Linux 4.15.6
[linux/fpc-iii.git] / kernel / sched / fair.c
blob26a71ebcd3c2ed94941673a07408a8656ea766c3
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
5 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
7 * Interactivity improvements by Mike Galbraith
8 * (C) 2007 Mike Galbraith <efault@gmx.de>
10 * Various enhancements by Dmitry Adamushko.
11 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
13 * Group scheduling enhancements by Srivatsa Vaddagiri
14 * Copyright IBM Corporation, 2007
15 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
17 * Scaled math optimizations by Thomas Gleixner
18 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
20 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
21 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
24 #include <linux/sched/mm.h>
25 #include <linux/sched/topology.h>
27 #include <linux/latencytop.h>
28 #include <linux/cpumask.h>
29 #include <linux/cpuidle.h>
30 #include <linux/slab.h>
31 #include <linux/profile.h>
32 #include <linux/interrupt.h>
33 #include <linux/mempolicy.h>
34 #include <linux/migrate.h>
35 #include <linux/task_work.h>
36 #include <linux/sched/isolation.h>
38 #include <trace/events/sched.h>
40 #include "sched.h"
43 * Targeted preemption latency for CPU-bound tasks:
45 * NOTE: this latency value is not the same as the concept of
46 * 'timeslice length' - timeslices in CFS are of variable length
47 * and have no persistent notion like in traditional, time-slice
48 * based scheduling concepts.
50 * (to see the precise effective timeslice length of your workload,
51 * run vmstat and monitor the context-switches (cs) field)
53 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
55 unsigned int sysctl_sched_latency = 6000000ULL;
56 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
59 * The initial- and re-scaling of tunables is configurable
61 * Options are:
63 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
64 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
65 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
67 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
69 enum sched_tunable_scaling sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
72 * Minimal preemption granularity for CPU-bound tasks:
74 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
76 unsigned int sysctl_sched_min_granularity = 750000ULL;
77 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
80 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
82 static unsigned int sched_nr_latency = 8;
85 * After fork, child runs first. If set to 0 (default) then
86 * parent will (try to) run first.
88 unsigned int sysctl_sched_child_runs_first __read_mostly;
91 * SCHED_OTHER wake-up granularity.
93 * This option delays the preemption effects of decoupled workloads
94 * and reduces their over-scheduling. Synchronous workloads will still
95 * have immediate wakeup/sleep latencies.
97 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
99 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
100 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
102 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
104 #ifdef CONFIG_SMP
106 * For asym packing, by default the lower numbered cpu has higher priority.
108 int __weak arch_asym_cpu_priority(int cpu)
110 return -cpu;
112 #endif
114 #ifdef CONFIG_CFS_BANDWIDTH
116 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
117 * each time a cfs_rq requests quota.
119 * Note: in the case that the slice exceeds the runtime remaining (either due
120 * to consumption or the quota being specified to be smaller than the slice)
121 * we will always only issue the remaining available time.
123 * (default: 5 msec, units: microseconds)
125 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
126 #endif
129 * The margin used when comparing utilization with CPU capacity:
130 * util * margin < capacity * 1024
132 * (default: ~20%)
134 unsigned int capacity_margin = 1280;
136 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
138 lw->weight += inc;
139 lw->inv_weight = 0;
142 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
144 lw->weight -= dec;
145 lw->inv_weight = 0;
148 static inline void update_load_set(struct load_weight *lw, unsigned long w)
150 lw->weight = w;
151 lw->inv_weight = 0;
155 * Increase the granularity value when there are more CPUs,
156 * because with more CPUs the 'effective latency' as visible
157 * to users decreases. But the relationship is not linear,
158 * so pick a second-best guess by going with the log2 of the
159 * number of CPUs.
161 * This idea comes from the SD scheduler of Con Kolivas:
163 static unsigned int get_update_sysctl_factor(void)
165 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
166 unsigned int factor;
168 switch (sysctl_sched_tunable_scaling) {
169 case SCHED_TUNABLESCALING_NONE:
170 factor = 1;
171 break;
172 case SCHED_TUNABLESCALING_LINEAR:
173 factor = cpus;
174 break;
175 case SCHED_TUNABLESCALING_LOG:
176 default:
177 factor = 1 + ilog2(cpus);
178 break;
181 return factor;
184 static void update_sysctl(void)
186 unsigned int factor = get_update_sysctl_factor();
188 #define SET_SYSCTL(name) \
189 (sysctl_##name = (factor) * normalized_sysctl_##name)
190 SET_SYSCTL(sched_min_granularity);
191 SET_SYSCTL(sched_latency);
192 SET_SYSCTL(sched_wakeup_granularity);
193 #undef SET_SYSCTL
196 void sched_init_granularity(void)
198 update_sysctl();
201 #define WMULT_CONST (~0U)
202 #define WMULT_SHIFT 32
204 static void __update_inv_weight(struct load_weight *lw)
206 unsigned long w;
208 if (likely(lw->inv_weight))
209 return;
211 w = scale_load_down(lw->weight);
213 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
214 lw->inv_weight = 1;
215 else if (unlikely(!w))
216 lw->inv_weight = WMULT_CONST;
217 else
218 lw->inv_weight = WMULT_CONST / w;
222 * delta_exec * weight / lw.weight
223 * OR
224 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
226 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
227 * we're guaranteed shift stays positive because inv_weight is guaranteed to
228 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
230 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
231 * weight/lw.weight <= 1, and therefore our shift will also be positive.
233 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
235 u64 fact = scale_load_down(weight);
236 int shift = WMULT_SHIFT;
238 __update_inv_weight(lw);
240 if (unlikely(fact >> 32)) {
241 while (fact >> 32) {
242 fact >>= 1;
243 shift--;
247 /* hint to use a 32x32->64 mul */
248 fact = (u64)(u32)fact * lw->inv_weight;
250 while (fact >> 32) {
251 fact >>= 1;
252 shift--;
255 return mul_u64_u32_shr(delta_exec, fact, shift);
259 const struct sched_class fair_sched_class;
261 /**************************************************************
262 * CFS operations on generic schedulable entities:
265 #ifdef CONFIG_FAIR_GROUP_SCHED
267 /* cpu runqueue to which this cfs_rq is attached */
268 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
270 return cfs_rq->rq;
273 /* An entity is a task if it doesn't "own" a runqueue */
274 #define entity_is_task(se) (!se->my_q)
276 static inline struct task_struct *task_of(struct sched_entity *se)
278 SCHED_WARN_ON(!entity_is_task(se));
279 return container_of(se, struct task_struct, se);
282 /* Walk up scheduling entities hierarchy */
283 #define for_each_sched_entity(se) \
284 for (; se; se = se->parent)
286 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
288 return p->se.cfs_rq;
291 /* runqueue on which this entity is (to be) queued */
292 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
294 return se->cfs_rq;
297 /* runqueue "owned" by this group */
298 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
300 return grp->my_q;
303 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
305 if (!cfs_rq->on_list) {
306 struct rq *rq = rq_of(cfs_rq);
307 int cpu = cpu_of(rq);
309 * Ensure we either appear before our parent (if already
310 * enqueued) or force our parent to appear after us when it is
311 * enqueued. The fact that we always enqueue bottom-up
312 * reduces this to two cases and a special case for the root
313 * cfs_rq. Furthermore, it also means that we will always reset
314 * tmp_alone_branch either when the branch is connected
315 * to a tree or when we reach the beg of the tree
317 if (cfs_rq->tg->parent &&
318 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
320 * If parent is already on the list, we add the child
321 * just before. Thanks to circular linked property of
322 * the list, this means to put the child at the tail
323 * of the list that starts by parent.
325 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
326 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
328 * The branch is now connected to its tree so we can
329 * reset tmp_alone_branch to the beginning of the
330 * list.
332 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
333 } else if (!cfs_rq->tg->parent) {
335 * cfs rq without parent should be put
336 * at the tail of the list.
338 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
339 &rq->leaf_cfs_rq_list);
341 * We have reach the beg of a tree so we can reset
342 * tmp_alone_branch to the beginning of the list.
344 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
345 } else {
347 * The parent has not already been added so we want to
348 * make sure that it will be put after us.
349 * tmp_alone_branch points to the beg of the branch
350 * where we will add parent.
352 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
353 rq->tmp_alone_branch);
355 * update tmp_alone_branch to points to the new beg
356 * of the branch
358 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
361 cfs_rq->on_list = 1;
365 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
367 if (cfs_rq->on_list) {
368 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
369 cfs_rq->on_list = 0;
373 /* Iterate thr' all leaf cfs_rq's on a runqueue */
374 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
375 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
376 leaf_cfs_rq_list)
378 /* Do the two (enqueued) entities belong to the same group ? */
379 static inline struct cfs_rq *
380 is_same_group(struct sched_entity *se, struct sched_entity *pse)
382 if (se->cfs_rq == pse->cfs_rq)
383 return se->cfs_rq;
385 return NULL;
388 static inline struct sched_entity *parent_entity(struct sched_entity *se)
390 return se->parent;
393 static void
394 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
396 int se_depth, pse_depth;
399 * preemption test can be made between sibling entities who are in the
400 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
401 * both tasks until we find their ancestors who are siblings of common
402 * parent.
