either appear before our parent (if already
292 * enqueued) or force our parent to appear after us when it is
293 * enqueued. The fact that we always enqueue bottom-up
294 * reduces this to two cases and a special case for the root
295 * cfs_rq. Furthermore, it also means that we will always reset
296 * tmp_alone_branch either when the branch is connected
297 * to a tree or when we reach the beg of the tree
298 */
299 if (cfs_rq->tg->parent &&
300 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
301 /*
302 * If parent is already on the list, we add the child
303 * just before. Thanks to circular linked property of
304 * the list, this means to put the child at the tail
305 * of the list that starts by parent.
306 */
307 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
308 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
309 /*
310 * The branch is now connected to its tree so we can
311 * reset tmp_alone_branch to the beginning of the
312 * list.
313 */
314 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
315 } else if (!cfs_rq->tg->parent) {
316 /*
317 * cfs rq without parent should be put
318 * at the tail of the list.
319 */
320 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
321 &rq->leaf_cfs_rq_list);
322 /*
323 * We have reach the beg of a tree so we can reset
324 * tmp_alone_branch to the beginning of the list.
325 */
326 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
327 } else {
328 /*
329 * The parent has not already been added so we want to
330 * make sure that it will be put after us.
331 * tmp_alone_branch points to the beg of the branch
332 * where we will add parent.
333 */
334 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
335 rq->tmp_alone_branch);
336 /*
337 * update tmp_alone_branch to points to the new beg
338 * of the branch
339 */
340 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
341 }
343 cfs_rq->on_list = 1;
344 }
345 }
347 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
348 {
349 if (cfs_rq->on_list) {
350 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
351 cfs_rq->on_list = 0;
352 }
353 }
355 /* Iterate thr' all leaf cfs_rq's on a runqueue */
356 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
357 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
358 leaf_cfs_rq_list)
360 /* Do the two (enqueued) entities belong to the same group ? */
361 static inline struct cfs_rq *
362 is_same_group(struct sched_entity *se, struct sched_entity *pse)
363 {
364 if (se->cfs_rq == pse->cfs_rq)
365 return se->cfs_rq;
367 return NULL;
368 }
370 static inline struct sched_entity *parent_entity(struct sched_entity *se)
371 {
372 return se->parent;
373 }
375 static void
376 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
377 {
378 int se_depth, pse_depth;
380 /*
381 * preemption test can be made between sibling entities who are in the
382 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
383 * both tasks until we find their ancestors who are siblings of common
384 * parent.
385 */
387 /* First walk up until both entities are at same depth */
388 se_depth = (*se)->depth;
389 pse_depth = (*pse)->depth;
391 while (se_depth > pse_depth) {
392 se_depth--;
393 *se = parent_entity(*se);
394 }
396 while (pse_depth > se_depth) {
397 pse_depth--;
398 *pse = parent_entity(*pse);
399 }
401 while (!is_same_group(*se, *pse)) {
402 *se = parent_entity(*se);
403 *pse = parent_entity(*pse);
404 }
405 }
407 #else /* !CONFIG_FAIR_GROUP_SCHED */
409 static inline struct task_struct *task_of(struct sched_entity *se)
410 {
411 return container_of(se, struct task_struct, se);
412 }
414 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
415 {
416 return container_of(cfs_rq, struct rq, cfs);
417 }
420 #define for_each_sched_entity(se) \
421 for (; se; se = NULL)
423 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
424 {
425 return &task_rq(p)->cfs;
426 }
428 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
429 {
430 struct task_struct *p = task_of(se);
431 struct rq *rq = task_rq(p);
433 return &rq->cfs;
434 }
436 /* runqueue "owned" by this group */
437 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
438 {
439 return NULL;
440 }
442 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
443 {
444 }
446 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
447 {
448 }
450 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
451 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
453 static inline struct sched_entity *parent_entity(struct sched_entity *se)
454 {
455 return NULL;
456 }
458 static inline void
459 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
460 {
461 }
463 #endif /* CONFIG_FAIR_GROUP_SCHED */
465 static __always_inline
466 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
468 /**************************************************************
469 * Scheduling class tree data structure manipulation methods:
470 */
472 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
473 {
474 s64 delta = (s64)(vruntime - max_vruntime);
475 if (delta > 0)
476 max_vruntime = vruntime;
478 return max_vruntime;
479 }
481 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
482 {
483 s64 delta = (s64)(vruntime - min_vruntime);
484 if (delta < 0)
485 min_vruntime = vruntime;
487 return min_vruntime;
488 }
490 static inline int entity_before(struct sched_entity *a,
491 struct sched_entity *b)
492 {
493 return (s64)(a->vruntime - b->vruntime) < 0;
494 }
496 static void update_min_vruntime(struct cfs_rq *cfs_rq)
497 {
498 struct sched_entity *curr = cfs_rq->curr;
499 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
501 u64 vruntime = cfs_rq->min_vruntime;
503 if (curr) {
504 if (curr->on_rq)
505 vruntime = curr->vruntime;
506 else
507 curr = NULL;
508 }
510 if (leftmost) { /* non-empty tree */
511 struct sched_entity *se;
512 se = rb_entry(leftmost, struct sched_entity, run_node);
514 if (!curr)
515 vruntime = se->vruntime;
516 else
517 vruntime = min_vruntime(vruntime, se->vruntime);
518 }
520 /* ensure we never gain time by being placed backwards. */
521 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
522 #ifndef CONFIG_64BIT
523 smp_wmb();
524 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
525 #endif
526 }
528 /*
529 * Enqueue an entity into the rb-tree:
530 */
531 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
532 {
533 struct rb_node **link = &cfs_rq->tasks_timeline.rb_root.rb_node;
534 struct rb_node *parent = NULL;
535 struct sched_entity *entry;
536 bool leftmost = true;
538 /*
539 * Find the right place in the rbtree:
540 */
541 while (*link) {
542 parent = *link;
543 entry = rb_entry(parent, struct sched_entity, run_node);
544 /*
545 * We dont care about collisions. Nodes with
546 * the same key stay together.
547 */
548 if (entity_before(se, entry)) {
549 link = &parent->rb_left;
550 } else {
551 link = &parent->rb_right;
552 leftmost = false;
553 }
554 }
556 rb_link_node(&se->run_node, parent, link);
557 rb_insert_color_cached(&se->run_node,
558 &cfs_rq->tasks_timeline, leftmost);
559 }
561 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
562 {
563 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
564 }
566 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
567 {
568 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
570 if (!left)
571 return NULL;
573 return rb_entry(left, struct sched_entity, run_node);
574 }
576 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
577 {
578 struct rb_node *next = rb_next(&se->run_node);
580 if (!next)
581 return NULL;
583 return rb_entry(next, struct sched_entity, run_node);
584 }
586 #ifdef CONFIG_SCHED_DEBUG
587 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
588 {
589 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
591 if (!last)
592 return NULL;
594 return rb_entry(last, struct sched_entity, run_node);
595 }
597 /**************************************************************
598 * Scheduling class statistics methods:
599 */
601 int sched_proc_update_handler(struct ctl_table *table, int write,
602 void __user *buffer, size_t *lenp,
603 loff_t *ppos)
604 {
605 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
606 unsigned int factor = get_update_sysctl_factor();
608 if (ret || !write)
609 return ret;
611 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
612 sysctl_sched_min_granularity);
614 #define WRT_SYSCTL(name) \
615 (normalized_sysctl_##name = sysctl_##name / (factor))
616 WRT_SYSCTL(sched_min_granularity);
617 WRT_SYSCTL(sched_latency);
618 WRT_SYSCTL(sched_wakeup_granularity);
619 #undef WRT_SYSCTL
621 return 0;
622 }
623 #endif
625 /*
626 * delta /= w
627 */
628 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
629 {
630 if (unlikely(se->load.weight != NICE_0_LOAD))
631 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
633 return delta;
634 }
636 /*
637 * The idea is to set a period in which each task runs once.
638 *
639 * When there are too many tasks (sched_nr_latency) we have to stretch
640 * this period because otherwise the slices get too small.
641 *
642 * p = (nr <= nl) ? l : l*nr/nl
643 */
644 static u64 __sched_period(unsigned long nr_running)
645 {
646 if (unlikely(nr_running > sched_nr_latency))
647 return nr_running * sysctl_sched_min_granularity;
648 else
649 return sysctl_sched_latency;
650 }
652 /*
653 * We calculate the wall-time slice from the period by taking a part
654 * proportional to the weight.
655 *
656 * s = p*P[w/rw]
657 */
658 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
659 {
660 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
662 for_each_sched_entity(se) {
663 struct load_weight *load;
664 struct load_weight lw;
666 cfs_rq = cfs_rq_of(se);
667 load = &cfs_rq->load;
669 if (unlikely(!se->on_rq)) {
670 lw = cfs_rq->load;
672 update_load_add(&lw, se->load.weight);
673 load = &lw;
674 }
675 slice = __calc_delta(slice, se->load.weight, load);
676 }
677 return slice;
678 }
680 /*
681 * We calculate the vruntime slice of a to-be-inserted task.
682 *
683 * vs = s/w
684 */
685 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
686 {
687 return calc_delta_fair(sched_slice(cfs_rq, se), se);
688 }
690 #ifdef CONFIG_SMP
691 #include "pelt.h"
692 #include "sched-pelt.h"
694 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
695 static unsigned long task_h_load(struct task_struct *p);
697 /* Give new sched_entity start runnable values to heavy its load in infant time */
698 void init_entity_runnable_average(struct sched_entity *se)
699 {
700 struct sched_avg *sa = &se->avg;
702 memset(sa, 0, sizeof(*sa));
704 /*
705 * Tasks are intialized with full load to be seen as heavy tasks until
706 * they get a chance to stabilize to their real load level.
707 * Group entities are intialized with zero load to reflect the fact that
708 * nothing has been attached to the task group yet.
709 */
710 if (entity_is_task(se))
711 sa->runnable_load_avg = sa->load_avg = scale_load_down(se->load.weight);
713 se->runnable_weight = se->load.weight;
715 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
716 }
718 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
719 static void attach_entity_cfs_rq(struct sched_entity *se);
721 /*
722 * With new tasks being created, their initial util_avgs are extrapolated
723 * based on the cfs_rq's current util_avg:
724 *
725 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
726 *
727 * However, in many cases, the above util_avg does not give a desired
728 * value. Moreover, the sum of the util_avgs may be divergent, such
729 * as when the series is a harmonic series.
730 *
731 * To solve this problem, we also cap the util_avg of successive tasks to
732 * only 1/2 of the left utilization budget:
733 *
734 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
735 *
736 * where n denotes the nth task and cpu_scale the CPU capacity.
737 *
738 * For example, for a CPU with 1024 of capacity, a simplest series from
739 * the beginning would be like:
740 *
741 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
742 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
743 *
744 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
745 * if util_avg > util_avg_cap.
746 */
747 void post_init_entity_util_avg(struct sched_entity *se)
748 {
749 struct cfs_rq *cfs_rq = cfs_rq_of(se);
750 struct sched_avg *sa = &se->avg;
751 long cpu_scale = arch_scale_cpu_capacity(NULL, cpu_of(rq_of(cfs_rq)));
752 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
754 if (cap > 0) {
755 if (cfs_rq->avg.util_avg != 0) {
756 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
757 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
759 if (sa->util_avg > cap)
760 sa->util_avg = cap;
761 } else {
762 sa->util_avg = cap;
763 }
764 }
766 if (entity_is_task(se)) {
767 struct task_struct *p = task_of(se);
768 if (p->sched_class != &fair_sched_class) {
769 /*
770 * For !fair tasks do:
771 *
772 update_cfs_rq_load_avg(now, cfs_rq);
773 attach_entity_load_avg(cfs_rq, se, 0);
774 switched_from_fair(rq, p);
775 *
776 * such that the next switched_to_fair() has the
777 * expected state.
778 */
779 se->avg.last_update_time = cfs_rq_clock_task(cfs_rq);
780 return;
781 }
782 }
784 attach_entity_cfs_rq(se);
785 }
787 #else /* !CONFIG_SMP */
788 void init_entity_runnable_average(struct sched_entity *se)
789 {
790 }
791 void post_init_entity_util_avg(struct sched_entity *se)
792 {
793 }
794 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
795 {
796 }
797 #endif /* CONFIG_SMP */
799 /*
800 * Update the current task's runtime statistics.
801 */
802 static void update_curr(struct cfs_rq *cfs_rq)
803 {
804 struct sched_entity *curr = cfs_rq->curr;
805 u64 now = rq_clock_task(rq_of(cfs_rq));
806 u64 delta_exec;
808 if (unlikely(!curr))
809 return;
811 delta_exec = now - curr->exec_start;
812 if (unlikely((s64)delta_exec <= 0))
813 return;
815 curr->exec_start = now;
817 schedstat_set(curr->statistics.exec_max,
818 max(delta_exec, curr->statistics.exec_max));
820 curr->sum_exec_runtime += delta_exec;
821 schedstat_add(cfs_rq->exec_clock, delta_exec);
823 curr->vruntime += calc_delta_fair(delta_exec, curr);
824 update_min_vruntime(cfs_rq);
826 if (entity_is_task(curr)) {
827 struct task_struct *curtask = task_of(curr);
829 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
830 cgroup_account_cputime(curtask, delta_exec);
831 account_group_exec_runtime(curtask, delta_exec);
832 }
834 account_cfs_rq_runtime(cfs_rq, delta_exec);
835 }
837 static void update_curr_fair(struct rq *rq)
838 {
839 update_curr(cfs_rq_of(&rq->curr->se));
840 }
842 static inline void
843 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
844 {
845 u64 wait_start, prev_wait_start;
847 if (!schedstat_enabled())
848 return;
850 wait_start = rq_clock(rq_of(cfs_rq));
851 prev_wait_start = schedstat_val(se->statistics.wait_start);
853 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
854 likely(wait_start > prev_wait_start))
855 wait_start -= prev_wait_start;
857 __schedstat_set(se->statistics.wait_start, wait_start);
858 }
860 static inline void
861 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
862 {
863 struct task_struct *p;
864 u64 delta;
866 if (!schedstat_enabled())
867 return;
869 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
871 if (entity_is_task(se)) {
872 p = task_of(se);
873 if (task_on_rq_migrating(p)) {
874 /*
875 * Preserve migrating task's wait time so wait_start
876 * time stamp can be adjusted to accumulate wait time
877 * prior to migration.
878 */
879 __schedstat_set(se->statistics.wait_start, delta);
880 return;
881 }
882 trace_sched_stat_wait(p, delta);
883 }
885 __schedstat_set(se->statistics.wait_max,
886 max(schedstat_val(se->statistics.wait_max), delta));
887 __schedstat_inc(se->statistics.wait_count);
888 __schedstat_add(se->statistics.wait_sum, delta);
889 __schedstat_set(se->statistics.wait_start, 0);
890 }
892 static inline void
893 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
894 {
895 struct task_struct *tsk = NULL;
896 u64 sleep_start, block_start;
898 if (!schedstat_enabled())
899 return;
901 sleep_start = schedstat_val(se->statistics.sleep_start);
902 block_start = schedstat_val(se->statistics.block_start);
904 if (entity_is_task(se))
905 tsk = task_of(se);
907 if (sleep_start) {
908 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
910 if ((s64)delta < 0)
911 delta = 0;
913 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
914 __schedstat_set(se->statistics.sleep_max, delta);
916 __schedstat_set(se->statistics.sleep_start, 0);
917 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
919 if (tsk) {
920 account_scheduler_latency(tsk, delta >> 10, 1);
921 trace_sched_stat_sleep(tsk, delta);
922 }
923 }
924 if (block_start) {
925 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
927 if ((s64)delta < 0)
928 delta = 0;
930 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
931 __schedstat_set(se->statistics.block_max, delta);
933 __schedstat_set(se->statistics.block_start, 0);
934 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
936 if (tsk) {
937 if (tsk->in_iowait) {
938 __schedstat_add(se->statistics.iowait_sum, delta);
939 __schedstat_inc(se->statistics.iowait_count);
940 trace_sched_stat_iowait(tsk, delta);
941 }
943 trace_sched_stat_blocked(tsk, delta);
945 /*
946 * Blocking time is in units of nanosecs, so shift by
947 * 20 to get a milliseconds-range estimation of the
948 * amount of time that the task spent sleeping:
949 */
950 if (unlikely(prof_on == SLEEP_PROFILING)) {
951 profile_hits(SLEEP_PROFILING,
952 (void *)get_wchan(tsk),
953 delta >> 20);
954 }
955 account_scheduler_latency(tsk, delta >> 10, 0);
956 }
957 }
958 }
960 /*
961 * Task is being enqueued - update stats:
962 */
963 static inline void
964 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
965 {
966 if (!schedstat_enabled())
967 return;
969 /*
970 * Are we enqueueing a waiting task? (for current tasks
971 * a dequeue/enqueue event is a NOP)
972 */
973 if (se != cfs_rq->curr)
974 update_stats_wait_start(cfs_rq, se);
976 if (flags & ENQUEUE_WAKEUP)
977 update_stats_enqueue_sleeper(cfs_rq, se);
978 }
980 static inline void
981 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
982 {
984 if (!schedstat_enabled())
985 return;
987 /*
988 * Mark the end of the wait period if dequeueing a
989 * waiting task:
990 */
991 if (se != cfs_rq->curr)
992 update_stats_wait_end(cfs_rq, se);
994 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
995 struct task_struct *tsk = task_of(se);
997 if (tsk->state & TASK_INTERRUPTIBLE)
998 __schedstat_set(se->statistics.sleep_start,
999 rq_clock(rq_of(cfs_rq)));
1000 if (tsk->state & TASK_UNINTERRUPTIBLE)
1001 __schedstat_set(se->statistics.block_start,
1002 rq_clock(rq_of(cfs_rq)));
1003 }
1004 }
1006 /*
1007 * We are picking a new current task - update its stats:
1008 */
1009 static inline void
1010 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1011 {
1012 /*
1013 * We are starting a new run period:
1014 */
1015 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1016 }
1018 /**************************************************
1019 * Scheduling class queueing methods:
1020 */
1022 #ifdef CONFIG_NUMA_BALANCING
1023 /*
1024 * Approximate time to scan a full NUMA task in ms. The task scan period is
1025 * calculated based on the tasks virtual memory size and
1026 * numa_balancing_scan_size.
1027 */
1028 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1029 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1031 /* Portion of address space to scan in MB */
1032 unsigned int sysctl_numa_balancing_scan_size = 256;
1034 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1035 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1037 struct numa_group {
1038 atomic_t refcount;
1040 spinlock_t lock; /* nr_tasks, tasks */
1041 int nr_tasks;
1042 pid_t gid;
1043 int active_nodes;
1045 struct rcu_head rcu;
1046 unsigned long total_faults;
1047 unsigned long max_faults_cpu;
1048 /*
1049 * Faults_cpu is used to decide whether memory should move
1050 * towards the CPU. As a consequence, these stats are weighted
1051 * more by CPU use than by memory faults.
1052 */
1053 unsigned long *faults_cpu;
1054 unsigned long faults[0];
1055 };
1057 static inline unsigned long group_faults_priv(struct numa_group *ng);
1058 static inline unsigned long group_faults_shared(struct numa_group *ng);
1060 static unsigned int task_nr_scan_windows(struct task_struct *p)
1061 {
1062 unsigned long rss = 0;
1063 unsigned long nr_scan_pages;
1065 /*
1066 * Calculations based on RSS as non-present and empty pages are skipped
1067 * by the PTE scanner and NUMA hinting faults should be trapped based
1068 * on resident pages
1069 */
1070 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1071 rss = get_mm_rss(p->mm);
1072 if (!rss)
1073 rss = nr_scan_pages;
1075 rss = round_up(rss, nr_scan_pages);
1076 return rss / nr_scan_pages;
1077 }
1079 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1080 #define MAX_SCAN_WINDOW 2560
1082 static unsigned int task_scan_min(struct task_struct *p)
1083 {
1084 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1085 unsigned int scan, floor;
1086 unsigned int windows = 1;
1088 if (scan_size < MAX_SCAN_WINDOW)
1089 windows = MAX_SCAN_WINDOW / scan_size;
1090 floor = 1000 / windows;
1092 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1093 return max_t(unsigned int, floor, scan);
1094 }
1096 static unsigned int task_scan_start(struct task_struct *p)
1097 {
1098 unsigned long smin = task_scan_min(p);
1099 unsigned long period = smin;
1101 /* Scale the maximum scan period with the amount of shared memory. */
1102 if (p->numa_group) {
1103 struct numa_group *ng = p->numa_group;
1104 unsigned long shared = group_faults_shared(ng);
1105 unsigned long private = group_faults_priv(ng);
1107 period *= atomic_read(&ng->refcount);
1108 period *= shared + 1;
1109 period /= private + shared + 1;
1110 }
1112 return max(smin, period);
1113 }
1115 static unsigned int task_scan_max(struct task_struct *p)
1116 {
1117 unsigned long smin = task_scan_min(p);
1118 unsigned long smax;
1120 /* Watch for min being lower than max due to floor calculations */
1121 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1123 /* Scale the maximum scan period with the amount of shared memory. */
1124 if (p->numa_group) {
1125 struct numa_group *ng = p->numa_group;
1126 unsigned long shared = group_faults_shared(ng);
1127 unsigned long private = group_faults_priv(ng);
1128 unsigned long period = smax;
1130 period *= atomic_read(&ng->refcount);
1131 period *= shared + 1;
1132 period /= private + shared + 1;
1134 smax = max(smax, period);
1135 }
1137 return max(smin, smax);
1138 }
1140 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
1141 {
1142 int mm_users = 0;
1143 struct mm_struct *mm = p->mm;
1145 if (mm) {
1146 mm_users = atomic_read(&mm->mm_users);
1147 if (mm_users == 1) {
1148 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1149 mm->numa_scan_seq = 0;
1150 }
1151 }
1152 p->node_stamp = 0;
1153 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
1154 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1155 p->numa_work.next = &p->numa_work;
1156 p->numa_faults = NULL;
1157 p->numa_group = NULL;
1158 p->last_task_numa_placement = 0;
1159 p->last_sum_exec_runtime = 0;
1161 /* New address space, reset the preferred nid */
1162 if (!(clone_flags & CLONE_VM)) {
1163 p->numa_preferred_nid = -1;
1164 return;
1165 }
1167 /*
1168 * New thread, keep existing numa_preferred_nid which should be copied
1169 * already by arch_dup_task_struct but stagger when scans start.
1170 */
1171 if (mm) {
1172 unsigned int delay;
1174 delay = min_t(unsigned int, task_scan_max(current),
1175 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
1176 delay += 2 * TICK_NSEC;
1177 p->node_stamp = delay;
1178 }
1179 }
1181 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1182 {
1183 rq->nr_numa_running += (p->numa_preferred_nid != -1);
1184 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1185 }
1187 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1188 {
1189 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1190 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1191 }
1193 /* Shared or private faults. */
1194 #define NR_NUMA_HINT_FAULT_TYPES 2
1196 /* Memory and CPU locality */
1197 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1199 /* Averaged statistics, and temporary buffers. */
1200 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1202 pid_t task_numa_group_id(struct task_struct *p)
1203 {
1204 return p->numa_group ? p->numa_group->gid : 0;
1205 }
1207 /*
1208 * The averaged statistics, shared & private, memory & CPU,
1209 * occupy the first half of the array. The second half of the
1210 * array is for current counters, which are averaged into the
1211 * first set by task_numa_placement.
1212 */
1213 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1214 {
1215 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1216 }
1218 static inline unsigned long task_faults(struct task_struct *p, int nid)
1219 {
1220 if (!p->numa_faults)
1221 return 0;
1223 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1224 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1225 }
1227 static inline unsigned long group_faults(struct task_struct *p, int nid)
1228 {
1229 if (!p->numa_group)
1230 return 0;
1232 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1233 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1234 }
1236 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1237 {
1238 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1239 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1240 }
1242 static inline unsigned long group_faults_priv(struct numa_group *ng)
1243 {
1244 unsigned long faults = 0;
1245 int node;
1247 for_each_online_node(node) {
1248 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1249 }
1251 return faults;
1252 }
1254 static inline unsigned long group_faults_shared(struct numa_group *ng)
1255 {
1256 unsigned long faults = 0;
1257 int node;
1259 for_each_online_node(node) {
1260 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1261 }
1263 return faults;
1264 }
1266 /*
1267 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1268 * considered part of a numa group's pseudo-interleaving set. Migrations
1269 * between these nodes are slowed down, to allow things to settle down.
1270 */
1271 #define ACTIVE_NODE_FRACTION 3
1273 static bool numa_is_active_node(int nid, struct numa_group *ng)
1274 {
1275 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1276 }
1278 /* Handle placement on systems where not all nodes are directly connected. */
1279 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1280 int maxdist, bool task)
1281 {
1282 unsigned long score = 0;
1283 int node;
1285 /*
1286 * All nodes are directly connected, and the same distance
1287 * from each other. No need for fancy placement algorithms.
1288 */
1289 if (sched_numa_topology_type == NUMA_DIRECT)
1290 return 0;
1292 /*
1293 * This code is called for each node, introducing N^2 complexity,
1294 * which should be ok given the number of nodes rarely exceeds 8.
1295 */
1296 for_each_online_node(node) {
1297 unsigned long faults;
1298 int dist = node_distance(nid, node);
1300 /*
1301 * The furthest away nodes in the system are not interesting
1302 * for placement; nid was already counted.
1303 */
1304 if (dist == sched_max_numa_distance || node == nid)
1305 continue;
1307 /*
1308 * On systems with a backplane NUMA topology, compare groups
1309 * of nodes, and move tasks towards the group with the most
1310 * memory accesses. When comparing two nodes at distance
1311 * "hoplimit", only nodes closer by than "hoplimit" are part
1312 * of each group. Skip other nodes.
1313 */
1314 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1315 dist >= maxdist)
1316 continue;
1318 /* Add up the faults from nearby nodes. */
1319 if (task)
1320 faults = task_faults(p, node);
1321 else
1322 faults = group_faults(p, node);
1324 /*
1325 * On systems with a glueless mesh NUMA topology, there are
1326 * no fixed "groups of nodes". Instead, nodes that are not
1327 * directly connected bounce traffic through intermediate
1328 * nodes; a numa_group can occupy any set of nodes.
1329 * The further away a node is, the less the faults count.
1330 * This seems to result in good task placement.
1331 */
1332 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1333 faults *= (sched_max_numa_distance - dist);
1334 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1335 }
1337 score += faults;
1338 }
1340 return score;
1341 }
1343 /*
1344 * These return the fraction of accesses done by a particular task, or
1345 * task group, on a particular numa node. The group weight is given a
1346 * larger multiplier, in order to group tasks together that are almost
1347 * evenly spread out between numa nodes.
1348 */
1349 static inline unsigned long task_weight(struct task_struct *p, int nid,
1350 int dist)
1351 {
1352 unsigned long faults, total_faults;
1354 if (!p->numa_faults)
1355 return 0;
1357 total_faults = p->total_numa_faults;
1359 if (!total_faults)
1360 return 0;
1362 faults = task_faults(p, nid);
1363 faults += score_nearby_nodes(p, nid, dist, true);
1365 return 1000 * faults / total_faults;
1366 }
1368 static inline unsigned long group_weight(struct task_struct *p, int nid,
1369 int dist)
1370 {
1371 unsigned long faults, total_faults;
1373 if (!p->numa_group)
1374 return 0;
1376 total_faults = p->numa_group->total_faults;
1378 if (!total_faults)
1379 return 0;
1381 faults = group_faults(p, nid);
1382 faults += score_nearby_nodes(p, nid, dist, false);
1384 return 1000 * faults / total_faults;
1385 }
1387 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1388 int src_nid, int dst_cpu)
1389 {
1390 struct numa_group *ng = p->numa_group;
1391 int dst_nid = cpu_to_node(dst_cpu);
1392 int last_cpupid, this_cpupid;
1394 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1395 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1397 /*
1398 * Allow first faults or private faults to migrate immediately early in
1399 * the lifetime of a task. The magic number 4 is based on waiting for
1400 * two full passes of the "multi-stage node selection" test that is
1401 * executed below.
1402 */
1403 if ((p->numa_preferred_nid == -1 || p->numa_scan_seq <= 4) &&
1404 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1405 return true;
1407 /*
1408 * Multi-stage node selection is used in conjunction with a periodic
1409 * migration fault to build a temporal task<->page relation. By using
1410 * a two-stage filter we remove short/unlikely relations.
1411 *
1412 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1413 * a task's usage of a particular page (n_p) per total usage of this
1414 * page (n_t) (in a given time-span) to a probability.
1415 *
1416 * Our periodic faults will sample this probability and getting the
1417 * same result twice in a row, given these samples are fully
1418 * independent, is then given by P(n)^2, provided our sample period
1419 * is sufficiently short compared to the usage pattern.
1420 *
1421 * This quadric squishes small probabilities, making it less likely we
1422 * act on an unlikely task<->page relation.
1423 */
1424 if (!cpupid_pid_unset(last_cpupid) &&
1425 cpupid_to_nid(last_cpupid) != dst_nid)
1426 return false;
1428 /* Always allow migrate on private faults */
1429 if (cpupid_match_pid(p, last_cpupid))
1430 return true;
1432 /* A shared fault, but p->numa_group has not been set up yet. */
1433 if (!ng)
1434 return true;
1436 /*
1437 * Destination node is much more heavily used than the source
1438 * node? Allow migration.
1439 */
1440 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1441 ACTIVE_NODE_FRACTION)
1442 return true;
1444 /*
1445 * Distribute memory according to CPU & memory use on each node,
1446 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1447 *
1448 * faults_cpu(dst) 3 faults_cpu(src)
1449 * --------------- * - > ---------------
1450 * faults_mem(dst) 4 faults_mem(src)
1451 */
1452 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1453 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1454 }
1456 static unsigned long weighted_cpuload(struct rq *rq);
1457 static unsigned long source_load(int cpu, int type);
1458 static unsigned long target_load(int cpu, int type);
1459 static unsigned long capacity_of(int cpu);
1461 /* Cached statistics for all CPUs within a node */
1462 struct numa_stats {
1463 unsigned long load;
1465 /* Total compute capacity of CPUs on a node */
1466 unsigned long compute_capacity;
1468 unsigned int nr_running;
1469 };
1471 /*
1472 * XXX borrowed from update_sg_lb_stats
1473 */
1474 static void update_numa_stats(struct numa_stats *ns, int nid)
1475 {
1476 int smt, cpu, cpus = 0;
1477 unsigned long capacity;
1479 memset(ns, 0, sizeof(*ns));
1480 for_each_cpu(cpu, cpumask_of_node(nid)) {
1481 struct rq *rq = cpu_rq(cpu);
1483 ns->nr_running += rq->nr_running;
1484 ns->load += weighted_cpuload(rq);
1485 ns->compute_capacity += capacity_of(cpu);
1487 cpus++;
1488 }
1490 /*
1491 * If we raced with hotplug and there are no CPUs left in our mask
1492 * the @ns structure is NULL'ed and task_numa_compare() will
1493 * not find this node attractive.
1494 *
1495 * We'll detect a huge imbalance and bail there.
1496 */
1497 if (!cpus)
1498 return;
1500 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1501 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1502 capacity = cpus / smt; /* cores */
1504 capacity = min_t(unsigned, capacity,
1505 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1506 }
1508 struct task_numa_env {
1509 struct task_struct *p;
1511 int src_cpu, src_nid;
1512 int dst_cpu, dst_nid;
1514 struct numa_stats src_stats, dst_stats;
1516 int imbalance_pct;
1517 int dist;
1519 struct task_struct *best_task;
1520 long best_imp;
1521 int best_cpu;
1522 };
1524 static void task_numa_assign(struct task_numa_env *env,
1525 struct task_struct *p, long imp)
1526 {
1527 struct rq *rq = cpu_rq(env->dst_cpu);
1529 /* Bail out if run-queue part of active NUMA balance. */
1530 if (xchg(&rq->numa_migrate_on, 1))
1531 return;
1533 /*
1534 * Clear previous best_cpu/rq numa-migrate flag, since task now
1535 * found a better CPU to move/swap.
1536 */
1537 if (env->best_cpu != -1) {
1538 rq = cpu_rq(env->best_cpu);
1539 WRITE_ONCE(rq->numa_migrate_on, 0);
1540 }
1542 if (env->best_task)
1543 put_task_struct(env->best_task);
1544 if (p)
1545 get_task_struct(p);
1547 env->best_task = p;
1548 env->best_imp = imp;
1549 env->best_cpu = env->dst_cpu;
1550 }
1552 static bool load_too_imbalanced(long src_load, long dst_load,
1553 struct task_numa_env *env)
1554 {
1555 long imb, old_imb;
1556 long orig_src_load, orig_dst_load;
1557 long src_capacity, dst_capacity;
1559 /*
1560 * The load is corrected for the CPU capacity available on each node.
1561 *
1562 * src_load dst_load
1563 * ------------ vs ---------
1564 * src_capacity dst_capacity
1565 */
1566 src_capacity = env->src_stats.compute_capacity;
1567 dst_capacity = env->dst_stats.compute_capacity;
1569 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1571 orig_src_load = env->src_stats.load;
1572 orig_dst_load = env->dst_stats.load;
1574 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1576 /* Would this change make things worse? */
1577 return (imb > old_imb);
1578 }
1580 /*
1581 * Maximum NUMA importance can be 1998 (2*999);
1582 * SMALLIMP @ 30 would be close to 1998/64.
1583 * Used to deter task migration.
1584 */
1585 #define SMALLIMP 30
1587 /*
1588 * This checks if the overall compute and NUMA accesses of the system would
1589 * be improved if the source tasks was migrated to the target dst_cpu taking
1590 * into account that it might be best if task running on the dst_cpu should
1591 * be exchanged with the source task
1592 */
1593 static void task_numa_compare(struct task_numa_env *env,
1594 long taskimp, long groupimp, bool maymove)
1595 {
1596 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1597 struct task_struct *cur;
1598 long src_load, dst_load;
1599 long load;
1600 long imp = env->p->numa_group ? groupimp : taskimp;
1601 long moveimp = imp;
1602 int dist = env->dist;
1604 if (READ_ONCE(dst_rq->numa_migrate_on))
1605 return;
1607 rcu_read_lock();
1608 cur = task_rcu_dereference(&dst_rq->curr);
1609 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1610 cur = NULL;
1612 /*
1613 * Because we have preemption enabled we can get migrated around and
1614 * end try selecting ourselves (current == env->p) as a swap candidate.
1615 */
1616 if (cur == env->p)
1617 goto unlock;
1619 if (!cur) {
1620 if (maymove && moveimp >= env->best_imp)
1621 goto assign;
1622 else
1623 goto unlock;
1624 }
1626 /*
1627 * "imp" is the fault differential for the source task between the
1628 * source and destination node. Calculate the total differential for
1629 * the source task and potential destination task. The more negative
1630 * the value is, the more remote accesses that would be expected to
1631 * be incurred if the tasks were swapped.
1632 */
1633 /* Skip this swap candidate if cannot move to the source cpu */
1634 if (!cpumask_test_cpu(env->src_cpu, &cur->cpus_allowed))
1635 goto unlock;
1637 /*
1638 * If dst and source tasks are in the same NUMA group, or not
1639 * in any group then look only at task weights.