405 /* First walk up until both entities are at same depth */
406 se_depth = (*se)->depth;
407 pse_depth = (*pse)->depth;
409 while (se_depth > pse_depth) {
410 se_depth--;
411 *se = parent_entity(*se);
414 while (pse_depth > se_depth) {
415 pse_depth--;
416 *pse = parent_entity(*pse);
419 while (!is_same_group(*se, *pse)) {
420 *se = parent_entity(*se);
421 *pse = parent_entity(*pse);
425 #else /* !CONFIG_FAIR_GROUP_SCHED */
427 static inline struct task_struct *task_of(struct sched_entity *se)
429 return container_of(se, struct task_struct, se);
432 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
434 return container_of(cfs_rq, struct rq, cfs);
437 #define entity_is_task(se) 1
439 #define for_each_sched_entity(se) \
440 for (; se; se = NULL)
442 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
444 return &task_rq(p)->cfs;
447 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
449 struct task_struct *p = task_of(se);
450 struct rq *rq = task_rq(p);
452 return &rq->cfs;
455 /* runqueue "owned" by this group */
456 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
458 return NULL;
461 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
465 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
469 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
470 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
472 static inline struct sched_entity *parent_entity(struct sched_entity *se)
474 return NULL;
477 static inline void
478 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
482 #endif /* CONFIG_FAIR_GROUP_SCHED */
484 static __always_inline
485 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
487 /**************************************************************
488 * Scheduling class tree data structure manipulation methods:
491 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
493 s64 delta = (s64)(vruntime - max_vruntime);
494 if (delta > 0)
495 max_vruntime = vruntime;
497 return max_vruntime;
500 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
502 s64 delta = (s64)(vruntime - min_vruntime);
503 if (delta < 0)
504 min_vruntime = vruntime;
506 return min_vruntime;
509 static inline int entity_before(struct sched_entity *a,
510 struct sched_entity *b)
512 return (s64)(a->vruntime - b->vruntime) < 0;
515 static void update_min_vruntime(struct cfs_rq *cfs_rq)
517 struct sched_entity *curr = cfs_rq->curr;
518 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
520 u64 vruntime = cfs_rq->min_vruntime;
522 if (curr) {
523 if (curr->on_rq)
524 vruntime = curr->vruntime;
525 else
526 curr = NULL;
529 if (leftmost) { /* non-empty tree */
530 struct sched_entity *se;
531 se = rb_entry(leftmost, struct sched_entity, run_node);
533 if (!curr)
534 vruntime = se->vruntime;
535 else
536 vruntime = min_vruntime(vruntime, se->vruntime);
539 /* ensure we never gain time by being placed backwards. */
540 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
541 #ifndef CONFIG_64BIT
542 smp_wmb();
543 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
544 #endif
548 * Enqueue an entity into the rb-tree:
550 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
552 struct rb_node **link = &cfs_rq->tasks_timeline.rb_root.rb_node;
553 struct rb_node *parent = NULL;
554 struct sched_entity *entry;
555 bool leftmost = true;
558 * Find the right place in the rbtree:
560 while (*link) {
561 parent = *link;
562 entry = rb_entry(parent, struct sched_entity, run_node);
564 * We dont care about collisions. Nodes with
565 * the same key stay together.
567 if (entity_before(se, entry)) {
568 link = &parent->rb_left;
569 } else {
570 link = &parent->rb_right;
571 leftmost = false;
575 rb_link_node(&se->run_node, parent, link);
576 rb_insert_color_cached(&se->run_node,
577 &cfs_rq->tasks_timeline, leftmost);
580 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
582 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
585 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
587 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
589 if (!left)
590 return NULL;
592 return rb_entry(left, struct sched_entity, run_node);
595 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
597 struct rb_node *next = rb_next(&se->run_node);
599 if (!next)
600 return NULL;
602 return rb_entry(next, struct sched_entity, run_node);
605 #ifdef CONFIG_SCHED_DEBUG
606 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
608 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
610 if (!last)
611 return NULL;
613 return rb_entry(last, struct sched_entity, run_node);
616 /**************************************************************
617 * Scheduling class statistics methods:
620 int sched_proc_update_handler(struct ctl_table *table, int write,
621 void __user *buffer, size_t *lenp,
622 loff_t *ppos)
624 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
625 unsigned int factor = get_update_sysctl_factor();
627 if (ret || !write)
628 return ret;
630 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
631 sysctl_sched_min_granularity);
633 #define WRT_SYSCTL(name) \
634 (normalized_sysctl_##name = sysctl_##name / (factor))
635 WRT_SYSCTL(sched_min_granularity);
636 WRT_SYSCTL(sched_latency);
637 WRT_SYSCTL(sched_wakeup_granularity);
638 #undef WRT_SYSCTL
640 return 0;
642 #endif
645 * delta /= w
647 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
649 if (unlikely(se->load.weight != NICE_0_LOAD))
650 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
652 return delta;
656 * The idea is to set a period in which each task runs once.
658 * When there are too many tasks (sched_nr_latency) we have to stretch
659 * this period because otherwise the slices get too small.
661 * p = (nr <= nl) ? l : l*nr/nl
663 static u64 __sched_period(unsigned long nr_running)
665 if (unlikely(nr_running > sched_nr_latency))
666 return nr_running * sysctl_sched_min_granularity;
667 else
668 return sysctl_sched_latency;
672 * We calculate the wall-time slice from the period by taking a part
673 * proportional to the weight.
675 * s = p*P[w/rw]
677 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
679 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
681 for_each_sched_entity(se) {
682 struct load_weight *load;
683 struct load_weight lw;
685 cfs_rq = cfs_rq_of(se);
686 load = &cfs_rq->load;
688 if (unlikely(!se->on_rq)) {
689 lw = cfs_rq->load;
691 update_load_add(&lw, se->load.weight);
692 load = &lw;
694 slice = __calc_delta(slice, se->load.weight, load);
696 return slice;
700 * We calculate the vruntime slice of a to-be-inserted task.
702 * vs = s/w
704 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
706 return calc_delta_fair(sched_slice(cfs_rq, se), se);
709 #ifdef CONFIG_SMP
711 #include "sched-pelt.h"
713 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
714 static unsigned long task_h_load(struct task_struct *p);
716 /* Give new sched_entity start runnable values to heavy its load in infant time */
717 void init_entity_runnable_average(struct sched_entity *se)
719 struct sched_avg *sa = &se->avg;
721 memset(sa, 0, sizeof(*sa));
724 * Tasks are intialized with full load to be seen as heavy tasks until
725 * they get a chance to stabilize to their real load level.
726 * Group entities are intialized with zero load to reflect the fact that
727 * nothing has been attached to the task group yet.
729 if (entity_is_task(se))
730 sa->runnable_load_avg = sa->load_avg = scale_load_down(se->load.weight);
732 se->runnable_weight = se->load.weight;
734 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
737 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
738 static void attach_entity_cfs_rq(struct sched_entity *se);
741 * With new tasks being created, their initial util_avgs are extrapolated
742 * based on the cfs_rq's current util_avg:
744 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
746 * However, in many cases, the above util_avg does not give a desired
747 * value. Moreover, the sum of the util_avgs may be divergent, such
748 * as when the series is a harmonic series.
750 * To solve this problem, we also cap the util_avg of successive tasks to
751 * only 1/2 of the left utilization budget:
753 * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
755 * where n denotes the nth task.
757 * For example, a simplest series from the beginning would be like:
759 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
760 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
762 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
763 * if util_avg > util_avg_cap.
765 void post_init_entity_util_avg(struct sched_entity *se)
767 struct cfs_rq *cfs_rq = cfs_rq_of(se);
768 struct sched_avg *sa = &se->avg;
769 long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
771 if (cap > 0) {
772 if (cfs_rq->avg.util_avg != 0) {
773 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
774 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
776 if (sa->util_avg > cap)
777 sa->util_avg = cap;
778 } else {
779 sa->util_avg = cap;
783 if (entity_is_task(se)) {
784 struct task_struct *p = task_of(se);
785 if (p->sched_class != &fair_sched_class) {
787 * For !fair tasks do:
789 update_cfs_rq_load_avg(now, cfs_rq);
790 attach_entity_load_avg(cfs_rq, se);
791 switched_from_fair(rq, p);
793 * such that the next switched_to_fair() has the
794 * expected state.
796 se->avg.last_update_time = cfs_rq_clock_task(cfs_rq);
797 return;
801 attach_entity_cfs_rq(se);
804 #else /* !CONFIG_SMP */
805 void init_entity_runnable_average(struct sched_entity *se)
808 void post_init_entity_util_avg(struct sched_entity *se)
811 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
814 #endif /* CONFIG_SMP */
817 * Update the current task's runtime statistics.
819 static void update_curr(struct cfs_rq *cfs_rq)
821 struct sched_entity *curr = cfs_rq->curr;
822 u64 now = rq_clock_task(rq_of(cfs_rq));
823 u64 delta_exec;
825 if (unlikely(!curr))
826 return;
828 delta_exec = now - curr->exec_start;
829 if (unlikely((s64)delta_exec <= 0))
830 return;
832 curr->exec_start = now;
834 schedstat_set(curr->statistics.exec_max,
835 max(delta_exec, curr->statistics.exec_max));
837 curr->sum_exec_runtime += delta_exec;
838 schedstat_add(cfs_rq->exec_clock, delta_exec);
840 curr->vruntime += calc_delta_fair(delta_exec, curr);
841 update_min_vruntime(cfs_rq);
843 if (entity_is_task(curr)) {
844 struct task_struct *curtask = task_of(curr);
846 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
847 cgroup_account_cputime(curtask, delta_exec);
848 account_group_exec_runtime(curtask, delta_exec);
851 account_cfs_rq_runtime(cfs_rq, delta_exec);
854 static void update_curr_fair(struct rq *rq)
856 update_curr(cfs_rq_of(&rq->curr->se));
859 static inline void
860 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
862 u64 wait_start, prev_wait_start;
864 if (!schedstat_enabled())
865 return;
867 wait_start = rq_clock(rq_of(cfs_rq));
868 prev_wait_start = schedstat_val(se->statistics.wait_start);
870 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
871 likely(wait_start > prev_wait_start))
872 wait_start -= prev_wait_start;
874 schedstat_set(se->statistics.wait_start, wait_start);
877 static inline void
878 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
880 struct task_struct *p;
881 u64 delta;
883 if (!schedstat_enabled())
884 return;
886 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
888 if (entity_is_task(se)) {
889 p = task_of(se);
890 if (task_on_rq_migrating(p)) {
892 * Preserve migrating task's wait time so wait_start
893 * time stamp can be adjusted to accumulate wait time
894 * prior to migration.
896 schedstat_set(se->statistics.wait_start, delta);
897 return;
899 trace_sched_stat_wait(p, delta);
902 schedstat_set(se->statistics.wait_max,
903 max(schedstat_val(se->statistics.wait_max), delta));
904 schedstat_inc(se->statistics.wait_count);
905 schedstat_add(se->statistics.wait_sum, delta);
906 schedstat_set(se->statistics.wait_start, 0);
909 static inline void
910 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
912 struct task_struct *tsk = NULL;
913 u64 sleep_start, block_start;
915 if (!schedstat_enabled())
916 return;
918 sleep_start = schedstat_val(se->statistics.sleep_start);
919 block_start = schedstat_val(se->statistics.block_start);
921 if (entity_is_task(se))
922 tsk = task_of(se);
924 if (sleep_start) {
925 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
927 if ((s64)delta < 0)
928 delta = 0;
930 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
931 schedstat_set(se->statistics.sleep_max, delta);
933 schedstat_set(se->statistics.sleep_start, 0);
934 schedstat_add(se->statistics.sum_sleep_runtime, delta);
936 if (tsk) {
937 account_scheduler_latency(tsk, delta >> 10, 1);
938 trace_sched_stat_sleep(tsk, delta);
941 if (block_start) {
942 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
944 if ((s64)delta < 0)
945 delta = 0;
947 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
948 schedstat_set(se->statistics.block_max, delta);
950 schedstat_set(se->statistics.block_start, 0);
951 schedstat_add(se->statistics.sum_sleep_runtime, delta);
953 if (tsk) {
954 if (tsk->in_iowait) {
955 schedstat_add(se->statistics.iowait_sum, delta);
956 schedstat_inc(se->statistics.iowait_count);
957 trace_sched_stat_iowait(tsk, delta);
960 trace_sched_stat_blocked(tsk, delta);
963 * Blocking time is in units of nanosecs, so shift by
964 * 20 to get a milliseconds-range estimation of the
965 * amount of time that the task spent sleeping:
967 if (unlikely(prof_on == SLEEP_PROFILING)) {
968 profile_hits(SLEEP_PROFILING,
969 (void *)get_wchan(tsk),
970 delta >> 20);
972 account_scheduler_latency(tsk, delta >> 10, 0);
978 * Task is being enqueued - update stats:
980 static inline void
981 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
983 if (!schedstat_enabled())
984 return;
987 * Are we enqueueing a waiting task? (for current tasks
988 * a dequeue/enqueue event is a NOP)
990 if (se != cfs_rq->curr)
991 update_stats_wait_start(cfs_rq, se);
993 if (flags & ENQUEUE_WAKEUP)
994 update_stats_enqueue_sleeper(cfs_rq, se);
997 static inline void
998 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1001 if (!schedstat_enabled())
1002 return;
1005 * Mark the end of the wait period if dequeueing a
1006 * waiting task:
1008 if (se != cfs_rq->curr)
1009 update_stats_wait_end(cfs_rq, se);
1011 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1012 struct task_struct *tsk = task_of(se);
1014 if (tsk->state & TASK_INTERRUPTIBLE)
1015 schedstat_set(se->statistics.sleep_start,
1016 rq_clock(rq_of(cfs_rq)));
1017 if (tsk->state & TASK_UNINTERRUPTIBLE)
1018 schedstat_set(se->statistics.block_start,
1019 rq_clock(rq_of(cfs_rq)));
1024 * We are picking a new current task - update its stats:
1026 static inline void
1027 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1030 * We are starting a new run period:
1032 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1035 /**************************************************
1036 * Scheduling class queueing methods:
1039 #ifdef CONFIG_NUMA_BALANCING
1041 * Approximate time to scan a full NUMA task in ms. The task scan period is
1042 * calculated based on the tasks virtual memory size and
1043 * numa_balancing_scan_size.
1045 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1046 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1048 /* Portion of address space to scan in MB */
1049 unsigned int sysctl_numa_balancing_scan_size = 256;
1051 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1052 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1054 struct numa_group {
1055 atomic_t refcount;
1057 spinlock_t lock; /* nr_tasks, tasks */
1058 int nr_tasks;
1059 pid_t gid;
1060 int active_nodes;
1062 struct rcu_head rcu;
1063 unsigned long total_faults;
1064 unsigned long max_faults_cpu;
1066 * Faults_cpu is used to decide whether memory should move
1067 * towards the CPU. As a consequence, these stats are weighted
1068 * more by CPU use than by memory faults.
1070 unsigned long *faults_cpu;
1071 unsigned long faults[0];
1074 static inline unsigned long group_faults_priv(struct numa_group *ng);
1075 static inline unsigned long group_faults_shared(struct numa_group *ng);
1077 static unsigned int task_nr_scan_windows(struct task_struct *p)
1079 unsigned long rss = 0;
1080 unsigned long nr_scan_pages;
1083 * Calculations based on RSS as non-present and empty pages are skipped
1084 * by the PTE scanner and NUMA hinting faults should be trapped based
1085 * on resident pages
1087 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1088 rss = get_mm_rss(p->mm);
1089 if (!rss)
1090 rss = nr_scan_pages;
1092 rss = round_up(rss, nr_scan_pages);
1093 return rss / nr_scan_pages;
1096 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1097 #define MAX_SCAN_WINDOW 2560
1099 static unsigned int task_scan_min(struct task_struct *p)
1101 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1102 unsigned int scan, floor;
1103 unsigned int windows = 1;
1105 if (scan_size < MAX_SCAN_WINDOW)
1106 windows = MAX_SCAN_WINDOW / scan_size;
1107 floor = 1000 / windows;
1109 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1110 return max_t(unsigned int, floor, scan);
1113 static unsigned int task_scan_start(struct task_struct *p)
1115 unsigned long smin = task_scan_min(p);
1116 unsigned long period = smin;
1118 /* Scale the maximum scan period with the amount of shared memory. */
1119 if (p->numa_group) {
1120 struct numa_group *ng = p->numa_group;
1121 unsigned long shared = group_faults_shared(ng);
1122 unsigned long private = group_faults_priv(ng);
1124 period *= atomic_read(&ng->refcount);
1125 period *= shared + 1;
1126 period /= private + shared + 1;
1129 return max(smin, period);
1132 static unsigned int task_scan_max(struct task_struct *p)
1134 unsigned long smin = task_scan_min(p);
1135 unsigned long smax;
1137 /* Watch for min being lower than max due to floor calculations */
1138 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1140 /* Scale the maximum scan period with the amount of shared memory. */
1141 if (p->numa_group) {
1142 struct numa_group *ng = p->numa_group;
1143 unsigned long shared = group_faults_shared(ng);
1144 unsigned long private = group_faults_priv(ng);
1145 unsigned long period = smax;
1147 period *= atomic_read(&ng->refcount);
1148 period *= shared + 1;
1149 period /= private + shared + 1;
1151 smax = max(smax, period);
1154 return max(smin, smax);
1157 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1159 rq->nr_numa_running += (p->numa_preferred_nid != -1);
1160 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1163 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1165 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1166 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1169 /* Shared or private faults. */
1170 #define NR_NUMA_HINT_FAULT_TYPES 2
1172 /* Memory and CPU locality */
1173 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1175 /* Averaged statistics, and temporary buffers. */
1176 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1178 pid_t task_numa_group_id(struct task_struct *p)
1180 return p->numa_group ? p->numa_group->gid : 0;
1184 * The averaged statistics, shared & private, memory & cpu,
1185 * occupy the first half of the array. The second half of the
1186 * array is for current counters, which are averaged into the
1187 * first set by task_numa_placement.
1189 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1191 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1194 static inline unsigned long task_faults(struct task_struct *p, int nid)
1196 if (!p->numa_faults)
1197 return 0;
1199 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1200 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1203 static inline unsigned long group_faults(struct task_struct *p, int nid)
1205 if (!p->numa_group)
1206 return 0;
1208 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1209 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1212 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1214 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1215 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1218 static inline unsigned long group_faults_priv(struct numa_group *ng)
1220 unsigned long faults = 0;
1221 int node;
1223 for_each_online_node(node) {
1224 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1227 return faults;
1230 static inline unsigned long group_faults_shared(struct numa_group *ng)
1232 unsigned long faults = 0;
1233 int node;
1235 for_each_online_node(node) {
1236 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1239 return faults;
1243 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1244 * considered part of a numa group's pseudo-interleaving set. Migrations
1245 * between these nodes are slowed down, to allow things to settle down.
1247 #define ACTIVE_NODE_FRACTION 3
1249 static bool numa_is_active_node(int nid, struct numa_group *ng)
1251 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1254 /* Handle placement on systems where not all nodes are directly connected. */
1255 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1256 int maxdist, bool task)
1258 unsigned long score = 0;
1259 int node;
1262 * All nodes are directly connected, and the same distance
1263 * from each other. No need for fancy placement algorithms.
1265 if (sched_numa_topology_type == NUMA_DIRECT)
1266 return 0;
1269 * This code is called for each node, introducing N^2 complexity,
1270 * which should be ok given the number of nodes rarely exceeds 8.
1272 for_each_online_node(node) {
1273 unsigned long faults;
1274 int dist = node_distance(nid, node);
1277 * The furthest away nodes in the system are not interesting
1278 * for placement; nid was already counted.
1280 if (dist == sched_max_numa_distance || node == nid)
1281 continue;
1284 * On systems with a backplane NUMA topology, compare groups
1285 * of nodes, and move tasks towards the group with the most
1286 * memory accesses. When comparing two nodes at distance
1287 * "hoplimit", only nodes closer by than "hoplimit" are part
1288 * of each group. Skip other nodes.
1290 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1291 dist > maxdist)
1292 continue;
1294 /* Add up the faults from nearby nodes. */
1295 if (task)
1296 faults = task_faults(p, node);
1297 else
1298 faults = group_faults(p, node);
1301 * On systems with a glueless mesh NUMA topology, there are
1302 * no fixed "groups of nodes". Instead, nodes that are not
1303 * directly connected bounce traffic through intermediate
1304 * nodes; a numa_group can occupy any set of nodes.
1305 * The further away a node is, the less the faults count.
1306 * This seems to result in good task placement.
1308 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1309 faults *= (sched_max_numa_distance - dist);
1310 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1313 score += faults;
1316 return score;
1320 * These return the fraction of accesses done by a particular task, or
1321 * task group, on a particular numa node. The group weight is given a
1322 * larger multiplier, in order to group tasks together that are almost
1323 * evenly spread out between numa nodes.
1325 static inline unsigned long task_weight(struct task_struct *p, int nid,
1326 int dist)
1328 unsigned long faults, total_faults;
1330 if (!p->numa_faults)
1331 return 0;
1333 total_faults = p->total_numa_faults;
1335 if (!total_faults)
1336 return 0;
1338 faults = task_faults(p, nid);
1339 faults += score_nearby_nodes(p, nid, dist, true);
1341 return 1000 * faults / total_faults;
1344 static inline unsigned long group_weight(struct task_struct *p, int nid,
1345 int dist)
1347 unsigned long faults, total_faults;
1349 if (!p->numa_group)
1350 return 0;
1352 total_faults = p->numa_group->total_faults;
1354 if (!total_faults)
1355 return 0;
1357 faults = group_faults(p, nid);
1358 faults += score_nearby_nodes(p, nid, dist, false);
1360 return 1000 * faults / total_faults;
1363 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1364 int src_nid, int dst_cpu)
1366 struct numa_group *ng = p->numa_group;
1367 int dst_nid = cpu_to_node(dst_cpu);
1368 int last_cpupid, this_cpupid;
1370 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1373 * Multi-stage node selection is used in conjunction with a periodic
1374 * migration fault to build a temporal task<->page relation. By using
1375 * a two-stage filter we remove short/unlikely relations.