1640 */
1641 if (cur->numa_group == env->p->numa_group) {
1642 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1643 task_weight(cur, env->dst_nid, dist);
1644 /*
1645 * Add some hysteresis to prevent swapping the
1646 * tasks within a group over tiny differences.
1647 */
1648 if (cur->numa_group)
1649 imp -= imp / 16;
1650 } else {
1651 /*
1652 * Compare the group weights. If a task is all by itself
1653 * (not part of a group), use the task weight instead.
1654 */
1655 if (cur->numa_group && env->p->numa_group)
1656 imp += group_weight(cur, env->src_nid, dist) -
1657 group_weight(cur, env->dst_nid, dist);
1658 else
1659 imp += task_weight(cur, env->src_nid, dist) -
1660 task_weight(cur, env->dst_nid, dist);
1661 }
1663 if (maymove && moveimp > imp && moveimp > env->best_imp) {
1664 imp = moveimp;
1665 cur = NULL;
1666 goto assign;
1667 }
1669 /*
1670 * If the NUMA importance is less than SMALLIMP,
1671 * task migration might only result in ping pong
1672 * of tasks and also hurt performance due to cache
1673 * misses.
1674 */
1675 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
1676 goto unlock;
1678 /*
1679 * In the overloaded case, try and keep the load balanced.
1680 */
1681 load = task_h_load(env->p) - task_h_load(cur);
1682 if (!load)
1683 goto assign;
1685 dst_load = env->dst_stats.load + load;
1686 src_load = env->src_stats.load - load;
1688 if (load_too_imbalanced(src_load, dst_load, env))
1689 goto unlock;
1691 assign:
1692 /*
1693 * One idle CPU per node is evaluated for a task numa move.
1694 * Call select_idle_sibling to maybe find a better one.
1695 */
1696 if (!cur) {
1697 /*
1698 * select_idle_siblings() uses an per-CPU cpumask that
1699 * can be used from IRQ context.
1700 */
1701 local_irq_disable();
1702 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1703 env->dst_cpu);
1704 local_irq_enable();
1705 }
1707 task_numa_assign(env, cur, imp);
1708 unlock:
1709 rcu_read_unlock();
1710 }
1712 static void task_numa_find_cpu(struct task_numa_env *env,
1713 long taskimp, long groupimp)
1714 {
1715 long src_load, dst_load, load;
1716 bool maymove = false;
1717 int cpu;
1719 load = task_h_load(env->p);
1720 dst_load = env->dst_stats.load + load;
1721 src_load = env->src_stats.load - load;
1723 /*
1724 * If the improvement from just moving env->p direction is better
1725 * than swapping tasks around, check if a move is possible.
1726 */
1727 maymove = !load_too_imbalanced(src_load, dst_load, env);
1729 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1730 /* Skip this CPU if the source task cannot migrate */
1731 if (!cpumask_test_cpu(cpu, &env->p->cpus_allowed))
1732 continue;
1734 env->dst_cpu = cpu;
1735 task_numa_compare(env, taskimp, groupimp, maymove);
1736 }
1737 }
1739 static int task_numa_migrate(struct task_struct *p)
1740 {
1741 struct task_numa_env env = {
1742 .p = p,
1744 .src_cpu = task_cpu(p),
1745 .src_nid = task_node(p),
1747 .imbalance_pct = 112,
1749 .best_task = NULL,
1750 .best_imp = 0,
1751 .best_cpu = -1,
1752 };
1753 struct sched_domain *sd;
1754 struct rq *best_rq;
1755 unsigned long taskweight, groupweight;
1756 int nid, ret, dist;
1757 long taskimp, groupimp;
1759 /*
1760 * Pick the lowest SD_NUMA domain, as that would have the smallest
1761 * imbalance and would be the first to start moving tasks about.
1762 *
1763 * And we want to avoid any moving of tasks about, as that would create
1764 * random movement of tasks -- counter the numa conditions we're trying
1765 * to satisfy here.
1766 */
1767 rcu_read_lock();
1768 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1769 if (sd)
1770 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1771 rcu_read_unlock();
1773 /*
1774 * Cpusets can break the scheduler domain tree into smaller
1775 * balance domains, some of which do not cross NUMA boundaries.
1776 * Tasks that are "trapped" in such domains cannot be migrated
1777 * elsewhere, so there is no point in (re)trying.
1778 */
1779 if (unlikely(!sd)) {
1780 sched_setnuma(p, task_node(p));
1781 return -EINVAL;
1782 }
1784 env.dst_nid = p->numa_preferred_nid;
1785 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1786 taskweight = task_weight(p, env.src_nid, dist);
1787 groupweight = group_weight(p, env.src_nid, dist);
1788 update_numa_stats(&env.src_stats, env.src_nid);
1789 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1790 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1791 update_numa_stats(&env.dst_stats, env.dst_nid);
1793 /* Try to find a spot on the preferred nid. */
1794 task_numa_find_cpu(&env, taskimp, groupimp);
1796 /*
1797 * Look at other nodes in these cases:
1798 * - there is no space available on the preferred_nid
1799 * - the task is part of a numa_group that is interleaved across
1800 * multiple NUMA nodes; in order to better consolidate the group,
1801 * we need to check other locations.
1802 */
1803 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1804 for_each_online_node(nid) {
1805 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1806 continue;
1808 dist = node_distance(env.src_nid, env.dst_nid);
1809 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1810 dist != env.dist) {
1811 taskweight = task_weight(p, env.src_nid, dist);
1812 groupweight = group_weight(p, env.src_nid, dist);
1813 }
1815 /* Only consider nodes where both task and groups benefit */
1816 taskimp = task_weight(p, nid, dist) - taskweight;
1817 groupimp = group_weight(p, nid, dist) - groupweight;
1818 if (taskimp < 0 && groupimp < 0)
1819 continue;
1821 env.dist = dist;
1822 env.dst_nid = nid;
1823 update_numa_stats(&env.dst_stats, env.dst_nid);
1824 task_numa_find_cpu(&env, taskimp, groupimp);
1825 }
1826 }
1828 /*
1829 * If the task is part of a workload that spans multiple NUMA nodes,
1830 * and is migrating into one of the workload's active nodes, remember
1831 * this node as the task's preferred numa node, so the workload can
1832 * settle down.
1833 * A task that migrated to a second choice node will be better off
1834 * trying for a better one later. Do not set the preferred node here.
1835 */
1836 if (p->numa_group) {
1837 if (env.best_cpu == -1)
1838 nid = env.src_nid;
1839 else
1840 nid = cpu_to_node(env.best_cpu);
1842 if (nid != p->numa_preferred_nid)
1843 sched_setnuma(p, nid);
1844 }
1846 /* No better CPU than the current one was found. */
1847 if (env.best_cpu == -1)
1848 return -EAGAIN;
1850 best_rq = cpu_rq(env.best_cpu);
1851 if (env.best_task == NULL) {
1852 ret = migrate_task_to(p, env.best_cpu);
1853 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1854 if (ret != 0)
1855 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1856 return ret;
1857 }
1859 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
1860 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1862 if (ret != 0)
1863 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1864 put_task_struct(env.best_task);
1865 return ret;
1866 }
1868 /* Attempt to migrate a task to a CPU on the preferred node. */
1869 static void numa_migrate_preferred(struct task_struct *p)
1870 {
1871 unsigned long interval = HZ;
1873 /* This task has no NUMA fault statistics yet */
1874 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1875 return;
1877 /* Periodically retry migrating the task to the preferred node */
1878 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1879 p->numa_migrate_retry = jiffies + interval;
1881 /* Success if task is already running on preferred CPU */
1882 if (task_node(p) == p->numa_preferred_nid)
1883 return;
1885 /* Otherwise, try migrate to a CPU on the preferred node */
1886 task_numa_migrate(p);
1887 }
1889 /*
1890 * Find out how many nodes on the workload is actively running on. Do this by
1891 * tracking the nodes from which NUMA hinting faults are triggered. This can
1892 * be different from the set of nodes where the workload's memory is currently
1893 * located.
1894 */
1895 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1896 {
1897 unsigned long faults, max_faults = 0;
1898 int nid, active_nodes = 0;
1900 for_each_online_node(nid) {
1901 faults = group_faults_cpu(numa_group, nid);
1902 if (faults > max_faults)
1903 max_faults = faults;
1904 }
1906 for_each_online_node(nid) {
1907 faults = group_faults_cpu(numa_group, nid);
1908 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1909 active_nodes++;
1910 }
1912 numa_group->max_faults_cpu = max_faults;
1913 numa_group->active_nodes = active_nodes;
1914 }
1916 /*
1917 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1918 * increments. The more local the fault statistics are, the higher the scan
1919 * period will be for the next scan window. If local/(local+remote) ratio is
1920 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1921 * the scan period will decrease. Aim for 70% local accesses.
1922 */
1923 #define NUMA_PERIOD_SLOTS 10
1924 #define NUMA_PERIOD_THRESHOLD 7
1926 /*
1927 * Increase the scan period (slow down scanning) if the majority of
1928 * our memory is already on our local node, or if the majority of
1929 * the page accesses are shared with other processes.
1930 * Otherwise, decrease the scan period.
1931 */
1932 static void update_task_scan_period(struct task_struct *p,
1933 unsigned long shared, unsigned long private)
1934 {
1935 unsigned int period_slot;
1936 int lr_ratio, ps_ratio;
1937 int diff;
1939 unsigned long remote = p->numa_faults_locality[0];
1940 unsigned long local = p->numa_faults_locality[1];
1942 /*
1943 * If there were no record hinting faults then either the task is
1944 * completely idle or all activity is areas that are not of interest
1945 * to automatic numa balancing. Related to that, if there were failed
1946 * migration then it implies we are migrating too quickly or the local
1947 * node is overloaded. In either case, scan slower
1948 */
1949 if (local + shared == 0 || p->numa_faults_locality[2]) {
1950 p->numa_scan_period = min(p->numa_scan_period_max,
1951 p->numa_scan_period << 1);
1953 p->mm->numa_next_scan = jiffies +
1954 msecs_to_jiffies(p->numa_scan_period);
1956 return;
1957 }
1959 /*
1960 * Prepare to scale scan period relative to the current period.
1961 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1962 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1963 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1964 */
1965 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1966 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1967 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
1969 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
1970 /*
1971 * Most memory accesses are local. There is no need to
1972 * do fast NUMA scanning, since memory is already local.
1973 */
1974 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
1975 if (!slot)
1976 slot = 1;
1977 diff = slot * period_slot;
1978 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
1979 /*
1980 * Most memory accesses are shared with other tasks.
1981 * There is no point in continuing fast NUMA scanning,
1982 * since other tasks may just move the memory elsewhere.
1983 */
1984 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
1985 if (!slot)
1986 slot = 1;
1987 diff = slot * period_slot;
1988 } else {
1989 /*
1990 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
1991 * yet they are not on the local NUMA node. Speed up
1992 * NUMA scanning to get the memory moved over.
1993 */
1994 int ratio = max(lr_ratio, ps_ratio);
1995 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1996 }
1998 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1999 task_scan_min(p), task_scan_max(p));
2000 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2001 }
2003 /*
2004 * Get the fraction of time the task has been running since the last
2005 * NUMA placement cycle. The scheduler keeps similar statistics, but
2006 * decays those on a 32ms period, which is orders of magnitude off
2007 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2008 * stats only if the task is so new there are no NUMA statistics yet.
2009 */
2010 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2011 {
2012 u64 runtime, delta, now;
2013 /* Use the start of this time slice to avoid calculations. */
2014 now = p->se.exec_start;
2015 runtime = p->se.sum_exec_runtime;
2017 if (p->last_task_numa_placement) {
2018 delta = runtime - p->last_sum_exec_runtime;
2019 *period = now - p->last_task_numa_placement;
2020 } else {
2021 delta = p->se.avg.load_sum;
2022 *period = LOAD_AVG_MAX;
2023 }
2025 p->last_sum_exec_runtime = runtime;
2026 p->last_task_numa_placement = now;
2028 return delta;
2029 }
2031 /*
2032 * Determine the preferred nid for a task in a numa_group. This needs to
2033 * be done in a way that produces consistent results with group_weight,
2034 * otherwise workloads might not converge.
2035 */
2036 static int preferred_group_nid(struct task_struct *p, int nid)
2037 {
2038 nodemask_t nodes;
2039 int dist;
2041 /* Direct connections between all NUMA nodes. */
2042 if (sched_numa_topology_type == NUMA_DIRECT)
2043 return nid;
2045 /*
2046 * On a system with glueless mesh NUMA topology, group_weight
2047 * scores nodes according to the number of NUMA hinting faults on
2048 * both the node itself, and on nearby nodes.
2049 */
2050 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2051 unsigned long score, max_score = 0;
2052 int node, max_node = nid;
2054 dist = sched_max_numa_distance;
2056 for_each_online_node(node) {
2057 score = group_weight(p, node, dist);
2058 if (score > max_score) {
2059 max_score = score;
2060 max_node = node;
2061 }
2062 }
2063 return max_node;
2064 }
2066 /*
2067 * Finding the preferred nid in a system with NUMA backplane
2068 * interconnect topology is more involved. The goal is to locate
2069 * tasks from numa_groups near each other in the system, and
2070 * untangle workloads from different sides of the system. This requires
2071 * searching down the hierarchy of node groups, recursively searching
2072 * inside the highest scoring group of nodes. The nodemask tricks
2073 * keep the complexity of the search down.
2074 */
2075 nodes = node_online_map;
2076 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2077 unsigned long max_faults = 0;
2078 nodemask_t max_group = NODE_MASK_NONE;
2079 int a, b;
2081 /* Are there nodes at this distance from each other? */
2082 if (!find_numa_distance(dist))
2083 continue;
2085 for_each_node_mask(a, nodes) {
2086 unsigned long faults = 0;
2087 nodemask_t this_group;
2088 nodes_clear(this_group);
2090 /* Sum group's NUMA faults; includes a==b case. */
2091 for_each_node_mask(b, nodes) {
2092 if (node_distance(a, b) < dist) {
2093 faults += group_faults(p, b);
2094 node_set(b, this_group);
2095 node_clear(b, nodes);
2096 }
2097 }
2099 /* Remember the top group. */
2100 if (faults > max_faults) {
2101 max_faults = faults;
2102 max_group = this_group;
2103 /*
2104 * subtle: at the smallest distance there is
2105 * just one node left in each "group", the
2106 * winner is the preferred nid.
2107 */
2108 nid = a;
2109 }
2110 }
2111 /* Next round, evaluate the nodes within max_group. */
2112 if (!max_faults)
2113 break;
2114 nodes = max_group;
2115 }
2116 return nid;
2117 }
2119 static void task_numa_placement(struct task_struct *p)
2120 {
2121 int seq, nid, max_nid = -1;
2122 unsigned long max_faults = 0;
2123 unsigned long fault_types[2] = { 0, 0 };
2124 unsigned long total_faults;
2125 u64 runtime, period;
2126 spinlock_t *group_lock = NULL;
2128 /*
2129 * The p->mm->numa_scan_seq field gets updated without
2130 * exclusive access. Use READ_ONCE() here to ensure
2131 * that the field is read in a single access:
2132 */
2133 seq = READ_ONCE(p->mm->numa_scan_seq);
2134 if (p->numa_scan_seq == seq)
2135 return;
2136 p->numa_scan_seq = seq;
2137 p->numa_scan_period_max = task_scan_max(p);
2139 total_faults = p->numa_faults_locality[0] +
2140 p->numa_faults_locality[1];
2141 runtime = numa_get_avg_runtime(p, &period);
2143 /* If the task is part of a group prevent parallel updates to group stats */
2144 if (p->numa_group) {
2145 group_lock = &p->numa_group->lock;
2146 spin_lock_irq(group_lock);
2147 }
2149 /* Find the node with the highest number of faults */
2150 for_each_online_node(nid) {
2151 /* Keep track of the offsets in numa_faults array */
2152 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2153 unsigned long faults = 0, group_faults = 0;
2154 int priv;
2156 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2157 long diff, f_diff, f_weight;
2159 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2160 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2161 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2162 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2164 /* Decay existing window, copy faults since last scan */
2165 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2166 fault_types[priv] += p->numa_faults[membuf_idx];
2167 p->numa_faults[membuf_idx] = 0;
2169 /*
2170 * Normalize the faults_from, so all tasks in a group
2171 * count according to CPU use, instead of by the raw
2172 * number of faults. Tasks with little runtime have
2173 * little over-all impact on throughput, and thus their
2174 * faults are less important.
2175 */
2176 f_weight = div64_u64(runtime << 16, period + 1);
2177 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2178 (total_faults + 1);
2179 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2180 p->numa_faults[cpubuf_idx] = 0;
2182 p->numa_faults[mem_idx] += diff;
2183 p->numa_faults[cpu_idx] += f_diff;
2184 faults += p->numa_faults[mem_idx];
2185 p->total_numa_faults += diff;
2186 if (p->numa_group) {
2187 /*
2188 * safe because we can only change our own group
2189 *
2190 * mem_idx represents the offset for a given
2191 * nid and priv in a specific region because it
2192 * is at the beginning of the numa_faults array.
2193 */
2194 p->numa_group->faults[mem_idx] += diff;
2195 p->numa_group->faults_cpu[mem_idx] += f_diff;
2196 p->numa_group->total_faults += diff;
2197 group_faults += p->numa_group->faults[mem_idx];
2198 }
2199 }
2201 if (!p->numa_group) {
2202 if (faults > max_faults) {
2203 max_faults = faults;
2204 max_nid = nid;
2205 }
2206 } else if (group_faults > max_faults) {
2207 max_faults = group_faults;
2208 max_nid = nid;
2209 }
2210 }
2212 if (p->numa_group) {
2213 numa_group_count_active_nodes(p->numa_group);
2214 spin_unlock_irq(group_lock);
2215 max_nid = preferred_group_nid(p, max_nid);
2216 }
2218 if (max_faults) {
2219 /* Set the new preferred node */
2220 if (max_nid != p->numa_preferred_nid)
2221 sched_setnuma(p, max_nid);
2222 }
2224 update_task_scan_period(p, fault_types[0], fault_types[1]);
2225 }
2227 static inline int get_numa_group(struct numa_group *grp)
2228 {
2229 return atomic_inc_not_zero(&grp->refcount);
2230 }
2232 static inline void put_numa_group(struct numa_group *grp)
2233 {
2234 if (atomic_dec_and_test(&grp->refcount))
2235 kfree_rcu(grp, rcu);
2236 }
2238 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2239 int *priv)
2240 {
2241 struct numa_group *grp, *my_grp;
2242 struct task_struct *tsk;
2243 bool join = false;
2244 int cpu = cpupid_to_cpu(cpupid);
2245 int i;
2247 if (unlikely(!p->numa_group)) {
2248 unsigned int size = sizeof(struct numa_group) +
2249 4*nr_node_ids*sizeof(unsigned long);
2251 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2252 if (!grp)
2253 return;
2255 atomic_set(&grp->refcount, 1);
2256 grp->active_nodes = 1;
2257 grp->max_faults_cpu = 0;
2258 spin_lock_init(&grp->lock);
2259 grp->gid = p->pid;
2260 /* Second half of the array tracks nids where faults happen */
2261 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2262 nr_node_ids;
2264 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2265 grp->faults[i] = p->numa_faults[i];
2267 grp->total_faults = p->total_numa_faults;
2269 grp->nr_tasks++;
2270 rcu_assign_pointer(p->numa_group, grp);
2271 }
2273 rcu_read_lock();
2274 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2276 if (!cpupid_match_pid(tsk, cpupid))
2277 goto no_join;
2279 grp = rcu_dereference(tsk->numa_group);
2280 if (!grp)
2281 goto no_join;
2283 my_grp = p->numa_group;
2284 if (grp == my_grp)
2285 goto no_join;
2287 /*
2288 * Only join the other group if its bigger; if we're the bigger group,
2289 * the other task will join us.
2290 */
2291 if (my_grp->nr_tasks > grp->nr_tasks)
2292 goto no_join;
2294 /*
2295 * Tie-break on the grp address.
2296 */
2297 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2298 goto no_join;
2300 /* Always join threads in the same process. */
2301 if (tsk->mm == current->mm)
2302 join = true;
2304 /* Simple filter to avoid false positives due to PID collisions */
2305 if (flags & TNF_SHARED)
2306 join = true;
2308 /* Update priv based on whether false sharing was detected */
2309 *priv = !join;
2311 if (join && !get_numa_group(grp))
2312 goto no_join;
2314 rcu_read_unlock();
2316 if (!join)
2317 return;
2319 BUG_ON(irqs_disabled());
2320 double_lock_irq(&my_grp->lock, &grp->lock);
2322 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2323 my_grp->faults[i] -= p->numa_faults[i];
2324 grp->faults[i] += p->numa_faults[i];
2325 }
2326 my_grp->total_faults -= p->total_numa_faults;
2327 grp->total_faults += p->total_numa_faults;
2329 my_grp->nr_tasks--;
2330 grp->nr_tasks++;
2332 spin_unlock(&my_grp->lock);
2333 spin_unlock_irq(&grp->lock);
2335 rcu_assign_pointer(p->numa_group, grp);
2337 put_numa_group(my_grp);
2338 return;
2340 no_join:
2341 rcu_read_unlock();
2342 return;
2343 }
2345 void task_numa_free(struct task_struct *p)
2346 {
2347 struct numa_group *grp = p->numa_group;
2348 void *numa_faults = p->numa_faults;
2349 unsigned long flags;
2350 int i;
2352 if (grp) {
2353 spin_lock_irqsave(&grp->lock, flags);
2354 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2355 grp->faults[i] -= p->numa_faults[i];
2356 grp->total_faults -= p->total_numa_faults;
2358 grp->nr_tasks--;
2359 spin_unlock_irqrestore(&grp->lock, flags);
2360 RCU_INIT_POINTER(p->numa_group, NULL);
2361 put_numa_group(grp);
2362 }
2364 p->numa_faults = NULL;
2365 kfree(numa_faults);
2366 }
2368 /*
2369 * Got a PROT_NONE fault for a page on @node.
2370 */
2371 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2372 {
2373 struct task_struct *p = current;
2374 bool migrated = flags & TNF_MIGRATED;
2375 int cpu_node = task_node(current);
2376 int local = !!(flags & TNF_FAULT_LOCAL);
2377 struct numa_group *ng;
2378 int priv;
2380 if (!static_branch_likely(&sched_numa_balancing))
2381 return;
2383 /* for example, ksmd faulting in a user's mm */
2384 if (!p->mm)
2385 return;
2387 /* Allocate buffer to track faults on a per-node basis */
2388 if (unlikely(!p->numa_faults)) {
2389 int size = sizeof(*p->numa_faults) *
2390 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2392 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2393 if (!p->numa_faults)
2394 return;
2396 p->total_numa_faults = 0;
2397 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2398 }
2400 /*
2401 * First accesses are treated as private, otherwise consider accesses
2402 * to be private if the accessing pid has not changed
2403 */
2404 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2405 priv = 1;
2406 } else {
2407 priv = cpupid_match_pid(p, last_cpupid);
2408 if (!priv && !(flags & TNF_NO_GROUP))
2409 task_numa_group(p, last_cpupid, flags, &priv);
2410 }
2412 /*
2413 * If a workload spans multiple NUMA nodes, a shared fault that
2414 * occurs wholly within the set of nodes that the workload is
2415 * actively using should be counted as local. This allows the
2416 * scan rate to slow down when a workload has settled down.
2417 */
2418 ng = p->numa_group;
2419 if (!priv && !local && ng && ng->active_nodes > 1 &&
2420 numa_is_active_node(cpu_node, ng) &&
2421 numa_is_active_node(mem_node, ng))
2422 local = 1;
2424 /*
2425 * Retry task to preferred node migration periodically, in case it
2426 * case it previously failed, or the scheduler moved us.
2427 */
2428 if (time_after(jiffies, p->numa_migrate_retry)) {
2429 task_numa_placement(p);
2430 numa_migrate_preferred(p);
2431 }
2433 if (migrated)
2434 p->numa_pages_migrated += pages;
2435 if (flags & TNF_MIGRATE_FAIL)
2436 p->numa_faults_locality[2] += pages;
2438 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2439 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2440 p->numa_faults_locality[local] += pages;
2441 }
2443 static void reset_ptenuma_scan(struct task_struct *p)
2444 {
2445 /*
2446 * We only did a read acquisition of the mmap sem, so
2447 * p->mm->numa_scan_seq is written to without exclusive access
2448 * and the update is not guaranteed to be atomic. That's not
2449 * much of an issue though, since this is just used for
2450 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2451 * expensive, to avoid any form of compiler optimizations:
2452 */
2453 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2454 p->mm->numa_scan_offset = 0;
2455 }
2457 /*
2458 * The expensive part of numa migration is done from task_work context.
2459 * Triggered from task_tick_numa().
2460 */
2461 void task_numa_work(struct callback_head *work)
2462 {
2463 unsigned long migrate, next_scan, now = jiffies;
2464 struct task_struct *p = current;
2465 struct mm_struct *mm = p->mm;
2466 u64 runtime = p->se.sum_exec_runtime;
2467 struct vm_area_struct *vma;
2468 unsigned long start, end;
2469 unsigned long nr_pte_updates = 0;
2470 long pages, virtpages;
2472 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2474 work->next = work; /* protect against double add */
2475 /*
2476 * Who cares about NUMA placement when they're dying.
2477 *
2478 * NOTE: make sure not to dereference p->mm before this check,
2479 * exit_task_work() happens _after_ exit_mm() so we could be called
2480 * without p->mm even though we still had it when we enqueued this
2481 * work.
2482 */
2483 if (p->flags & PF_EXITING)
2484 return;
2486 if (!mm->numa_next_scan) {
2487 mm->numa_next_scan = now +
2488 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2489 }
2491 /*
2492 * Enforce maximal scan/migration frequency..
2493 */
2494 migrate = mm->numa_next_scan;
2495 if (time_before(now, migrate))
2496 return;
2498 if (p->numa_scan_period == 0) {
2499 p->numa_scan_period_max = task_scan_max(p);
2500 p->numa_scan_period = task_scan_start(p);
2501 }
2503 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2504 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2505 return;
2507 /*
2508 * Delay this task enough that another task of this mm will likely win
2509 * the next time around.
2510 */
2511 p->node_stamp += 2 * TICK_NSEC;
2513 start = mm->numa_scan_offset;
2514 pages = sysctl_numa_balancing_scan_size;
2515 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2516 virtpages = pages * 8; /* Scan up to this much virtual space */
2517 if (!pages)
2518 return;
2521 if (!down_read_trylock(&mm->mmap_sem))
2522 return;
2523 vma = find_vma(mm, start);
2524 if (!vma) {
2525 reset_ptenuma_scan(p);
2526 start = 0;
2527 vma = mm->mmap;
2528 }
2529 for (; vma; vma = vma->vm_next) {
2530 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2531 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2532 continue;
2533 }
2535 /*
2536 * Shared library pages mapped by multiple processes are not
2537 * migrated as it is expected they are cache replicated. Avoid
2538 * hinting faults in read-only file-backed mappings or the vdso
2539 * as migrating the pages will be of marginal benefit.
2540 */
2541 if (!vma->vm_mm ||
2542 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2543 continue;
2545 /*
2546 * Skip inaccessible VMAs to avoid any confusion between
2547 * PROT_NONE and NUMA hinting ptes
2548 */
2549 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2550 continue;
2552 do {
2553 start = max(start, vma->vm_start);
2554 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2555 end = min(end, vma->vm_end);
2556 nr_pte_updates = change_prot_numa(vma, start, end);
2558 /*
2559 * Try to scan sysctl_numa_balancing_size worth of
2560 * hpages that have at least one present PTE that
2561 * is not already pte-numa. If the VMA contains
2562 * areas that are unused or already full of prot_numa
2563 * PTEs, scan up to virtpages, to skip through those
2564 * areas faster.
2565 */
2566 if (nr_pte_updates)
2567 pages -= (end - start) >> PAGE_SHIFT;
2568 virtpages -= (end - start) >> PAGE_SHIFT;
2570 start = end;
2571 if (pages <= 0 || virtpages <= 0)
2572 goto out;
2574 cond_resched();
2575 } while (end != vma->vm_end);
2576 }
2578 out:
2579 /*
2580 * It is possible to reach the end of the VMA list but the last few
2581 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2582 * would find the !migratable VMA on the next scan but not reset the
2583 * scanner to the start so check it now.
2584 */
2585 if (vma)
2586 mm->numa_scan_offset = start;
2587 else
2588 reset_ptenuma_scan(p);
2589 up_read(&mm->mmap_sem);
2591 /*
2592 * Make sure tasks use at least 32x as much time to run other code
2593 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2594 * Usually update_task_scan_period slows down scanning enough; on an
2595 * overloaded system we need to limit overhead on a per task basis.
2596 */
2597 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2598 u64 diff = p->se.sum_exec_runtime - runtime;
2599 p->node_stamp += 32 * diff;
2600 }
2601 }
2603 /*
2604 * Drive the periodic memory faults..
2605 */
2606 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2607 {
2608 struct callback_head *work = &curr->numa_work;
2609 u64 period, now;
2611 /*
2612 * We don't care about NUMA placement if we don't have memory.
2613 */
2614 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2615 return;
2617 /*
2618 * Using runtime rather than walltime has the dual advantage that
2619 * we (mostly) drive the selection from busy threads and that the
2620 * task needs to have done some actual work before we bother with
2621 * NUMA placement.
2622 */
2623 now = curr->se.sum_exec_runtime;
2624 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2626 if (now > curr->node_stamp + period) {
2627 if (!curr->node_stamp)
2628 curr->numa_scan_period = task_scan_start(curr);
2629 curr->node_stamp += period;
2631 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2632 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2633 task_work_add(curr, work, true);
2634 }
2635 }
2636 }
2638 static void update_scan_period(struct task_struct *p, int new_cpu)
2639 {
2640 int src_nid = cpu_to_node(task_cpu(p));
2641 int dst_nid = cpu_to_node(new_cpu);
2643 if (!static_branch_likely(&sched_numa_balancing))
2644 return;
2646 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
2647 return;
2649 if (src_nid == dst_nid)
2650 return;
2652 /*
2653 * Allow resets if faults have been trapped before one scan
2654 * has completed. This is most likely due to a new task that
2655 * is pulled cross-node due to wakeups or load balancing.
2656 */
2657 if (p->numa_scan_seq) {
2658 /*
2659 * Avoid scan adjustments if moving to the preferred
2660 * node or if the task was not previously running on
2661 * the preferred node.
2662 */
2663 if (dst_nid == p->numa_preferred_nid ||
2664 (p->numa_preferred_nid != -1 && src_nid != p->numa_preferred_nid))
2665 return;
2666 }
2668 p->numa_scan_period = task_scan_start(p);
2669 }
2671 #else
2672 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2673 {
2674 }
2676 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2677 {
2678 }
2680 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2681 {
2682 }
2684 static inline void update_scan_period(struct task_struct *p, int new_cpu)
2685 {
2686 }
2688 #endif /* CONFIG_NUMA_BALANCING */
2690 static void
2691 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2692 {
2693 update_load_add(&cfs_rq->load, se->load.weight);
2694 if (!parent_entity(se))
2695 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2696 #ifdef CONFIG_SMP
2697 if (entity_is_task(se)) {
2698 struct rq *rq = rq_of(cfs_rq);
2700 account_numa_enqueue(rq, task_of(se));
2701 list_add(&se->group_node, &rq->cfs_tasks);
2702 }
2703 #endif
2704 cfs_rq->nr_running++;
2705 }
2707 static void
2708 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2709 {
2710 update_load_sub(&cfs_rq->load, se->load.weight);
2711 if (!parent_entity(se))
2712 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2713 #ifdef CONFIG_SMP
2714 if (entity_is_task(se)) {
2715 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2716 list_del_init(&se->group_node);
2717 }
2718 #endif
2719 cfs_rq->nr_running--;
2720 }
2722 /*
2723 * Signed add and clamp on underflow.
2724 *
2725 * Explicitly do a load-store to ensure the intermediate value never hits
2726 * memory. This allows lockless observations without ever seeing the negative
2727 * values.
2728 */
2729 #define add_positive(_ptr, _val) do { \
2730 typeof(_ptr) ptr = (_ptr); \
2731 typeof(_val) val = (_val); \
2732 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2733 \
2734 res = var + val; \
2735 \
2736 if (val < 0 && res > var) \
2737 res = 0; \
2738 \
2739 WRITE_ONCE(*ptr, res); \
2740 } while (0)
2742 /*
2743 * Unsigned subtract and clamp on underflow.
2744 *
2745 * Explicitly do a load-store to ensure the intermediate value never hits
2746 * memory. This allows lockless observations without ever seeing the negative
2747 * values.
2748 */
2749 #define sub_positive(_ptr, _val) do { \
2750 typeof(_ptr) ptr = (_ptr); \
2751 typeof(*ptr) val = (_val); \
2752 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2753 res = var - val; \
2754 if (res > var) \
2755 res = 0; \
2756 WRITE_ONCE(*ptr, res); \
2757 } while (0)
2759 #ifdef CONFIG_SMP
2760 static inline void
2761 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2762 {
2763 cfs_rq->runnable_weight += se->runnable_weight;
2765 cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg;
2766 cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum;
2767 }
2769 static inline void
2770 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2771 {
2772 cfs_rq->runnable_weight -= se->runnable_weight;
2774 sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg);
2775 sub_positive(&cfs_rq->avg.runnable_load_sum,
2776 se_runnable(se) * se->avg.runnable_load_sum);
2777 }
2779 static inline void
2780 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2781 {
2782 cfs_rq->avg.load_avg += se->avg.load_avg;
2783 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
2784 }
2786 static inline void
2787 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2788 {
2789 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2790 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
2791 }
2792 #else
2793 static inline void
2794 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2795 static inline void
2796 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2797 static inline void
2798 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2799 static inline void
2800 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2801 #endif
2803 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2804 unsigned long weight, unsigned long runnable)
2805 {
2806 if (se->on_rq) {
2807 /* commit outstanding execution time */
2808 if (cfs_rq->curr == se)
2809 update_curr(cfs_rq);
2810 account_entity_dequeue(cfs_rq, se);
2811 dequeue_runnable_load_avg(cfs_rq, se);
2812 }
2813 dequeue_load_avg(cfs_rq, se);
2815 se->runnable_weight = runnable;
2816 update_load_set(&se->load, weight);
2818 #ifdef CONFIG_SMP
2819 do {
2820 u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
2822 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
2823 se->avg.runnable_load_avg =
2824 div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider);
2825 } while (0);
2826 #endif
2828 enqueue_load_avg(cfs_rq, se);
2829 if (se->on_rq) {
2830 account_entity_enqueue(cfs_rq, se);
2831 enqueue_runnable_load_avg(cfs_rq, se);
2832 }
2833 }
2835 void reweight_task(struct task_struct *p, int prio)
2836 {
2837 struct sched_entity *se = &p->se;
2838 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2839 struct load_weight *load = &se->load;
2840 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
2842 reweight_entity(cfs_rq, se, weight, weight);
2843 load->inv_weight = sched_prio_to_wmult[prio];
2844 }
2846 #ifdef CONFIG_FAIR_GROUP_SCHED
2847 #ifdef CONFIG_SMP
2848 /*
2849 * All this does is approximate the hierarchical proportion which includes that
2850 * global sum we all love to hate.