1377 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1378 * a task's usage of a particular page (n_p) per total usage of this
1379 * page (n_t) (in a given time-span) to a probability.
1381 * Our periodic faults will sample this probability and getting the
1382 * same result twice in a row, given these samples are fully
1383 * independent, is then given by P(n)^2, provided our sample period
1384 * is sufficiently short compared to the usage pattern.
1386 * This quadric squishes small probabilities, making it less likely we
1387 * act on an unlikely task<->page relation.
1389 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1390 if (!cpupid_pid_unset(last_cpupid) &&
1391 cpupid_to_nid(last_cpupid) != dst_nid)
1392 return false;
1394 /* Always allow migrate on private faults */
1395 if (cpupid_match_pid(p, last_cpupid))
1396 return true;
1398 /* A shared fault, but p->numa_group has not been set up yet. */
1399 if (!ng)
1400 return true;
1403 * Destination node is much more heavily used than the source
1404 * node? Allow migration.
1406 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1407 ACTIVE_NODE_FRACTION)
1408 return true;
1411 * Distribute memory according to CPU & memory use on each node,
1412 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1414 * faults_cpu(dst) 3 faults_cpu(src)
1415 * --------------- * - > ---------------
1416 * faults_mem(dst) 4 faults_mem(src)
1418 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1419 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1422 static unsigned long weighted_cpuload(struct rq *rq);
1423 static unsigned long source_load(int cpu, int type);
1424 static unsigned long target_load(int cpu, int type);
1425 static unsigned long capacity_of(int cpu);
1427 /* Cached statistics for all CPUs within a node */
1428 struct numa_stats {
1429 unsigned long nr_running;
1430 unsigned long load;
1432 /* Total compute capacity of CPUs on a node */
1433 unsigned long compute_capacity;
1435 /* Approximate capacity in terms of runnable tasks on a node */
1436 unsigned long task_capacity;
1437 int has_free_capacity;
1441 * XXX borrowed from update_sg_lb_stats
1443 static void update_numa_stats(struct numa_stats *ns, int nid)
1445 int smt, cpu, cpus = 0;
1446 unsigned long capacity;
1448 memset(ns, 0, sizeof(*ns));
1449 for_each_cpu(cpu, cpumask_of_node(nid)) {
1450 struct rq *rq = cpu_rq(cpu);
1452 ns->nr_running += rq->nr_running;
1453 ns->load += weighted_cpuload(rq);
1454 ns->compute_capacity += capacity_of(cpu);
1456 cpus++;
1460 * If we raced with hotplug and there are no CPUs left in our mask
1461 * the @ns structure is NULL'ed and task_numa_compare() will
1462 * not find this node attractive.
1464 * We'll either bail at !has_free_capacity, or we'll detect a huge
1465 * imbalance and bail there.
1467 if (!cpus)
1468 return;
1470 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1471 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1472 capacity = cpus / smt; /* cores */
1474 ns->task_capacity = min_t(unsigned, capacity,
1475 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1476 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1479 struct task_numa_env {
1480 struct task_struct *p;
1482 int src_cpu, src_nid;
1483 int dst_cpu, dst_nid;
1485 struct numa_stats src_stats, dst_stats;
1487 int imbalance_pct;
1488 int dist;
1490 struct task_struct *best_task;
1491 long best_imp;
1492 int best_cpu;
1495 static void task_numa_assign(struct task_numa_env *env,
1496 struct task_struct *p, long imp)
1498 if (env->best_task)
1499 put_task_struct(env->best_task);
1500 if (p)
1501 get_task_struct(p);
1503 env->best_task = p;
1504 env->best_imp = imp;
1505 env->best_cpu = env->dst_cpu;
1508 static bool load_too_imbalanced(long src_load, long dst_load,
1509 struct task_numa_env *env)
1511 long imb, old_imb;
1512 long orig_src_load, orig_dst_load;
1513 long src_capacity, dst_capacity;
1516 * The load is corrected for the CPU capacity available on each node.
1518 * src_load dst_load
1519 * ------------ vs ---------
1520 * src_capacity dst_capacity
1522 src_capacity = env->src_stats.compute_capacity;
1523 dst_capacity = env->dst_stats.compute_capacity;
1525 /* We care about the slope of the imbalance, not the direction. */
1526 if (dst_load < src_load)
1527 swap(dst_load, src_load);
1529 /* Is the difference below the threshold? */
1530 imb = dst_load * src_capacity * 100 -
1531 src_load * dst_capacity * env->imbalance_pct;
1532 if (imb <= 0)
1533 return false;
1536 * The imbalance is above the allowed threshold.
1537 * Compare it with the old imbalance.
1539 orig_src_load = env->src_stats.load;
1540 orig_dst_load = env->dst_stats.load;
1542 if (orig_dst_load < orig_src_load)
1543 swap(orig_dst_load, orig_src_load);
1545 old_imb = orig_dst_load * src_capacity * 100 -
1546 orig_src_load * dst_capacity * env->imbalance_pct;
1548 /* Would this change make things worse? */
1549 return (imb > old_imb);
1553 * This checks if the overall compute and NUMA accesses of the system would
1554 * be improved if the source tasks was migrated to the target dst_cpu taking
1555 * into account that it might be best if task running on the dst_cpu should
1556 * be exchanged with the source task
1558 static void task_numa_compare(struct task_numa_env *env,
1559 long taskimp, long groupimp)
1561 struct rq *src_rq = cpu_rq(env->src_cpu);
1562 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1563 struct task_struct *cur;
1564 long src_load, dst_load;
1565 long load;
1566 long imp = env->p->numa_group ? groupimp : taskimp;
1567 long moveimp = imp;
1568 int dist = env->dist;
1570 rcu_read_lock();
1571 cur = task_rcu_dereference(&dst_rq->curr);
1572 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1573 cur = NULL;
1576 * Because we have preemption enabled we can get migrated around and
1577 * end try selecting ourselves (current == env->p) as a swap candidate.
1579 if (cur == env->p)
1580 goto unlock;
1583 * "imp" is the fault differential for the source task between the
1584 * source and destination node. Calculate the total differential for
1585 * the source task and potential destination task. The more negative
1586 * the value is, the more rmeote accesses that would be expected to
1587 * be incurred if the tasks were swapped.
1589 if (cur) {
1590 /* Skip this swap candidate if cannot move to the source cpu */
1591 if (!cpumask_test_cpu(env->src_cpu, &cur->cpus_allowed))
1592 goto unlock;
1595 * If dst and source tasks are in the same NUMA group, or not
1596 * in any group then look only at task weights.
1598 if (cur->numa_group == env->p->numa_group) {
1599 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1600 task_weight(cur, env->dst_nid, dist);
1602 * Add some hysteresis to prevent swapping the
1603 * tasks within a group over tiny differences.
1605 if (cur->numa_group)
1606 imp -= imp/16;
1607 } else {
1609 * Compare the group weights. If a task is all by
1610 * itself (not part of a group), use the task weight
1611 * instead.
1613 if (cur->numa_group)
1614 imp += group_weight(cur, env->src_nid, dist) -
1615 group_weight(cur, env->dst_nid, dist);
1616 else
1617 imp += task_weight(cur, env->src_nid, dist) -
1618 task_weight(cur, env->dst_nid, dist);
1622 if (imp <= env->best_imp && moveimp <= env->best_imp)
1623 goto unlock;
1625 if (!cur) {
1626 /* Is there capacity at our destination? */
1627 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1628 !env->dst_stats.has_free_capacity)
1629 goto unlock;
1631 goto balance;
1634 /* Balance doesn't matter much if we're running a task per cpu */
1635 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1636 dst_rq->nr_running == 1)
1637 goto assign;
1640 * In the overloaded case, try and keep the load balanced.
1642 balance:
1643 load = task_h_load(env->p);
1644 dst_load = env->dst_stats.load + load;
1645 src_load = env->src_stats.load - load;
1647 if (moveimp > imp && moveimp > env->best_imp) {
1649 * If the improvement from just moving env->p direction is
1650 * better than swapping tasks around, check if a move is
1651 * possible. Store a slightly smaller score than moveimp,
1652 * so an actually idle CPU will win.
1654 if (!load_too_imbalanced(src_load, dst_load, env)) {
1655 imp = moveimp - 1;
1656 cur = NULL;
1657 goto assign;
1661 if (imp <= env->best_imp)
1662 goto unlock;
1664 if (cur) {
1665 load = task_h_load(cur);
1666 dst_load -= load;
1667 src_load += load;
1670 if (load_too_imbalanced(src_load, dst_load, env))
1671 goto unlock;
1674 * One idle CPU per node is evaluated for a task numa move.
1675 * Call select_idle_sibling to maybe find a better one.
1677 if (!cur) {
1679 * select_idle_siblings() uses an per-cpu cpumask that
1680 * can be used from IRQ context.
1682 local_irq_disable();
1683 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1684 env->dst_cpu);
1685 local_irq_enable();
1688 assign:
1689 task_numa_assign(env, cur, imp);
1690 unlock:
1691 rcu_read_unlock();
1694 static void task_numa_find_cpu(struct task_numa_env *env,
1695 long taskimp, long groupimp)
1697 int cpu;
1699 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1700 /* Skip this CPU if the source task cannot migrate */
1701 if (!cpumask_test_cpu(cpu, &env->p->cpus_allowed))
1702 continue;
1704 env->dst_cpu = cpu;
1705 task_numa_compare(env, taskimp, groupimp);
1709 /* Only move tasks to a NUMA node less busy than the current node. */
1710 static bool numa_has_capacity(struct task_numa_env *env)
1712 struct numa_stats *src = &env->src_stats;
1713 struct numa_stats *dst = &env->dst_stats;
1715 if (src->has_free_capacity && !dst->has_free_capacity)
1716 return false;
1719 * Only consider a task move if the source has a higher load
1720 * than the destination, corrected for CPU capacity on each node.
1722 * src->load dst->load
1723 * --------------------- vs ---------------------
1724 * src->compute_capacity dst->compute_capacity
1726 if (src->load * dst->compute_capacity * env->imbalance_pct >
1728 dst->load * src->compute_capacity * 100)
1729 return true;
1731 return false;
1734 static int task_numa_migrate(struct task_struct *p)
1736 struct task_numa_env env = {
1737 .p = p,
1739 .src_cpu = task_cpu(p),
1740 .src_nid = task_node(p),
1742 .imbalance_pct = 112,
1744 .best_task = NULL,
1745 .best_imp = 0,
1746 .best_cpu = -1,
1748 struct sched_domain *sd;
1749 unsigned long taskweight, groupweight;
1750 int nid, ret, dist;
1751 long taskimp, groupimp;
1754 * Pick the lowest SD_NUMA domain, as that would have the smallest
1755 * imbalance and would be the first to start moving tasks about.
1757 * And we want to avoid any moving of tasks about, as that would create
1758 * random movement of tasks -- counter the numa conditions we're trying
1759 * to satisfy here.
1761 rcu_read_lock();
1762 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1763 if (sd)
1764 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1765 rcu_read_unlock();
1768 * Cpusets can break the scheduler domain tree into smaller
1769 * balance domains, some of which do not cross NUMA boundaries.
1770 * Tasks that are "trapped" in such domains cannot be migrated
1771 * elsewhere, so there is no point in (re)trying.
1773 if (unlikely(!sd)) {
1774 p->numa_preferred_nid = task_node(p);
1775 return -EINVAL;
1778 env.dst_nid = p->numa_preferred_nid;
1779 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1780 taskweight = task_weight(p, env.src_nid, dist);
1781 groupweight = group_weight(p, env.src_nid, dist);
1782 update_numa_stats(&env.src_stats, env.src_nid);
1783 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1784 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1785 update_numa_stats(&env.dst_stats, env.dst_nid);
1787 /* Try to find a spot on the preferred nid. */
1788 if (numa_has_capacity(&env))
1789 task_numa_find_cpu(&env, taskimp, groupimp);
1792 * Look at other nodes in these cases:
1793 * - there is no space available on the preferred_nid
1794 * - the task is part of a numa_group that is interleaved across
1795 * multiple NUMA nodes; in order to better consolidate the group,
1796 * we need to check other locations.
1798 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1799 for_each_online_node(nid) {
1800 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1801 continue;
1803 dist = node_distance(env.src_nid, env.dst_nid);
1804 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1805 dist != env.dist) {
1806 taskweight = task_weight(p, env.src_nid, dist);
1807 groupweight = group_weight(p, env.src_nid, dist);
1810 /* Only consider nodes where both task and groups benefit */
1811 taskimp = task_weight(p, nid, dist) - taskweight;
1812 groupimp = group_weight(p, nid, dist) - groupweight;
1813 if (taskimp < 0 && groupimp < 0)
1814 continue;
1816 env.dist = dist;
1817 env.dst_nid = nid;
1818 update_numa_stats(&env.dst_stats, env.dst_nid);
1819 if (numa_has_capacity(&env))
1820 task_numa_find_cpu(&env, taskimp, groupimp);
1825 * If the task is part of a workload that spans multiple NUMA nodes,
1826 * and is migrating into one of the workload's active nodes, remember
1827 * this node as the task's preferred numa node, so the workload can
1828 * settle down.
1829 * A task that migrated to a second choice node will be better off
1830 * trying for a better one later. Do not set the preferred node here.
1832 if (p->numa_group) {
1833 struct numa_group *ng = p->numa_group;
1835 if (env.best_cpu == -1)
1836 nid = env.src_nid;
1837 else
1838 nid = env.dst_nid;
1840 if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng))
1841 sched_setnuma(p, env.dst_nid);
1844 /* No better CPU than the current one was found. */
1845 if (env.best_cpu == -1)
1846 return -EAGAIN;
1849 * Reset the scan period if the task is being rescheduled on an
1850 * alternative node to recheck if the tasks is now properly placed.
1852 p->numa_scan_period = task_scan_start(p);
1854 if (env.best_task == NULL) {
1855 ret = migrate_task_to(p, env.best_cpu);
1856 if (ret != 0)
1857 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1858 return ret;
1861 ret = migrate_swap(p, env.best_task);
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;
1868 /* Attempt to migrate a task to a CPU on the preferred node. */
1869 static void numa_migrate_preferred(struct task_struct *p)
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);
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.
1895 static void numa_group_count_active_nodes(struct numa_group *numa_group)
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;
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++;
1912 numa_group->max_faults_cpu = max_faults;
1913 numa_group->active_nodes = active_nodes;
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.
1923 #define NUMA_PERIOD_SLOTS 10
1924 #define NUMA_PERIOD_THRESHOLD 7
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.
1932 static void update_task_scan_period(struct task_struct *p,
1933 unsigned long shared, unsigned long private)
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];
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
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;
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)
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) {
1971 * Most memory accesses are local. There is no need to
1972 * do fast NUMA scanning, since memory is already local.
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) {
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.
1984 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
1985 if (!slot)
1986 slot = 1;
1987 diff = slot * period_slot;
1988 } else {
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.
1994 int ratio = max(lr_ratio, ps_ratio);
1995 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
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));
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.
2010 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
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;
2025 p->last_sum_exec_runtime = runtime;
2026 p->last_task_numa_placement = now;
2028 return delta;
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.
2036 static int preferred_group_nid(struct task_struct *p, int nid)
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;
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.
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;
2063 return max_node;
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.
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);
2099 /* Remember the top group. */
2100 if (faults > max_faults) {
2101 max_faults = faults;
2102 max_group = this_group;
2104 * subtle: at the smallest distance there is
2105 * just one node left in each "group", the
2106 * winner is the preferred nid.
2108 nid = a;
2111 /* Next round, evaluate the nodes within max_group. */
2112 if (!max_faults)
2113 break;
2114 nodes = max_group;
2116 return nid;
2119 static void task_numa_placement(struct task_struct *p)
2121 int seq, nid, max_nid = -1, max_group_nid = -1;
2122 unsigned long max_faults = 0, max_group_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;
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:
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);
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;
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.
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) {
2188 * safe because we can only change our own group
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.
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];
2201 if (faults > max_faults) {
2202 max_faults = faults;
2203 max_nid = nid;
2206 if (group_faults > max_group_faults) {
2207 max_group_faults = group_faults;
2208 max_group_nid = nid;
2212 update_task_scan_period(p, fault_types[0], fault_types[1]);
2214 if (p->numa_group) {
2215 numa_group_count_active_nodes(p->numa_group);
2216 spin_unlock_irq(group_lock);
2217 max_nid = preferred_group_nid(p, max_group_nid);
2220 if (max_faults) {
2221 /* Set the new preferred node */
2222 if (max_nid != p->numa_preferred_nid)
2223 sched_setnuma(p, max_nid);
2225 if (task_node(p) != p->numa_preferred_nid)
2226 numa_migrate_preferred(p);
2230 static inline int get_numa_group(struct numa_group *grp)
2232 return atomic_inc_not_zero(&grp->refcount);
2235 static inline void put_numa_group(struct numa_group *grp)
2237 if (atomic_dec_and_test(&grp->refcount))
2238 kfree_rcu(grp, rcu);
2241 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2242 int *priv)
2244 struct numa_group *grp, *my_grp;
2245 struct task_struct *tsk;
2246 bool join = false;
2247 int cpu = cpupid_to_cpu(cpupid);
2248 int i;
2250 if (unlikely(!p->numa_group)) {
2251 unsigned int size = sizeof(struct numa_group) +
2252 4*nr_node_ids*sizeof(unsigned long);
2254 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2255 if (!grp)
2256 return;
2258 atomic_set(&grp->refcount, 1);
2259 grp->active_nodes = 1;
2260 grp->max_faults_cpu = 0;
2261 spin_lock_init(&grp->lock);
2262 grp->gid = p->pid;
2263 /* Second half of the array tracks nids where faults happen */
2264 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2265 nr_node_ids;
2267 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2268 grp->faults[i] = p->numa_faults[i];
2270 grp->total_faults = p->total_numa_faults;
2272 grp->nr_tasks++;
2273 rcu_assign_pointer(p->numa_group, grp);
2276 rcu_read_lock();
2277 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2279 if (!cpupid_match_pid(tsk, cpupid))
2280 goto no_join;
2282 grp = rcu_dereference(tsk->numa_group);
2283 if (!grp)
2284 goto no_join;
2286 my_grp = p->numa_group;
2287 if (grp == my_grp)
2288 goto no_join;
2291 * Only join the other group if its bigger; if we're the bigger group,
2292 * the other task will join us.
2294 if (my_grp->nr_tasks > grp->nr_tasks)
2295 goto no_join;
2298 * Tie-break on the grp address.
2300 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2301 goto no_join;
2303 /* Always join threads in the same process. */
2304 if (tsk->mm == current->mm)
2305 join = true;
2307 /* Simple filter to avoid false positives due to PID collisions */
2308 if (flags & TNF_SHARED)
2309 join = true;
2311 /* Update priv based on whether false sharing was detected */
2312 *priv = !join;
2314 if (join && !get_numa_group(grp))
2315 goto no_join;
2317 rcu_read_unlock();
2319 if (!join)
2320 return;
2322 BUG_ON(irqs_disabled());
2323 double_lock_irq(&my_grp->lock, &grp->lock);
2325 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2326 my_grp->faults[i] -= p->numa_faults[i];
2327 grp->faults[i] += p->numa_faults[i];
2329 my_grp->total_faults -= p->total_numa_faults;
2330 grp->total_faults += p->total_numa_faults;
2332 my_grp->nr_tasks--;
2333 grp->nr_tasks++;
2335 spin_unlock(&my_grp->lock);
2336 spin_unlock_irq(&grp->lock);
2338 rcu_assign_pointer(p->numa_group, grp);
2340 put_numa_group(my_grp);
2341 return;
2343 no_join:
2344 rcu_read_unlock();
2345 return;
2348 void task_numa_free(struct task_struct *p)
2350 struct numa_group *grp = p->numa_group;
2351 void *numa_faults = p->numa_faults;
2352 unsigned long flags;
2353 int i;
2355 if (grp) {
2356 spin_lock_irqsave(&grp->lock, flags);
2357 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2358 grp->faults[i] -= p->numa_faults[i];
2359 grp->total_faults -= p->total_numa_faults;
2361 grp->nr_tasks--;
2362 spin_unlock_irqrestore(&grp->lock, flags);
2363 RCU_INIT_POINTER(p->numa_group, NULL);
2364 put_numa_group(grp);
2367 p->numa_faults = NULL;
2368 kfree(numa_faults);
2372 * Got a PROT_NONE fault for a page on @node.