2851 *
2852 * That is, the weight of a group entity, is the proportional share of the
2853 * group weight based on the group runqueue weights. That is:
2854 *
2855 * tg->weight * grq->load.weight
2856 * ge->load.weight = ----------------------------- (1)
2857 * \Sum grq->load.weight
2858 *
2859 * Now, because computing that sum is prohibitively expensive to compute (been
2860 * there, done that) we approximate it with this average stuff. The average
2861 * moves slower and therefore the approximation is cheaper and more stable.
2862 *
2863 * So instead of the above, we substitute:
2864 *
2865 * grq->load.weight -> grq->avg.load_avg (2)
2866 *
2867 * which yields the following:
2868 *
2869 * tg->weight * grq->avg.load_avg
2870 * ge->load.weight = ------------------------------ (3)
2871 * tg->load_avg
2872 *
2873 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2874 *
2875 * That is shares_avg, and it is right (given the approximation (2)).
2876 *
2877 * The problem with it is that because the average is slow -- it was designed
2878 * to be exactly that of course -- this leads to transients in boundary
2879 * conditions. In specific, the case where the group was idle and we start the
2880 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2881 * yielding bad latency etc..
2882 *
2883 * Now, in that special case (1) reduces to:
2884 *
2885 * tg->weight * grq->load.weight
2886 * ge->load.weight = ----------------------------- = tg->weight (4)
2887 * grp->load.weight
2888 *
2889 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2890 *
2891 * So what we do is modify our approximation (3) to approach (4) in the (near)
2892 * UP case, like:
2893 *
2894 * ge->load.weight =
2895 *
2896 * tg->weight * grq->load.weight
2897 * --------------------------------------------------- (5)
2898 * tg->load_avg - grq->avg.load_avg + grq->load.weight
2899 *
2900 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
2901 * we need to use grq->avg.load_avg as its lower bound, which then gives:
2902 *
2903 *
2904 * tg->weight * grq->load.weight
2905 * ge->load.weight = ----------------------------- (6)
2906 * tg_load_avg'
2907 *
2908 * Where:
2909 *
2910 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
2911 * max(grq->load.weight, grq->avg.load_avg)
2912 *
2913 * And that is shares_weight and is icky. In the (near) UP case it approaches
2914 * (4) while in the normal case it approaches (3). It consistently
2915 * overestimates the ge->load.weight and therefore:
2916 *
2917 * \Sum ge->load.weight >= tg->weight
2918 *
2919 * hence icky!
2920 */
2921 static long calc_group_shares(struct cfs_rq *cfs_rq)
2922 {
2923 long tg_weight, tg_shares, load, shares;
2924 struct task_group *tg = cfs_rq->tg;
2926 tg_shares = READ_ONCE(tg->shares);
2928 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
2930 tg_weight = atomic_long_read(&tg->load_avg);
2932 /* Ensure tg_weight >= load */
2933 tg_weight -= cfs_rq->tg_load_avg_contrib;
2934 tg_weight += load;
2936 shares = (tg_shares * load);
2937 if (tg_weight)
2938 shares /= tg_weight;
2940 /*
2941 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2942 * of a group with small tg->shares value. It is a floor value which is
2943 * assigned as a minimum load.weight to the sched_entity representing
2944 * the group on a CPU.
2945 *
2946 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2947 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2948 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2949 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2950 * instead of 0.
2951 */
2952 return clamp_t(long, shares, MIN_SHARES, tg_shares);
2953 }
2955 /*
2956 * This calculates the effective runnable weight for a group entity based on
2957 * the group entity weight calculated above.
2958 *
2959 * Because of the above approximation (2), our group entity weight is
2960 * an load_avg based ratio (3). This means that it includes blocked load and
2961 * does not represent the runnable weight.
2962 *
2963 * Approximate the group entity's runnable weight per ratio from the group
2964 * runqueue:
2965 *
2966 * grq->avg.runnable_load_avg
2967 * ge->runnable_weight = ge->load.weight * -------------------------- (7)
2968 * grq->avg.load_avg
2969 *
2970 * However, analogous to above, since the avg numbers are slow, this leads to
2971 * transients in the from-idle case. Instead we use:
2972 *
2973 * ge->runnable_weight = ge->load.weight *
2974 *
2975 * max(grq->avg.runnable_load_avg, grq->runnable_weight)
2976 * ----------------------------------------------------- (8)
2977 * max(grq->avg.load_avg, grq->load.weight)
2978 *
2979 * Where these max() serve both to use the 'instant' values to fix the slow
2980 * from-idle and avoid the /0 on to-idle, similar to (6).
2981 */
2982 static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares)
2983 {
2984 long runnable, load_avg;
2986 load_avg = max(cfs_rq->avg.load_avg,
2987 scale_load_down(cfs_rq->load.weight));
2989 runnable = max(cfs_rq->avg.runnable_load_avg,
2990 scale_load_down(cfs_rq->runnable_weight));
2992 runnable *= shares;
2993 if (load_avg)
2994 runnable /= load_avg;
2996 return clamp_t(long, runnable, MIN_SHARES, shares);
2997 }
2998 #endif /* CONFIG_SMP */
3000 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3002 /*
3003 * Recomputes the group entity based on the current state of its group
3004 * runqueue.
3005 */
3006 static void update_cfs_group(struct sched_entity *se)
3007 {
3008 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3009 long shares, runnable;
3011 if (!gcfs_rq)
3012 return;
3014 if (throttled_hierarchy(gcfs_rq))
3015 return;
3017 #ifndef CONFIG_SMP
3018 runnable = shares = READ_ONCE(gcfs_rq->tg->shares);
3020 if (likely(se->load.weight == shares))
3021 return;
3022 #else
3023 shares = calc_group_shares(gcfs_rq);
3024 runnable = calc_group_runnable(gcfs_rq, shares);
3025 #endif
3027 reweight_entity(cfs_rq_of(se), se, shares, runnable);
3028 }
3030 #else /* CONFIG_FAIR_GROUP_SCHED */
3031 static inline void update_cfs_group(struct sched_entity *se)
3032 {
3033 }
3034 #endif /* CONFIG_FAIR_GROUP_SCHED */
3036 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3037 {
3038 struct rq *rq = rq_of(cfs_rq);
3040 if (&rq->cfs == cfs_rq || (flags & SCHED_CPUFREQ_MIGRATION)) {
3041 /*
3042 * There are a few boundary cases this might miss but it should
3043 * get called often enough that that should (hopefully) not be
3044 * a real problem.
3045 *
3046 * It will not get called when we go idle, because the idle
3047 * thread is a different class (!fair), nor will the utilization
3048 * number include things like RT tasks.
3049 *
3050 * As is, the util number is not freq-invariant (we'd have to
3051 * implement arch_scale_freq_capacity() for that).
3052 *
3053 * See cpu_util().
3054 */
3055 cpufreq_update_util(rq, flags);
3056 }
3057 }
3059 #ifdef CONFIG_SMP
3060 #ifdef CONFIG_FAIR_GROUP_SCHED
3061 /**
3062 * update_tg_load_avg - update the tg's load avg
3063 * @cfs_rq: the cfs_rq whose avg changed
3064 * @force: update regardless of how small the difference
3065 *
3066 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3067 * However, because tg->load_avg is a global value there are performance
3068 * considerations.
3069 *
3070 * In order to avoid having to look at the other cfs_rq's, we use a
3071 * differential update where we store the last value we propagated. This in
3072 * turn allows skipping updates if the differential is 'small'.
3073 *
3074 * Updating tg's load_avg is necessary before update_cfs_share().
3075 */
3076 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3077 {
3078 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3080 /*
3081 * No need to update load_avg for root_task_group as it is not used.
3082 */
3083 if (cfs_rq->tg == &root_task_group)
3084 return;
3086 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3087 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3088 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3089 }
3090 }
3092 /*
3093 * Called within set_task_rq() right before setting a task's CPU. The
3094 * caller only guarantees p->pi_lock is held; no other assumptions,
3095 * including the state of rq->lock, should be made.
3096 */
3097 void set_task_rq_fair(struct sched_entity *se,
3098 struct cfs_rq *prev, struct cfs_rq *next)
3099 {
3100 u64 p_last_update_time;
3101 u64 n_last_update_time;
3103 if (!sched_feat(ATTACH_AGE_LOAD))
3104 return;
3106 /*
3107 * We are supposed to update the task to "current" time, then its up to
3108 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3109 * getting what current time is, so simply throw away the out-of-date
3110 * time. This will result in the wakee task is less decayed, but giving
3111 * the wakee more load sounds not bad.
3112 */
3113 if (!(se->avg.last_update_time && prev))
3114 return;
3116 #ifndef CONFIG_64BIT
3117 {
3118 u64 p_last_update_time_copy;
3119 u64 n_last_update_time_copy;
3121 do {
3122 p_last_update_time_copy = prev->load_last_update_time_copy;
3123 n_last_update_time_copy = next->load_last_update_time_copy;
3125 smp_rmb();
3127 p_last_update_time = prev->avg.last_update_time;
3128 n_last_update_time = next->avg.last_update_time;
3130 } while (p_last_update_time != p_last_update_time_copy ||
3131 n_last_update_time != n_last_update_time_copy);
3132 }
3133 #else
3134 p_last_update_time = prev->avg.last_update_time;
3135 n_last_update_time = next->avg.last_update_time;
3136 #endif
3137 __update_load_avg_blocked_se(p_last_update_time, cpu_of(rq_of(prev)), se);
3138 se->avg.last_update_time = n_last_update_time;
3139 }
3142 /*
3143 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3144 * propagate its contribution. The key to this propagation is the invariant
3145 * that for each group:
3146 *
3147 * ge->avg == grq->avg (1)
3148 *
3149 * _IFF_ we look at the pure running and runnable sums. Because they
3150 * represent the very same entity, just at different points in the hierarchy.
3151 *
3152 * Per the above update_tg_cfs_util() is trivial and simply copies the running
3153 * sum over (but still wrong, because the group entity and group rq do not have
3154 * their PELT windows aligned).
3155 *
3156 * However, update_tg_cfs_runnable() is more complex. So we have:
3157 *
3158 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3159 *
3160 * And since, like util, the runnable part should be directly transferable,
3161 * the following would _appear_ to be the straight forward approach:
3162 *
3163 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3164 *
3165 * And per (1) we have:
3166 *
3167 * ge->avg.runnable_avg == grq->avg.runnable_avg
3168 *
3169 * Which gives:
3170 *
3171 * ge->load.weight * grq->avg.load_avg
3172 * ge->avg.load_avg = ----------------------------------- (4)
3173 * grq->load.weight
3174 *
3175 * Except that is wrong!
3176 *
3177 * Because while for entities historical weight is not important and we
3178 * really only care about our future and therefore can consider a pure
3179 * runnable sum, runqueues can NOT do this.
3180 *
3181 * We specifically want runqueues to have a load_avg that includes
3182 * historical weights. Those represent the blocked load, the load we expect
3183 * to (shortly) return to us. This only works by keeping the weights as
3184 * integral part of the sum. We therefore cannot decompose as per (3).
3185 *
3186 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3187 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3188 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3189 * runnable section of these tasks overlap (or not). If they were to perfectly
3190 * align the rq as a whole would be runnable 2/3 of the time. If however we
3191 * always have at least 1 runnable task, the rq as a whole is always runnable.
3192 *
3193 * So we'll have to approximate.. :/
3194 *
3195 * Given the constraint:
3196 *
3197 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3198 *
3199 * We can construct a rule that adds runnable to a rq by assuming minimal
3200 * overlap.
3201 *
3202 * On removal, we'll assume each task is equally runnable; which yields:
3203 *
3204 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3205 *
3206 * XXX: only do this for the part of runnable > running ?
3207 *
3208 */
3210 static inline void
3211 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3212 {
3213 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3215 /* Nothing to update */
3216 if (!delta)
3217 return;
3219 /*
3220 * The relation between sum and avg is:
3221 *
3222 * LOAD_AVG_MAX - 1024 + sa->period_contrib
3223 *
3224 * however, the PELT windows are not aligned between grq and gse.
3225 */
3227 /* Set new sched_entity's utilization */
3228 se->avg.util_avg = gcfs_rq->avg.util_avg;
3229 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3231 /* Update parent cfs_rq utilization */
3232 add_positive(&cfs_rq->avg.util_avg, delta);
3233 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3234 }
3236 static inline void
3237 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3238 {
3239 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3240 unsigned long runnable_load_avg, load_avg;
3241 u64 runnable_load_sum, load_sum = 0;
3242 s64 delta_sum;
3244 if (!runnable_sum)
3245 return;
3247 gcfs_rq->prop_runnable_sum = 0;
3249 if (runnable_sum >= 0) {
3250 /*
3251 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3252 * the CPU is saturated running == runnable.
3253 */
3254 runnable_sum += se->avg.load_sum;
3255 runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
3256 } else {
3257 /*
3258 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3259 * assuming all tasks are equally runnable.
3260 */
3261 if (scale_load_down(gcfs_rq->load.weight)) {
3262 load_sum = div_s64(gcfs_rq->avg.load_sum,
3263 scale_load_down(gcfs_rq->load.weight));
3264 }
3266 /* But make sure to not inflate se's runnable */
3267 runnable_sum = min(se->avg.load_sum, load_sum);
3268 }
3270 /*
3271 * runnable_sum can't be lower than running_sum
3272 * As running sum is scale with CPU capacity wehreas the runnable sum
3273 * is not we rescale running_sum 1st
3274 */
3275 running_sum = se->avg.util_sum /
3276 arch_scale_cpu_capacity(NULL, cpu_of(rq_of(cfs_rq)));
3277 runnable_sum = max(runnable_sum, running_sum);
3279 load_sum = (s64)se_weight(se) * runnable_sum;
3280 load_avg = div_s64(load_sum, LOAD_AVG_MAX);
3282 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3283 delta_avg = load_avg - se->avg.load_avg;
3285 se->avg.load_sum = runnable_sum;
3286 se->avg.load_avg = load_avg;
3287 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3288 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3290 runnable_load_sum = (s64)se_runnable(se) * runnable_sum;
3291 runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX);
3292 delta_sum = runnable_load_sum - se_weight(se) * se->avg.runnable_load_sum;
3293 delta_avg = runnable_load_avg - se->avg.runnable_load_avg;
3295 se->avg.runnable_load_sum = runnable_sum;
3296 se->avg.runnable_load_avg = runnable_load_avg;
3298 if (se->on_rq) {
3299 add_positive(&cfs_rq->avg.runnable_load_avg, delta_avg);
3300 add_positive(&cfs_rq->avg.runnable_load_sum, delta_sum);
3301 }
3302 }
3304 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3305 {
3306 cfs_rq->propagate = 1;
3307 cfs_rq->prop_runnable_sum += runnable_sum;
3308 }
3310 /* Update task and its cfs_rq load average */
3311 static inline int propagate_entity_load_avg(struct sched_entity *se)
3312 {
3313 struct cfs_rq *cfs_rq, *gcfs_rq;
3315 if (entity_is_task(se))
3316 return 0;
3318 gcfs_rq = group_cfs_rq(se);
3319 if (!gcfs_rq->propagate)
3320 return 0;
3322 gcfs_rq->propagate = 0;
3324 cfs_rq = cfs_rq_of(se);
3326 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3328 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3329 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3331 return 1;
3332 }
3334 /*
3335 * Check if we need to update the load and the utilization of a blocked
3336 * group_entity:
3337 */
3338 static inline bool skip_blocked_update(struct sched_entity *se)
3339 {
3340 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3342 /*
3343 * If sched_entity still have not zero load or utilization, we have to
3344 * decay it:
3345 */
3346 if (se->avg.load_avg || se->avg.util_avg)
3347 return false;
3349 /*
3350 * If there is a pending propagation, we have to update the load and
3351 * the utilization of the sched_entity:
3352 */
3353 if (gcfs_rq->propagate)
3354 return false;
3356 /*
3357 * Otherwise, the load and the utilization of the sched_entity is
3358 * already zero and there is no pending propagation, so it will be a
3359 * waste of time to try to decay it:
3360 */
3361 return true;
3362 }
3364 #else /* CONFIG_FAIR_GROUP_SCHED */
3366 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3368 static inline int propagate_entity_load_avg(struct sched_entity *se)
3369 {
3370 return 0;
3371 }
3373 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3375 #endif /* CONFIG_FAIR_GROUP_SCHED */
3377 /**
3378 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3379 * @now: current time, as per cfs_rq_clock_task()
3380 * @cfs_rq: cfs_rq to update
3381 *
3382 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3383 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3384 * post_init_entity_util_avg().
3385 *
3386 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3387 *
3388 * Returns true if the load decayed or we removed load.
3389 *
3390 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3391 * call update_tg_load_avg() when this function returns true.
3392 */
3393 static inline int
3394 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3395 {
3396 unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0;
3397 struct sched_avg *sa = &cfs_rq->avg;
3398 int decayed = 0;
3400 if (cfs_rq->removed.nr) {
3401 unsigned long r;
3402 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3404 raw_spin_lock(&cfs_rq->removed.lock);
3405 swap(cfs_rq->removed.util_avg, removed_util);
3406 swap(cfs_rq->removed.load_avg, removed_load);
3407 swap(cfs_rq->removed.runnable_sum, removed_runnable_sum);
3408 cfs_rq->removed.nr = 0;
3409 raw_spin_unlock(&cfs_rq->removed.lock);
3411 r = removed_load;
3412 sub_positive(&sa->load_avg, r);
3413 sub_positive(&sa->load_sum, r * divider);
3415 r = removed_util;
3416 sub_positive(&sa->util_avg, r);
3417 sub_positive(&sa->util_sum, r * divider);
3419 add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum);
3421 decayed = 1;
3422 }
3424 decayed |= __update_load_avg_cfs_rq(now, cpu_of(rq_of(cfs_rq)), cfs_rq);
3426 #ifndef CONFIG_64BIT
3427 smp_wmb();
3428 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3429 #endif
3431 if (decayed)
3432 cfs_rq_util_change(cfs_rq, 0);
3434 return decayed;
3435 }
3437 /**
3438 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3439 * @cfs_rq: cfs_rq to attach to
3440 * @se: sched_entity to attach
3441 * @flags: migration hints
3442 *
3443 * Must call update_cfs_rq_load_avg() before this, since we rely on
3444 * cfs_rq->avg.last_update_time being current.
3445 */
3446 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3447 {
3448 u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3450 /*
3451 * When we attach the @se to the @cfs_rq, we must align the decay
3452 * window because without that, really weird and wonderful things can
3453 * happen.
3454 *
3455 * XXX illustrate
3456 */
3457 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3458 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3460 /*
3461 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3462 * period_contrib. This isn't strictly correct, but since we're
3463 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3464 * _sum a little.
3465 */
3466 se->avg.util_sum = se->avg.util_avg * divider;
3468 se->avg.load_sum = divider;
3469 if (se_weight(se)) {
3470 se->avg.load_sum =
3471 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3472 }
3474 se->avg.runnable_load_sum = se->avg.load_sum;
3476 enqueue_load_avg(cfs_rq, se);
3477 cfs_rq->avg.util_avg += se->avg.util_avg;
3478 cfs_rq->avg.util_sum += se->avg.util_sum;
3480 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3482 cfs_rq_util_change(cfs_rq, flags);
3483 }
3485 /**
3486 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3487 * @cfs_rq: cfs_rq to detach from
3488 * @se: sched_entity to detach
3489 *
3490 * Must call update_cfs_rq_load_avg() before this, since we rely on
3491 * cfs_rq->avg.last_update_time being current.
3492 */
3493 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3494 {
3495 dequeue_load_avg(cfs_rq, se);
3496 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3497 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3499 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3501 cfs_rq_util_change(cfs_rq, 0);
3502 }
3504 /*
3505 * Optional action to be done while updating the load average
3506 */
3507 #define UPDATE_TG 0x1
3508 #define SKIP_AGE_LOAD 0x2
3509 #define DO_ATTACH 0x4
3511 /* Update task and its cfs_rq load average */
3512 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3513 {
3514 u64 now = cfs_rq_clock_task(cfs_rq);
3515 struct rq *rq = rq_of(cfs_rq);
3516 int cpu = cpu_of(rq);
3517 int decayed;
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, cpu, cfs_rq, se);
3526 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3527 decayed |= propagate_entity_load_avg(se);
3529 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
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);
3541 } else if (decayed && (flags & UPDATE_TG))
3542 update_tg_load_avg(cfs_rq, 0);
3543 }
3545 #ifndef CONFIG_64BIT
3546 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3547 {
3548 u64 last_update_time_copy;
3549 u64 last_update_time;
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);
3557 return last_update_time;
3558 }
3559 #else
3560 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3561 {
3562 return cfs_rq->avg.last_update_time;
3563 }
3564 #endif
3566 /*
3567 * Synchronize entity load avg of dequeued entity without locking
3568 * the previous rq.
3569 */
3570 void sync_entity_load_avg(struct sched_entity *se)
3571 {
3572 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3573 u64 last_update_time;
3575 last_update_time = cfs_rq_last_update_time(cfs_rq);
3576 __update_load_avg_blocked_se(last_update_time, cpu_of(rq_of(cfs_rq)), se);
3577 }
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 */
3583 void remove_entity_load_avg(struct sched_entity *se)
3584 {
3585 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3586 unsigned long flags;
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 * Similarly for groups, they will have passed through
3594 * post_init_entity_util_avg() before unregister_sched_fair_group()
3595 * calls this.
3596 */
3598 sync_entity_load_avg(se);
3600 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3601 ++cfs_rq->removed.nr;
3602 cfs_rq->removed.util_avg += se->avg.util_avg;
3603 cfs_rq->removed.load_avg += se->avg.load_avg;
3604 cfs_rq->removed.runnable_sum += se->avg.load_sum; /* == runnable_sum */
3605 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3606 }
3608 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3609 {
3610 return cfs_rq->avg.runnable_load_avg;
3611 }
3613 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3614 {
3615 return cfs_rq->avg.load_avg;
3616 }
3618 static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3620 static inline unsigned long task_util(struct task_struct *p)
3621 {
3622 return READ_ONCE(p->se.avg.util_avg);
3623 }
3625 static inline unsigned long _task_util_est(struct task_struct *p)
3626 {
3627 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3629 return max(ue.ewma, ue.enqueued);
3630 }
3632 static inline unsigned long task_util_est(struct task_struct *p)
3633 {
3634 return max(task_util(p), _task_util_est(p));
3635 }
3637 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3638 struct task_struct *p)
3639 {
3640 unsigned int enqueued;
3642 if (!sched_feat(UTIL_EST))
3643 return;
3645 /* Update root cfs_rq's estimated utilization */
3646 enqueued = cfs_rq->avg.util_est.enqueued;
3647 enqueued += (_task_util_est(p) | UTIL_AVG_UNCHANGED);
3648 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3649 }
3651 /*
3652 * Check if a (signed) value is within a specified (unsigned) margin,
3653 * based on the observation that:
3654 *
3655 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3656 *
3657 * NOTE: this only works when value + maring < INT_MAX.
3658 */
3659 static inline bool within_margin(int value, int margin)
3660 {
3661 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3662 }
3664 static void
3665 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep)
3666 {
3667 long last_ewma_diff;
3668 struct util_est ue;
3670 if (!sched_feat(UTIL_EST))
3671 return;
3673 /* Update root cfs_rq's estimated utilization */
3674 ue.enqueued = cfs_rq->avg.util_est.enqueued;
3675 ue.enqueued -= min_t(unsigned int, ue.enqueued,
3676 (_task_util_est(p) | UTIL_AVG_UNCHANGED));
3677 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued);
3679 /*
3680 * Skip update of task's estimated utilization when the task has not
3681 * yet completed an activation, e.g. being migrated.
3682 */
3683 if (!task_sleep)
3684 return;
3686 /*
3687 * If the PELT values haven't changed since enqueue time,
3688 * skip the util_est update.
3689 */
3690 ue = p->se.avg.util_est;
3691 if (ue.enqueued & UTIL_AVG_UNCHANGED)
3692 return;
3694 /*
3695 * Skip update of task's estimated utilization when its EWMA is
3696 * already ~1% close to its last activation value.
3697 */
3698 ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
3699 last_ewma_diff = ue.enqueued - ue.ewma;
3700 if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
3701 return;
3703 /*
3704 * Update Task's estimated utilization
3705 *
3706 * When *p completes an activation we can consolidate another sample
3707 * of the task size. This is done by storing the current PELT value
3708 * as ue.enqueued and by using this value to update the Exponential
3709 * Weighted Moving Average (EWMA):
3710 *
3711 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
3712 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
3713 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
3714 * = w * ( last_ewma_diff ) + ewma(t-1)
3715 * = w * (last_ewma_diff + ewma(t-1) / w)
3716 *
3717 * Where 'w' is the weight of new samples, which is configured to be
3718 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
3719 */
3720 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
3721 ue.ewma += last_ewma_diff;
3722 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
3723 WRITE_ONCE(p->se.avg.util_est, ue);
3724 }
3726 #else /* CONFIG_SMP */
3728 #define UPDATE_TG 0x0
3729 #define SKIP_AGE_LOAD 0x0
3730 #define DO_ATTACH 0x0
3732 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
3733 {
3734 cfs_rq_util_change(cfs_rq, 0);
3735 }
3737 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3739 static inline void
3740 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) {}
3741 static inline void
3742 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3744 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3745 {
3746 return 0;
3747 }
3749 static inline void
3750 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
3752 static inline void
3753 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p,
3754 bool task_sleep) {}
3756 #endif /* CONFIG_SMP */
3758 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3759 {
3760 #ifdef CONFIG_SCHED_DEBUG
3761 s64 d = se->vruntime - cfs_rq->min_vruntime;
3763 if (d < 0)
3764 d = -d;
3766 if (d > 3*sysctl_sched_latency)
3767 schedstat_inc(cfs_rq->nr_spread_over);
3768 #endif
3769 }
3771 static void
3772 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3773 {
3774 u64 vruntime = cfs_rq->min_vruntime;
3776 /*
3777 * The 'current' period is already promised to the current tasks,
3778 * however the extra weight of the new task will slow them down a
3779 * little, place the new task so that it fits in the slot that
3780 * stays open at the end.
3781 */
3782 if (initial && sched_feat(START_DEBIT))
3783 vruntime += sched_vslice(cfs_rq, se);
3785 /* sleeps up to a single latency don't count. */
3786 if (!initial) {
3787 unsigned long thresh = sysctl_sched_latency;
3789 /*
3790 * Halve their sleep time's effect, to allow
3791 * for a gentler effect of sleepers:
3792 */
3793 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3794 thresh >>= 1;
3796 vruntime -= thresh;
3797 }
3799 /* ensure we never gain time by being placed backwards. */
3800 se->vruntime = max_vruntime(se->vruntime, vruntime);
3801 }
3803 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3805 static inline void check_schedstat_required(void)
3806 {
3807 #ifdef CONFIG_SCHEDSTATS
3808 if (schedstat_enabled())
3809 return;
3811 /* Force schedstat enabled if a dependent tracepoint is active */
3812 if (trace_sched_stat_wait_enabled() ||
3813 trace_sched_stat_sleep_enabled() ||
3814 trace_sched_stat_iowait_enabled() ||
3815 trace_sched_stat_blocked_enabled() ||
3816 trace_sched_stat_runtime_enabled()) {
3817 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3818 "stat_blocked and stat_runtime require the "
3819 "kernel parameter schedstats=enable or "
3820 "kernel.sched_schedstats=1\n");
3821 }
3822 #endif
3823 }
3826 /*
3827 * MIGRATION
3828 *
3829 * dequeue
3830 * update_curr()
3831 * update_min_vruntime()
3832 * vruntime -= min_vruntime
3833 *
3834 * enqueue
3835 * update_curr()
3836 * update_min_vruntime()
3837 * vruntime += min_vruntime
3838 *
3839 * this way the vruntime transition between RQs is done when both
3840 * min_vruntime are up-to-date.
3841 *
3842 * WAKEUP (remote)
3843 *
3844 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3845 * vruntime -= min_vruntime
3846 *
3847 * enqueue
3848 * update_curr()
3849 * update_min_vruntime()
3850 * vruntime += min_vruntime
3851 *
3852 * this way we don't have the most up-to-date min_vruntime on the originating
3853 * CPU and an up-to-date min_vruntime on the destination CPU.
3854 */
3856 static void
3857 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3858 {
3859 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3860 bool curr = cfs_rq->curr == se;
3862 /*
3863 * If we're the current task, we must renormalise before calling
3864 * update_curr().
3865 */
3866 if (renorm && curr)
3867 se->vruntime += cfs_rq->min_vruntime;
3869 update_curr(cfs_rq);
3871 /*
3872 * Otherwise, renormalise after, such that we're placed at the current
3873 * moment in time, instead of some random moment in the past. Being
3874 * placed in the past could significantly boost this task to the
3875 * fairness detriment of existing tasks.
3876 */
3877 if (renorm && !curr)
3878 se->vruntime += cfs_rq->min_vruntime;
3880 /*
3881 * When enqueuing a sched_entity, we must:
3882 * - Update loads to have both entity and cfs_rq synced with now.
3883 * - Add its load to cfs_rq->runnable_avg
3884 * - For group_entity, update its weight to reflect the new share of
3885 * its group cfs_rq
3886 * - Add its new weight to cfs_rq->load.weight
3887 */
3888 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
3889 update_cfs_group(se);
3890 enqueue_runnable_load_avg(cfs_rq, se);
3891 account_entity_enqueue(cfs_rq, se);
3893 if (flags & ENQUEUE_WAKEUP)
3894 place_entity(cfs_rq, se, 0);
3896 check_schedstat_required();
3897 update_stats_enqueue(cfs_rq, se, flags);
3898 check_spread(cfs_rq, se);
3899 if (!curr)
3900 __enqueue_entity(cfs_rq, se);
3901 se->on_rq = 1;
3903 if (cfs_rq->nr_running == 1) {
3904 list_add_leaf_cfs_rq(cfs_rq);
3905 check_enqueue_throttle(cfs_rq);
3906 }
3907 }
3909 static void __clear_buddies_last(struct sched_entity *se)
3910 {
3911 for_each_sched_entity(se) {
3912 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3913 if (cfs_rq->last != se)
3914 break;
3916 cfs_rq->last = NULL;
3917 }
3918 }
3920 static void __clear_buddies_next(struct sched_entity *se)
3921 {
3922 for_each_sched_entity(se) {
3923 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3924 if (cfs_rq->next != se)
3925 break;
3927 cfs_rq->next = NULL;
3928 }
3929 }
3931 static void __clear_buddies_skip(struct sched_entity *se)
3932 {
3933 for_each_sched_entity(se) {
3934 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3935 if (cfs_rq->skip != se)
3936 break;
3938 cfs_rq->skip = NULL;
3939 }
3940 }
3942 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3943 {
3944 if (cfs_rq->last == se)
3945 __clear_buddies_last(se);
3947 if (cfs_rq->next == se)
3948 __clear_buddies_next(se);
3950 if (cfs_rq->skip == se)
3951 __clear_buddies_skip(se);
3952 }
3954 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3956 static void
3957 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3958 {
3959 /*
3960 * Update run-time statistics of the 'current'.
3961 */
3962 update_curr(cfs_rq);
3964 /*
3965 * When dequeuing a sched_entity, we must:
3966 * - Update loads to have both entity and cfs_rq synced with now.
3967 * - Substract its load from the cfs_rq->runnable_avg.
3968 * - Substract its previous weight from cfs_rq->load.weight.
3969 * - For group entity, update its weight to reflect the new share
3970 * of its group cfs_rq.
3971 */
3972 update_load_avg(cfs_rq, se, UPDATE_TG);
3973 dequeue_runnable_load_avg(cfs_rq, se);
3975 update_stats_dequeue(cfs_rq, se, flags);
3977 clear_buddies(cfs_rq, se);
3979 if (se != cfs_rq->curr)
3980 __dequeue_entity(cfs_rq, se);
3981 se->on_rq = 0;
3982 account_entity_dequeue(cfs_rq, se);
3984 /*
3985 * Normalize after update_curr(); which will also have moved
3986 * min_vruntime if @se is the one holding it back. But before doing
3987 * update_min_vruntime() again, which will discount @se's position and
3988 * can move min_vruntime forward still more.
3989 */
3990 if (!(flags & DEQUEUE_SLEEP))
3991 se->vruntime -= cfs_rq->min_vruntime;
3993 /* return excess runtime on last dequeue */
3994 return_cfs_rq_runtime(cfs_rq);
3996 update_cfs_group(se);
3998 /*
3999 * Now advance min_vruntime if @se was the entity holding it back,
4000 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4001 * put back on, and if we advance min_vruntime, we'll be placed back
4002 * further than we started -- ie. we'll be penalized.
4003 */
4004 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
4005 update_min_vruntime(cfs_rq);
4006 }
4008 /*
4009 * Preempt the current task with a newly woken task if needed:
4010 */
4011 static void
4012 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4013 {
4014 unsigned long ideal_runtime, delta_exec;
4015 struct sched_entity *se;
4016 s64 delta;
4018 ideal_runtime = sched_slice(cfs_rq, curr);
4019 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4020 if (delta_exec > ideal_runtime) {
4021 resched_curr(rq_of(cfs_rq));
4022 /*
4023 * The current task ran long enough, ensure it doesn't get
4024 * re-elected due to buddy favours.
4025 */
4026 clear_buddies(cfs_rq, curr);
4027 return;
4028 }
4030 /*
4031 * Ensure that a task that missed wakeup preemption by a
4032 * narrow margin doesn't have to wait for a full slice.
4033 * This also mitigates buddy induced latencies under load.
4034 */
4035 if (delta_exec < sysctl_sched_min_granularity)
4036 return;
4038 se = __pick_first_entity(cfs_rq);
4039 delta = curr->vruntime - se->vruntime;
4041 if (delta < 0)
4042 return;
4044 if (delta > ideal_runtime)
4045 resched_curr(rq_of(cfs_rq));
4046 }
4048 static void
4049 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4050 {
4051 /* 'current' is not kept within the tree. */
4052 if (se->on_rq) {
4053 /*
4054 * Any task has to be enqueued before it get to execute on
4055 * a CPU. So account for the time it spent waiting on the
4056 * runqueue.
4057 */
4058 update_stats_wait_end(cfs_rq, se);
4059 __dequeue_entity(cfs_rq, se);
4060 update_load_avg(cfs_rq, se, UPDATE_TG);
4061 }
4063 update_stats_curr_start(cfs_rq, se);
4064 cfs_rq->curr = se;
4066 /*
4067 * Track our maximum slice length, if the CPU's load is at
4068 * least twice that of our own weight (i.e. dont track it
4069 * when there are only lesser-weight tasks around):
4070 */
4071 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
4072 schedstat_set(se->statistics.slice_max,
4073 max((u64)schedstat_val(se->statistics.slice_max),
4074 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4075 }
4077 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4078 }
4080 static int
4081 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4083 /*
4084 * Pick the next process, keeping these things in mind, in this order:
4085 * 1) keep things fair between processes/task groups
4086 * 2) pick the "next" process, since someone really wants that to run
4087 * 3) pick the "last" process, for cache locality
4088 * 4) do not run the "skip" process, if something else is available
4089 */
4090 static struct sched_entity *
4091 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4092 {
4093 struct sched_entity *left = __pick_first_entity(cfs_rq);
4094 struct sched_entity *se;
4096 /*
4097 * If curr is set we have to see if its left of the leftmost entity
4098 * still in the tree, provided there was anything in the tree at all.
4099 */
4100 if (!left || (curr && entity_before(curr, left)))
4101 left = curr;
4103 se = left; /* ideally we run the leftmost entity */
4105 /*
4106 * Avoid running the skip buddy, if running something else can
4107 * be done without getting too unfair.
4108 */
4109 if (cfs_rq->skip == se) {
4110 struct sched_entity *second;
4112 if (se == curr) {
4113 second = __pick_first_entity(cfs_rq);
4114 } else {
4115 second = __pick_next_entity(se);
4116 if (!second || (curr && entity_before(curr, second)))
4117 second = curr;
4118 }
4120 if (second && wakeup_preempt_entity(second, left) < 1)
4121 se = second;
4122 }
4124 /*
4125 * Prefer last buddy, try to return the CPU to a preempted task.
4126 */
4127 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4128 se = cfs_rq->last;
4130 /*
4131 * Someone really wants this to run. If it's not unfair, run it.
4132 */
4133 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4134 se = cfs_rq->next;
4136 clear_buddies(cfs_rq, se);
4138 return se;
4139 }
4141 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4143 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4144 {
4145 /*
4146 * If still on the runqueue then deactivate_task()
4147 * was not called and update_curr() has to be done:
4148 */
4149 if (prev->on_rq)
4150 update_curr(cfs_rq);
4152 /* throttle cfs_rqs exceeding runtime */
4153 check_cfs_rq_runtime(cfs_rq);
4155 check_spread(cfs_rq, prev);
4157 if (prev->on_rq) {
4158 update_stats_wait_start(cfs_rq, prev);
4159 /* Put 'current' back into the tree. */
4160 __enqueue_entity(cfs_rq, prev);
4161 /* in !on_rq case, update occurred at dequeue */
4162 update_load_avg(cfs_rq, prev, 0);
4163 }
4164 cfs_rq->curr = NULL;
4165 }
4167 static void
4168 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4169 {
4170 /*
4171 * Update run-time statistics of the 'current'.
4172 */
4173 update_curr(cfs_rq);
4175 /*
4176 * Ensure that runnable average is periodically updated.
4177 */
4178 update_load_avg(cfs_rq, curr, UPDATE_TG);
4179 update_cfs_group(curr);
4181 #ifdef CONFIG_SCHED_HRTICK
4182 /*
4183 * queued ticks are scheduled to match the slice, so don't bother
4184 * validating it and just reschedule.
4185 */
4186 if (queued) {
4187 resched_curr(rq_of(cfs_rq));
4188 return;
4189 }
4190 /*
4191 * don't let the period tick interfere with the hrtick preemption
4192 */
4193 if (!sched_feat(DOUBLE_TICK) &&
4194 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4195 return;
4196 #endif
4198 if (cfs_rq->nr_running > 1)
4199 check_preempt_tick(cfs_rq, curr);
4200 }
4203 /**************************************************
4204 * CFS bandwidth control machinery
4205 */
4207 #ifdef CONFIG_CFS_BANDWIDTH
4209 #ifdef HAVE_JUMP_LABEL
4210 static struct static_key __cfs_bandwidth_used;
4212 static inline bool cfs_bandwidth_used(void)
4213 {
4214 return static_key_false(&__cfs_bandwidth_used);
4215 }
4217 void cfs_bandwidth_usage_inc(void)
4218 {
4219 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4220 }
4222 void cfs_bandwidth_usage_dec(void)
4223 {
4224 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4225 }
4226 #else /* HAVE_JUMP_LABEL */
4227 static bool cfs_bandwidth_used(void)
4228 {
4229 return true;
4230 }
4232 void cfs_bandwidth_usage_inc(void) {}
4233 void cfs_bandwidth_usage_dec(void) {}
4234 #endif /* HAVE_JUMP_LABEL */
4236 /*
4237 * default period for cfs group bandwidth.
4238 * default: 0.1s, units: nanoseconds
4239 */
4240 static inline u64 default_cfs_period(void)
4241 {
4242 return 100000000ULL;
4243 }
4245 static inline u64 sched_cfs_bandwidth_slice(void)
4246 {
4247 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4248 }
4250 /*
4251 * Replenish runtime according to assigned quota and update expiration time.
4252 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4253 * additional synchronization around rq->lock.
4254 *
4255 * requires cfs_b->lock
4256 */
4257 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4258 {
4259 u64 now;
4261 if (cfs_b->quota == RUNTIME_INF)
4262 return;
4264 now = sched_clock_cpu(smp_processor_id());
4265 cfs_b->runtime = cfs_b->quota;
4266 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
4267 cfs_b->expires_seq++;
4268 }
4270 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4271 {
4272 return &tg->cfs_bandwidth;
4273 }
4275 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4276 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4277 {
4278 if (unlikely(cfs_rq->throttle_count))
4279 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4281 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4282 }
4284 /* returns 0 on failure to allocate runtime */
4285 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4286 {
4287 struct task_group *tg = cfs_rq->tg;
4288 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4289 u64 amount = 0, min_amount, expires;
4290 int expires_seq;
4292 /* note: this is a positive sum as runtime_remaining <= 0 */
4293 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4295 raw_spin_lock(&cfs_b->lock);
4296 if (cfs_b->quota == RUNTIME_INF)
4297 amount = min_amount;
4298 else {
4299 start_cfs_bandwidth(cfs_b);
4301 if (cfs_b->runtime > 0) {
4302 amount = min(cfs_b->runtime, min_amount);
4303 cfs_b->runtime -= amount;
4304 cfs_b->idle = 0;
4305 }
4306 }
4307 expires_seq = cfs_b->expires_seq;
4308 expires = cfs_b->runtime_expires;
4309 raw_spin_unlock(&cfs_b->lock);
4311 cfs_rq->runtime_remaining += amount;
4312 /*
4313 * we may have advanced our local expiration to account for allowed
4314 * spread between our sched_clock and the one on which runtime was
4315 * issued.
4316 */
4317 if (cfs_rq->expires_seq != expires_seq) {
4318 cfs_rq->expires_seq = expires_seq;
4319 cfs_rq->runtime_expires = expires;
4320 }
4322 return cfs_rq->runtime_remaining > 0;
4323 }
4325 /*
4326 * Note: This depends on the synchronization provided by sched_clock and the
4327 * fact that rq->clock snapshots this value.
4328 */
4329 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4330 {
4331 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4333 /* if the deadline is ahead of our clock, nothing to do */
4334 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
4335 return;
4337 if (cfs_rq->runtime_remaining < 0)
4338 return;
4340 /*
4341 * If the local deadline has passed we have to consider the
4342 * possibility that our sched_clock is 'fast' and the global deadline
4343 * has not truly expired.
4344 *
4345 * Fortunately we can check determine whether this the case by checking
4346 * whether the global deadline(cfs_b->expires_seq) has advanced.
4347 */
4348 if (cfs_rq->expires_seq == cfs_b->expires_seq) {
4349 /* extend local deadline, drift is bounded above by 2 ticks */
4350 cfs_rq->runtime_expires += TICK_NSEC;
4351 } else {
4352 /* global deadline is ahead, expiration has passed */
4353 cfs_rq->runtime_remaining = 0;
4354 }
4355 }
4357 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4358 {
4359 /* dock delta_exec before expiring quota (as it could span periods) */
4360 cfs_rq->runtime_remaining -= delta_exec;
4361 expire_cfs_rq_runtime(cfs_rq);
4363 if (likely(cfs_rq->runtime_remaining > 0))
4364 return;
4366 /*
4367 * if we're unable to extend our runtime we resched so that the active
4368 * hierarchy can be throttled
4369 */
4370 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4371 resched_curr(rq_of(cfs_rq));
4372 }
4374 static __always_inline
4375 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4376 {
4377 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4378 return;
4380 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4381 }
4383 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4384 {
4385 return cfs_bandwidth_used() && cfs_rq->throttled;
4386 }
4388 /* check whether cfs_rq, or any parent, is throttled */
4389 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4390 {
4391 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4392 }
4394 /*
4395 * Ensure that neither of the group entities corresponding to src_cpu or
4396 * dest_cpu are members of a throttled hierarchy when performing group
4397 * load-balance operations.
4398 */
4399 static inline int throttled_lb_pair(struct task_group *tg,
4400 int src_cpu, int dest_cpu)
4401 {
4402 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4404 src_cfs_rq = tg->cfs_rq[src_cpu];
4405 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4407 return throttled_hierarchy(src_cfs_rq) ||
4408 throttled_hierarchy(dest_cfs_rq);
4409 }
4411 static int tg_unthrottle_up(struct task_group *tg, void *data)
4412 {
4413 struct rq *rq = data;
4414 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4416 cfs_rq->throttle_count--;
4417 if (!cfs_rq->throttle_count) {
4418 /* adjust cfs_rq_clock_task() */
4419 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4420 cfs_rq->throttled_clock_task;
4421 }
4423 return 0;
4424 }
4426 static int tg_throttle_down(struct task_group *tg, void *data)
4427 {
4428 struct rq *rq = data;
4429 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4431 /* group is entering throttled state, stop time */
4432 if (!cfs_rq->throttle_count)
4433 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4434 cfs_rq->throttle_count++;
4436 return 0;
4437 }
4439 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4440 {
4441 struct rq *rq = rq_of(cfs_rq);
4442 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4443 struct sched_entity *se;
4444 long task_delta, dequeue = 1;
4445 bool empty;
4447 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4449 /* freeze hierarchy runnable averages while throttled */
4450 rcu_read_lock();
4451 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4452 rcu_read_unlock();
4454 task_delta = cfs_rq->h_nr_running;
4455 for_each_sched_entity(se) {
4456 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4457 /* throttled entity or throttle-on-deactivate */
4458 if (!se->on_rq)
4459 break;
4461 if (dequeue)
4462 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4463 qcfs_rq->h_nr_running -= task_delta;
4465 if (qcfs_rq->load.weight)
4466 dequeue = 0;
4467 }
4469 if (!se)
4470 sub_nr_running(rq, task_delta);
4472 cfs_rq->throttled = 1;
4473 cfs_rq->throttled_clock = rq_clock(rq);
4474 raw_spin_lock(&cfs_b->lock);
4475 empty = list_empty(&cfs_b->throttled_cfs_rq);
4477 /*
4478 * Add to the _head_ of the list, so that an already-started
4479 * distribute_cfs_runtime will not see us
4480 */
4481 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4483 /*
4484 * If we're the first throttled task, make sure the bandwidth
4485 * timer is running.
4486 */
4487 if (empty)
4488 start_cfs_bandwidth(cfs_b);
4490 raw_spin_unlock(&cfs_b->lock);
4491 }
4493 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4494 {
4495 struct rq *rq = rq_of(cfs_rq);
4496 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4497 struct sched_entity *se;
4498 int enqueue = 1;
4499 long task_delta;
4501 se = cfs_rq->tg->se[cpu_of(rq)];
4503 cfs_rq->throttled = 0;
4505 update_rq_clock(rq);
4507 raw_spin_lock(&cfs_b->lock);
4508 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4509 list_del_rcu(&cfs_rq->throttled_list);
4510 raw_spin_unlock(&cfs_b->lock);
4512 /* update hierarchical throttle state */
4513 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4515 if (!cfs_rq->load.weight)
4516 return;
4518 task_delta = cfs_rq->h_nr_running;
4519 for_each_sched_entity(se) {
4520 if (se->on_rq)
4521 enqueue = 0;
4523 cfs_rq = cfs_rq_of(se);
4524 if (enqueue)
4525 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4526 cfs_rq->h_nr_running += task_delta;
4528 if (cfs_rq_throttled(cfs_rq))
4529 break;
4530 }
4532 if (!se)
4533 add_nr_running(rq, task_delta);
4535 /* Determine whether we need to wake up potentially idle CPU: */
4536 if (rq->curr == rq->idle && rq->cfs.nr_running)
4537 resched_curr(rq);
4538 }
4540 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4541 u64 remaining, u64 expires)
4542 {
4543 struct cfs_rq *cfs_rq;
4544 u64 runtime;
4545 u64 starting_runtime = remaining;
4547 rcu_read_lock();
4548 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4549 throttled_list) {
4550 struct rq *rq = rq_of(cfs_rq);
4551 struct rq_flags rf;
4553 rq_lock(rq, &rf);
4554 if (!cfs_rq_throttled(cfs_rq))
4555 goto next;
4557 runtime = -cfs_rq->runtime_remaining + 1;
4558 if (runtime > remaining)
4559 runtime = remaining;
4560 remaining -= runtime;
4562 cfs_rq->runtime_remaining += runtime;
4563 cfs_rq->runtime_expires = expires;
4565 /* we check whether we're throttled above */
4566 if (cfs_rq->runtime_remaining > 0)
4567 unthrottle_cfs_rq(cfs_rq);
4569 next:
4570 rq_unlock(rq, &rf);
4572 if (!remaining)
4573 break;
4574 }
4575 rcu_read_unlock();
4577 return starting_runtime - remaining;
4578 }
4580 /*
4581 * Responsible for refilling a task_group's bandwidth and unthrottling its
4582 * cfs_rqs as appropriate. If there has been no activity within the last
4583 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4584 * used to track this state.
4585 */
4586 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4587 {
4588 u64 runtime, runtime_expires;
4589 int throttled;
4591 /* no need to continue the timer with no bandwidth constraint */
4592 if (cfs_b->quota == RUNTIME_INF)
4593 goto out_deactivate;
4595 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4596 cfs_b->nr_periods += overrun;
4598 /*
4599 * idle depends on !throttled (for the case of a large deficit), and if
4600 * we're going inactive then everything else can be deferred
4601 */
4602 if (cfs_b->idle && !throttled)
4603 goto out_deactivate;
4605 __refill_cfs_bandwidth_runtime(cfs_b);
4607 if (!throttled) {
4608 /* mark as potentially idle for the upcoming period */
4609 cfs_b->idle = 1;
4610 return 0;
4611 }
4613 /* account preceding periods in which throttling occurred */
4614 cfs_b->nr_throttled += overrun;
4616 runtime_expires = cfs_b->runtime_expires;
4618 /*
4619 * This check is repeated as we are holding onto the new bandwidth while
4620 * we unthrottle. This can potentially race with an unthrottled group
4621 * trying to acquire new bandwidth from the global pool. This can result
4622 * in us over-using our runtime if it is all used during this loop, but
4623 * only by limited amounts in that extreme case.
4624 */
4625 while (throttled && cfs_b->runtime > 0) {
4626 runtime = cfs_b->runtime;
4627 raw_spin_unlock(&cfs_b->lock);
4628 /* we can't nest cfs_b->lock while distributing bandwidth */
4629 runtime = distribute_cfs_runtime(cfs_b, runtime,
4630 runtime_expires);
4631 raw_spin_lock(&cfs_b->lock);
4633 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4635 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4636 }
4638 /*
4639 * While we are ensured activity in the period following an
4640 * unthrottle, this also covers the case in which the new bandwidth is
4641 * insufficient to cover the existing bandwidth deficit. (Forcing the
4642 * timer to remain active while there are any throttled entities.)
4643 */
4644 cfs_b->idle = 0;
4646 return 0;
4648 out_deactivate:
4649 return 1;
4650 }
4652 /* a cfs_rq won't donate quota below this amount */
4653 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4654 /* minimum remaining period time to redistribute slack quota */
4655 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4656 /* how long we wait to gather additional slack before distributing */
4657 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4659 /*
4660 * Are we near the end of the current quota period?
4661 *
4662 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4663 * hrtimer base being cleared by hrtimer_start. In the case of
4664 * migrate_hrtimers, base is never cleared, so we are fine.
4665 */
4666 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4667 {
4668 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4669 u64 remaining;
4671 /* if the call-back is running a quota refresh is already occurring */
4672 if (hrtimer_callback_running(refresh_timer))
4673 return 1;
4675 /* is a quota refresh about to occur? */
4676 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4677 if (remaining < min_expire)
4678 return 1;
4680 return 0;
4681 }
4683 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4684 {
4685 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4687 /* if there's a quota refresh soon don't bother with slack */
4688 if (runtime_refresh_within(cfs_b, min_left))
4689 return;
4691 hrtimer_start(&cfs_b->slack_timer,
4692 ns_to_ktime(cfs_bandwidth_slack_period),
4693 HRTIMER_MODE_REL);
4694 }
4696 /* we know any runtime found here is valid as update_curr() precedes return */
4697 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4698 {
4699 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4700 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4702 if (slack_runtime <= 0)
4703 return;
4705 raw_spin_lock(&cfs_b->lock);
4706 if (cfs_b->quota != RUNTIME_INF &&
4707 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4708 cfs_b->runtime += slack_runtime;
4710 /* we are under rq->lock, defer unthrottling using a timer */
4711 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4712 !list_empty(&cfs_b->throttled_cfs_rq))
4713 start_cfs_slack_bandwidth(cfs_b);
4714 }
4715 raw_spin_unlock(&cfs_b->lock);
4717 /* even if it's not valid for return we don't want to try again */
4718 cfs_rq->runtime_remaining -= slack_runtime;
4719 }
4721 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4722 {
4723 if (!cfs_bandwidth_used())
4724 return;
4726 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4727 return;
4729 __return_cfs_rq_runtime(cfs_rq);
4730 }
4732 /*
4733 * This is done with a timer (instead of inline with bandwidth return) since
4734 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4735 */
4736 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4737 {
4738 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4739 u64 expires;
4741 /* confirm we're still not at a refresh boundary */
4742 raw_spin_lock(&cfs_b->lock);
4743 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4744 raw_spin_unlock(&cfs_b->lock);
4745 return;
4746 }
4748 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4749 runtime = cfs_b->runtime;
4751 expires = cfs_b->runtime_expires;
4752 raw_spin_unlock(&cfs_b->lock);
4754 if (!runtime)
4755 return;
4757 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4759 raw_spin_lock(&cfs_b->lock);
4760 if (expires == cfs_b->runtime_expires)
4761 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4762 raw_spin_unlock(&cfs_b->lock);
4763 }
4765 /*
4766 * When a group wakes up we want to make sure that its quota is not already
4767 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4768 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4769 */
4770 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4771 {
4772 if (!cfs_bandwidth_used())
4773 return;
4775 /* an active group must be handled by the update_curr()->put() path */
4776 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4777 return;
4779 /* ensure the group is not already throttled */
4780 if (cfs_rq_throttled(cfs_rq))
4781 return;
4783 /* update runtime allocation */
4784 account_cfs_rq_runtime(cfs_rq, 0);
4785 if (cfs_rq->runtime_remaining <= 0)
4786 throttle_cfs_rq(cfs_rq);
4787 }
4789 static void sync_throttle(struct task_group *tg, int cpu)
4790 {
4791 struct cfs_rq *pcfs_rq, *cfs_rq;
4793 if (!cfs_bandwidth_used())
4794 return;
4796 if (!tg->parent)
4797 return;
4799 cfs_rq = tg->cfs_rq[cpu];
4800 pcfs_rq = tg->parent->cfs_rq[cpu];
4802 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4803 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4804 }
4806 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4807 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4808 {
4809 if (!cfs_bandwidth_used())
4810 return false;
4812 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4813 return false;
4815 /*
4816 * it's possible for a throttled entity to be forced into a running
4817 * state (e.g. set_curr_task), in this case we're finished.
4818 */
4819 if (cfs_rq_throttled(cfs_rq))
4820 return true;
4822 throttle_cfs_rq(cfs_rq);
4823 return true;
4824 }
4826 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4827 {
4828 struct cfs_bandwidth *cfs_b =
4829 container_of(timer, struct cfs_bandwidth, slack_timer);
4831 do_sched_cfs_slack_timer(cfs_b);
4833 return HRTIMER_NORESTART;
4834 }
4836 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4837 {
4838 struct cfs_bandwidth *cfs_b =
4839 container_of(timer, struct cfs_bandwidth, period_timer);
4840 int overrun;
4841 int idle = 0;
4843 raw_spin_lock(&cfs_b->lock);
4844 for (;;) {
4845 overrun = hrtimer_forward_now(timer, cfs_b->period);
4846 if (!overrun)
4847 break;
4849 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4850 }
4851 if (idle)
4852 cfs_b->period_active = 0;
4853 raw_spin_unlock(&cfs_b->lock);
4855 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4856 }
4858 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4859 {
4860 raw_spin_lock_init(&cfs_b->lock);
4861 cfs_b->runtime = 0;
4862 cfs_b->quota = RUNTIME_INF;
4863 cfs_b->period = ns_to_ktime(default_cfs_period());
4865 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4866 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4867 cfs_b->period_timer.function = sched_cfs_period_timer;
4868 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4869 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4870 }
4872 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4873 {
4874 cfs_rq->runtime_enabled = 0;
4875 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4876 }
4878 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4879 {
4880 u64 overrun;
4882 lockdep_assert_held(&cfs_b->lock);
4884 if (cfs_b->period_active)
4885 return;
4887 cfs_b->period_active = 1;
4888 overrun = hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4889 cfs_b->runtime_expires += (overrun + 1) * ktime_to_ns(cfs_b->period);
4890 cfs_b->expires_seq++;
4891 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4892 }
4894 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4895 {
4896 /* init_cfs_bandwidth() was not called */
4897 if (!cfs_b->throttled_cfs_rq.next)
4898 return;
4900 hrtimer_cancel(&cfs_b->period_timer);
4901 hrtimer_cancel(&cfs_b->slack_timer);
4902 }
4904 /*
4905 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
4906 *
4907 * The race is harmless, since modifying bandwidth settings of unhooked group
4908 * bits doesn't do much.
4909 */
4911 /* cpu online calback */
4912 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4913 {
4914 struct task_group *tg;
4916 lockdep_assert_held(&rq->lock);
4918 rcu_read_lock();
4919 list_for_each_entry_rcu(tg, &task_groups, list) {
4920 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
4921 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4923 raw_spin_lock(&cfs_b->lock);
4924 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4925 raw_spin_unlock(&cfs_b->lock);
4926 }
4927 rcu_read_unlock();
4928 }
4930 /* cpu offline callback */
4931 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4932 {
4933 struct task_group *tg;
4935 lockdep_assert_held(&rq->lock);
4937 rcu_read_lock();
4938 list_for_each_entry_rcu(tg, &task_groups, list) {
4939 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4941 if (!cfs_rq->runtime_enabled)
4942 continue;
4944 /*
4945 * clock_task is not advancing so we just need to make sure
4946 * there's some valid quota amount
4947 */
4948 cfs_rq->runtime_remaining = 1;
4949 /*
4950 * Offline rq is schedulable till CPU is completely disabled
4951 * in take_cpu_down(), so we prevent new cfs throttling here.
4952 */
4953 cfs_rq->runtime_enabled = 0;
4955 if (cfs_rq_throttled(cfs_rq))
4956 unthrottle_cfs_rq(cfs_rq);
4957 }
4958 rcu_read_unlock();
4959 }
4961 #else /* CONFIG_CFS_BANDWIDTH */
4962 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4963 {
4964 return rq_clock_task(rq_of(cfs_rq));
4965 }
4967 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4968 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4969 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4970 static inline void sync_throttle(struct task_group *tg, int cpu) {}
4971 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4973 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4974 {
4975 return 0;
4976 }
4978 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4979 {
4980 return 0;
4981 }
4983 static inline int throttled_lb_pair(struct task_group *tg,
4984 int src_cpu, int dest_cpu)
4985 {
4986 return 0;
4987 }
4989 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4991 #ifdef CONFIG_FAIR_GROUP_SCHED
4992 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4993 #endif
4995 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4996 {
4997 return NULL;
4998 }
4999 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5000 static inline void update_runtime_enabled(struct rq *rq) {}
5001 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5003 #endif /* CONFIG_CFS_BANDWIDTH */
5005 /**************************************************
5006 * CFS operations on tasks:
5007 */
5009 #ifdef CONFIG_SCHED_HRTICK
5010 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5011 {
5012 struct sched_entity *se = &p->se;
5013 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5015 SCHED_WARN_ON(task_rq(p) != rq);
5017 if (rq->cfs.h_nr_running > 1) {
5018 u64 slice = sched_slice(cfs_rq, se);
5019 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5020 s64 delta = slice - ran;
5022 if (delta < 0) {
5023 if (rq->curr == p)
5024 resched_curr(rq);
5025 return;
5026 }
5027 hrtick_start(rq, delta);
5028 }
5029 }
5031 /*
5032 * called from enqueue/dequeue and updates the hrtick when the
5033 * current task is from our class and nr_running is low enough
5034 * to matter.
5035 */
5036 static void hrtick_update(struct rq *rq)
5037 {
5038 struct task_struct *curr = rq->curr;
5040 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
5041 return;
5043 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5044 hrtick_start_fair(rq, curr);
5045 }
5046 #else /* !CONFIG_SCHED_HRTICK */
5047 static inline void
5048 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5049 {
5050 }
5052 static inline void hrtick_update(struct rq *rq)
5053 {
5054 }
5055 #endif
5057 /*
5058 * The enqueue_task method is called before nr_running is
5059 * increased. Here we update the fair scheduling stats and
5060 * then put the task into the rbtree:
5061 */
5062 static void
5063 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5064 {
5065 struct cfs_rq *cfs_rq;
5066 struct sched_entity *se = &p->se;
5068 /*
5069 * The code below (indirectly) updates schedutil which looks at
5070 * the cfs_rq utilization to select a frequency.
5071 * Let's add the task's estimated utilization to the cfs_rq's
5072 * estimated utilization, before we update schedutil.
5073 */
5074 util_est_enqueue(&rq->cfs, p);
5076 /*
5077 * If in_iowait is set, the code below may not trigger any cpufreq
5078 * utilization updates, so do it here explicitly with the IOWAIT flag
5079 * passed.
5080 */
5081 if (p->in_iowait)
5082 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5084 for_each_sched_entity(se) {
5085 if (se->on_rq)
5086 break;
5087 cfs_rq = cfs_rq_of(se);
5088 enqueue_entity(cfs_rq, se, flags);
5090 /*
5091 * end evaluation on encountering a throttled cfs_rq
5092 *
5093 * note: in the case of encountering a throttled cfs_rq we will
5094 * post the final h_nr_running increment below.
5095 */
5096 if (cfs_rq_throttled(cfs_rq))
5097 break;
5098 cfs_rq->h_nr_running++;
5100 flags = ENQUEUE_WAKEUP;
5101 }
5103 for_each_sched_entity(se) {
5104 cfs_rq = cfs_rq_of(se);
5105 cfs_rq->h_nr_running++;
5107 if (cfs_rq_throttled(cfs_rq))
5108 break;
5110 update_load_avg(cfs_rq, se, UPDATE_TG);
5111 update_cfs_group(se);
5112 }
5114 if (!se)
5115 add_nr_running(rq, 1);
5117 hrtick_update(rq);
5118 }
5120 static void set_next_buddy(struct sched_entity *se);
5122 /*
5123 * The dequeue_task method is called before nr_running is
5124 * decreased. We remove the task from the rbtree and
5125 * update the fair scheduling stats:
5126 */
5127 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5128 {
5129 struct cfs_rq *cfs_rq;
5130 struct sched_entity *se = &p->se;
5131 int task_sleep = flags & DEQUEUE_SLEEP;
5133 for_each_sched_entity(se) {
5134 cfs_rq = cfs_rq_of(se);
5135 dequeue_entity(cfs_rq, se, flags);
5137 /*
5138 * end evaluation on encountering a throttled cfs_rq
5139 *
5140 * note: in the case of encountering a throttled cfs_rq we will
5141 * post the final h_nr_running decrement below.
5142 */
5143 if (cfs_rq_throttled(cfs_rq))
5144 break;
5145 cfs_rq->h_nr_running--;
5147 /* Don't dequeue parent if it has other entities besides us */
5148 if (cfs_rq->load.weight) {
5149 /* Avoid re-evaluating load for this entity: */
5150 se = parent_entity(se);
5151 /*
5152 * Bias pick_next to pick a task from this cfs_rq, as
5153 * p is sleeping when it is within its sched_slice.
5154 */
5155 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5156 set_next_buddy(se);
5157 break;
5158 }
5159 flags |= DEQUEUE_SLEEP;
5160 }
5162 for_each_sched_entity(se) {
5163 cfs_rq = cfs_rq_of(se);
5164 cfs_rq->h_nr_running--;
5166 if (cfs_rq_throttled(cfs_rq))
5167 break;
5169 update_load_avg(cfs_rq, se, UPDATE_TG);
5170 update_cfs_group(se);
5171 }
5173 if (!se)
5174 sub_nr_running(rq, 1);
5176 util_est_dequeue(&rq->cfs, p, task_sleep);
5177 hrtick_update(rq);
5178 }
5180 #ifdef CONFIG_SMP
5182 /* Working cpumask for: load_balance, load_balance_newidle. */
5183 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5184 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5186 #ifdef CONFIG_NO_HZ_COMMON
5187 /*
5188 * per rq 'load' arrray crap; XXX kill this.
5189 */
5191 /*
5192 * The exact cpuload calculated at every tick would be:
5193 *
5194 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
5195 *
5196 * If a CPU misses updates for n ticks (as it was idle) and update gets
5197 * called on the n+1-th tick when CPU may be busy, then we have:
5198 *
5199 * load_n = (1 - 1/2^i)^n * load_0
5200 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
5201 *
5202 * decay_load_missed() below does efficient calculation of
5203 *
5204 * load' = (1 - 1/2^i)^n * load
5205 *
5206 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
5207 * This allows us to precompute the above in said factors, thereby allowing the
5208 * reduction of an arbitrary n in O(log_2 n) steps. (See also
5209 * fixed_power_int())
5210 *
5211 * The calculation is approximated on a 128 point scale.
5212 */
5213 #define DEGRADE_SHIFT 7
5215 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
5216 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
5217 { 0, 0, 0, 0, 0, 0, 0, 0 },
5218 { 64, 32, 8, 0, 0, 0, 0, 0 },
5219 { 96, 72, 40, 12, 1, 0, 0, 0 },
5220 { 112, 98, 75, 43, 15, 1, 0, 0 },
5221 { 120, 112, 98, 76, 45, 16, 2, 0 }
5222 };
5224 /*
5225 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
5226 * would be when CPU is idle and so we just decay the old load without
5227 * adding any new load.
5228 */
5229 static unsigned long
5230 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
5231 {
5232 int j = 0;
5234 if (!missed_updates)
5235 return load;
5237 if (missed_updates >= degrade_zero_ticks[idx])
5238 return 0;
5240 if (idx == 1)
5241 return load >> missed_updates;
5243 while (missed_updates) {
5244 if (missed_updates % 2)
5245 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
5247 missed_updates >>= 1;
5248 j++;
5249 }
5250 return load;
5251 }
5253 static struct {
5254 cpumask_var_t idle_cpus_mask;
5255 atomic_t nr_cpus;
5256 int has_blocked; /* Idle CPUS has blocked load */
5257 unsigned long next_balance; /* in jiffy units */
5258 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5259 } nohz ____cacheline_aligned;
5261 #endif /* CONFIG_NO_HZ_COMMON */
5263 /**
5264 * __cpu_load_update - update the rq->cpu_load[] statistics
5265 * @this_rq: The rq to update statistics for
5266 * @this_load: The current load
5267 * @pending_updates: The number of missed updates
5268 *
5269 * Update rq->cpu_load[] statistics. This function is usually called every
5270 * scheduler tick (TICK_NSEC).
5271 *
5272 * This function computes a decaying average:
5273 *
5274 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
5275 *
5276 * Because of NOHZ it might not get called on every tick which gives need for
5277 * the @pending_updates argument.
5278 *
5279 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
5280 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
5281 * = A * (A * load[i]_n-2 + B) + B
5282 * = A * (A * (A * load[i]_n-3 + B) + B) + B
5283 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5284 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5285 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5286 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5287 *
5288 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5289 * any change in load would have resulted in the tick being turned back on.
5290 *
5291 * For regular NOHZ, this reduces to:
5292 *
5293 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
5294 *
5295 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5296 * term.
5297 */
5298 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
5299 unsigned long pending_updates)
5300 {
5301 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
5302 int i, scale;
5304 this_rq->nr_load_updates++;
5306 /* Update our load: */
5307 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
5308 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
5309 unsigned long old_load, new_load;
5311 /* scale is effectively 1 << i now, and >> i divides by scale */
5313 old_load = this_rq->cpu_load[i];
5314 #ifdef CONFIG_NO_HZ_COMMON
5315 old_load = decay_load_missed(old_load, pending_updates - 1, i);
5316 if (tickless_load) {
5317 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
5318 /*
5319 * old_load can never be a negative value because a
5320 * decayed tickless_load cannot be greater than the
5321 * original tickless_load.
5322 */
5323 old_load += tickless_load;
5324 }
5325 #endif
5326 new_load = this_load;
5327 /*
5328 * Round up the averaging division if load is increasing. This
5329 * prevents us from getting stuck on 9 if the load is 10, for
5330 * example.
5331 */
5332 if (new_load > old_load)
5333 new_load += scale - 1;
5335 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
5336 }
5337 }
5339 /* Used instead of source_load when we know the type == 0 */
5340 static unsigned long weighted_cpuload(struct rq *rq)
5341 {
5342 return cfs_rq_runnable_load_avg(&rq->cfs);
5343 }
5345 #ifdef CONFIG_NO_HZ_COMMON
5346 /*
5347 * There is no sane way to deal with nohz on smp when using jiffies because the
5348 * CPU doing the jiffies update might drift wrt the CPU doing the jiffy reading
5349 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5350 *
5351 * Therefore we need to avoid the delta approach from the regular tick when
5352 * possible since that would seriously skew the load calculation. This is why we
5353 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5354 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5355 * loop exit, nohz_idle_balance, nohz full exit...)
5356 *
5357 * This means we might still be one tick off for nohz periods.
5358 */
5360 static void cpu_load_update_nohz(struct rq *this_rq,
5361 unsigned long curr_jiffies,
5362 unsigned long load)
5363 {
5364 unsigned long pending_updates;
5366 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5367 if (pending_updates) {
5368 this_rq->last_load_update_tick = curr_jiffies;
5369 /*
5370 * In the regular NOHZ case, we were idle, this means load 0.
5371 * In the NOHZ_FULL case, we were non-idle, we should consider
5372 * its weighted load.
5373 */
5374 cpu_load_update(this_rq, load, pending_updates);
5375 }
5376 }
5378 /*
5379 * Called from nohz_idle_balance() to update the load ratings before doing the
5380 * idle balance.
5381 */
5382 static void cpu_load_update_idle(struct rq *this_rq)
5383 {
5384 /*
5385 * bail if there's load or we're actually up-to-date.
5386 */
5387 if (weighted_cpuload(this_rq))
5388 return;
5390 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
5391 }
5393 /*
5394 * Record CPU load on nohz entry so we know the tickless load to account
5395 * on nohz exit. cpu_load[0] happens then to be updated more frequently
5396 * than other cpu_load[idx] but it should be fine as cpu_load readers
5397 * shouldn't rely into synchronized cpu_load[*] updates.