2374 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2376 struct task_struct *p = current;
2377 bool migrated = flags & TNF_MIGRATED;
2378 int cpu_node = task_node(current);
2379 int local = !!(flags & TNF_FAULT_LOCAL);
2380 struct numa_group *ng;
2381 int priv;
2383 if (!static_branch_likely(&sched_numa_balancing))
2384 return;
2386 /* for example, ksmd faulting in a user's mm */
2387 if (!p->mm)
2388 return;
2390 /* Allocate buffer to track faults on a per-node basis */
2391 if (unlikely(!p->numa_faults)) {
2392 int size = sizeof(*p->numa_faults) *
2393 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2395 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2396 if (!p->numa_faults)
2397 return;
2399 p->total_numa_faults = 0;
2400 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2404 * First accesses are treated as private, otherwise consider accesses
2405 * to be private if the accessing pid has not changed
2407 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2408 priv = 1;
2409 } else {
2410 priv = cpupid_match_pid(p, last_cpupid);
2411 if (!priv && !(flags & TNF_NO_GROUP))
2412 task_numa_group(p, last_cpupid, flags, &priv);
2416 * If a workload spans multiple NUMA nodes, a shared fault that
2417 * occurs wholly within the set of nodes that the workload is
2418 * actively using should be counted as local. This allows the
2419 * scan rate to slow down when a workload has settled down.
2421 ng = p->numa_group;
2422 if (!priv && !local && ng && ng->active_nodes > 1 &&
2423 numa_is_active_node(cpu_node, ng) &&
2424 numa_is_active_node(mem_node, ng))
2425 local = 1;
2427 task_numa_placement(p);
2430 * Retry task to preferred node migration periodically, in case it
2431 * case it previously failed, or the scheduler moved us.
2433 if (time_after(jiffies, p->numa_migrate_retry))
2434 numa_migrate_preferred(p);
2436 if (migrated)
2437 p->numa_pages_migrated += pages;
2438 if (flags & TNF_MIGRATE_FAIL)
2439 p->numa_faults_locality[2] += pages;
2441 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2442 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2443 p->numa_faults_locality[local] += pages;
2446 static void reset_ptenuma_scan(struct task_struct *p)
2449 * We only did a read acquisition of the mmap sem, so
2450 * p->mm->numa_scan_seq is written to without exclusive access
2451 * and the update is not guaranteed to be atomic. That's not
2452 * much of an issue though, since this is just used for
2453 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2454 * expensive, to avoid any form of compiler optimizations:
2456 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2457 p->mm->numa_scan_offset = 0;
2461 * The expensive part of numa migration is done from task_work context.
2462 * Triggered from task_tick_numa().
2464 void task_numa_work(struct callback_head *work)
2466 unsigned long migrate, next_scan, now = jiffies;
2467 struct task_struct *p = current;
2468 struct mm_struct *mm = p->mm;
2469 u64 runtime = p->se.sum_exec_runtime;
2470 struct vm_area_struct *vma;
2471 unsigned long start, end;
2472 unsigned long nr_pte_updates = 0;
2473 long pages, virtpages;
2475 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2477 work->next = work; /* protect against double add */
2479 * Who cares about NUMA placement when they're dying.
2481 * NOTE: make sure not to dereference p->mm before this check,
2482 * exit_task_work() happens _after_ exit_mm() so we could be called
2483 * without p->mm even though we still had it when we enqueued this
2484 * work.
2486 if (p->flags & PF_EXITING)
2487 return;
2489 if (!mm->numa_next_scan) {
2490 mm->numa_next_scan = now +
2491 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2495 * Enforce maximal scan/migration frequency..
2497 migrate = mm->numa_next_scan;
2498 if (time_before(now, migrate))
2499 return;
2501 if (p->numa_scan_period == 0) {
2502 p->numa_scan_period_max = task_scan_max(p);
2503 p->numa_scan_period = task_scan_start(p);
2506 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2507 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2508 return;
2511 * Delay this task enough that another task of this mm will likely win
2512 * the next time around.
2514 p->node_stamp += 2 * TICK_NSEC;
2516 start = mm->numa_scan_offset;
2517 pages = sysctl_numa_balancing_scan_size;
2518 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2519 virtpages = pages * 8; /* Scan up to this much virtual space */
2520 if (!pages)
2521 return;
2524 if (!down_read_trylock(&mm->mmap_sem))
2525 return;
2526 vma = find_vma(mm, start);
2527 if (!vma) {
2528 reset_ptenuma_scan(p);
2529 start = 0;
2530 vma = mm->mmap;
2532 for (; vma; vma = vma->vm_next) {
2533 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2534 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2535 continue;
2539 * Shared library pages mapped by multiple processes are not
2540 * migrated as it is expected they are cache replicated. Avoid
2541 * hinting faults in read-only file-backed mappings or the vdso
2542 * as migrating the pages will be of marginal benefit.
2544 if (!vma->vm_mm ||
2545 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2546 continue;
2549 * Skip inaccessible VMAs to avoid any confusion between
2550 * PROT_NONE and NUMA hinting ptes
2552 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2553 continue;
2555 do {
2556 start = max(start, vma->vm_start);
2557 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2558 end = min(end, vma->vm_end);
2559 nr_pte_updates = change_prot_numa(vma, start, end);
2562 * Try to scan sysctl_numa_balancing_size worth of
2563 * hpages that have at least one present PTE that
2564 * is not already pte-numa. If the VMA contains
2565 * areas that are unused or already full of prot_numa
2566 * PTEs, scan up to virtpages, to skip through those
2567 * areas faster.
2569 if (nr_pte_updates)
2570 pages -= (end - start) >> PAGE_SHIFT;
2571 virtpages -= (end - start) >> PAGE_SHIFT;
2573 start = end;
2574 if (pages <= 0 || virtpages <= 0)
2575 goto out;
2577 cond_resched();
2578 } while (end != vma->vm_end);
2581 out:
2583 * It is possible to reach the end of the VMA list but the last few
2584 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2585 * would find the !migratable VMA on the next scan but not reset the
2586 * scanner to the start so check it now.
2588 if (vma)
2589 mm->numa_scan_offset = start;
2590 else
2591 reset_ptenuma_scan(p);
2592 up_read(&mm->mmap_sem);
2595 * Make sure tasks use at least 32x as much time to run other code
2596 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2597 * Usually update_task_scan_period slows down scanning enough; on an
2598 * overloaded system we need to limit overhead on a per task basis.
2600 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2601 u64 diff = p->se.sum_exec_runtime - runtime;
2602 p->node_stamp += 32 * diff;
2607 * Drive the periodic memory faults..
2609 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2611 struct callback_head *work = &curr->numa_work;
2612 u64 period, now;
2615 * We don't care about NUMA placement if we don't have memory.
2617 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2618 return;
2621 * Using runtime rather than walltime has the dual advantage that
2622 * we (mostly) drive the selection from busy threads and that the
2623 * task needs to have done some actual work before we bother with
2624 * NUMA placement.
2626 now = curr->se.sum_exec_runtime;
2627 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2629 if (now > curr->node_stamp + period) {
2630 if (!curr->node_stamp)
2631 curr->numa_scan_period = task_scan_start(curr);
2632 curr->node_stamp += period;
2634 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2635 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2636 task_work_add(curr, work, true);
2641 #else
2642 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2646 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2650 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2654 #endif /* CONFIG_NUMA_BALANCING */
2656 static void
2657 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2659 update_load_add(&cfs_rq->load, se->load.weight);
2660 if (!parent_entity(se))
2661 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2662 #ifdef CONFIG_SMP
2663 if (entity_is_task(se)) {
2664 struct rq *rq = rq_of(cfs_rq);
2666 account_numa_enqueue(rq, task_of(se));
2667 list_add(&se->group_node, &rq->cfs_tasks);
2669 #endif
2670 cfs_rq->nr_running++;
2673 static void
2674 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2676 update_load_sub(&cfs_rq->load, se->load.weight);
2677 if (!parent_entity(se))
2678 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2679 #ifdef CONFIG_SMP
2680 if (entity_is_task(se)) {
2681 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2682 list_del_init(&se->group_node);
2684 #endif
2685 cfs_rq->nr_running--;
2689 * Signed add and clamp on underflow.
2691 * Explicitly do a load-store to ensure the intermediate value never hits
2692 * memory. This allows lockless observations without ever seeing the negative
2693 * values.
2695 #define add_positive(_ptr, _val) do { \
2696 typeof(_ptr) ptr = (_ptr); \
2697 typeof(_val) val = (_val); \
2698 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2700 res = var + val; \
2702 if (val < 0 && res > var) \
2703 res = 0; \
2705 WRITE_ONCE(*ptr, res); \
2706 } while (0)
2709 * Unsigned subtract and clamp on underflow.
2711 * Explicitly do a load-store to ensure the intermediate value never hits
2712 * memory. This allows lockless observations without ever seeing the negative
2713 * values.
2715 #define sub_positive(_ptr, _val) do { \
2716 typeof(_ptr) ptr = (_ptr); \
2717 typeof(*ptr) val = (_val); \
2718 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2719 res = var - val; \
2720 if (res > var) \
2721 res = 0; \
2722 WRITE_ONCE(*ptr, res); \
2723 } while (0)
2725 #ifdef CONFIG_SMP
2727 * XXX we want to get rid of these helpers and use the full load resolution.