5398 */
5399 void cpu_load_update_nohz_start(void)
5400 {
5401 struct rq *this_rq = this_rq();
5403 /*
5404 * This is all lockless but should be fine. If weighted_cpuload changes
5405 * concurrently we'll exit nohz. And cpu_load write can race with
5406 * cpu_load_update_idle() but both updater would be writing the same.
5407 */
5408 this_rq->cpu_load[0] = weighted_cpuload(this_rq);
5409 }
5411 /*
5412 * Account the tickless load in the end of a nohz frame.
5413 */
5414 void cpu_load_update_nohz_stop(void)
5415 {
5416 unsigned long curr_jiffies = READ_ONCE(jiffies);
5417 struct rq *this_rq = this_rq();
5418 unsigned long load;
5419 struct rq_flags rf;
5421 if (curr_jiffies == this_rq->last_load_update_tick)
5422 return;
5424 load = weighted_cpuload(this_rq);
5425 rq_lock(this_rq, &rf);
5426 update_rq_clock(this_rq);
5427 cpu_load_update_nohz(this_rq, curr_jiffies, load);
5428 rq_unlock(this_rq, &rf);
5429 }
5430 #else /* !CONFIG_NO_HZ_COMMON */
5431 static inline void cpu_load_update_nohz(struct rq *this_rq,
5432 unsigned long curr_jiffies,
5433 unsigned long load) { }
5434 #endif /* CONFIG_NO_HZ_COMMON */
5436 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
5437 {
5438 #ifdef CONFIG_NO_HZ_COMMON
5439 /* See the mess around cpu_load_update_nohz(). */
5440 this_rq->last_load_update_tick = READ_ONCE(jiffies);
5441 #endif
5442 cpu_load_update(this_rq, load, 1);
5443 }
5445 /*
5446 * Called from scheduler_tick()
5447 */
5448 void cpu_load_update_active(struct rq *this_rq)
5449 {
5450 unsigned long load = weighted_cpuload(this_rq);
5452 if (tick_nohz_tick_stopped())
5453 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
5454 else
5455 cpu_load_update_periodic(this_rq, load);
5456 }
5458 /*
5459 * Return a low guess at the load of a migration-source CPU weighted
5460 * according to the scheduling class and "nice" value.
5461 *
5462 * We want to under-estimate the load of migration sources, to
5463 * balance conservatively.
5464 */
5465 static unsigned long source_load(int cpu, int type)
5466 {
5467 struct rq *rq = cpu_rq(cpu);
5468 unsigned long total = weighted_cpuload(rq);
5470 if (type == 0 || !sched_feat(LB_BIAS))
5471 return total;
5473 return min(rq->cpu_load[type-1], total);
5474 }
5476 /*
5477 * Return a high guess at the load of a migration-target CPU weighted
5478 * according to the scheduling class and "nice" value.
5479 */
5480 static unsigned long target_load(int cpu, int type)
5481 {
5482 struct rq *rq = cpu_rq(cpu);
5483 unsigned long total = weighted_cpuload(rq);
5485 if (type == 0 || !sched_feat(LB_BIAS))
5486 return total;
5488 return max(rq->cpu_load[type-1], total);
5489 }
5491 static unsigned long capacity_of(int cpu)
5492 {
5493 return cpu_rq(cpu)->cpu_capacity;
5494 }
5496 static unsigned long capacity_orig_of(int cpu)
5497 {
5498 return cpu_rq(cpu)->cpu_capacity_orig;
5499 }
5501 static unsigned long cpu_avg_load_per_task(int cpu)
5502 {
5503 struct rq *rq = cpu_rq(cpu);
5504 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5505 unsigned long load_avg = weighted_cpuload(rq);
5507 if (nr_running)
5508 return load_avg / nr_running;
5510 return 0;
5511 }
5513 static void record_wakee(struct task_struct *p)
5514 {
5515 /*
5516 * Only decay a single time; tasks that have less then 1 wakeup per
5517 * jiffy will not have built up many flips.
5518 */
5519 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5520 current->wakee_flips >>= 1;
5521 current->wakee_flip_decay_ts = jiffies;
5522 }
5524 if (current->last_wakee != p) {
5525 current->last_wakee = p;
5526 current->wakee_flips++;
5527 }
5528 }
5530 /*
5531 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5532 *
5533 * A waker of many should wake a different task than the one last awakened
5534 * at a frequency roughly N times higher than one of its wakees.
5535 *
5536 * In order to determine whether we should let the load spread vs consolidating
5537 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5538 * partner, and a factor of lls_size higher frequency in the other.
5539 *
5540 * With both conditions met, we can be relatively sure that the relationship is
5541 * non-monogamous, with partner count exceeding socket size.
5542 *
5543 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5544 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5545 * socket size.
5546 */
5547 static int wake_wide(struct task_struct *p)
5548 {
5549 unsigned int master = current->wakee_flips;
5550 unsigned int slave = p->wakee_flips;
5551 int factor = this_cpu_read(sd_llc_size);
5553 if (master < slave)
5554 swap(master, slave);
5555 if (slave < factor || master < slave * factor)
5556 return 0;
5557 return 1;
5558 }
5560 /*
5561 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5562 * soonest. For the purpose of speed we only consider the waking and previous
5563 * CPU.
5564 *
5565 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5566 * cache-affine and is (or will be) idle.
5567 *
5568 * wake_affine_weight() - considers the weight to reflect the average
5569 * scheduling latency of the CPUs. This seems to work
5570 * for the overloaded case.
5571 */
5572 static int
5573 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5574 {
5575 /*
5576 * If this_cpu is idle, it implies the wakeup is from interrupt
5577 * context. Only allow the move if cache is shared. Otherwise an
5578 * interrupt intensive workload could force all tasks onto one
5579 * node depending on the IO topology or IRQ affinity settings.
5580 *
5581 * If the prev_cpu is idle and cache affine then avoid a migration.
5582 * There is no guarantee that the cache hot data from an interrupt
5583 * is more important than cache hot data on the prev_cpu and from
5584 * a cpufreq perspective, it's better to have higher utilisation
5585 * on one CPU.
5586 */
5587 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5588 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5590 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5591 return this_cpu;
5593 return nr_cpumask_bits;
5594 }
5596 static int
5597 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5598 int this_cpu, int prev_cpu, int sync)
5599 {
5600 s64 this_eff_load, prev_eff_load;
5601 unsigned long task_load;
5603 this_eff_load = target_load(this_cpu, sd->wake_idx);
5605 if (sync) {
5606 unsigned long current_load = task_h_load(current);
5608 if (current_load > this_eff_load)
5609 return this_cpu;
5611 this_eff_load -= current_load;
5612 }
5614 task_load = task_h_load(p);
5616 this_eff_load += task_load;
5617 if (sched_feat(WA_BIAS))
5618 this_eff_load *= 100;
5619 this_eff_load *= capacity_of(prev_cpu);
5621 prev_eff_load = source_load(prev_cpu, sd->wake_idx);
5622 prev_eff_load -= task_load;
5623 if (sched_feat(WA_BIAS))
5624 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5625 prev_eff_load *= capacity_of(this_cpu);
5627 /*
5628 * If sync, adjust the weight of prev_eff_load such that if
5629 * prev_eff == this_eff that select_idle_sibling() will consider
5630 * stacking the wakee on top of the waker if no other CPU is
5631 * idle.
5632 */
5633 if (sync)
5634 prev_eff_load += 1;
5636 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5637 }
5639 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5640 int this_cpu, int prev_cpu, int sync)
5641 {
5642 int target = nr_cpumask_bits;
5644 if (sched_feat(WA_IDLE))
5645 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5647 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5648 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5650 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5651 if (target == nr_cpumask_bits)
5652 return prev_cpu;
5654 schedstat_inc(sd->ttwu_move_affine);
5655 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5656 return target;
5657 }
5659 static unsigned long cpu_util_wake(int cpu, struct task_struct *p);
5661 static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
5662 {
5663 return max_t(long, capacity_of(cpu) - cpu_util_wake(cpu, p), 0);
5664 }
5666 /*
5667 * find_idlest_group finds and returns the least busy CPU group within the
5668 * domain.
5669 *
5670 * Assumes p is allowed on at least one CPU in sd.
5671 */
5672 static struct sched_group *
5673 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5674 int this_cpu, int sd_flag)
5675 {
5676 struct sched_group *idlest = NULL, *group = sd->groups;
5677 struct sched_group *most_spare_sg = NULL;
5678 unsigned long min_runnable_load = ULONG_MAX;
5679 unsigned long this_runnable_load = ULONG_MAX;
5680 unsigned long min_avg_load = ULONG_MAX, this_avg_load = ULONG_MAX;
5681 unsigned long most_spare = 0, this_spare = 0;
5682 int load_idx = sd->forkexec_idx;
5683 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5684 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5685 (sd->imbalance_pct-100) / 100;
5687 if (sd_flag & SD_BALANCE_WAKE)
5688 load_idx = sd->wake_idx;
5690 do {
5691 unsigned long load, avg_load, runnable_load;
5692 unsigned long spare_cap, max_spare_cap;
5693 int local_group;
5694 int i;
5696 /* Skip over this group if it has no CPUs allowed */
5697 if (!cpumask_intersects(sched_group_span(group),
5698 &p->cpus_allowed))
5699 continue;
5701 local_group = cpumask_test_cpu(this_cpu,
5702 sched_group_span(group));
5704 /*
5705 * Tally up the load of all CPUs in the group and find
5706 * the group containing the CPU with most spare capacity.
5707 */
5708 avg_load = 0;
5709 runnable_load = 0;
5710 max_spare_cap = 0;
5712 for_each_cpu(i, sched_group_span(group)) {
5713 /* Bias balancing toward CPUs of our domain */
5714 if (local_group)
5715 load = source_load(i, load_idx);
5716 else
5717 load = target_load(i, load_idx);
5719 runnable_load += load;
5721 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5723 spare_cap = capacity_spare_wake(i, p);
5725 if (spare_cap > max_spare_cap)
5726 max_spare_cap = spare_cap;
5727 }
5729 /* Adjust by relative CPU capacity of the group */
5730 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5731 group->sgc->capacity;
5732 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5733 group->sgc->capacity;
5735 if (local_group) {
5736 this_runnable_load = runnable_load;
5737 this_avg_load = avg_load;
5738 this_spare = max_spare_cap;
5739 } else {
5740 if (min_runnable_load > (runnable_load + imbalance)) {
5741 /*
5742 * The runnable load is significantly smaller
5743 * so we can pick this new CPU:
5744 */
5745 min_runnable_load = runnable_load;
5746 min_avg_load = avg_load;
5747 idlest = group;
5748 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5749 (100*min_avg_load > imbalance_scale*avg_load)) {
5750 /*
5751 * The runnable loads are close so take the
5752 * blocked load into account through avg_load:
5753 */
5754 min_avg_load = avg_load;
5755 idlest = group;
5756 }
5758 if (most_spare < max_spare_cap) {
5759 most_spare = max_spare_cap;
5760 most_spare_sg = group;
5761 }
5762 }
5763 } while (group = group->next, group != sd->groups);
5765 /*
5766 * The cross-over point between using spare capacity or least load
5767 * is too conservative for high utilization tasks on partially
5768 * utilized systems if we require spare_capacity > task_util(p),
5769 * so we allow for some task stuffing by using
5770 * spare_capacity > task_util(p)/2.
5771 *
5772 * Spare capacity can't be used for fork because the utilization has
5773 * not been set yet, we must first select a rq to compute the initial
5774 * utilization.
5775 */
5776 if (sd_flag & SD_BALANCE_FORK)
5777 goto skip_spare;
5779 if (this_spare > task_util(p) / 2 &&
5780 imbalance_scale*this_spare > 100*most_spare)
5781 return NULL;
5783 if (most_spare > task_util(p) / 2)
5784 return most_spare_sg;
5786 skip_spare:
5787 if (!idlest)
5788 return NULL;
5790 /*
5791 * When comparing groups across NUMA domains, it's possible for the
5792 * local domain to be very lightly loaded relative to the remote
5793 * domains but "imbalance" skews the comparison making remote CPUs
5794 * look much more favourable. When considering cross-domain, add
5795 * imbalance to the runnable load on the remote node and consider
5796 * staying local.
5797 */
5798 if ((sd->flags & SD_NUMA) &&
5799 min_runnable_load + imbalance >= this_runnable_load)
5800 return NULL;
5802 if (min_runnable_load > (this_runnable_load + imbalance))
5803 return NULL;
5805 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5806 (100*this_avg_load < imbalance_scale*min_avg_load))
5807 return NULL;
5809 return idlest;
5810 }
5812 /*
5813 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5814 */
5815 static int
5816 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5817 {
5818 unsigned long load, min_load = ULONG_MAX;
5819 unsigned int min_exit_latency = UINT_MAX;
5820 u64 latest_idle_timestamp = 0;
5821 int least_loaded_cpu = this_cpu;
5822 int shallowest_idle_cpu = -1;
5823 int i;
5825 /* Check if we have any choice: */
5826 if (group->group_weight == 1)
5827 return cpumask_first(sched_group_span(group));
5829 /* Traverse only the allowed CPUs */
5830 for_each_cpu_and(i, sched_group_span(group), &p->cpus_allowed) {
5831 if (available_idle_cpu(i)) {
5832 struct rq *rq = cpu_rq(i);
5833 struct cpuidle_state *idle = idle_get_state(rq);
5834 if (idle && idle->exit_latency < min_exit_latency) {
5835 /*
5836 * We give priority to a CPU whose idle state
5837 * has the smallest exit latency irrespective
5838 * of any idle timestamp.
5839 */
5840 min_exit_latency = idle->exit_latency;
5841 latest_idle_timestamp = rq->idle_stamp;
5842 shallowest_idle_cpu = i;
5843 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5844 rq->idle_stamp > latest_idle_timestamp) {
5845 /*
5846 * If equal or no active idle state, then
5847 * the most recently idled CPU might have
5848 * a warmer cache.
5849 */
5850 latest_idle_timestamp = rq->idle_stamp;
5851 shallowest_idle_cpu = i;
5852 }
5853 } else if (shallowest_idle_cpu == -1) {
5854 load = weighted_cpuload(cpu_rq(i));
5855 if (load < min_load) {
5856 min_load = load;
5857 least_loaded_cpu = i;
5858 }
5859 }
5860 }
5862 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5863 }
5865 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
5866 int cpu, int prev_cpu, int sd_flag)
5867 {
5868 int new_cpu = cpu;
5870 if (!cpumask_intersects(sched_domain_span(sd), &p->cpus_allowed))
5871 return prev_cpu;
5873 /*
5874 * We need task's util for capacity_spare_wake, sync it up to prev_cpu's
5875 * last_update_time.
5876 */
5877 if (!(sd_flag & SD_BALANCE_FORK))
5878 sync_entity_load_avg(&p->se);
5880 while (sd) {
5881 struct sched_group *group;
5882 struct sched_domain *tmp;
5883 int weight;
5885 if (!(sd->flags & sd_flag)) {
5886 sd = sd->child;
5887 continue;
5888 }
5890 group = find_idlest_group(sd, p, cpu, sd_flag);
5891 if (!group) {
5892 sd = sd->child;
5893 continue;
5894 }
5896 new_cpu = find_idlest_group_cpu(group, p, cpu);
5897 if (new_cpu == cpu) {
5898 /* Now try balancing at a lower domain level of 'cpu': */
5899 sd = sd->child;
5900 continue;
5901 }
5903 /* Now try balancing at a lower domain level of 'new_cpu': */
5904 cpu = new_cpu;
5905 weight = sd->span_weight;
5906 sd = NULL;
5907 for_each_domain(cpu, tmp) {
5908 if (weight <= tmp->span_weight)
5909 break;
5910 if (tmp->flags & sd_flag)
5911 sd = tmp;
5912 }
5913 }
5915 return new_cpu;
5916 }
5918 #ifdef CONFIG_SCHED_SMT
5919 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5921 static inline void set_idle_cores(int cpu, int val)
5922 {
5923 struct sched_domain_shared *sds;
5925 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5926 if (sds)
5927 WRITE_ONCE(sds->has_idle_cores, val);
5928 }
5930 static inline bool test_idle_cores(int cpu, bool def)
5931 {
5932 struct sched_domain_shared *sds;
5934 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5935 if (sds)
5936 return READ_ONCE(sds->has_idle_cores);
5938 return def;
5939 }
5941 /*
5942 * Scans the local SMT mask to see if the entire core is idle, and records this
5943 * information in sd_llc_shared->has_idle_cores.
5944 *
5945 * Since SMT siblings share all cache levels, inspecting this limited remote
5946 * state should be fairly cheap.
5947 */
5948 void __update_idle_core(struct rq *rq)
5949 {
5950 int core = cpu_of(rq);
5951 int cpu;
5953 rcu_read_lock();
5954 if (test_idle_cores(core, true))
5955 goto unlock;
5957 for_each_cpu(cpu, cpu_smt_mask(core)) {
5958 if (cpu == core)
5959 continue;
5961 if (!available_idle_cpu(cpu))
5962 goto unlock;
5963 }
5965 set_idle_cores(core, 1);
5966 unlock:
5967 rcu_read_unlock();
5968 }
5970 /*
5971 * Scan the entire LLC domain for idle cores; this dynamically switches off if
5972 * there are no idle cores left in the system; tracked through
5973 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5974 */
5975 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5976 {
5977 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5978 int core, cpu;
5980 if (!static_branch_likely(&sched_smt_present))
5981 return -1;
5983 if (!test_idle_cores(target, false))
5984 return -1;
5986 cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
5988 for_each_cpu_wrap(core, cpus, target) {
5989 bool idle = true;
5991 for_each_cpu(cpu, cpu_smt_mask(core)) {
5992 cpumask_clear_cpu(cpu, cpus);
5993 if (!available_idle_cpu(cpu))
5994 idle = false;
5995 }
5997 if (idle)
5998 return core;
5999 }
6001 /*
6002 * Failed to find an idle core; stop looking for one.
6003 */
6004 set_idle_cores(target, 0);
6006 return -1;
6007 }
6009 /*
6010 * Scan the local SMT mask for idle CPUs.
6011 */
6012 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6013 {
6014 int cpu;
6016 if (!static_branch_likely(&sched_smt_present))
6017 return -1;
6019 for_each_cpu(cpu, cpu_smt_mask(target)) {
6020 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6021 continue;
6022 if (available_idle_cpu(cpu))
6023 return cpu;
6024 }
6026 return -1;
6027 }
6029 #else /* CONFIG_SCHED_SMT */
6031 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6032 {
6033 return -1;
6034 }
6036 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6037 {
6038 return -1;
6039 }
6041 #endif /* CONFIG_SCHED_SMT */
6043 /*
6044 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6045 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6046 * average idle time for this rq (as found in rq->avg_idle).
6047 */
6048 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
6049 {
6050 struct sched_domain *this_sd;
6051 u64 avg_cost, avg_idle;
6052 u64 time, cost;
6053 s64 delta;
6054 int cpu, nr = INT_MAX;
6056 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6057 if (!this_sd)
6058 return -1;
6060 /*
6061 * Due to large variance we need a large fuzz factor; hackbench in
6062 * particularly is sensitive here.
6063 */
6064 avg_idle = this_rq()->avg_idle / 512;
6065 avg_cost = this_sd->avg_scan_cost + 1;
6067 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
6068 return -1;
6070 if (sched_feat(SIS_PROP)) {
6071 u64 span_avg = sd->span_weight * avg_idle;
6072 if (span_avg > 4*avg_cost)
6073 nr = div_u64(span_avg, avg_cost);
6074 else
6075 nr = 4;
6076 }
6078 time = local_clock();
6080 for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
6081 if (!--nr)
6082 return -1;
6083 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6084 continue;
6085 if (available_idle_cpu(cpu))
6086 break;
6087 }
6089 time = local_clock() - time;
6090 cost = this_sd->avg_scan_cost;
6091 delta = (s64)(time - cost) / 8;
6092 this_sd->avg_scan_cost += delta;
6094 return cpu;
6095 }
6097 /*
6098 * Try and locate an idle core/thread in the LLC cache domain.
6099 */
6100 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6101 {
6102 struct sched_domain *sd;
6103 int i, recent_used_cpu;
6105 if (available_idle_cpu(target))
6106 return target;
6108 /*
6109 * If the previous CPU is cache affine and idle, don't be stupid:
6110 */
6111 if (prev != target && cpus_share_cache(prev, target) && available_idle_cpu(prev))
6112 return prev;
6114 /* Check a recently used CPU as a potential idle candidate: */
6115 recent_used_cpu = p->recent_used_cpu;
6116 if (recent_used_cpu != prev &&
6117 recent_used_cpu != target &&
6118 cpus_share_cache(recent_used_cpu, target) &&
6119 available_idle_cpu(recent_used_cpu) &&
6120 cpumask_test_cpu(p->recent_used_cpu, &p->cpus_allowed)) {
6121 /*
6122 * Replace recent_used_cpu with prev as it is a potential
6123 * candidate for the next wake:
6124 */
6125 p->recent_used_cpu = prev;
6126 return recent_used_cpu;
6127 }
6129 sd = rcu_dereference(per_cpu(sd_llc, target));
6130 if (!sd)
6131 return target;
6133 i = select_idle_core(p, sd, target);
6134 if ((unsigned)i < nr_cpumask_bits)
6135 return i;
6137 i = select_idle_cpu(p, sd, target);
6138 if ((unsigned)i < nr_cpumask_bits)
6139 return i;
6141 i = select_idle_smt(p, sd, target);
6142 if ((unsigned)i < nr_cpumask_bits)
6143 return i;
6145 return target;
6146 }
6148 /**
6149 * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
6150 * @cpu: the CPU to get the utilization of
6151 *
6152 * The unit of the return value must be the one of capacity so we can compare
6153 * the utilization with the capacity of the CPU that is available for CFS task
6154 * (ie cpu_capacity).
6155 *
6156 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6157 * recent utilization of currently non-runnable tasks on a CPU. It represents
6158 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6159 * capacity_orig is the cpu_capacity available at the highest frequency
6160 * (arch_scale_freq_capacity()).
6161 * The utilization of a CPU converges towards a sum equal to or less than the
6162 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6163 * the running time on this CPU scaled by capacity_curr.
6164 *
6165 * The estimated utilization of a CPU is defined to be the maximum between its
6166 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6167 * currently RUNNABLE on that CPU.
6168 * This allows to properly represent the expected utilization of a CPU which
6169 * has just got a big task running since a long sleep period. At the same time
6170 * however it preserves the benefits of the "blocked utilization" in
6171 * describing the potential for other tasks waking up on the same CPU.
6172 *
6173 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6174 * higher than capacity_orig because of unfortunate rounding in
6175 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6176 * the average stabilizes with the new running time. We need to check that the
6177 * utilization stays within the range of [0..capacity_orig] and cap it if
6178 * necessary. Without utilization capping, a group could be seen as overloaded
6179 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6180 * available capacity. We allow utilization to overshoot capacity_curr (but not
6181 * capacity_orig) as it useful for predicting the capacity required after task
6182 * migrations (scheduler-driven DVFS).
6183 *
6184 * Return: the (estimated) utilization for the specified CPU
6185 */
6186 static inline unsigned long cpu_util(int cpu)
6187 {
6188 struct cfs_rq *cfs_rq;
6189 unsigned int util;
6191 cfs_rq = &cpu_rq(cpu)->cfs;
6192 util = READ_ONCE(cfs_rq->avg.util_avg);
6194 if (sched_feat(UTIL_EST))
6195 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6197 return min_t(unsigned long, util, capacity_orig_of(cpu));
6198 }
6200 /*
6201 * cpu_util_wake: Compute CPU utilization with any contributions from
6202 * the waking task p removed.
6203 */
6204 static unsigned long cpu_util_wake(int cpu, struct task_struct *p)
6205 {
6206 struct cfs_rq *cfs_rq;
6207 unsigned int util;
6209 /* Task has no contribution or is new */
6210 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6211 return cpu_util(cpu);
6213 cfs_rq = &cpu_rq(cpu)->cfs;
6214 util = READ_ONCE(cfs_rq->avg.util_avg);
6216 /* Discount task's blocked util from CPU's util */
6217 util -= min_t(unsigned int, util, task_util(p));
6219 /*
6220 * Covered cases:
6221 *
6222 * a) if *p is the only task sleeping on this CPU, then:
6223 * cpu_util (== task_util) > util_est (== 0)
6224 * and thus we return:
6225 * cpu_util_wake = (cpu_util - task_util) = 0
6226 *
6227 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6228 * IDLE, then:
6229 * cpu_util >= task_util
6230 * cpu_util > util_est (== 0)
6231 * and thus we discount *p's blocked utilization to return:
6232 * cpu_util_wake = (cpu_util - task_util) >= 0
6233 *
6234 * c) if other tasks are RUNNABLE on that CPU and
6235 * util_est > cpu_util
6236 * then we use util_est since it returns a more restrictive
6237 * estimation of the spare capacity on that CPU, by just
6238 * considering the expected utilization of tasks already
6239 * runnable on that CPU.
6240 *
6241 * Cases a) and b) are covered by the above code, while case c) is
6242 * covered by the following code when estimated utilization is
6243 * enabled.
6244 */
6245 if (sched_feat(UTIL_EST))
6246 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6248 /*
6249 * Utilization (estimated) can exceed the CPU capacity, thus let's
6250 * clamp to the maximum CPU capacity to ensure consistency with
6251 * the cpu_util call.
6252 */
6253 return min_t(unsigned long, util, capacity_orig_of(cpu));
6254 }
6256 /*
6257 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6258 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6259 *
6260 * In that case WAKE_AFFINE doesn't make sense and we'll let
6261 * BALANCE_WAKE sort things out.
6262 */
6263 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6264 {
6265 long min_cap, max_cap;
6267 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6268 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
6270 /* Minimum capacity is close to max, no need to abort wake_affine */
6271 if (max_cap - min_cap < max_cap >> 3)
6272 return 0;
6274 /* Bring task utilization in sync with prev_cpu */
6275 sync_entity_load_avg(&p->se);
6277 return min_cap * 1024 < task_util(p) * capacity_margin;
6278 }
6280 /*
6281 * select_task_rq_fair: Select target runqueue for the waking task in domains
6282 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6283 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6284 *
6285 * Balances load by selecting the idlest CPU in the idlest group, or under
6286 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6287 *
6288 * Returns the target CPU number.
6289 *
6290 * preempt must be disabled.
6291 */
6292 static int
6293 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6294 {
6295 struct sched_domain *tmp, *sd = NULL;
6296 int cpu = smp_processor_id();
6297 int new_cpu = prev_cpu;
6298 int want_affine = 0;
6299 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6301 if (sd_flag & SD_BALANCE_WAKE) {
6302 record_wakee(p);
6303 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
6304 && cpumask_test_cpu(cpu, &p->cpus_allowed);
6305 }
6307 rcu_read_lock();
6308 for_each_domain(cpu, tmp) {
6309 if (!(tmp->flags & SD_LOAD_BALANCE))
6310 break;
6312 /*
6313 * If both 'cpu' and 'prev_cpu' are part of this domain,
6314 * cpu is a valid SD_WAKE_AFFINE target.
6315 */
6316 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6317 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6318 if (cpu != prev_cpu)
6319 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6321 sd = NULL; /* Prefer wake_affine over balance flags */
6322 break;
6323 }
6325 if (tmp->flags & sd_flag)
6326 sd = tmp;
6327 else if (!want_affine)
6328 break;
6329 }
6331 if (unlikely(sd)) {
6332 /* Slow path */
6333 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6334 } else if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */
6335 /* Fast path */
6337 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6339 if (want_affine)
6340 current->recent_used_cpu = cpu;
6341 }
6342 rcu_read_unlock();
6344 return new_cpu;
6345 }
6347 static void detach_entity_cfs_rq(struct sched_entity *se);
6349 /*
6350 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6351 * cfs_rq_of(p) references at time of call are still valid and identify the
6352 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6353 */
6354 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
6355 {
6356 /*
6357 * As blocked tasks retain absolute vruntime the migration needs to
6358 * deal with this by subtracting the old and adding the new
6359 * min_vruntime -- the latter is done by enqueue_entity() when placing
6360 * the task on the new runqueue.
6361 */
6362 if (p->state == TASK_WAKING) {
6363 struct sched_entity *se = &p->se;
6364 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6365 u64 min_vruntime;
6367 #ifndef CONFIG_64BIT
6368 u64 min_vruntime_copy;
6370 do {
6371 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6372 smp_rmb();
6373 min_vruntime = cfs_rq->min_vruntime;
6374 } while (min_vruntime != min_vruntime_copy);
6375 #else
6376 min_vruntime = cfs_rq->min_vruntime;
6377 #endif
6379 se->vruntime -= min_vruntime;
6380 }
6382 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6383 /*
6384 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6385 * rq->lock and can modify state directly.
6386 */
6387 lockdep_assert_held(&task_rq(p)->lock);
6388 detach_entity_cfs_rq(&p->se);
6390 } else {
6391 /*
6392 * We are supposed to update the task to "current" time, then
6393 * its up to date and ready to go to new CPU/cfs_rq. But we
6394 * have difficulty in getting what current time is, so simply
6395 * throw away the out-of-date time. This will result in the
6396 * wakee task is less decayed, but giving the wakee more load
6397 * sounds not bad.
6398 */
6399 remove_entity_load_avg(&p->se);
6400 }
6402 /* Tell new CPU we are migrated */
6403 p->se.avg.last_update_time = 0;
6405 /* We have migrated, no longer consider this task hot */
6406 p->se.exec_start = 0;
6408 update_scan_period(p, new_cpu);
6409 }
6411 static void task_dead_fair(struct task_struct *p)
6412 {
6413 remove_entity_load_avg(&p->se);
6414 }
6415 #endif /* CONFIG_SMP */
6417 static unsigned long wakeup_gran(struct sched_entity *se)
6418 {
6419 unsigned long gran = sysctl_sched_wakeup_granularity;
6421 /*
6422 * Since its curr running now, convert the gran from real-time
6423 * to virtual-time in his units.
6424 *
6425 * By using 'se' instead of 'curr' we penalize light tasks, so
6426 * they get preempted easier. That is, if 'se' < 'curr' then
6427 * the resulting gran will be larger, therefore penalizing the
6428 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6429 * be smaller, again penalizing the lighter task.
6430 *
6431 * This is especially important for buddies when the leftmost
6432 * task is higher priority than the buddy.
6433 */
6434 return calc_delta_fair(gran, se);
6435 }
6437 /*
6438 * Should 'se' preempt 'curr'.
6439 *
6440 * |s1
6441 * |s2
6442 * |s3
6443 * g
6444 * |<--->|c
6445 *
6446 * w(c, s1) = -1
6447 * w(c, s2) = 0
6448 * w(c, s3) = 1
6449 *
6450 */
6451 static int
6452 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6453 {
6454 s64 gran, vdiff = curr->vruntime - se->vruntime;
6456 if (vdiff <= 0)
6457 return -1;
6459 gran = wakeup_gran(se);
6460 if (vdiff > gran)
6461 return 1;
6463 return 0;
6464 }
6466 static void set_last_buddy(struct sched_entity *se)
6467 {
6468 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6469 return;
6471 for_each_sched_entity(se) {
6472 if (SCHED_WARN_ON(!se->on_rq))
6473 return;
6474 cfs_rq_of(se)->last = se;
6475 }
6476 }
6478 static void set_next_buddy(struct sched_entity *se)
6479 {
6480 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6481 return;
6483 for_each_sched_entity(se) {
6484 if (SCHED_WARN_ON(!se->on_rq))
6485 return;
6486 cfs_rq_of(se)->next = se;
6487 }
6488 }
6490 static void set_skip_buddy(struct sched_entity *se)
6491 {
6492 for_each_sched_entity(se)
6493 cfs_rq_of(se)->skip = se;
6494 }
6496 /*
6497 * Preempt the current task with a newly woken task if needed:
6498 */
6499 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6500 {
6501 struct task_struct *curr = rq->curr;
6502 struct sched_entity *se = &curr->se, *pse = &p->se;
6503 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6504 int scale = cfs_rq->nr_running >= sched_nr_latency;
6505 int next_buddy_marked = 0;
6507 if (unlikely(se == pse))
6508 return;
6510 /*
6511 * This is possible from callers such as attach_tasks(), in which we
6512 * unconditionally check_prempt_curr() after an enqueue (which may have
6513 * lead to a throttle). This both saves work and prevents false
6514 * next-buddy nomination below.
6515 */
6516 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6517 return;
6519 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6520 set_next_buddy(pse);
6521 next_buddy_marked = 1;
6522 }
6524 /*
6525 * We can come here with TIF_NEED_RESCHED already set from new task
6526 * wake up path.
6527 *
6528 * Note: this also catches the edge-case of curr being in a throttled
6529 * group (e.g. via set_curr_task), since update_curr() (in the
6530 * enqueue of curr) will have resulted in resched being set. This
6531 * prevents us from potentially nominating it as a false LAST_BUDDY
6532 * below.
6533 */
6534 if (test_tsk_need_resched(curr))
6535 return;
6537 /* Idle tasks are by definition preempted by non-idle tasks. */
6538 if (unlikely(curr->policy == SCHED_IDLE) &&
6539 likely(p->policy != SCHED_IDLE))
6540 goto preempt;
6542 /*
6543 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6544 * is driven by the tick):
6545 */
6546 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6547 return;
6549 find_matching_se(&se, &pse);
6550 update_curr(cfs_rq_of(se));
6551 BUG_ON(!pse);
6552 if (wakeup_preempt_entity(se, pse) == 1) {
6553 /*
6554 * Bias pick_next to pick the sched entity that is
6555 * triggering this preemption.
6556 */
6557 if (!next_buddy_marked)
6558 set_next_buddy(pse);
6559 goto preempt;
6560 }
6562 return;
6564 preempt:
6565 resched_curr(rq);
6566 /*
6567 * Only set the backward buddy when the current task is still
6568 * on the rq. This can happen when a wakeup gets interleaved
6569 * with schedule on the ->pre_schedule() or idle_balance()
6570 * point, either of which can * drop the rq lock.
6571 *
6572 * Also, during early boot the idle thread is in the fair class,
6573 * for obvious reasons its a bad idea to schedule back to it.
6574 */
6575 if (unlikely(!se->on_rq || curr == rq->idle))
6576 return;
6578 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6579 set_last_buddy(se);
6580 }
6582 static struct task_struct *
6583 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6584 {
6585 struct cfs_rq *cfs_rq = &rq->cfs;
6586 struct sched_entity *se;
6587 struct task_struct *p;
6588 int new_tasks;
6590 again:
6591 if (!cfs_rq->nr_running)
6592 goto idle;
6594 #ifdef CONFIG_FAIR_GROUP_SCHED
6595 if (prev->sched_class != &fair_sched_class)
6596 goto simple;
6598 /*
6599 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6600 * likely that a next task is from the same cgroup as the current.
6601 *
6602 * Therefore attempt to avoid putting and setting the entire cgroup
6603 * hierarchy, only change the part that actually changes.
6604 */
6606 do {
6607 struct sched_entity *curr = cfs_rq->curr;
6609 /*
6610 * Since we got here without doing put_prev_entity() we also
6611 * have to consider cfs_rq->curr. If it is still a runnable
6612 * entity, update_curr() will update its vruntime, otherwise
6613 * forget we've ever seen it.
6614 */
6615 if (curr) {
6616 if (curr->on_rq)
6617 update_curr(cfs_rq);
6618 else
6619 curr = NULL;
6621 /*
6622 * This call to check_cfs_rq_runtime() will do the
6623 * throttle and dequeue its entity in the parent(s).
6624 * Therefore the nr_running test will indeed
6625 * be correct.
6626 */
6627 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6628 cfs_rq = &rq->cfs;
6630 if (!cfs_rq->nr_running)
6631 goto idle;
6633 goto simple;
6634 }
6635 }
6637 se = pick_next_entity(cfs_rq, curr);
6638 cfs_rq = group_cfs_rq(se);
6639 } while (cfs_rq);
6641 p = task_of(se);
6643 /*
6644 * Since we haven't yet done put_prev_entity and if the selected task
6645 * is a different task than we started out with, try and touch the
6646 * least amount of cfs_rqs.
6647 */
6648 if (prev != p) {
6649 struct sched_entity *pse = &prev->se;
6651 while (!(cfs_rq = is_same_group(se, pse))) {
6652 int se_depth = se->depth;
6653 int pse_depth = pse->depth;
6655 if (se_depth <= pse_depth) {
6656 put_prev_entity(cfs_rq_of(pse), pse);
6657 pse = parent_entity(pse);
6658 }
6659 if (se_depth >= pse_depth) {
6660 set_next_entity(cfs_rq_of(se), se);
6661 se = parent_entity(se);
6662 }
6663 }
6665 put_prev_entity(cfs_rq, pse);
6666 set_next_entity(cfs_rq, se);
6667 }
6669 goto done;
6670 simple:
6671 #endif
6673 put_prev_task(rq, prev);
6675 do {
6676 se = pick_next_entity(cfs_rq, NULL);
6677 set_next_entity(cfs_rq, se);
6678 cfs_rq = group_cfs_rq(se);
6679 } while (cfs_rq);
6681 p = task_of(se);
6683 done: __maybe_unused;
6684 #ifdef CONFIG_SMP
6685 /*
6686 * Move the next running task to the front of
6687 * the list, so our cfs_tasks list becomes MRU
6688 * one.
6689 */
6690 list_move(&p->se.group_node, &rq->cfs_tasks);
6691 #endif
6693 if (hrtick_enabled(rq))
6694 hrtick_start_fair(rq, p);
6696 return p;
6698 idle:
6699 new_tasks = idle_balance(rq, rf);
6701 /*
6702 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6703 * possible for any higher priority task to appear. In that case we
6704 * must re-start the pick_next_entity() loop.
6705 */
6706 if (new_tasks < 0)
6707 return RETRY_TASK;
6709 if (new_tasks > 0)
6710 goto again;
6712 return NULL;
6713 }
6715 /*
6716 * Account for a descheduled task:
6717 */
6718 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6719 {
6720 struct sched_entity *se = &prev->se;
6721 struct cfs_rq *cfs_rq;
6723 for_each_sched_entity(se) {
6724 cfs_rq = cfs_rq_of(se);
6725 put_prev_entity(cfs_rq, se);
6726 }
6727 }
6729 /*
6730 * sched_yield() is very simple
6731 *
6732 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6733 */
6734 static void yield_task_fair(struct rq *rq)
6735 {
6736 struct task_struct *curr = rq->curr;
6737 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6738 struct sched_entity *se = &curr->se;
6740 /*
6741 * Are we the only task in the tree?
6742 */
6743 if (unlikely(rq->nr_running == 1))
6744 return;
6746 clear_buddies(cfs_rq, se);
6748 if (curr->policy != SCHED_BATCH) {
6749 update_rq_clock(rq);
6750 /*
6751 * Update run-time statistics of the 'current'.
6752 */
6753 update_curr(cfs_rq);
6754 /*
6755 * Tell update_rq_clock() that we've just updated,
6756 * so we don't do microscopic update in schedule()
6757 * and double the fastpath cost.
6758 */
6759 rq_clock_skip_update(rq);
6760 }
6762 set_skip_buddy(se);
6763 }
6765 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6766 {
6767 struct sched_entity *se = &p->se;
6769 /* throttled hierarchies are not runnable */
6770 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6771 return false;
6773 /* Tell the scheduler that we'd really like pse to run next. */
6774 set_next_buddy(se);
6776 yield_task_fair(rq);
6778 return true;
6779 }
6781 #ifdef CONFIG_SMP
6782 /**************************************************
6783 * Fair scheduling class load-balancing methods.
6784 *
6785 * BASICS
6786 *
6787 * The purpose of load-balancing is to achieve the same basic fairness the
6788 * per-CPU scheduler provides, namely provide a proportional amount of compute
6789 * time to each task. This is expressed in the following equation:
6790 *
6791 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6792 *
6793 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
6794 * W_i,0 is defined as:
6795 *
6796 * W_i,0 = \Sum_j w_i,j (2)
6797 *
6798 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
6799 * is derived from the nice value as per sched_prio_to_weight[].
6800 *
6801 * The weight average is an exponential decay average of the instantaneous
6802 * weight:
6803 *
6804 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6805 *
6806 * C_i is the compute capacity of CPU i, typically it is the
6807 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6808 * can also include other factors [XXX].
6809 *
6810 * To achieve this balance we define a measure of imbalance which follows
6811 * directly from (1):
6812 *
6813 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6814 *
6815 * We them move tasks around to minimize the imbalance. In the continuous
6816 * function space it is obvious this converges, in the discrete case we get
6817 * a few fun cases generally called infeasible weight scenarios.
6818 *
6819 * [XXX expand on:
6820 * - infeasible weights;
6821 * - local vs global optima in the discrete case. ]
6822 *
6823 *
6824 * SCHED DOMAINS
6825 *
6826 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6827 * for all i,j solution, we create a tree of CPUs that follows the hardware
6828 * topology where each level pairs two lower groups (or better). This results
6829 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
6830 * tree to only the first of the previous level and we decrease the frequency
6831 * of load-balance at each level inv. proportional to the number of CPUs in
6832 * the groups.
6833 *
6834 * This yields:
6835 *
6836 * log_2 n 1 n
6837 * \Sum { --- * --- * 2^i } = O(n) (5)
6838 * i = 0 2^i 2^i
6839 * `- size of each group
6840 * | | `- number of CPUs doing load-balance
6841 * | `- freq
6842 * `- sum over all levels
6843 *
6844 * Coupled with a limit on how many tasks we can migrate every balance pass,
6845 * this makes (5) the runtime complexity of the balancer.
6846 *
6847 * An important property here is that each CPU is still (indirectly) connected
6848 * to every other CPU in at most O(log n) steps:
6849 *
6850 * The adjacency matrix of the resulting graph is given by:
6851 *
6852 * log_2 n
6853 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6854 * k = 0
6855 *
6856 * And you'll find that:
6857 *
6858 * A^(log_2 n)_i,j != 0 for all i,j (7)
6859 *
6860 * Showing there's indeed a path between every CPU in at most O(log n) steps.
6861 * The task movement gives a factor of O(m), giving a convergence complexity
6862 * of:
6863 *
6864 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6865 *
6866 *
6867 * WORK CONSERVING
6868 *
6869 * In order to avoid CPUs going idle while there's still work to do, new idle
6870 * balancing is more aggressive and has the newly idle CPU iterate up the domain
6871 * tree itself instead of relying on other CPUs to bring it work.
6872 *
6873 * This adds some complexity to both (5) and (8) but it reduces the total idle
6874 * time.
6875 *
6876 * [XXX more?]
6877 *
6878 *
6879 * CGROUPS
6880 *
6881 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6882 *
6883 * s_k,i
6884 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6885 * S_k
6886 *
6887 * Where
6888 *
6889 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6890 *
6891 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
6892 *
6893 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6894 * property.
6895 *
6896 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6897 * rewrite all of this once again.]
6898 */
6900 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6902 enum fbq_type { regular, remote, all };
6904 #define LBF_ALL_PINNED 0x01
6905 #define LBF_NEED_BREAK 0x02
6906 #define LBF_DST_PINNED 0x04
6907 #define LBF_SOME_PINNED 0x08
6908 #define LBF_NOHZ_STATS 0x10
6909 #define LBF_NOHZ_AGAIN 0x20
6911 struct lb_env {
6912 struct sched_domain *sd;
6914 struct rq *src_rq;
6915 int src_cpu;
6917 int dst_cpu;
6918 struct rq *dst_rq;
6920 struct cpumask *dst_grpmask;
6921 int new_dst_cpu;
6922 enum cpu_idle_type idle;
6923 long imbalance;
6924 /* The set of CPUs under consideration for load-balancing */
6925 struct cpumask *cpus;
6927 unsigned int flags;
6929 unsigned int loop;
6930 unsigned int loop_break;
6931 unsigned int loop_max;
6933 enum fbq_type fbq_type;
6934 struct list_head tasks;
6935 };
6937 /*
6938 * Is this task likely cache-hot:
6939 */
6940 static int task_hot(struct task_struct *p, struct lb_env *env)
6941 {
6942 s64 delta;
6944 lockdep_assert_held(&env->src_rq->lock);
6946 if (p->sched_class != &fair_sched_class)
6947 return 0;
6949 if (unlikely(p->policy == SCHED_IDLE))
6950 return 0;
6952 /*
6953 * Buddy candidates are cache hot:
6954 */
6955 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6956 (&p->se == cfs_rq_of(&p->se)->next ||
6957 &p->se == cfs_rq_of(&p->se)->last))
6958 return 1;
6960 if (sysctl_sched_migration_cost == -1)
6961 return 1;
6962 if (sysctl_sched_migration_cost == 0)
6963 return 0;
6965 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6967 return delta < (s64)sysctl_sched_migration_cost;
6968 }
6970 #ifdef CONFIG_NUMA_BALANCING
6971 /*
6972 * Returns 1, if task migration degrades locality
6973 * Returns 0, if task migration improves locality i.e migration preferred.
6974 * Returns -1, if task migration is not affected by locality.
6975 */
6976 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6977 {
6978 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6979 unsigned long src_weight, dst_weight;
6980 int src_nid, dst_nid, dist;
6982 if (!static_branch_likely(&sched_numa_balancing))
6983 return -1;
6985 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6986 return -1;
6988 src_nid = cpu_to_node(env->src_cpu);
6989 dst_nid = cpu_to_node(env->dst_cpu);
6991 if (src_nid == dst_nid)
6992 return -1;
6994 /* Migrating away from the preferred node is always bad. */
6995 if (src_nid == p->numa_preferred_nid) {
6996 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6997 return 1;
6998 else
6999 return -1;
7000 }
7002 /* Encourage migration to the preferred node. */
7003 if (dst_nid == p->numa_preferred_nid)
7004 return 0;
7006 /* Leaving a core idle is often worse than degrading locality. */
7007 if (env->idle == CPU_IDLE)
7008 return -1;
7010 dist = node_distance(src_nid, dst_nid);
7011 if (numa_group) {
7012 src_weight = group_weight(p, src_nid, dist);
7013 dst_weight = group_weight(p, dst_nid, dist);
7014 } else {
7015 src_weight = task_weight(p, src_nid, dist);
7016 dst_weight = task_weight(p, dst_nid, dist);
7017 }
7019 return dst_weight < src_weight;
7020 }
7022 #else
7023 static inline int migrate_degrades_locality(struct task_struct *p,
7024 struct lb_env *env)
7025 {
7026 return -1;
7027 }
7028 #endif
7030 /*
7031 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7032 */
7033 static
7034 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7035 {
7036 int tsk_cache_hot;
7038 lockdep_assert_held(&env->src_rq->lock);
7040 /*
7041 * We do not migrate tasks that are:
7042 * 1) throttled_lb_pair, or
7043 * 2) cannot be migrated to this CPU due to cpus_allowed, or
7044 * 3) running (obviously), or
7045 * 4) are cache-hot on their current CPU.
7046 */
7047 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7048 return 0;
7050 if (!cpumask_test_cpu(env->dst_cpu, &p->cpus_allowed)) {
7051 int cpu;
7053 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7055 env->flags |= LBF_SOME_PINNED;
7057 /*
7058 * Remember if this task can be migrated to any other CPU in
7059 * our sched_group. We may want to revisit it if we couldn't
7060 * meet load balance goals by pulling other tasks on src_cpu.
7061 *
7062 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7063 * already computed one in current iteration.
7064 */
7065 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
7066 return 0;
7068 /* Prevent to re-select dst_cpu via env's CPUs: */
7069 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7070 if (cpumask_test_cpu(cpu, &p->cpus_allowed)) {
7071 env->flags |= LBF_DST_PINNED;
7072 env->new_dst_cpu = cpu;
7073 break;
7074 }
7075 }
7077 return 0;
7078 }
7080 /* Record that we found atleast one task that could run on dst_cpu */
7081 env->flags &= ~LBF_ALL_PINNED;
7083 if (task_running(env->src_rq, p)) {
7084 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7085 return 0;
7086 }
7088 /*
7089 * Aggressive migration if:
7090 * 1) destination numa is preferred
7091 * 2) task is cache cold, or
7092 * 3) too many balance attempts have failed.
7093 */
7094 tsk_cache_hot = migrate_degrades_locality(p, env);
7095 if (tsk_cache_hot == -1)
7096 tsk_cache_hot = task_hot(p, env);
7098 if (tsk_cache_hot <= 0 ||
7099 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7100 if (tsk_cache_hot == 1) {
7101 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7102 schedstat_inc(p->se.statistics.nr_forced_migrations);
7103 }
7104 return 1;
7105 }
7107 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7108 return 0;
7109 }
7111 /*
7112 * detach_task() -- detach the task for the migration specified in env
7113 */
7114 static void detach_task(struct task_struct *p, struct lb_env *env)
7115 {
7116 lockdep_assert_held(&env->src_rq->lock);
7118 p->on_rq = TASK_ON_RQ_MIGRATING;
7119 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7120 set_task_cpu(p, env->dst_cpu);
7121 }
7123 /*
7124 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7125 * part of active balancing operations within "domain".
7126 *
7127 * Returns a task if successful and NULL otherwise.
7128 */
7129 static struct task_struct *detach_one_task(struct lb_env *env)
7130 {
7131 struct task_struct *p;
7133 lockdep_assert_held(&env->src_rq->lock);
7135 list_for_each_entry_reverse(p,
7136 &env->src_rq->cfs_tasks, se.group_node) {
7137 if (!can_migrate_task(p, env))
7138 continue;
7140 detach_task(p, env);
7142 /*
7143 * Right now, this is only the second place where
7144 * lb_gained[env->idle] is updated (other is detach_tasks)
7145 * so we can safely collect stats here rather than
7146 * inside detach_tasks().
7147 */
7148 schedstat_inc(env->sd->lb_gained[env->idle]);
7149 return p;
7150 }
7151 return NULL;
7152 }
7154 static const unsigned int sched_nr_migrate_break = 32;
7156 /*
7157 * detach_tasks() -- tries to detach up to imbalance weighted load from
7158 * busiest_rq, as part of a balancing operation within domain "sd".
7159 *
7160 * Returns number of detached tasks if successful and 0 otherwise.
7161 */
7162 static int detach_tasks(struct lb_env *env)
7163 {
7164 struct list_head *tasks = &env->src_rq->cfs_tasks;
7165 struct task_struct *p;
7166 unsigned long load;
7167 int detached = 0;
7169 lockdep_assert_held(&env->src_rq->lock);
7171 if (env->imbalance <= 0)
7172 return 0;
7174 while (!list_empty(tasks)) {
7175 /*
7176 * We don't want to steal all, otherwise we may be treated likewise,
7177 * which could at worst lead to a livelock crash.
7178 */
7179 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7180 break;
7182 p = list_last_entry(tasks, struct task_struct, se.group_node);
7184 env->loop++;
7185 /* We've more or less seen every task there is, call it quits */
7186 if (env->loop > env->loop_max)
7187 break;
7189 /* take a breather every nr_migrate tasks */
7190 if (env->loop > env->loop_break) {
7191 env->loop_break += sched_nr_migrate_break;
7192 env->flags |= LBF_NEED_BREAK;
7193 break;
7194 }
7196 if (!can_migrate_task(p, env))
7197 goto next;
7199 load = task_h_load(p);
7201 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7202 goto next;
7204 if ((load / 2) > env->imbalance)
7205 goto next;
7207 detach_task(p, env);
7208 list_add(&p->se.group_node, &env->tasks);
7210 detached++;
7211 env->imbalance -= load;
7213 #ifdef CONFIG_PREEMPT
7214 /*
7215 * NEWIDLE balancing is a source of latency, so preemptible
7216 * kernels will stop after the first task is detached to minimize
7217 * the critical section.
7218 */
7219 if (env->idle == CPU_NEWLY_IDLE)
7220 break;
7221 #endif
7223 /*
7224 * We only want to steal up to the prescribed amount of
7225 * weighted load.
7226 */
7227 if (env->imbalance <= 0)
7228 break;
7230 continue;
7231 next:
7232 list_move(&p->se.group_node, tasks);
7233 }
7235 /*
7236 * Right now, this is one of only two places we collect this stat
7237 * so we can safely collect detach_one_task() stats here rather
7238 * than inside detach_one_task().
7239 */
7240 schedstat_add(env->sd->lb_gained[env->idle], detached);
7242 return detached;
7243 }
7245 /*
7246 * attach_task() -- attach the task detached by detach_task() to its new rq.
7247 */
7248 static void attach_task(struct rq *rq, struct task_struct *p)
7249 {
7250 lockdep_assert_held(&rq->lock);
7252 BUG_ON(task_rq(p) != rq);
7253 activate_task(rq, p, ENQUEUE_NOCLOCK);
7254 p->on_rq = TASK_ON_RQ_QUEUED;
7255 check_preempt_curr(rq, p, 0);
7256 }
7258 /*
7259 * attach_one_task() -- attaches the task returned from detach_one_task() to
7260 * its new rq.
7261 */
7262 static void attach_one_task(struct rq *rq, struct task_struct *p)
7263 {
7264 struct rq_flags rf;
7266 rq_lock(rq, &rf);
7267 update_rq_clock(rq);
7268 attach_task(rq, p);
7269 rq_unlock(rq, &rf);
7270 }
7272 /*
7273 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7274 * new rq.
7275 */
7276 static void attach_tasks(struct lb_env *env)
7277 {
7278 struct list_head *tasks = &env->tasks;
7279 struct task_struct *p;
7280 struct rq_flags rf;
7282 rq_lock(env->dst_rq, &rf);
7283 update_rq_clock(env->dst_rq);
7285 while (!list_empty(tasks)) {
7286 p = list_first_entry(tasks, struct task_struct, se.group_node);
7287 list_del_init(&p->se.group_node);
7289 attach_task(env->dst_rq, p);
7290 }
7292 rq_unlock(env->dst_rq, &rf);
7293 }
7295 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
7296 {
7297 if (cfs_rq->avg.load_avg)
7298 return true;
7300 if (cfs_rq->avg.util_avg)
7301 return true;
7303 return false;
7304 }
7306 static inline bool others_have_blocked(struct rq *rq)
7307 {
7308 if (READ_ONCE(rq->avg_rt.util_avg))
7309 return true;
7311 if (READ_ONCE(rq->avg_dl.util_avg))
7312 return true;
7314 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
7315 if (READ_ONCE(rq->avg_irq.util_avg))
7316 return true;
7317 #endif
7319 return false;
7320 }
7322 #ifdef CONFIG_FAIR_GROUP_SCHED
7324 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7325 {
7326 if (cfs_rq->load.weight)
7327 return false;
7329 if (cfs_rq->avg.load_sum)
7330 return false;
7332 if (cfs_rq->avg.util_sum)
7333 return false;
7335 if (cfs_rq->avg.runnable_load_sum)
7336 return false;
7338 return true;
7339 }
7341 static void update_blocked_averages(int cpu)
7342 {
7343 struct rq *rq = cpu_rq(cpu);
7344 struct cfs_rq *cfs_rq, *pos;
7345 const struct sched_class *curr_class;
7346 struct rq_flags rf;
7347 bool done = true;
7349 rq_lock_irqsave(rq, &rf);
7350 update_rq_clock(rq);
7352 /*
7353 * Iterates the task_group tree in a bottom up fashion, see
7354 * list_add_leaf_cfs_rq() for details.
7355 */
7356 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7357 struct sched_entity *se;
7359 /* throttled entities do not contribute to load */
7360 if (throttled_hierarchy(cfs_rq))
7361 continue;
7363 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
7364 update_tg_load_avg(cfs_rq, 0);
7366 /* Propagate pending load changes to the parent, if any: */
7367 se = cfs_rq->tg->se[cpu];
7368 if (se && !skip_blocked_update(se))
7369 update_load_avg(cfs_rq_of(se), se, 0);
7371 /*
7372 * There can be a lot of idle CPU cgroups. Don't let fully
7373 * decayed cfs_rqs linger on the list.
7374 */
7375 if (cfs_rq_is_decayed(cfs_rq))
7376 list_del_leaf_cfs_rq(cfs_rq);
7378 /* Don't need periodic decay once load/util_avg are null */
7379 if (cfs_rq_has_blocked(cfs_rq))
7380 done = false;
7381 }
7383 curr_class = rq->curr->sched_class;
7384 update_rt_rq_load_avg(rq_clock_task(rq), rq, curr_class == &rt_sched_class);
7385 update_dl_rq_load_avg(rq_clock_task(rq), rq, curr_class == &dl_sched_class);
7386 update_irq_load_avg(rq, 0);
7387 /* Don't need periodic decay once load/util_avg are null */
7388 if (others_have_blocked(rq))
7389 done = false;
7391 #ifdef CONFIG_NO_HZ_COMMON
7392 rq->last_blocked_load_update_tick = jiffies;
7393 if (done)
7394 rq->has_blocked_load = 0;
7395 #endif
7396 rq_unlock_irqrestore(rq, &rf);
7397 }
7399 /*
7400 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7401 * This needs to be done in a top-down fashion because the load of a child
7402 * group is a fraction of its parents load.
7403 */
7404 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7405 {
7406 struct rq *rq = rq_of(cfs_rq);
7407 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7408 unsigned long now = jiffies;
7409 unsigned long load;
7411 if (cfs_rq->last_h_load_update == now)
7412 return;
7414 cfs_rq->h_load_next = NULL;
7415 for_each_sched_entity(se) {
7416 cfs_rq = cfs_rq_of(se);
7417 cfs_rq->h_load_next = se;
7418 if (cfs_rq->last_h_load_update == now)
7419 break;
7420 }
7422 if (!se) {
7423 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7424 cfs_rq->last_h_load_update = now;
7425 }
7427 while ((se = cfs_rq->h_load_next) != NULL) {
7428 load = cfs_rq->h_load;
7429 load = div64_ul(load * se->avg.load_avg,
7430 cfs_rq_load_avg(cfs_rq) + 1);
7431 cfs_rq = group_cfs_rq(se);
7432 cfs_rq->h_load = load;
7433 cfs_rq->last_h_load_update = now;
7434 }
7435 }
7437 static unsigned long task_h_load(struct task_struct *p)
7438 {
7439 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7441 update_cfs_rq_h_load(cfs_rq);
7442 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7443 cfs_rq_load_avg(cfs_rq) + 1);
7444 }
7445 #else
7446 static inline void update_blocked_averages(int cpu)
7447 {
7448 struct rq *rq = cpu_rq(cpu);
7449 struct cfs_rq *cfs_rq = &rq->cfs;
7450 const struct sched_class *curr_class;
7451 struct rq_flags rf;
7453 rq_lock_irqsave(rq, &rf);
7454 update_rq_clock(rq);
7455 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
7457 curr_class = rq->curr->sched_class;
7458 update_rt_rq_load_avg(rq_clock_task(rq), rq, curr_class == &rt_sched_class);
7459 update_dl_rq_load_avg(rq_clock_task(rq), rq, curr_class == &dl_sched_class);
7460 update_irq_load_avg(rq, 0);
7461 #ifdef CONFIG_NO_HZ_COMMON
7462 rq->last_blocked_load_update_tick = jiffies;
7463 if (!cfs_rq_has_blocked(cfs_rq) && !others_have_blocked(rq))
7464 rq->has_blocked_load = 0;
7465 #endif
7466 rq_unlock_irqrestore(rq, &rf);
7467 }
7469 static unsigned long task_h_load(struct task_struct *p)
7470 {
7471 return p->se.avg.load_avg;
7472 }
7473 #endif
7475 /********** Helpers for find_busiest_group ************************/
7477 enum group_type {
7478 group_other = 0,
7479 group_imbalanced,
7480 group_overloaded,
7481 };
7483 /*
7484 * sg_lb_stats - stats of a sched_group required for load_balancing
7485 */
7486 struct sg_lb_stats {
7487 unsigned long avg_load; /*Avg load across the CPUs of the group */
7488 unsigned long group_load; /* Total load over the CPUs of the group */
7489 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7490 unsigned long load_per_task;
7491 unsigned long group_capacity;
7492 unsigned long group_util; /* Total utilization of the group */
7493 unsigned int sum_nr_running; /* Nr tasks running in the group */
7494 unsigned int idle_cpus;
7495 unsigned int group_weight;
7496 enum group_type group_type;
7497 int group_no_capacity;
7498 #ifdef CONFIG_NUMA_BALANCING
7499 unsigned int nr_numa_running;
7500 unsigned int nr_preferred_running;
7501 #endif
7502 };
7504 /*
7505 * sd_lb_stats - Structure to store the statistics of a sched_domain
7506 * during load balancing.
7507 */
7508 struct sd_lb_stats {
7509 struct sched_group *busiest; /* Busiest group in this sd */
7510 struct sched_group *local; /* Local group in this sd */
7511 unsigned long total_running;
7512 unsigned long total_load; /* Total load of all groups in sd */
7513 unsigned long total_capacity; /* Total capacity of all groups in sd */
7514 unsigned long avg_load; /* Average load across all groups in sd */
7516 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7517 struct sg_lb_stats local_stat; /* Statistics of the local group */
7518 };
7520 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7521 {
7522 /*
7523 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7524 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7525 * We must however clear busiest_stat::avg_load because
7526 * update_sd_pick_busiest() reads this before assignment.
7527 */
7528 *sds = (struct sd_lb_stats){
7529 .busiest = NULL,
7530 .local = NULL,
7531 .total_running = 0UL,
7532 .total_load = 0UL,
7533 .total_capacity = 0UL,
7534 .busiest_stat = {
7535 .avg_load = 0UL,
7536 .sum_nr_running = 0,
7537 .group_type = group_other,
7538 },
7539 };
7540 }
7542 /**
7543 * get_sd_load_idx - Obtain the load index for a given sched domain.
7544 * @sd: The sched_domain whose load_idx is to be obtained.
7545 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7546 *
7547 * Return: The load index.
7548 */
7549 static inline int get_sd_load_idx(struct sched_domain *sd,
7550 enum cpu_idle_type idle)
7551 {
7552 int load_idx;
7554 switch (idle) {
7555 case CPU_NOT_IDLE:
7556 load_idx = sd->busy_idx;
7557 break;
7559 case CPU_NEWLY_IDLE:
7560 load_idx = sd->newidle_idx;
7561 break;
7562 default:
7563 load_idx = sd->idle_idx;
7564 break;
7565 }
7567 return load_idx;
7568 }
7570 static unsigned long scale_rt_capacity(struct sched_domain *sd, int cpu)
7571 {
7572 struct rq *rq = cpu_rq(cpu);
7573 unsigned long max = arch_scale_cpu_capacity(sd, cpu);
7574 unsigned long used, free;
7575 unsigned long irq;
7577 irq = cpu_util_irq(rq);
7579 if (unlikely(irq >= max))
7580 return 1;
7582 used = READ_ONCE(rq->avg_rt.util_avg);
7583 used += READ_ONCE(rq->avg_dl.util_avg);
7585 if (unlikely(used >= max))
7586 return 1;
7588 free = max - used;
7590 return scale_irq_capacity(free, irq, max);
7591 }
7593 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7594 {
7595 unsigned long capacity = scale_rt_capacity(sd, cpu);
7596 struct sched_group *sdg = sd->groups;
7598 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(sd, cpu);
7600 if (!capacity)
7601 capacity = 1;
7603 cpu_rq(cpu)->cpu_capacity = capacity;
7604 sdg->sgc->capacity = capacity;
7605 sdg->sgc->min_capacity = capacity;
7606 }
7608 void update_group_capacity(struct sched_domain *sd, int cpu)
7609 {
7610 struct sched_domain *child = sd->child;
7611 struct sched_group *group, *sdg = sd->groups;
7612 unsigned long capacity, min_capacity;
7613 unsigned long interval;
7615 interval = msecs_to_jiffies(sd->balance_interval);
7616 interval = clamp(interval, 1UL, max_load_balance_interval);
7617 sdg->sgc->next_update = jiffies + interval;
7619 if (!child) {
7620 update_cpu_capacity(sd, cpu);
7621 return;
7622 }
7624 capacity = 0;
7625 min_capacity = ULONG_MAX;
7627 if (child->flags & SD_OVERLAP) {
7628 /*
7629 * SD_OVERLAP domains cannot assume that child groups
7630 * span the current group.
7631 */
7633 for_each_cpu(cpu, sched_group_span(sdg)) {
7634 struct sched_group_capacity *sgc;
7635 struct rq *rq = cpu_rq(cpu);
7637 /*
7638 * build_sched_domains() -> init_sched_groups_capacity()
7639 * gets here before we've attached the domains to the
7640 * runqueues.
7641 *
7642 * Use capacity_of(), which is set irrespective of domains
7643 * in update_cpu_capacity().
7644 *
7645 * This avoids capacity from being 0 and
7646 * causing divide-by-zero issues on boot.
7647 */
7648 if (unlikely(!rq->sd)) {
7649 capacity += capacity_of(cpu);
7650 } else {
7651 sgc = rq->sd->groups->sgc;
7652 capacity += sgc->capacity;
7653 }
7655 min_capacity = min(capacity, min_capacity);
7656 }
7657 } else {
7658 /*
7659 * !SD_OVERLAP domains can assume that child groups
7660 * span the current group.
7661 */
7663 group = child->groups;
7664 do {
7665 struct sched_group_capacity *sgc = group->sgc;
7667 capacity += sgc->capacity;
7668 min_capacity = min(sgc->min_capacity, min_capacity);
7669 group = group->next;
7670 } while (group != child->groups);
7671 }
7673 sdg->sgc->capacity = capacity;
7674 sdg->sgc->min_capacity = min_capacity;
7675 }
7677 /*
7678 * Check whether the capacity of the rq has been noticeably reduced by side
7679 * activity. The imbalance_pct is used for the threshold.
7680 * Return true is the capacity is reduced
7681 */
7682 static inline int
7683 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7684 {
7685 return ((rq->cpu_capacity * sd->imbalance_pct) <
7686 (rq->cpu_capacity_orig * 100));
7687 }
7689 /*
7690 * Group imbalance indicates (and tries to solve) the problem where balancing
7691 * groups is inadequate due to ->cpus_allowed constraints.
7692 *
7693 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
7694 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
7695 * Something like:
7696 *
7697 * { 0 1 2 3 } { 4 5 6 7 }
7698 * * * * *
7699 *
7700 * If we were to balance group-wise we'd place two tasks in the first group and
7701 * two tasks in the second group. Clearly this is undesired as it will overload
7702 * cpu 3 and leave one of the CPUs in the second group unused.
7703 *
7704 * The current solution to this issue is detecting the skew in the first group
7705 * by noticing the lower domain failed to reach balance and had difficulty
7706 * moving tasks due to affinity constraints.
7707 *
7708 * When this is so detected; this group becomes a candidate for busiest; see
7709 * update_sd_pick_busiest(). And calculate_imbalance() and
7710 * find_busiest_group() avoid some of the usual balance conditions to allow it
7711 * to create an effective group imbalance.
7712 *
7713 * This is a somewhat tricky proposition since the next run might not find the
7714 * group imbalance and decide the groups need to be balanced again. A most
7715 * subtle and fragile situation.
7716 */
7718 static inline int sg_imbalanced(struct sched_group *group)
7719 {
7720 return group->sgc->imbalance;
7721 }
7723 /*
7724 * group_has_capacity returns true if the group has spare capacity that could
7725 * be used by some tasks.
7726 * We consider that a group has spare capacity if the * number of task is
7727 * smaller than the number of CPUs or if the utilization is lower than the
7728 * available capacity for CFS tasks.
7729 * For the latter, we use a threshold to stabilize the state, to take into
7730 * account the variance of the tasks' load and to return true if the available
7731 * capacity in meaningful for the load balancer.
7732 * As an example, an available capacity of 1% can appear but it doesn't make
7733 * any benefit for the load balance.
7734 */
7735 static inline bool
7736 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7737 {
7738 if (sgs->sum_nr_running < sgs->group_weight)
7739 return true;
7741 if ((sgs->group_capacity * 100) >
7742 (sgs->group_util * env->sd->imbalance_pct))
7743 return true;
7745 return false;
7746 }
7748 /*
7749 * group_is_overloaded returns true if the group has more tasks than it can
7750 * handle.
7751 * group_is_overloaded is not equals to !group_has_capacity because a group
7752 * with the exact right number of tasks, has no more spare capacity but is not
7753 * overloaded so both group_has_capacity and group_is_overloaded return
7754 * false.
7755 */
7756 static inline bool
7757 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7758 {
7759 if (sgs->sum_nr_running <= sgs->group_weight)
7760 return false;
7762 if ((sgs->group_capacity * 100) <
7763 (sgs->group_util * env->sd->imbalance_pct))
7764 return true;
7766 return false;
7767 }
7769 /*
7770 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7771 * per-CPU capacity than sched_group ref.
7772 */
7773 static inline bool
7774 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7775 {
7776 return sg->sgc->min_capacity * capacity_margin <
7777 ref->sgc->min_capacity * 1024;
7778 }
7780 static inline enum
7781 group_type group_classify(struct sched_group *group,
7782 struct sg_lb_stats *sgs)
7783 {
7784 if (sgs->group_no_capacity)
7785 return group_overloaded;
7787 if (sg_imbalanced(group))
7788 return group_imbalanced;
7790 return group_other;
7791 }
7793 static bool update_nohz_stats(struct rq *rq, bool force)
7794 {
7795 #ifdef CONFIG_NO_HZ_COMMON
7796 unsigned int cpu = rq->cpu;
7798 if (!rq->has_blocked_load)
7799 return false;
7801 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
7802 return false;
7804 if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick))
7805 return true;
7807 update_blocked_averages(cpu);
7809 return rq->has_blocked_load;
7810 #else
7811 return false;
7812 #endif
7813 }
7815 /**
7816 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7817 * @env: The load balancing environment.
7818 * @group: sched_group whose statistics are to be updated.
7819 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7820 * @local_group: Does group contain this_cpu.
7821 * @sgs: variable to hold the statistics for this group.
7822 * @overload: Indicate more than one runnable task for any CPU.