2729 static inline long se_weight(struct sched_entity *se)
2731 return scale_load_down(se->load.weight);
2734 static inline long se_runnable(struct sched_entity *se)
2736 return scale_load_down(se->runnable_weight);
2739 static inline void
2740 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2742 cfs_rq->runnable_weight += se->runnable_weight;
2744 cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg;
2745 cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum;
2748 static inline void
2749 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2751 cfs_rq->runnable_weight -= se->runnable_weight;
2753 sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg);
2754 sub_positive(&cfs_rq->avg.runnable_load_sum,
2755 se_runnable(se) * se->avg.runnable_load_sum);
2758 static inline void
2759 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2761 cfs_rq->avg.load_avg += se->avg.load_avg;
2762 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
2765 static inline void
2766 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2768 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2769 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
2771 #else
2772 static inline void
2773 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2774 static inline void
2775 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2776 static inline void
2777 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2778 static inline void
2779 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2780 #endif
2782 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2783 unsigned long weight, unsigned long runnable)
2785 if (se->on_rq) {
2786 /* commit outstanding execution time */
2787 if (cfs_rq->curr == se)
2788 update_curr(cfs_rq);
2789 account_entity_dequeue(cfs_rq, se);
2790 dequeue_runnable_load_avg(cfs_rq, se);
2792 dequeue_load_avg(cfs_rq, se);
2794 se->runnable_weight = runnable;
2795 update_load_set(&se->load, weight);
2797 #ifdef CONFIG_SMP
2798 do {
2799 u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
2801 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
2802 se->avg.runnable_load_avg =
2803 div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider);
2804 } while (0);
2805 #endif
2807 enqueue_load_avg(cfs_rq, se);
2808 if (se->on_rq) {
2809 account_entity_enqueue(cfs_rq, se);
2810 enqueue_runnable_load_avg(cfs_rq, se);
2814 void reweight_task(struct task_struct *p, int prio)
2816 struct sched_entity *se = &p->se;
2817 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2818 struct load_weight *load = &se->load;
2819 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
2821 reweight_entity(cfs_rq, se, weight, weight);
2822 load->inv_weight = sched_prio_to_wmult[prio];
2825 #ifdef CONFIG_FAIR_GROUP_SCHED
2826 # ifdef CONFIG_SMP
2828 * All this does is approximate the hierarchical proportion which includes that
2829 * global sum we all love to hate.
2831 * That is, the weight of a group entity, is the proportional share of the
2832 * group weight based on the group runqueue weights. That is:
2834 * tg->weight * grq->load.weight
2835 * ge->load.weight = ----------------------------- (1)
2836 * \Sum grq->load.weight
2838 * Now, because computing that sum is prohibitively expensive to compute (been
2839 * there, done that) we approximate it with this average stuff. The average
2840 * moves slower and therefore the approximation is cheaper and more stable.
2842 * So instead of the above, we substitute:
2844 * grq->load.weight -> grq->avg.load_avg (2)
2846 * which yields the following:
2848 * tg->weight * grq->avg.load_avg
2849 * ge->load.weight = ------------------------------ (3)
2850 * tg->load_avg
2852 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2854 * That is shares_avg, and it is right (given the approximation (2)).
2856 * The problem with it is that because the average is slow -- it was designed
2857 * to be exactly that of course -- this leads to transients in boundary
2858 * conditions. In specific, the case where the group was idle and we start the
2859 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2860 * yielding bad latency etc..
2862 * Now, in that special case (1) reduces to:
2864 * tg->weight * grq->load.weight
2865 * ge->load.weight = ----------------------------- = tg->weight (4)
2866 * grp->load.weight
2868 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2870 * So what we do is modify our approximation (3) to approach (4) in the (near)
2871 * UP case, like:
2873 * ge->load.weight =
2875 * tg->weight * grq->load.weight
2876 * --------------------------------------------------- (5)
2877 * tg->load_avg - grq->avg.load_avg + grq->load.weight
2879 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
2880 * we need to use grq->avg.load_avg as its lower bound, which then gives:
2883 * tg->weight * grq->load.weight
2884 * ge->load.weight = ----------------------------- (6)
2885 * tg_load_avg'
2887 * Where:
2889 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
2890 * max(grq->load.weight, grq->avg.load_avg)
2892 * And that is shares_weight and is icky. In the (near) UP case it approaches
2893 * (4) while in the normal case it approaches (3). It consistently
2894 * overestimates the ge->load.weight and therefore:
2896 * \Sum ge->load.weight >= tg->weight
2898 * hence icky!
2900 static long calc_group_shares(struct cfs_rq *cfs_rq)
2902 long tg_weight, tg_shares, load, shares;
2903 struct task_group *tg = cfs_rq->tg;
2905 tg_shares = READ_ONCE(tg->shares);
2907 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
2909 tg_weight = atomic_long_read(&tg->load_avg);
2911 /* Ensure tg_weight >= load */
2912 tg_weight -= cfs_rq->tg_load_avg_contrib;
2913 tg_weight += load;
2915 shares = (tg_shares * load);
2916 if (tg_weight)
2917 shares /= tg_weight;
2920 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2921 * of a group with small tg->shares value. It is a floor value which is
2922 * assigned as a minimum load.weight to the sched_entity representing
2923 * the group on a CPU.
2925 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2926 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2927 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2928 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2929 * instead of 0.
2931 return clamp_t(long, shares, MIN_SHARES, tg_shares);
2935 * This calculates the effective runnable weight for a group entity based on
2936 * the group entity weight calculated above.
2938 * Because of the above approximation (2), our group entity weight is
2939 * an load_avg based ratio (3). This means that it includes blocked load and
2940 * does not represent the runnable weight.
2942 * Approximate the group entity's runnable weight per ratio from the group
2943 * runqueue:
2945 * grq->avg.runnable_load_avg
2946 * ge->runnable_weight = ge->load.weight * -------------------------- (7)
2947 * grq->avg.load_avg
2949 * However, analogous to above, since the avg numbers are slow, this leads to
2950 * transients in the from-idle case. Instead we use:
2952 * ge->runnable_weight = ge->load.weight *
2954 * max(grq->avg.runnable_load_avg, grq->runnable_weight)
2955 * ----------------------------------------------------- (8)
2956 * max(grq->avg.load_avg, grq->load.weight)
2958 * Where these max() serve both to use the 'instant' values to fix the slow
2959 * from-idle and avoid the /0 on to-idle, similar to (6).
2961 static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares)
2963 long runnable, load_avg;
2965 load_avg = max(cfs_rq->avg.load_avg,
2966 scale_load_down(cfs_rq->load.weight));
2968 runnable = max(cfs_rq->avg.runnable_load_avg,
2969 scale_load_down(cfs_rq->runnable_weight));
2971 runnable *= shares;
2972 if (load_avg)
2973 runnable /= load_avg;
2975 return clamp_t(long, runnable, MIN_SHARES, shares);
2977 # endif /* CONFIG_SMP */
2979 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2982 * Recomputes the group entity based on the current state of its group
2983 * runqueue.
2985 static void update_cfs_group(struct sched_entity *se)
2987 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
2988 long shares, runnable;
2990 if (!gcfs_rq)
2991 return;
2993 if (throttled_hierarchy(gcfs_rq))
2994 return;
2996 #ifndef CONFIG_SMP
2997 runnable = shares = READ_ONCE(gcfs_rq->tg->shares);
2999 if (likely(se->load.weight == shares))
3000 return;
3001 #else
3002 shares = calc_group_shares(gcfs_rq);
3003 runnable = calc_group_runnable(gcfs_rq, shares);
3004 #endif
3006 reweight_entity(cfs_rq_of(se), se, shares, runnable);
3009 #else /* CONFIG_FAIR_GROUP_SCHED */
3010 static inline void update_cfs_group(struct sched_entity *se)
3013 #endif /* CONFIG_FAIR_GROUP_SCHED */
3015 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
3017 struct rq *rq = rq_of(cfs_rq);
3019 if (&rq->cfs == cfs_rq) {
3021 * There are a few boundary cases this might miss but it should
3022 * get called often enough that that should (hopefully) not be
3023 * a real problem -- added to that it only calls on the local
3024 * CPU, so if we enqueue remotely we'll miss an update, but
3025 * the next tick/schedule should update.
3027 * It will not get called when we go idle, because the idle
3028 * thread is a different class (!fair), nor will the utilization
3029 * number include things like RT tasks.
3031 * As is, the util number is not freq-invariant (we'd have to
3032 * implement arch_scale_freq_capacity() for that).
3034 * See cpu_util().
3036 cpufreq_update_util(rq, 0);
3040 #ifdef CONFIG_SMP
3042 * Approximate:
3043 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
3045 static u64 decay_load(u64 val, u64 n)
3047 unsigned int local_n;
3049 if (unlikely(n > LOAD_AVG_PERIOD * 63))
3050 return 0;
3052 /* after bounds checking we can collapse to 32-bit */
3053 local_n = n;
3056 * As y^PERIOD = 1/2, we can combine
3057 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
3058 * With a look-up table which covers y^n (n<PERIOD)
3060 * To achieve constant time decay_load.
3062 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
3063 val >>= local_n / LOAD_AVG_PERIOD;
3064 local_n %= LOAD_AVG_PERIOD;
3067 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
3068 return val;
3071 static u32 __accumulate_pelt_segments(u64 periods, u32 d1, u32 d3)
3073 u32 c1, c2, c3 = d3; /* y^0 == 1 */
3076 * c1 = d1 y^p
3078 c1 = decay_load((u64)d1, periods);
3081 * p-1
3082 * c2 = 1024 \Sum y^n
3083 * n=1
3085 * inf inf
3086 * = 1024 ( \Sum y^n - \Sum y^n - y^0 )
3087 * n=0 n=p
3089 c2 = LOAD_AVG_MAX - decay_load(LOAD_AVG_MAX, periods) - 1024;
3091 return c1 + c2 + c3;
3094 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
3097 * Accumulate the three separate parts of the sum; d1 the remainder
3098 * of the last (incomplete) period, d2 the span of full periods and d3
3099 * the remainder of the (incomplete) current period.
3101 * d1 d2 d3
3102 * ^ ^ ^
3103 * | | |
3104 * |<->|<----------------->|<--->|
3105 * ... |---x---|------| ... |------|-----x (now)
3107 * p-1
3108 * u' = (u + d1) y^p + 1024 \Sum y^n + d3 y^0
3109 * n=1
3111 * = u y^p + (Step 1)
3113 * p-1
3114 * d1 y^p + 1024 \Sum y^n + d3 y^0 (Step 2)
3115 * n=1
3117 static __always_inline u32
3118 accumulate_sum(u64 delta, int cpu, struct sched_avg *sa,
3119 unsigned long load, unsigned long runnable, int running)
3121 unsigned long scale_freq, scale_cpu;
3122 u32 contrib = (u32)delta; /* p == 0 -> delta < 1024 */
3123 u64 periods;
3125 scale_freq = arch_scale_freq_capacity(NULL, cpu);
3126 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
3128 delta += sa->period_contrib;
3129 periods = delta / 1024; /* A period is 1024us (~1ms) */
3132 * Step 1: decay old *_sum if we crossed period boundaries.