7823 */
7824 static inline void update_sg_lb_stats(struct lb_env *env,
7825 struct sched_group *group, int load_idx,
7826 int local_group, struct sg_lb_stats *sgs,
7827 bool *overload)
7828 {
7829 unsigned long load;
7830 int i, nr_running;
7832 memset(sgs, 0, sizeof(*sgs));
7834 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
7835 struct rq *rq = cpu_rq(i);
7837 if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false))
7838 env->flags |= LBF_NOHZ_AGAIN;
7840 /* Bias balancing toward CPUs of our domain: */
7841 if (local_group)
7842 load = target_load(i, load_idx);
7843 else
7844 load = source_load(i, load_idx);
7846 sgs->group_load += load;
7847 sgs->group_util += cpu_util(i);
7848 sgs->sum_nr_running += rq->cfs.h_nr_running;
7850 nr_running = rq->nr_running;
7851 if (nr_running > 1)
7852 *overload = true;
7854 #ifdef CONFIG_NUMA_BALANCING
7855 sgs->nr_numa_running += rq->nr_numa_running;
7856 sgs->nr_preferred_running += rq->nr_preferred_running;
7857 #endif
7858 sgs->sum_weighted_load += weighted_cpuload(rq);
7859 /*
7860 * No need to call idle_cpu() if nr_running is not 0
7861 */
7862 if (!nr_running && idle_cpu(i))
7863 sgs->idle_cpus++;
7864 }
7866 /* Adjust by relative CPU capacity of the group */
7867 sgs->group_capacity = group->sgc->capacity;
7868 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7870 if (sgs->sum_nr_running)
7871 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7873 sgs->group_weight = group->group_weight;
7875 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7876 sgs->group_type = group_classify(group, sgs);
7877 }
7879 /**
7880 * update_sd_pick_busiest - return 1 on busiest group
7881 * @env: The load balancing environment.
7882 * @sds: sched_domain statistics
7883 * @sg: sched_group candidate to be checked for being the busiest
7884 * @sgs: sched_group statistics
7885 *
7886 * Determine if @sg is a busier group than the previously selected
7887 * busiest group.
7888 *
7889 * Return: %true if @sg is a busier group than the previously selected
7890 * busiest group. %false otherwise.
7891 */
7892 static bool update_sd_pick_busiest(struct lb_env *env,
7893 struct sd_lb_stats *sds,
7894 struct sched_group *sg,
7895 struct sg_lb_stats *sgs)
7896 {
7897 struct sg_lb_stats *busiest = &sds->busiest_stat;
7899 if (sgs->group_type > busiest->group_type)
7900 return true;
7902 if (sgs->group_type < busiest->group_type)
7903 return false;
7905 if (sgs->avg_load <= busiest->avg_load)
7906 return false;
7908 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
7909 goto asym_packing;
7911 /*
7912 * Candidate sg has no more than one task per CPU and
7913 * has higher per-CPU capacity. Migrating tasks to less
7914 * capable CPUs may harm throughput. Maximize throughput,
7915 * power/energy consequences are not considered.
7916 */
7917 if (sgs->sum_nr_running <= sgs->group_weight &&
7918 group_smaller_cpu_capacity(sds->local, sg))
7919 return false;
7921 asym_packing:
7922 /* This is the busiest node in its class. */
7923 if (!(env->sd->flags & SD_ASYM_PACKING))
7924 return true;
7926 /* No ASYM_PACKING if target CPU is already busy */
7927 if (env->idle == CPU_NOT_IDLE)
7928 return true;
7929 /*
7930 * ASYM_PACKING needs to move all the work to the highest
7931 * prority CPUs in the group, therefore mark all groups
7932 * of lower priority than ourself as busy.
7933 */
7934 if (sgs->sum_nr_running &&
7935 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
7936 if (!sds->busiest)
7937 return true;
7939 /* Prefer to move from lowest priority CPU's work */
7940 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
7941 sg->asym_prefer_cpu))
7942 return true;
7943 }
7945 return false;
7946 }
7948 #ifdef CONFIG_NUMA_BALANCING
7949 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7950 {
7951 if (sgs->sum_nr_running > sgs->nr_numa_running)
7952 return regular;
7953 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7954 return remote;
7955 return all;
7956 }
7958 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7959 {
7960 if (rq->nr_running > rq->nr_numa_running)
7961 return regular;
7962 if (rq->nr_running > rq->nr_preferred_running)
7963 return remote;
7964 return all;
7965 }
7966 #else
7967 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7968 {
7969 return all;
7970 }
7972 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7973 {
7974 return regular;
7975 }
7976 #endif /* CONFIG_NUMA_BALANCING */
7978 /**
7979 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7980 * @env: The load balancing environment.
7981 * @sds: variable to hold the statistics for this sched_domain.
7982 */
7983 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7984 {
7985 struct sched_domain *child = env->sd->child;
7986 struct sched_group *sg = env->sd->groups;
7987 struct sg_lb_stats *local = &sds->local_stat;
7988 struct sg_lb_stats tmp_sgs;
7989 int load_idx, prefer_sibling = 0;
7990 bool overload = false;
7992 if (child && child->flags & SD_PREFER_SIBLING)
7993 prefer_sibling = 1;
7995 #ifdef CONFIG_NO_HZ_COMMON
7996 if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked))
7997 env->flags |= LBF_NOHZ_STATS;
7998 #endif
8000 load_idx = get_sd_load_idx(env->sd, env->idle);
8002 do {
8003 struct sg_lb_stats *sgs = &tmp_sgs;
8004 int local_group;
8006 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
8007 if (local_group) {
8008 sds->local = sg;
8009 sgs = local;
8011 if (env->idle != CPU_NEWLY_IDLE ||
8012 time_after_eq(jiffies, sg->sgc->next_update))
8013 update_group_capacity(env->sd, env->dst_cpu);
8014 }
8016 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
8017 &overload);
8019 if (local_group)
8020 goto next_group;
8022 /*
8023 * In case the child domain prefers tasks go to siblings
8024 * first, lower the sg capacity so that we'll try
8025 * and move all the excess tasks away. We lower the capacity
8026 * of a group only if the local group has the capacity to fit
8027 * these excess tasks. The extra check prevents the case where
8028 * you always pull from the heaviest group when it is already
8029 * under-utilized (possible with a large weight task outweighs
8030 * the tasks on the system).
8031 */
8032 if (prefer_sibling && sds->local &&
8033 group_has_capacity(env, local) &&
8034 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
8035 sgs->group_no_capacity = 1;
8036 sgs->group_type = group_classify(sg, sgs);
8037 }
8039 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
8040 sds->busiest = sg;
8041 sds->busiest_stat = *sgs;
8042 }
8044 next_group:
8045 /* Now, start updating sd_lb_stats */
8046 sds->total_running += sgs->sum_nr_running;
8047 sds->total_load += sgs->group_load;
8048 sds->total_capacity += sgs->group_capacity;
8050 sg = sg->next;
8051 } while (sg != env->sd->groups);
8053 #ifdef CONFIG_NO_HZ_COMMON
8054 if ((env->flags & LBF_NOHZ_AGAIN) &&
8055 cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) {
8057 WRITE_ONCE(nohz.next_blocked,
8058 jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD));
8059 }
8060 #endif
8062 if (env->sd->flags & SD_NUMA)
8063 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8065 if (!env->sd->parent) {
8066 /* update overload indicator if we are at root domain */
8067 if (env->dst_rq->rd->overload != overload)
8068 env->dst_rq->rd->overload = overload;
8069 }
8070 }
8072 /**
8073 * check_asym_packing - Check to see if the group is packed into the
8074 * sched domain.
8075 *
8076 * This is primarily intended to used at the sibling level. Some
8077 * cores like POWER7 prefer to use lower numbered SMT threads. In the
8078 * case of POWER7, it can move to lower SMT modes only when higher
8079 * threads are idle. When in lower SMT modes, the threads will
8080 * perform better since they share less core resources. Hence when we
8081 * have idle threads, we want them to be the higher ones.
8082 *
8083 * This packing function is run on idle threads. It checks to see if
8084 * the busiest CPU in this domain (core in the P7 case) has a higher
8085 * CPU number than the packing function is being run on. Here we are
8086 * assuming lower CPU number will be equivalent to lower a SMT thread
8087 * number.
8088 *
8089 * Return: 1 when packing is required and a task should be moved to
8090 * this CPU. The amount of the imbalance is returned in env->imbalance.
8091 *
8092 * @env: The load balancing environment.
8093 * @sds: Statistics of the sched_domain which is to be packed
8094 */
8095 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
8096 {
8097 int busiest_cpu;
8099 if (!(env->sd->flags & SD_ASYM_PACKING))
8100 return 0;
8102 if (env->idle == CPU_NOT_IDLE)
8103 return 0;
8105 if (!sds->busiest)
8106 return 0;
8108 busiest_cpu = sds->busiest->asym_prefer_cpu;
8109 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
8110 return 0;
8112 env->imbalance = DIV_ROUND_CLOSEST(
8113 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
8114 SCHED_CAPACITY_SCALE);
8116 return 1;
8117 }
8119 /**
8120 * fix_small_imbalance - Calculate the minor imbalance that exists
8121 * amongst the groups of a sched_domain, during
8122 * load balancing.
8123 * @env: The load balancing environment.
8124 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
8125 */
8126 static inline
8127 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8128 {
8129 unsigned long tmp, capa_now = 0, capa_move = 0;
8130 unsigned int imbn = 2;
8131 unsigned long scaled_busy_load_per_task;
8132 struct sg_lb_stats *local, *busiest;
8134 local = &sds->local_stat;
8135 busiest = &sds->busiest_stat;
8137 if (!local->sum_nr_running)
8138 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
8139 else if (busiest->load_per_task > local->load_per_task)
8140 imbn = 1;
8142 scaled_busy_load_per_task =
8143 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8144 busiest->group_capacity;
8146 if (busiest->avg_load + scaled_busy_load_per_task >=
8147 local->avg_load + (scaled_busy_load_per_task * imbn)) {
8148 env->imbalance = busiest->load_per_task;
8149 return;
8150 }
8152 /*
8153 * OK, we don't have enough imbalance to justify moving tasks,
8154 * however we may be able to increase total CPU capacity used by
8155 * moving them.
8156 */
8158 capa_now += busiest->group_capacity *
8159 min(busiest->load_per_task, busiest->avg_load);
8160 capa_now += local->group_capacity *
8161 min(local->load_per_task, local->avg_load);
8162 capa_now /= SCHED_CAPACITY_SCALE;
8164 /* Amount of load we'd subtract */
8165 if (busiest->avg_load > scaled_busy_load_per_task) {
8166 capa_move += busiest->group_capacity *
8167 min(busiest->load_per_task,
8168 busiest->avg_load - scaled_busy_load_per_task);
8169 }
8171 /* Amount of load we'd add */
8172 if (busiest->avg_load * busiest->group_capacity <
8173 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
8174 tmp = (busiest->avg_load * busiest->group_capacity) /
8175 local->group_capacity;
8176 } else {
8177 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8178 local->group_capacity;
8179 }
8180 capa_move += local->group_capacity *
8181 min(local->load_per_task, local->avg_load + tmp);
8182 capa_move /= SCHED_CAPACITY_SCALE;
8184 /* Move if we gain throughput */
8185 if (capa_move > capa_now)
8186 env->imbalance = busiest->load_per_task;
8187 }
8189 /**
8190 * calculate_imbalance - Calculate the amount of imbalance present within the
8191 * groups of a given sched_domain during load balance.
8192 * @env: load balance environment
8193 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8194 */
8195 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8196 {
8197 unsigned long max_pull, load_above_capacity = ~0UL;
8198 struct sg_lb_stats *local, *busiest;
8200 local = &sds->local_stat;
8201 busiest = &sds->busiest_stat;
8203 if (busiest->group_type == group_imbalanced) {
8204 /*
8205 * In the group_imb case we cannot rely on group-wide averages
8206 * to ensure CPU-load equilibrium, look at wider averages. XXX
8207 */
8208 busiest->load_per_task =
8209 min(busiest->load_per_task, sds->avg_load);
8210 }
8212 /*
8213 * Avg load of busiest sg can be less and avg load of local sg can
8214 * be greater than avg load across all sgs of sd because avg load
8215 * factors in sg capacity and sgs with smaller group_type are
8216 * skipped when updating the busiest sg:
8217 */
8218 if (busiest->avg_load <= sds->avg_load ||
8219 local->avg_load >= sds->avg_load) {
8220 env->imbalance = 0;
8221 return fix_small_imbalance(env, sds);
8222 }
8224 /*
8225 * If there aren't any idle CPUs, avoid creating some.
8226 */
8227 if (busiest->group_type == group_overloaded &&
8228 local->group_type == group_overloaded) {
8229 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
8230 if (load_above_capacity > busiest->group_capacity) {
8231 load_above_capacity -= busiest->group_capacity;
8232 load_above_capacity *= scale_load_down(NICE_0_LOAD);
8233 load_above_capacity /= busiest->group_capacity;
8234 } else
8235 load_above_capacity = ~0UL;
8236 }
8238 /*
8239 * We're trying to get all the CPUs to the average_load, so we don't
8240 * want to push ourselves above the average load, nor do we wish to
8241 * reduce the max loaded CPU below the average load. At the same time,
8242 * we also don't want to reduce the group load below the group
8243 * capacity. Thus we look for the minimum possible imbalance.
8244 */
8245 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8247 /* How much load to actually move to equalise the imbalance */
8248 env->imbalance = min(
8249 max_pull * busiest->group_capacity,
8250 (sds->avg_load - local->avg_load) * local->group_capacity
8251 ) / SCHED_CAPACITY_SCALE;
8253 /*
8254 * if *imbalance is less than the average load per runnable task
8255 * there is no guarantee that any tasks will be moved so we'll have
8256 * a think about bumping its value to force at least one task to be
8257 * moved
8258 */
8259 if (env->imbalance < busiest->load_per_task)
8260 return fix_small_imbalance(env, sds);
8261 }
8263 /******* find_busiest_group() helpers end here *********************/
8265 /**
8266 * find_busiest_group - Returns the busiest group within the sched_domain
8267 * if there is an imbalance.
8268 *
8269 * Also calculates the amount of weighted load which should be moved
8270 * to restore balance.
8271 *
8272 * @env: The load balancing environment.
8273 *
8274 * Return: - The busiest group if imbalance exists.
8275 */
8276 static struct sched_group *find_busiest_group(struct lb_env *env)
8277 {
8278 struct sg_lb_stats *local, *busiest;
8279 struct sd_lb_stats sds;
8281 init_sd_lb_stats(&sds);
8283 /*
8284 * Compute the various statistics relavent for load balancing at
8285 * this level.
8286 */
8287 update_sd_lb_stats(env, &sds);
8288 local = &sds.local_stat;
8289 busiest = &sds.busiest_stat;
8291 /* ASYM feature bypasses nice load balance check */
8292 if (check_asym_packing(env, &sds))
8293 return sds.busiest;
8295 /* There is no busy sibling group to pull tasks from */
8296 if (!sds.busiest || busiest->sum_nr_running == 0)
8297 goto out_balanced;
8299 /* XXX broken for overlapping NUMA groups */
8300 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8301 / sds.total_capacity;
8303 /*
8304 * If the busiest group is imbalanced the below checks don't
8305 * work because they assume all things are equal, which typically
8306 * isn't true due to cpus_allowed constraints and the like.
8307 */
8308 if (busiest->group_type == group_imbalanced)
8309 goto force_balance;
8311 /*
8312 * When dst_cpu is idle, prevent SMP nice and/or asymmetric group
8313 * capacities from resulting in underutilization due to avg_load.
8314 */
8315 if (env->idle != CPU_NOT_IDLE && group_has_capacity(env, local) &&
8316 busiest->group_no_capacity)
8317 goto force_balance;
8319 /*
8320 * If the local group is busier than the selected busiest group
8321 * don't try and pull any tasks.
8322 */
8323 if (local->avg_load >= busiest->avg_load)
8324 goto out_balanced;
8326 /*
8327 * Don't pull any tasks if this group is already above the domain
8328 * average load.
8329 */
8330 if (local->avg_load >= sds.avg_load)
8331 goto out_balanced;
8333 if (env->idle == CPU_IDLE) {
8334 /*
8335 * This CPU is idle. If the busiest group is not overloaded
8336 * and there is no imbalance between this and busiest group
8337 * wrt idle CPUs, it is balanced. The imbalance becomes
8338 * significant if the diff is greater than 1 otherwise we
8339 * might end up to just move the imbalance on another group
8340 */
8341 if ((busiest->group_type != group_overloaded) &&
8342 (local->idle_cpus <= (busiest->idle_cpus + 1)))
8343 goto out_balanced;
8344 } else {
8345 /*
8346 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8347 * imbalance_pct to be conservative.
8348 */
8349 if (100 * busiest->avg_load <=
8350 env->sd->imbalance_pct * local->avg_load)
8351 goto out_balanced;
8352 }
8354 force_balance:
8355 /* Looks like there is an imbalance. Compute it */
8356 calculate_imbalance(env, &sds);
8357 return env->imbalance ? sds.busiest : NULL;
8359 out_balanced:
8360 env->imbalance = 0;
8361 return NULL;
8362 }
8364 /*
8365 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
8366 */
8367 static struct rq *find_busiest_queue(struct lb_env *env,
8368 struct sched_group *group)
8369 {
8370 struct rq *busiest = NULL, *rq;
8371 unsigned long busiest_load = 0, busiest_capacity = 1;
8372 int i;
8374 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8375 unsigned long capacity, wl;
8376 enum fbq_type rt;
8378 rq = cpu_rq(i);
8379 rt = fbq_classify_rq(rq);
8381 /*
8382 * We classify groups/runqueues into three groups:
8383 * - regular: there are !numa tasks
8384 * - remote: there are numa tasks that run on the 'wrong' node
8385 * - all: there is no distinction
8386 *
8387 * In order to avoid migrating ideally placed numa tasks,
8388 * ignore those when there's better options.
8389 *
8390 * If we ignore the actual busiest queue to migrate another
8391 * task, the next balance pass can still reduce the busiest
8392 * queue by moving tasks around inside the node.
8393 *
8394 * If we cannot move enough load due to this classification
8395 * the next pass will adjust the group classification and
8396 * allow migration of more tasks.
8397 *
8398 * Both cases only affect the total convergence complexity.
8399 */
8400 if (rt > env->fbq_type)
8401 continue;
8403 capacity = capacity_of(i);
8405 wl = weighted_cpuload(rq);
8407 /*
8408 * When comparing with imbalance, use weighted_cpuload()
8409 * which is not scaled with the CPU capacity.
8410 */
8412 if (rq->nr_running == 1 && wl > env->imbalance &&
8413 !check_cpu_capacity(rq, env->sd))
8414 continue;
8416 /*
8417 * For the load comparisons with the other CPU's, consider
8418 * the weighted_cpuload() scaled with the CPU capacity, so
8419 * that the load can be moved away from the CPU that is
8420 * potentially running at a lower capacity.
8421 *
8422 * Thus we're looking for max(wl_i / capacity_i), crosswise
8423 * multiplication to rid ourselves of the division works out
8424 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8425 * our previous maximum.
8426 */
8427 if (wl * busiest_capacity > busiest_load * capacity) {
8428 busiest_load = wl;
8429 busiest_capacity = capacity;
8430 busiest = rq;
8431 }
8432 }
8434 return busiest;
8435 }
8437 /*
8438 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8439 * so long as it is large enough.
8440 */
8441 #define MAX_PINNED_INTERVAL 512
8443 static int need_active_balance(struct lb_env *env)
8444 {
8445 struct sched_domain *sd = env->sd;
8447 if (env->idle == CPU_NEWLY_IDLE) {
8449 /*
8450 * ASYM_PACKING needs to force migrate tasks from busy but
8451 * lower priority CPUs in order to pack all tasks in the
8452 * highest priority CPUs.
8453 */
8454 if ((sd->flags & SD_ASYM_PACKING) &&
8455 sched_asym_prefer(env->dst_cpu, env->src_cpu))
8456 return 1;
8457 }
8459 /*
8460 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8461 * It's worth migrating the task if the src_cpu's capacity is reduced
8462 * because of other sched_class or IRQs if more capacity stays
8463 * available on dst_cpu.
8464 */
8465 if ((env->idle != CPU_NOT_IDLE) &&
8466 (env->src_rq->cfs.h_nr_running == 1)) {
8467 if ((check_cpu_capacity(env->src_rq, sd)) &&
8468 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8469 return 1;
8470 }
8472 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8473 }
8475 static int active_load_balance_cpu_stop(void *data);
8477 static int should_we_balance(struct lb_env *env)
8478 {
8479 struct sched_group *sg = env->sd->groups;
8480 int cpu, balance_cpu = -1;
8482 /*
8483 * Ensure the balancing environment is consistent; can happen
8484 * when the softirq triggers 'during' hotplug.
8485 */
8486 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
8487 return 0;
8489 /*
8490 * In the newly idle case, we will allow all the CPUs
8491 * to do the newly idle load balance.
8492 */
8493 if (env->idle == CPU_NEWLY_IDLE)
8494 return 1;
8496 /* Try to find first idle CPU */
8497 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8498 if (!idle_cpu(cpu))
8499 continue;
8501 balance_cpu = cpu;
8502 break;
8503 }
8505 if (balance_cpu == -1)
8506 balance_cpu = group_balance_cpu(sg);
8508 /*
8509 * First idle CPU or the first CPU(busiest) in this sched group
8510 * is eligible for doing load balancing at this and above domains.
8511 */
8512 return balance_cpu == env->dst_cpu;
8513 }
8515 /*
8516 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8517 * tasks if there is an imbalance.
8518 */
8519 static int load_balance(int this_cpu, struct rq *this_rq,
8520 struct sched_domain *sd, enum cpu_idle_type idle,
8521 int *continue_balancing)
8522 {
8523 int ld_moved, cur_ld_moved, active_balance = 0;
8524 struct sched_domain *sd_parent = sd->parent;
8525 struct sched_group *group;
8526 struct rq *busiest;
8527 struct rq_flags rf;
8528 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8530 struct lb_env env = {
8531 .sd = sd,
8532 .dst_cpu = this_cpu,
8533 .dst_rq = this_rq,
8534 .dst_grpmask = sched_group_span(sd->groups),
8535 .idle = idle,
8536 .loop_break = sched_nr_migrate_break,
8537 .cpus = cpus,
8538 .fbq_type = all,
8539 .tasks = LIST_HEAD_INIT(env.tasks),
8540 };
8542 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8544 schedstat_inc(sd->lb_count[idle]);
8546 redo:
8547 if (!should_we_balance(&env)) {
8548 *continue_balancing = 0;
8549 goto out_balanced;
8550 }
8552 group = find_busiest_group(&env);
8553 if (!group) {
8554 schedstat_inc(sd->lb_nobusyg[idle]);
8555 goto out_balanced;
8556 }
8558 busiest = find_busiest_queue(&env, group);
8559 if (!busiest) {
8560 schedstat_inc(sd->lb_nobusyq[idle]);
8561 goto out_balanced;
8562 }
8564 BUG_ON(busiest == env.dst_rq);
8566 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8568 env.src_cpu = busiest->cpu;
8569 env.src_rq = busiest;
8571 ld_moved = 0;
8572 if (busiest->nr_running > 1) {
8573 /*
8574 * Attempt to move tasks. If find_busiest_group has found
8575 * an imbalance but busiest->nr_running <= 1, the group is
8576 * still unbalanced. ld_moved simply stays zero, so it is
8577 * correctly treated as an imbalance.
8578 */
8579 env.flags |= LBF_ALL_PINNED;
8580 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8582 more_balance:
8583 rq_lock_irqsave(busiest, &rf);
8584 update_rq_clock(busiest);
8586 /*
8587 * cur_ld_moved - load moved in current iteration
8588 * ld_moved - cumulative load moved across iterations
8589 */
8590 cur_ld_moved = detach_tasks(&env);
8592 /*
8593 * We've detached some tasks from busiest_rq. Every
8594 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8595 * unlock busiest->lock, and we are able to be sure
8596 * that nobody can manipulate the tasks in parallel.
8597 * See task_rq_lock() family for the details.
8598 */
8600 rq_unlock(busiest, &rf);
8602 if (cur_ld_moved) {
8603 attach_tasks(&env);
8604 ld_moved += cur_ld_moved;
8605 }
8607 local_irq_restore(rf.flags);
8609 if (env.flags & LBF_NEED_BREAK) {
8610 env.flags &= ~LBF_NEED_BREAK;
8611 goto more_balance;
8612 }
8614 /*
8615 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8616 * us and move them to an alternate dst_cpu in our sched_group
8617 * where they can run. The upper limit on how many times we
8618 * iterate on same src_cpu is dependent on number of CPUs in our
8619 * sched_group.
8620 *
8621 * This changes load balance semantics a bit on who can move
8622 * load to a given_cpu. In addition to the given_cpu itself
8623 * (or a ilb_cpu acting on its behalf where given_cpu is
8624 * nohz-idle), we now have balance_cpu in a position to move
8625 * load to given_cpu. In rare situations, this may cause
8626 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8627 * _independently_ and at _same_ time to move some load to
8628 * given_cpu) causing exceess load to be moved to given_cpu.
8629 * This however should not happen so much in practice and
8630 * moreover subsequent load balance cycles should correct the
8631 * excess load moved.
8632 */
8633 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8635 /* Prevent to re-select dst_cpu via env's CPUs */
8636 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8638 env.dst_rq = cpu_rq(env.new_dst_cpu);
8639 env.dst_cpu = env.new_dst_cpu;
8640 env.flags &= ~LBF_DST_PINNED;
8641 env.loop = 0;
8642 env.loop_break = sched_nr_migrate_break;
8644 /*
8645 * Go back to "more_balance" rather than "redo" since we
8646 * need to continue with same src_cpu.
8647 */
8648 goto more_balance;
8649 }
8651 /*
8652 * We failed to reach balance because of affinity.
8653 */
8654 if (sd_parent) {
8655 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8657 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8658 *group_imbalance = 1;
8659 }
8661 /* All tasks on this runqueue were pinned by CPU affinity */
8662 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8663 cpumask_clear_cpu(cpu_of(busiest), cpus);
8664 /*
8665 * Attempting to continue load balancing at the current
8666 * sched_domain level only makes sense if there are
8667 * active CPUs remaining as possible busiest CPUs to
8668 * pull load from which are not contained within the
8669 * destination group that is receiving any migrated
8670 * load.
8671 */
8672 if (!cpumask_subset(cpus, env.dst_grpmask)) {
8673 env.loop = 0;
8674 env.loop_break = sched_nr_migrate_break;
8675 goto redo;
8676 }
8677 goto out_all_pinned;
8678 }
8679 }
8681 if (!ld_moved) {
8682 schedstat_inc(sd->lb_failed[idle]);
8683 /*
8684 * Increment the failure counter only on periodic balance.
8685 * We do not want newidle balance, which can be very
8686 * frequent, pollute the failure counter causing
8687 * excessive cache_hot migrations and active balances.
8688 */
8689 if (idle != CPU_NEWLY_IDLE)
8690 sd->nr_balance_failed++;
8692 if (need_active_balance(&env)) {
8693 unsigned long flags;
8695 raw_spin_lock_irqsave(&busiest->lock, flags);
8697 /*
8698 * Don't kick the active_load_balance_cpu_stop,
8699 * if the curr task on busiest CPU can't be
8700 * moved to this_cpu:
8701 */
8702 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
8703 raw_spin_unlock_irqrestore(&busiest->lock,
8704 flags);
8705 env.flags |= LBF_ALL_PINNED;
8706 goto out_one_pinned;
8707 }
8709 /*
8710 * ->active_balance synchronizes accesses to
8711 * ->active_balance_work. Once set, it's cleared
8712 * only after active load balance is finished.
8713 */
8714 if (!busiest->active_balance) {
8715 busiest->active_balance = 1;
8716 busiest->push_cpu = this_cpu;
8717 active_balance = 1;
8718 }
8719 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8721 if (active_balance) {
8722 stop_one_cpu_nowait(cpu_of(busiest),
8723 active_load_balance_cpu_stop, busiest,
8724 &busiest->active_balance_work);
8725 }
8727 /* We've kicked active balancing, force task migration. */
8728 sd->nr_balance_failed = sd->cache_nice_tries+1;
8729 }
8730 } else
8731 sd->nr_balance_failed = 0;
8733 if (likely(!active_balance)) {
8734 /* We were unbalanced, so reset the balancing interval */
8735 sd->balance_interval = sd->min_interval;
8736 } else {
8737 /*
8738 * If we've begun active balancing, start to back off. This
8739 * case may not be covered by the all_pinned logic if there
8740 * is only 1 task on the busy runqueue (because we don't call
8741 * detach_tasks).
8742 */
8743 if (sd->balance_interval < sd->max_interval)
8744 sd->balance_interval *= 2;
8745 }
8747 goto out;
8749 out_balanced:
8750 /*
8751 * We reach balance although we may have faced some affinity
8752 * constraints. Clear the imbalance flag if it was set.
8753 */
8754 if (sd_parent) {
8755 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8757 if (*group_imbalance)
8758 *group_imbalance = 0;
8759 }
8761 out_all_pinned:
8762 /*
8763 * We reach balance because all tasks are pinned at this level so
8764 * we can't migrate them. Let the imbalance flag set so parent level
8765 * can try to migrate them.
8766 */
8767 schedstat_inc(sd->lb_balanced[idle]);
8769 sd->nr_balance_failed = 0;
8771 out_one_pinned:
8772 /* tune up the balancing interval */
8773 if (((env.flags & LBF_ALL_PINNED) &&
8774 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8775 (sd->balance_interval < sd->max_interval))
8776 sd->balance_interval *= 2;
8778 ld_moved = 0;
8779 out:
8780 return ld_moved;
8781 }
8783 static inline unsigned long
8784 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8785 {
8786 unsigned long interval = sd->balance_interval;
8788 if (cpu_busy)
8789 interval *= sd->busy_factor;
8791 /* scale ms to jiffies */
8792 interval = msecs_to_jiffies(interval);
8793 interval = clamp(interval, 1UL, max_load_balance_interval);
8795 return interval;
8796 }
8798 static inline void
8799 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
8800 {
8801 unsigned long interval, next;
8803 /* used by idle balance, so cpu_busy = 0 */
8804 interval = get_sd_balance_interval(sd, 0);
8805 next = sd->last_balance + interval;
8807 if (time_after(*next_balance, next))
8808 *next_balance = next;
8809 }
8811 /*
8812 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
8813 * running tasks off the busiest CPU onto idle CPUs. It requires at
8814 * least 1 task to be running on each physical CPU where possible, and
8815 * avoids physical / logical imbalances.
8816 */
8817 static int active_load_balance_cpu_stop(void *data)
8818 {
8819 struct rq *busiest_rq = data;
8820 int busiest_cpu = cpu_of(busiest_rq);
8821 int target_cpu = busiest_rq->push_cpu;
8822 struct rq *target_rq = cpu_rq(target_cpu);
8823 struct sched_domain *sd;
8824 struct task_struct *p = NULL;
8825 struct rq_flags rf;
8827 rq_lock_irq(busiest_rq, &rf);
8828 /*
8829 * Between queueing the stop-work and running it is a hole in which
8830 * CPUs can become inactive. We should not move tasks from or to
8831 * inactive CPUs.
8832 */
8833 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
8834 goto out_unlock;
8836 /* Make sure the requested CPU hasn't gone down in the meantime: */
8837 if (unlikely(busiest_cpu != smp_processor_id() ||
8838 !busiest_rq->active_balance))
8839 goto out_unlock;
8841 /* Is there any task to move? */
8842 if (busiest_rq->nr_running <= 1)
8843 goto out_unlock;
8845 /*
8846 * This condition is "impossible", if it occurs
8847 * we need to fix it. Originally reported by
8848 * Bjorn Helgaas on a 128-CPU setup.
8849 */
8850 BUG_ON(busiest_rq == target_rq);
8852 /* Search for an sd spanning us and the target CPU. */
8853 rcu_read_lock();
8854 for_each_domain(target_cpu, sd) {
8855 if ((sd->flags & SD_LOAD_BALANCE) &&
8856 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8857 break;
8858 }
8860 if (likely(sd)) {
8861 struct lb_env env = {
8862 .sd = sd,
8863 .dst_cpu = target_cpu,
8864 .dst_rq = target_rq,
8865 .src_cpu = busiest_rq->cpu,
8866 .src_rq = busiest_rq,
8867 .idle = CPU_IDLE,
8868 /*
8869 * can_migrate_task() doesn't need to compute new_dst_cpu
8870 * for active balancing. Since we have CPU_IDLE, but no
8871 * @dst_grpmask we need to make that test go away with lying
8872 * about DST_PINNED.
8873 */
8874 .flags = LBF_DST_PINNED,
8875 };
8877 schedstat_inc(sd->alb_count);
8878 update_rq_clock(busiest_rq);
8880 p = detach_one_task(&env);
8881 if (p) {
8882 schedstat_inc(sd->alb_pushed);
8883 /* Active balancing done, reset the failure counter. */
8884 sd->nr_balance_failed = 0;
8885 } else {
8886 schedstat_inc(sd->alb_failed);
8887 }
8888 }
8889 rcu_read_unlock();
8890 out_unlock:
8891 busiest_rq->active_balance = 0;
8892 rq_unlock(busiest_rq, &rf);
8894 if (p)
8895 attach_one_task(target_rq, p);
8897 local_irq_enable();
8899 return 0;
8900 }
8902 static DEFINE_SPINLOCK(balancing);
8904 /*
8905 * Scale the max load_balance interval with the number of CPUs in the system.
8906 * This trades load-balance latency on larger machines for less cross talk.
8907 */
8908 void update_max_interval(void)
8909 {
8910 max_load_balance_interval = HZ*num_online_cpus()/10;
8911 }
8913 /*
8914 * It checks each scheduling domain to see if it is due to be balanced,
8915 * and initiates a balancing operation if so.
8916 *
8917 * Balancing parameters are set up in init_sched_domains.
8918 */
8919 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8920 {
8921 int continue_balancing = 1;
8922 int cpu = rq->cpu;
8923 unsigned long interval;
8924 struct sched_domain *sd;
8925 /* Earliest time when we have to do rebalance again */
8926 unsigned long next_balance = jiffies + 60*HZ;
8927 int update_next_balance = 0;
8928 int need_serialize, need_decay = 0;
8929 u64 max_cost = 0;
8931 rcu_read_lock();
8932 for_each_domain(cpu, sd) {
8933 /*
8934 * Decay the newidle max times here because this is a regular
8935 * visit to all the domains. Decay ~1% per second.
8936 */
8937 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8938 sd->max_newidle_lb_cost =
8939 (sd->max_newidle_lb_cost * 253) / 256;
8940 sd->next_decay_max_lb_cost = jiffies + HZ;
8941 need_decay = 1;
8942 }
8943 max_cost += sd->max_newidle_lb_cost;
8945 if (!(sd->flags & SD_LOAD_BALANCE))
8946 continue;
8948 /*
8949 * Stop the load balance at this level. There is another
8950 * CPU in our sched group which is doing load balancing more
8951 * actively.
8952 */
8953 if (!continue_balancing) {
8954 if (need_decay)
8955 continue;
8956 break;
8957 }
8959 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8961 need_serialize = sd->flags & SD_SERIALIZE;
8962 if (need_serialize) {
8963 if (!spin_trylock(&balancing))
8964 goto out;
8965 }
8967 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8968 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8969 /*
8970 * The LBF_DST_PINNED logic could have changed
8971 * env->dst_cpu, so we can't know our idle
8972 * state even if we migrated tasks. Update it.
8973 */
8974 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8975 }
8976 sd->last_balance = jiffies;
8977 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8978 }
8979 if (need_serialize)
8980 spin_unlock(&balancing);
8981 out:
8982 if (time_after(next_balance, sd->last_balance + interval)) {
8983 next_balance = sd->last_balance + interval;
8984 update_next_balance = 1;
8985 }
8986 }
8987 if (need_decay) {
8988 /*
8989 * Ensure the rq-wide value also decays but keep it at a
8990 * reasonable floor to avoid funnies with rq->avg_idle.
8991 */
8992 rq->max_idle_balance_cost =
8993 max((u64)sysctl_sched_migration_cost, max_cost);
8994 }
8995 rcu_read_unlock();
8997 /*
8998 * next_balance will be updated only when there is a need.
8999 * When the cpu is attached to null domain for ex, it will not be
9000 * updated.
9001 */
9002 if (likely(update_next_balance)) {
9003 rq->next_balance = next_balance;
9005 #ifdef CONFIG_NO_HZ_COMMON
9006 /*
9007 * If this CPU has been elected to perform the nohz idle
9008 * balance. Other idle CPUs have already rebalanced with
9009 * nohz_idle_balance() and nohz.next_balance has been
9010 * updated accordingly. This CPU is now running the idle load
9011 * balance for itself and we need to update the
9012 * nohz.next_balance accordingly.
9013 */
9014 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9015 nohz.next_balance = rq->next_balance;
9016 #endif
9017 }
9018 }
9020 static inline int on_null_domain(struct rq *rq)
9021 {
9022 return unlikely(!rcu_dereference_sched(rq->sd));
9023 }
9025 #ifdef CONFIG_NO_HZ_COMMON
9026 /*
9027 * idle load balancing details
9028 * - When one of the busy CPUs notice that there may be an idle rebalancing
9029 * needed, they will kick the idle load balancer, which then does idle
9030 * load balancing for all the idle CPUs.
9031 */
9033 static inline int find_new_ilb(void)
9034 {
9035 int ilb = cpumask_first(nohz.idle_cpus_mask);
9037 if (ilb < nr_cpu_ids && idle_cpu(ilb))
9038 return ilb;
9040 return nr_cpu_ids;
9041 }
9043 /*
9044 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
9045 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
9046 * CPU (if there is one).
9047 */
9048 static void kick_ilb(unsigned int flags)
9049 {
9050 int ilb_cpu;
9052 nohz.next_balance++;
9054 ilb_cpu = find_new_ilb();
9056 if (ilb_cpu >= nr_cpu_ids)
9057 return;
9059 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
9060 if (flags & NOHZ_KICK_MASK)
9061 return;
9063 /*
9064 * Use smp_send_reschedule() instead of resched_cpu().
9065 * This way we generate a sched IPI on the target CPU which
9066 * is idle. And the softirq performing nohz idle load balance
9067 * will be run before returning from the IPI.
9068 */
9069 smp_send_reschedule(ilb_cpu);
9070 }
9072 /*
9073 * Current heuristic for kicking the idle load balancer in the presence
9074 * of an idle cpu in the system.
9075 * - This rq has more than one task.
9076 * - This rq has at least one CFS task and the capacity of the CPU is
9077 * significantly reduced because of RT tasks or IRQs.
9078 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
9079 * multiple busy cpu.
9080 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
9081 * domain span are idle.
9082 */
9083 static void nohz_balancer_kick(struct rq *rq)
9084 {
9085 unsigned long now = jiffies;
9086 struct sched_domain_shared *sds;
9087 struct sched_domain *sd;
9088 int nr_busy, i, cpu = rq->cpu;
9089 unsigned int flags = 0;
9091 if (unlikely(rq->idle_balance))
9092 return;
9094 /*
9095 * We may be recently in ticked or tickless idle mode. At the first
9096 * busy tick after returning from idle, we will update the busy stats.
9097 */
9098 nohz_balance_exit_idle(rq);
9100 /*
9101 * None are in tickless mode and hence no need for NOHZ idle load
9102 * balancing.
9103 */
9104 if (likely(!atomic_read(&nohz.nr_cpus)))
9105 return;
9107 if (READ_ONCE(nohz.has_blocked) &&
9108 time_after(now, READ_ONCE(nohz.next_blocked)))
9109 flags = NOHZ_STATS_KICK;
9111 if (time_before(now, nohz.next_balance))
9112 goto out;
9114 if (rq->nr_running >= 2) {
9115 flags = NOHZ_KICK_MASK;
9116 goto out;
9117 }
9119 rcu_read_lock();
9120 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
9121 if (sds) {
9122 /*
9123 * XXX: write a coherent comment on why we do this.
9124 * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com
9125 */
9126 nr_busy = atomic_read(&sds->nr_busy_cpus);
9127 if (nr_busy > 1) {
9128 flags = NOHZ_KICK_MASK;
9129 goto unlock;
9130 }
9132 }
9134 sd = rcu_dereference(rq->sd);
9135 if (sd) {
9136 if ((rq->cfs.h_nr_running >= 1) &&
9137 check_cpu_capacity(rq, sd)) {
9138 flags = NOHZ_KICK_MASK;
9139 goto unlock;
9140 }
9141 }
9143 sd = rcu_dereference(per_cpu(sd_asym, cpu));
9144 if (sd) {
9145 for_each_cpu(i, sched_domain_span(sd)) {
9146 if (i == cpu ||
9147 !cpumask_test_cpu(i, nohz.idle_cpus_mask))
9148 continue;
9150 if (sched_asym_prefer(i, cpu)) {
9151 flags = NOHZ_KICK_MASK;
9152 goto unlock;
9153 }
9154 }
9155 }
9156 unlock:
9157 rcu_read_unlock();
9158 out:
9159 if (flags)
9160 kick_ilb(flags);
9161 }
9163 static void set_cpu_sd_state_busy(int cpu)
9164 {
9165 struct sched_domain *sd;
9167 rcu_read_lock();
9168 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9170 if (!sd || !sd->nohz_idle)
9171 goto unlock;
9172 sd->nohz_idle = 0;
9174 atomic_inc(&sd->shared->nr_busy_cpus);
9175 unlock:
9176 rcu_read_unlock();
9177 }
9179 void nohz_balance_exit_idle(struct rq *rq)
9180 {
9181 SCHED_WARN_ON(rq != this_rq());
9183 if (likely(!rq->nohz_tick_stopped))
9184 return;
9186 rq->nohz_tick_stopped = 0;
9187 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
9188 atomic_dec(&nohz.nr_cpus);
9190 set_cpu_sd_state_busy(rq->cpu);
9191 }
9193 static void set_cpu_sd_state_idle(int cpu)
9194 {
9195 struct sched_domain *sd;
9197 rcu_read_lock();
9198 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9200 if (!sd || sd->nohz_idle)
9201 goto unlock;
9202 sd->nohz_idle = 1;
9204 atomic_dec(&sd->shared->nr_busy_cpus);
9205 unlock:
9206 rcu_read_unlock();
9207 }
9209 /*
9210 * This routine will record that the CPU is going idle with tick stopped.
9211 * This info will be used in performing idle load balancing in the future.
9212 */
9213 void nohz_balance_enter_idle(int cpu)
9214 {
9215 struct rq *rq = cpu_rq(cpu);
9217 SCHED_WARN_ON(cpu != smp_processor_id());
9219 /* If this CPU is going down, then nothing needs to be done: */
9220 if (!cpu_active(cpu))
9221 return;
9223 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
9224 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
9225 return;
9227 /*
9228 * Can be set safely without rq->lock held
9229 * If a clear happens, it will have evaluated last additions because
9230 * rq->lock is held during the check and the clear
9231 */
9232 rq->has_blocked_load = 1;
9234 /*
9235 * The tick is still stopped but load could have been added in the
9236 * meantime. We set the nohz.has_blocked flag to trig a check of the
9237 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
9238 * of nohz.has_blocked can only happen after checking the new load
9239 */
9240 if (rq->nohz_tick_stopped)
9241 goto out;
9243 /* If we're a completely isolated CPU, we don't play: */
9244 if (on_null_domain(rq))
9245 return;
9247 rq->nohz_tick_stopped = 1;
9249 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9250 atomic_inc(&nohz.nr_cpus);
9252 /*
9253 * Ensures that if nohz_idle_balance() fails to observe our
9254 * @idle_cpus_mask store, it must observe the @has_blocked
9255 * store.
9256 */
9257 smp_mb__after_atomic();
9259 set_cpu_sd_state_idle(cpu);
9261 out:
9262 /*
9263 * Each time a cpu enter idle, we assume that it has blocked load and
9264 * enable the periodic update of the load of idle cpus
9265 */
9266 WRITE_ONCE(nohz.has_blocked, 1);
9267 }
9269 /*
9270 * Internal function that runs load balance for all idle cpus. The load balance
9271 * can be a simple update of blocked load or a complete load balance with
9272 * tasks movement depending of flags.
9273 * The function returns false if the loop has stopped before running
9274 * through all idle CPUs.
9275 */
9276 static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
9277 enum cpu_idle_type idle)
9278 {
9279 /* Earliest time when we have to do rebalance again */
9280 unsigned long now = jiffies;
9281 unsigned long next_balance = now + 60*HZ;
9282 bool has_blocked_load = false;
9283 int update_next_balance = 0;
9284 int this_cpu = this_rq->cpu;
9285 int balance_cpu;
9286 int ret = false;
9287 struct rq *rq;
9289 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
9291 /*
9292 * We assume there will be no idle load after this update and clear
9293 * the has_blocked flag. If a cpu enters idle in the mean time, it will
9294 * set the has_blocked flag and trig another update of idle load.
9295 * Because a cpu that becomes idle, is added to idle_cpus_mask before
9296 * setting the flag, we are sure to not clear the state and not
9297 * check the load of an idle cpu.
9298 */
9299 WRITE_ONCE(nohz.has_blocked, 0);
9301 /*
9302 * Ensures that if we miss the CPU, we must see the has_blocked
9303 * store from nohz_balance_enter_idle().
9304 */
9305 smp_mb();
9307 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9308 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9309 continue;
9311 /*
9312 * If this CPU gets work to do, stop the load balancing
9313 * work being done for other CPUs. Next load
9314 * balancing owner will pick it up.
9315 */
9316 if (need_resched()) {
9317 has_blocked_load = true;
9318 goto abort;
9319 }
9321 rq = cpu_rq(balance_cpu);
9323 has_blocked_load |= update_nohz_stats(rq, true);
9325 /*
9326 * If time for next balance is due,
9327 * do the balance.
9328 */
9329 if (time_after_eq(jiffies, rq->next_balance)) {
9330 struct rq_flags rf;
9332 rq_lock_irqsave(rq, &rf);
9333 update_rq_clock(rq);
9334 cpu_load_update_idle(rq);
9335 rq_unlock_irqrestore(rq, &rf);
9337 if (flags & NOHZ_BALANCE_KICK)
9338 rebalance_domains(rq, CPU_IDLE);
9339 }
9341 if (time_after(next_balance, rq->next_balance)) {
9342 next_balance = rq->next_balance;
9343 update_next_balance = 1;
9344 }
9345 }
9347 /* Newly idle CPU doesn't need an update */
9348 if (idle != CPU_NEWLY_IDLE) {
9349 update_blocked_averages(this_cpu);
9350 has_blocked_load |= this_rq->has_blocked_load;
9351 }
9353 if (flags & NOHZ_BALANCE_KICK)
9354 rebalance_domains(this_rq, CPU_IDLE);
9356 WRITE_ONCE(nohz.next_blocked,
9357 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
9359 /* The full idle balance loop has been done */
9360 ret = true;
9362 abort:
9363 /* There is still blocked load, enable periodic update */
9364 if (has_blocked_load)
9365 WRITE_ONCE(nohz.has_blocked, 1);
9367 /*
9368 * next_balance will be updated only when there is a need.
9369 * When the CPU is attached to null domain for ex, it will not be
9370 * updated.
9371 */
9372 if (likely(update_next_balance))
9373 nohz.next_balance = next_balance;
9375 return ret;
9376 }
9378 /*
9379 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9380 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9381 */
9382 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9383 {
9384 int this_cpu = this_rq->cpu;
9385 unsigned int flags;
9387 if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK))
9388 return false;
9390 if (idle != CPU_IDLE) {
9391 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9392 return false;
9393 }
9395 /*
9396 * barrier, pairs with nohz_balance_enter_idle(), ensures ...
9397 */
9398 flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9399 if (!(flags & NOHZ_KICK_MASK))
9400 return false;
9402 _nohz_idle_balance(this_rq, flags, idle);
9404 return true;
9405 }
9407 static void nohz_newidle_balance(struct rq *this_rq)
9408 {
9409 int this_cpu = this_rq->cpu;
9411 /*
9412 * This CPU doesn't want to be disturbed by scheduler
9413 * housekeeping
9414 */
9415 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
9416 return;
9418 /* Will wake up very soon. No time for doing anything else*/
9419 if (this_rq->avg_idle < sysctl_sched_migration_cost)
9420 return;
9422 /* Don't need to update blocked load of idle CPUs*/
9423 if (!READ_ONCE(nohz.has_blocked) ||
9424 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
9425 return;
9427 raw_spin_unlock(&this_rq->lock);
9428 /*
9429 * This CPU is going to be idle and blocked load of idle CPUs
9430 * need to be updated. Run the ilb locally as it is a good
9431 * candidate for ilb instead of waking up another idle CPU.
9432 * Kick an normal ilb if we failed to do the update.
9433 */
9434 if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE))
9435 kick_ilb(NOHZ_STATS_KICK);
9436 raw_spin_lock(&this_rq->lock);
9437 }
9439 #else /* !CONFIG_NO_HZ_COMMON */
9440 static inline void nohz_balancer_kick(struct rq *rq) { }
9442 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9443 {
9444 return false;
9445 }
9447 static inline void nohz_newidle_balance(struct rq *this_rq) { }
9448 #endif /* CONFIG_NO_HZ_COMMON */
9450 /*
9451 * idle_balance is called by schedule() if this_cpu is about to become
9452 * idle. Attempts to pull tasks from other CPUs.
9453 */
9454 static int idle_balance(struct rq *this_rq, struct rq_flags *rf)
9455 {
9456 unsigned long next_balance = jiffies + HZ;
9457 int this_cpu = this_rq->cpu;
9458 struct sched_domain *sd;
9459 int pulled_task = 0;
9460 u64 curr_cost = 0;
9462 /*
9463 * We must set idle_stamp _before_ calling idle_balance(), such that we
9464 * measure the duration of idle_balance() as idle time.
9465 */
9466 this_rq->idle_stamp = rq_clock(this_rq);
9468 /*
9469 * Do not pull tasks towards !active CPUs...
9470 */
9471 if (!cpu_active(this_cpu))
9472 return 0;
9474 /*
9475 * This is OK, because current is on_cpu, which avoids it being picked
9476 * for load-balance and preemption/IRQs are still disabled avoiding
9477 * further scheduler activity on it and we're being very careful to
9478 * re-start the picking loop.
9479 */
9480 rq_unpin_lock(this_rq, rf);
9482 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
9483 !this_rq->rd->overload) {
9485 rcu_read_lock();
9486 sd = rcu_dereference_check_sched_domain(this_rq->sd);
9487 if (sd)
9488 update_next_balance(sd, &next_balance);
9489 rcu_read_unlock();
9491 nohz_newidle_balance(this_rq);
9493 goto out;
9494 }
9496 raw_spin_unlock(&this_rq->lock);
9498 update_blocked_averages(this_cpu);
9499 rcu_read_lock();
9500 for_each_domain(this_cpu, sd) {
9501 int continue_balancing = 1;
9502 u64 t0, domain_cost;
9504 if (!(sd->flags & SD_LOAD_BALANCE))
9505 continue;
9507 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
9508 update_next_balance(sd, &next_balance);
9509 break;
9510 }
9512 if (sd->flags & SD_BALANCE_NEWIDLE) {
9513 t0 = sched_clock_cpu(this_cpu);
9515 pulled_task = load_balance(this_cpu, this_rq,
9516 sd, CPU_NEWLY_IDLE,
9517 &continue_balancing);
9519 domain_cost = sched_clock_cpu(this_cpu) - t0;
9520 if (domain_cost > sd->max_newidle_lb_cost)
9521 sd->max_newidle_lb_cost = domain_cost;
9523 curr_cost += domain_cost;
9524 }
9526 update_next_balance(sd, &next_balance);
9528 /*
9529 * Stop searching for tasks to pull if there are
9530 * now runnable tasks on this rq.
9531 */
9532 if (pulled_task || this_rq->nr_running > 0)
9533 break;
9534 }
9535 rcu_read_unlock();
9537 raw_spin_lock(&this_rq->lock);
9539 if (curr_cost > this_rq->max_idle_balance_cost)
9540 this_rq->max_idle_balance_cost = curr_cost;
9542 out:
9543 /*
9544 * While browsing the domains, we released the rq lock, a task could
9545 * have been enqueued in the meantime. Since we're not going idle,
9546 * pretend we pulled a task.
9547 */
9548 if (this_rq->cfs.h_nr_running && !pulled_task)
9549 pulled_task = 1;
9551 /* Move the next balance forward */
9552 if (time_after(this_rq->next_balance, next_balance))
9553 this_rq->next_balance = next_balance;
9555 /* Is there a task of a high priority class? */
9556 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
9557 pulled_task = -1;
9559 if (pulled_task)
9560 this_rq->idle_stamp = 0;
9562 rq_repin_lock(this_rq, rf);
9564 return pulled_task;
9565 }
9567 /*
9568 * run_rebalance_domains is triggered when needed from the scheduler tick.
9569 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9570 */
9571 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
9572 {
9573 struct rq *this_rq = this_rq();
9574 enum cpu_idle_type idle = this_rq->idle_balance ?
9575 CPU_IDLE : CPU_NOT_IDLE;
9577 /*
9578 * If this CPU has a pending nohz_balance_kick, then do the
9579 * balancing on behalf of the other idle CPUs whose ticks are
9580 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9581 * give the idle CPUs a chance to load balance. Else we may
9582 * load balance only within the local sched_domain hierarchy
9583 * and abort nohz_idle_balance altogether if we pull some load.
9584 */
9585 if (nohz_idle_balance(this_rq, idle))
9586 return;
9588 /* normal load balance */
9589 update_blocked_averages(this_rq->cpu);
9590 rebalance_domains(this_rq, idle);
9591 }
9593 /*
9594 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9595 */
9596 void trigger_load_balance(struct rq *rq)
9597 {
9598 /* Don't need to rebalance while attached to NULL domain */
9599 if (unlikely(on_null_domain(rq)))
9600 return;
9602 if (time_after_eq(jiffies, rq->next_balance))
9603 raise_softirq(SCHED_SOFTIRQ);
9605 nohz_balancer_kick(rq);
9606 }
9608 static void rq_online_fair(struct rq *rq)
9609 {
9610 update_sysctl();
9612 update_runtime_enabled(rq);
9613 }
9615 static void rq_offline_fair(struct rq *rq)
9616 {
9617 update_sysctl();
9619 /* Ensure any throttled groups are reachable by pick_next_task */
9620 unthrottle_offline_cfs_rqs(rq);
9621 }
9623 #endif /* CONFIG_SMP */
9625 /*
9626 * scheduler tick hitting a task of our scheduling class.
9627 *
9628 * NOTE: This function can be called remotely by the tick offload that
9629 * goes along full dynticks. Therefore no local assumption can be made
9630 * and everything must be accessed through the @rq and @curr passed in
9631 * parameters.
9632 */
9633 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9634 {
9635 struct cfs_rq *cfs_rq;
9636 struct sched_entity *se = &curr->se;
9638 for_each_sched_entity(se) {
9639 cfs_rq = cfs_rq_of(se);
9640 entity_tick(cfs_rq, se, queued);
9641 }
9643 if (static_branch_unlikely(&sched_numa_balancing))
9644 task_tick_numa(rq, curr);
9645 }
9647 /*
9648 * called on fork with the child task as argument from the parent's context
9649 * - child not yet on the tasklist
9650 * - preemption disabled
9651 */
9652 static void task_fork_fair(struct task_struct *p)
9653 {
9654 struct cfs_rq *cfs_rq;
9655 struct sched_entity *se = &p->se, *curr;
9656 struct rq *rq = this_rq();
9657 struct rq_flags rf;
9659 rq_lock(rq, &rf);
9660 update_rq_clock(rq);
9662 cfs_rq = task_cfs_rq(current);
9663 curr = cfs_rq->curr;
9664 if (curr) {
9665 update_curr(cfs_rq);
9666 se->vruntime = curr->vruntime;
9667 }
9668 place_entity(cfs_rq, se, 1);
9670 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9671 /*
9672 * Upon rescheduling, sched_class::put_prev_task() will place
9673 * 'current' within the tree based on its new key value.
9674 */
9675 swap(curr->vruntime, se->vruntime);
9676 resched_curr(rq);
9677 }
9679 se->vruntime -= cfs_rq->min_vruntime;
9680 rq_unlock(rq, &rf);
9681 }
9683 /*
9684 * Priority of the task has changed. Check to see if we preempt
9685 * the current task.
9686 */
9687 static void
9688 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9689 {
9690 if (!task_on_rq_queued(p))
9691 return;
9693 /*
9694 * Reschedule if we are currently running on this runqueue and
9695 * our priority decreased, or if we are not currently running on
9696 * this runqueue and our priority is higher than the current's
9697 */
9698 if (rq->curr == p) {
9699 if (p->prio > oldprio)
9700 resched_curr(rq);
9701 } else
9702 check_preempt_curr(rq, p, 0);
9703 }
9705 static inline bool vruntime_normalized(struct task_struct *p)
9706 {
9707 struct sched_entity *se = &p->se;
9709 /*
9710 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9711 * the dequeue_entity(.flags=0) will already have normalized the
9712 * vruntime.
9713 */
9714 if (p->on_rq)
9715 return true;
9717 /*
9718 * When !on_rq, vruntime of the task has usually NOT been normalized.
9719 * But there are some cases where it has already been normalized:
9720 *
9721 * - A forked child which is waiting for being woken up by
9722 * wake_up_new_task().
9723 * - A task which has been woken up by try_to_wake_up() and
9724 * waiting for actually being woken up by sched_ttwu_pending().
9725 */
9726 if (!se->sum_exec_runtime ||
9727 (p->state == TASK_WAKING && p->sched_remote_wakeup))
9728 return true;
9730 return false;
9731 }
9733 #ifdef CONFIG_FAIR_GROUP_SCHED
9734 /*
9735 * Propagate the changes of the sched_entity across the tg tree to make it
9736 * visible to the root
9737 */
9738 static void propagate_entity_cfs_rq(struct sched_entity *se)
9739 {
9740 struct cfs_rq *cfs_rq;
9742 /* Start to propagate at parent */
9743 se = se->parent;
9745 for_each_sched_entity(se) {
9746 cfs_rq = cfs_rq_of(se);
9748 if (cfs_rq_throttled(cfs_rq))
9749 break;
9751 update_load_avg(cfs_rq, se, UPDATE_TG);
9752 }
9753 }
9754 #else
9755 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
9756 #endif
9758 static void detach_entity_cfs_rq(struct sched_entity *se)
9759 {
9760 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9762 /* Catch up with the cfs_rq and remove our load when we leave */
9763 update_load_avg(cfs_rq, se, 0);
9764 detach_entity_load_avg(cfs_rq, se);
9765 update_tg_load_avg(cfs_rq, false);
9766 propagate_entity_cfs_rq(se);
9767 }
9769 static void attach_entity_cfs_rq(struct sched_entity *se)
9770 {
9771 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9773 #ifdef CONFIG_FAIR_GROUP_SCHED
9774 /*
9775 * Since the real-depth could have been changed (only FAIR
9776 * class maintain depth value), reset depth properly.
9777 */
9778 se->depth = se->parent ? se->parent->depth + 1 : 0;
9779 #endif
9781 /* Synchronize entity with its cfs_rq */
9782 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
9783 attach_entity_load_avg(cfs_rq, se, 0);
9784 update_tg_load_avg(cfs_rq, false);
9785 propagate_entity_cfs_rq(se);
9786 }
9788 static void detach_task_cfs_rq(struct task_struct *p)
9789 {
9790 struct sched_entity *se = &p->se;
9791 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9793 if (!vruntime_normalized(p)) {
9794 /*
9795 * Fix up our vruntime so that the current sleep doesn't
9796 * cause 'unlimited' sleep bonus.
9797 */
9798 place_entity(cfs_rq, se, 0);
9799 se->vruntime -= cfs_rq->min_vruntime;
9800 }
9802 detach_entity_cfs_rq(se);
9803 }
9805 static void attach_task_cfs_rq(struct task_struct *p)
9806 {
9807 struct sched_entity *se = &p->se;
9808 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9810 attach_entity_cfs_rq(se);
9812 if (!vruntime_normalized(p))
9813 se->vruntime += cfs_rq->min_vruntime;
9814 }
9816 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9817 {
9818 detach_task_cfs_rq(p);
9819 }
9821 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9822 {
9823 attach_task_cfs_rq(p);
9825 if (task_on_rq_queued(p)) {
9826 /*
9827 * We were most likely switched from sched_rt, so
9828 * kick off the schedule if running, otherwise just see
9829 * if we can still preempt the current task.
9830 */
9831 if (rq->curr == p)
9832 resched_curr(rq);
9833 else
9834 check_preempt_curr(rq, p, 0);
9835 }
9836 }
9838 /* Account for a task changing its policy or group.
9839 *
9840 * This routine is mostly called to set cfs_rq->curr field when a task
9841 * migrates between groups/classes.
9842 */
9843 static void set_curr_task_fair(struct rq *rq)
9844 {
9845 struct sched_entity *se = &rq->curr->se;
9847 for_each_sched_entity(se) {
9848 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9850 set_next_entity(cfs_rq, se);
9851 /* ensure bandwidth has been allocated on our new cfs_rq */
9852 account_cfs_rq_runtime(cfs_rq, 0);
9853 }
9854 }
9856 void init_cfs_rq(struct cfs_rq *cfs_rq)
9857 {
9858 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
9859 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9860 #ifndef CONFIG_64BIT
9861 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9862 #endif
9863 #ifdef CONFIG_SMP
9864 raw_spin_lock_init(&cfs_rq->removed.lock);
9865 #endif
9866 }
9868 #ifdef CONFIG_FAIR_GROUP_SCHED
9869 static void task_set_group_fair(struct task_struct *p)
9870 {
9871 struct sched_entity *se = &p->se;
9873 set_task_rq(p, task_cpu(p));
9874 se->depth = se->parent ? se->parent->depth + 1 : 0;
9875 }
9877 static void task_move_group_fair(struct task_struct *p)
9878 {
9879 detach_task_cfs_rq(p);
9880 set_task_rq(p, task_cpu(p));
9882 #ifdef CONFIG_SMP
9883 /* Tell se's cfs_rq has been changed -- migrated */
9884 p->se.avg.last_update_time = 0;
9885 #endif
9886 attach_task_cfs_rq(p);
9887 }
9889 static void task_change_group_fair(struct task_struct *p, int type)
9890 {
9891 switch (type) {
9892 case TASK_SET_GROUP:
9893 task_set_group_fair(p);
9894 break;
9896 case TASK_MOVE_GROUP:
9897 task_move_group_fair(p);
9898 break;
9899 }
9900 }
9902 void free_fair_sched_group(struct task_group *tg)
9903 {
9904 int i;
9906 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9908 for_each_possible_cpu(i) {
9909 if (tg->cfs_rq)
9910 kfree(tg->cfs_rq[i]);
9911 if (tg->se)
9912 kfree(tg->se[i]);
9913 }
9915 kfree(tg->cfs_rq);
9916 kfree(tg->se);
9917 }
9919 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9920 {
9921 struct sched_entity *se;
9922 struct cfs_rq *cfs_rq;
9923 int i;
9925 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
9926 if (!tg->cfs_rq)
9927 goto err;
9928 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
9929 if (!tg->se)
9930 goto err;
9932 tg->shares = NICE_0_LOAD;
9934 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9936 for_each_possible_cpu(i) {
9937 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9938 GFP_KERNEL, cpu_to_node(i));
9939 if (!cfs_rq)
9940 goto err;
9942 se = kzalloc_node(sizeof(struct sched_entity),
9943 GFP_KERNEL, cpu_to_node(i));
9944 if (!se)
9945 goto err_free_rq;
9947 init_cfs_rq(cfs_rq);
9948 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9949 init_entity_runnable_average(se);
9950 }
9952 return 1;
9954 err_free_rq:
9955 kfree(cfs_rq);
9956 err:
9957 return 0;
9958 }
9960 void online_fair_sched_group(struct task_group *tg)
9961 {
9962 struct sched_entity *se;
9963 struct rq *rq;
9964 int i;
9966 for_each_possible_cpu(i) {
9967 rq = cpu_rq(i);
9968 se = tg->se[i];
9970 raw_spin_lock_irq(&rq->lock);
9971 update_rq_clock(rq);
9972 attach_entity_cfs_rq(se);
9973 sync_throttle(tg, i);
9974 raw_spin_unlock_irq(&rq->lock);
9975 }
9976 }
9978 void unregister_fair_sched_group(struct task_group *tg)
9979 {
9980 unsigned long flags;
9981 struct rq *rq;
9982 int cpu;
9984 for_each_possible_cpu(cpu) {
9985 if (tg->se[cpu])
9986 remove_entity_load_avg(tg->se[cpu]);
9988 /*
9989 * Only empty task groups can be destroyed; so we can speculatively
9990 * check on_list without danger of it being re-added.
9991 */
9992 if (!tg->cfs_rq[cpu]->on_list)
9993 continue;
9995 rq = cpu_rq(cpu);
9997 raw_spin_lock_irqsave(&rq->lock, flags);
9998 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9999 raw_spin_unlock_irqrestore(&rq->lock, flags);
10000 }
10001 }
10003 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
10004 struct sched_entity *se, int cpu,
10005 struct sched_entity *parent)
10006 {
10007 struct rq *rq = cpu_rq(cpu);
10009 cfs_rq->tg = tg;
10010 cfs_rq->rq = rq;
10011 init_cfs_rq_runtime(cfs_rq);
10013 tg->cfs_rq[cpu] = cfs_rq;
10014 tg->se[cpu] = se;
10016 /* se could be NULL for root_task_group */
10017 if (!se)
10018 return;
10020 if (!parent) {
10021 se->cfs_rq = &rq->cfs;
10022 se->depth = 0;
10023 } else {
10024 se->cfs_rq = parent->my_q;
10025 se->depth = parent->depth + 1;
10026 }
10028 se->my_q = cfs_rq;
10029 /* guarantee group entities always have weight */
10030 update_load_set(&se->load, NICE_0_LOAD);
10031 se->parent = parent;
10032 }
10034 static DEFINE_MUTEX(shares_mutex);
10036 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10037 {
10038 int i;
10040 /*
10041 * We can't change the weight of the root cgroup.
10042 */
10043 if (!tg->se[0])
10044 return -EINVAL;
10046 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
10048 mutex_lock(&shares_mutex);
10049 if (tg->shares == shares)
10050 goto done;
10052 tg->shares = shares;
10053 for_each_possible_cpu(i) {
10054 struct rq *rq = cpu_rq(i);
10055 struct sched_entity *se = tg->se[i];
10056 struct rq_flags rf;
10058 /* Propagate contribution to hierarchy */
10059 rq_lock_irqsave(rq, &rf);
10060 update_rq_clock(rq);
10061 for_each_sched_entity(se) {
10062 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
10063 update_cfs_group(se);
10064 }
10065 rq_unlock_irqrestore(rq, &rf);
10066 }
10068 done:
10069 mutex_unlock(&shares_mutex);
10070 return 0;
10071 }
10072 #else /* CONFIG_FAIR_GROUP_SCHED */
10074 void free_fair_sched_group(struct task_group *tg) { }
10076 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10077 {
10078 return 1;
10079 }
10081 void online_fair_sched_group(struct task_group *tg) { }
10083 void unregister_fair_sched_group(struct task_group *tg) { }
10085 #endif /* CONFIG_FAIR_GROUP_SCHED */
10088 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
10089 {
10090 struct sched_entity *se = &task->se;
10091 unsigned int rr_interval = 0;
10093 /*
10094 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
10095 * idle runqueue:
10096 */
10097 if (rq->cfs.load.weight)
10098 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
10100 return rr_interval;
10101 }
10103 /*
10104 * All the scheduling class methods:
10105 */
10106 const struct sched_class fair_sched_class = {
10107 .next = &idle_sched_class,
10108 .enqueue_task = enqueue_task_fair,
10109 .dequeue_task = dequeue_task_fair,
10110 .yield_task = yield_task_fair,
10111 .yield_to_task = yield_to_task_fair,
10113 .check_preempt_curr = check_preempt_wakeup,
10115 .pick_next_task = pick_next_task_fair,
10116 .put_prev_task = put_prev_task_fair,
10118 #ifdef CONFIG_SMP
10119 .select_task_rq = select_task_rq_fair,
10120 .migrate_task_rq = migrate_task_rq_fair,
10122 .rq_online = rq_online_fair,
10123 .rq_offline = rq_offline_fair,
10125 .task_dead = task_dead_fair,
10126 .set_cpus_allowed = set_cpus_allowed_common,
10127 #endif
10129 .set_curr_task = set_curr_task_fair,
10130 .task_tick = task_tick_fair,
10131 .task_fork = task_fork_fair,
10133 .prio_changed = prio_changed_fair,
10134 .switched_from = switched_from_fair,
10135 .switched_to = switched_to_fair,
10137 .get_rr_interval = get_rr_interval_fair,
10139 .update_curr = update_curr_fair,
10141 #ifdef CONFIG_FAIR_GROUP_SCHED
10142 .task_change_group = task_change_group_fair,
10143 #endif
10144 };
10146 #ifdef CONFIG_SCHED_DEBUG
10147 void print_cfs_stats(struct seq_file *m, int cpu)
10148 {
10149 struct cfs_rq *cfs_rq, *pos;
10151 rcu_read_lock();
10152 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
10153 print_cfs_rq(m, cpu, cfs_rq);
10154 rcu_read_unlock();
10155 }
10157 #ifdef CONFIG_NUMA_BALANCING
10158 void show_numa_stats(struct task_struct *p, struct seq_file *m)
10159 {
10160 int node;
10161 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
10163 for_each_online_node(node) {
10164 if (p->numa_faults) {
10165 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
10166 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
10167 }
10168 if (p->numa_group) {
10169 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
10170 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
10171 }
10172 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
10173 }
10174 }
10175 #endif /* CONFIG_NUMA_BALANCING */
10176 #endif /* CONFIG_SCHED_DEBUG */
10178 __init void init_sched_fair_class(void)
10179 {
10180 #ifdef CONFIG_SMP
10181 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
10183 #ifdef CONFIG_NO_HZ_COMMON
10184 nohz.next_balance = jiffies;
10185 nohz.next_blocked = jiffies;
10186 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
10187 #endif
10188 #endif /* SMP */
10190 }