3134 if (periods) {
3135 sa->load_sum = decay_load(sa->load_sum, periods);
3136 sa->runnable_load_sum =
3137 decay_load(sa->runnable_load_sum, periods);
3138 sa->util_sum = decay_load((u64)(sa->util_sum), periods);
3141 * Step 2
3143 delta %= 1024;
3144 contrib = __accumulate_pelt_segments(periods,
3145 1024 - sa->period_contrib, delta);
3147 sa->period_contrib = delta;
3149 contrib = cap_scale(contrib, scale_freq);
3150 if (load)
3151 sa->load_sum += load * contrib;
3152 if (runnable)
3153 sa->runnable_load_sum += runnable * contrib;
3154 if (running)
3155 sa->util_sum += contrib * scale_cpu;
3157 return periods;
3161 * We can represent the historical contribution to runnable average as the
3162 * coefficients of a geometric series. To do this we sub-divide our runnable
3163 * history into segments of approximately 1ms (1024us); label the segment that
3164 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
3166 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
3167 * p0 p1 p2
3168 * (now) (~1ms ago) (~2ms ago)
3170 * Let u_i denote the fraction of p_i that the entity was runnable.
3172 * We then designate the fractions u_i as our co-efficients, yielding the
3173 * following representation of historical load:
3174 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
3176 * We choose y based on the with of a reasonably scheduling period, fixing:
3177 * y^32 = 0.5
3179 * This means that the contribution to load ~32ms ago (u_32) will be weighted
3180 * approximately half as much as the contribution to load within the last ms
3181 * (u_0).
3183 * When a period "rolls over" and we have new u_0`, multiplying the previous
3184 * sum again by y is sufficient to update:
3185 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
3186 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
3188 static __always_inline int
3189 ___update_load_sum(u64 now, int cpu, struct sched_avg *sa,
3190 unsigned long load, unsigned long runnable, int running)
3192 u64 delta;
3194 delta = now - sa->last_update_time;
3196 * This should only happen when time goes backwards, which it
3197 * unfortunately does during sched clock init when we swap over to TSC.
3199 if ((s64)delta < 0) {
3200 sa->last_update_time = now;
3201 return 0;
3205 * Use 1024ns as the unit of measurement since it's a reasonable
3206 * approximation of 1us and fast to compute.
3208 delta >>= 10;
3209 if (!delta)
3210 return 0;
3212 sa->last_update_time += delta << 10;
3215 * running is a subset of runnable (weight) so running can't be set if
3216 * runnable is clear. But there are some corner cases where the current
3217 * se has been already dequeued but cfs_rq->curr still points to it.
3218 * This means that weight will be 0 but not running for a sched_entity
3219 * but also for a cfs_rq if the latter becomes idle. As an example,
3220 * this happens during idle_balance() which calls
3221 * update_blocked_averages()
3223 if (!load)
3224 runnable = running = 0;
3227 * Now we know we crossed measurement unit boundaries. The *_avg
3228 * accrues by two steps:
3230 * Step 1: accumulate *_sum since last_update_time. If we haven't
3231 * crossed period boundaries, finish.
3233 if (!accumulate_sum(delta, cpu, sa, load, runnable, running))
3234 return 0;
3236 return 1;
3239 static __always_inline void
3240 ___update_load_avg(struct sched_avg *sa, unsigned long load, unsigned long runnable)
3242 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3245 * Step 2: update *_avg.
3247 sa->load_avg = div_u64(load * sa->load_sum, divider);
3248 sa->runnable_load_avg = div_u64(runnable * sa->runnable_load_sum, divider);
3249 sa->util_avg = sa->util_sum / divider;
3253 * sched_entity:
3255 * task:
3256 * se_runnable() == se_weight()
3258 * group: [ see update_cfs_group() ]
3259 * se_weight() = tg->weight * grq->load_avg / tg->load_avg
3260 * se_runnable() = se_weight(se) * grq->runnable_load_avg / grq->load_avg
3262 * load_sum := runnable_sum
3263 * load_avg = se_weight(se) * runnable_avg
3265 * runnable_load_sum := runnable_sum
3266 * runnable_load_avg = se_runnable(se) * runnable_avg
3268 * XXX collapse load_sum and runnable_load_sum
3270 * cfq_rs:
3272 * load_sum = \Sum se_weight(se) * se->avg.load_sum
3273 * load_avg = \Sum se->avg.load_avg
3275 * runnable_load_sum = \Sum se_runnable(se) * se->avg.runnable_load_sum
3276 * runnable_load_avg = \Sum se->avg.runable_load_avg
3279 static int
3280 __update_load_avg_blocked_se(u64 now, int cpu, struct sched_entity *se)
3282 if (entity_is_task(se))
3283 se->runnable_weight = se->load.weight;
3285 if (___update_load_sum(now, cpu, &se->avg, 0, 0, 0)) {
3286 ___update_load_avg(&se->avg, se_weight(se), se_runnable(se));
3287 return 1;
3290 return 0;
3293 static int
3294 __update_load_avg_se(u64 now, int cpu, struct cfs_rq *cfs_rq, struct sched_entity *se)
3296 if (entity_is_task(se))
3297 se->runnable_weight = se->load.weight;
3299 if (___update_load_sum(now, cpu, &se->avg, !!se->on_rq, !!se->on_rq,
3300 cfs_rq->curr == se)) {
3302 ___update_load_avg(&se->avg, se_weight(se), se_runnable(se));
3303 return 1;
3306 return 0;
3309 static int
3310 __update_load_avg_cfs_rq(u64 now, int cpu, struct cfs_rq *cfs_rq)
3312 if (___update_load_sum(now, cpu, &cfs_rq->avg,
3313 scale_load_down(cfs_rq->load.weight),
3314 scale_load_down(cfs_rq->runnable_weight),
3315 cfs_rq->curr != NULL)) {
3317 ___update_load_avg(&cfs_rq->avg, 1, 1);
3318 return 1;
3321 return 0;
3324 #ifdef CONFIG_FAIR_GROUP_SCHED
3326 * update_tg_load_avg - update the tg's load avg
3327 * @cfs_rq: the cfs_rq whose avg changed
3328 * @force: update regardless of how small the difference
3330 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3331 * However, because tg->load_avg is a global value there are performance
3332 * considerations.
3334 * In order to avoid having to look at the other cfs_rq's, we use a
3335 * differential update where we store the last value we propagated. This in
3336 * turn allows skipping updates if the differential is 'small'.
3338 * Updating tg's load_avg is necessary before update_cfs_share().
3340 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3342 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3345 * No need to update load_avg for root_task_group as it is not used.
3347 if (cfs_rq->tg == &root_task_group)
3348 return;
3350 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3351 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3352 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3357 * Called within set_task_rq() right before setting a task's cpu. The
3358 * caller only guarantees p->pi_lock is held; no other assumptions,
3359 * including the state of rq->lock, should be made.
3361 void set_task_rq_fair(struct sched_entity *se,
3362 struct cfs_rq *prev, struct cfs_rq *next)
3364 u64 p_last_update_time;
3365 u64 n_last_update_time;
3367 if (!sched_feat(ATTACH_AGE_LOAD))
3368 return;
3371 * We are supposed to update the task to "current" time, then its up to
3372 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3373 * getting what current time is, so simply throw away the out-of-date
3374 * time. This will result in the wakee task is less decayed, but giving
3375 * the wakee more load sounds not bad.
3377 if (!(se->avg.last_update_time && prev))
3378 return;
3380 #ifndef CONFIG_64BIT
3382 u64 p_last_update_time_copy;
3383 u64 n_last_update_time_copy;
3385 do {
3386 p_last_update_time_copy = prev->load_last_update_time_copy;
3387 n_last_update_time_copy = next->load_last_update_time_copy;
3389 smp_rmb();
3391 p_last_update_time = prev->avg.last_update_time;
3392 n_last_update_time = next->avg.last_update_time;
3394 } while (p_last_update_time != p_last_update_time_copy ||
3395 n_last_update_time != n_last_update_time_copy);
3397 #else
3398 p_last_update_time = prev->avg.last_update_time;
3399 n_last_update_time = next->avg.last_update_time;
3400 #endif
3401 __update_load_avg_blocked_se(p_last_update_time, cpu_of(rq_of(prev)), se);
3402 se->avg.last_update_time = n_last_update_time;
3407 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3408 * propagate its contribution. The key to this propagation is the invariant
3409 * that for each group:
3411 * ge->avg == grq->avg (1)
3413 * _IFF_ we look at the pure running and runnable sums. Because they
3414 * represent the very same entity, just at different points in the hierarchy.
3416 * Per the above update_tg_cfs_util() is trivial and simply copies the running
3417 * sum over (but still wrong, because the group entity and group rq do not have
3418 * their PELT windows aligned).
3420 * However, update_tg_cfs_runnable() is more complex. So we have:
3422 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3424 * And since, like util, the runnable part should be directly transferable,
3425 * the following would _appear_ to be the straight forward approach:
3427 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3429 * And per (1) we have:
3431 * ge->avg.runnable_avg == grq->avg.runnable_avg
3433 * Which gives:
3435 * ge->load.weight * grq->avg.load_avg
3436 * ge->avg.load_avg = ----------------------------------- (4)
3437 * grq->load.weight
3439 * Except that is wrong!
3441 * Because while for entities historical weight is not important and we
3442 * really only care about our future and therefore can consider a pure
3443 * runnable sum, runqueues can NOT do this.
3445 * We specifically want runqueues to have a load_avg that includes
3446 * historical weights. Those represent the blocked load, the load we expect
3447 * to (shortly) return to us. This only works by keeping the weights as
3448 * integral part of the sum. We therefore cannot decompose as per (3).
3450 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3451 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3452 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3453 * runnable section of these tasks overlap (or not). If they were to perfectly
3454 * align the rq as a whole would be runnable 2/3 of the time. If however we
3455 * always have at least 1 runnable task, the rq as a whole is always runnable.
3457 * So we'll have to approximate.. :/
3459 * Given the constraint:
3461 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3463 * We can construct a rule that adds runnable to a rq by assuming minimal
3464 * overlap.
3466 * On removal, we'll assume each task is equally runnable; which yields:
3468 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3470 * XXX: only do this for the part of runnable > running ?
3474 static inline void
3475 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3477 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3479 /* Nothing to update */
3480 if (!delta)
3481 return;
3484 * The relation between sum and avg is:
3486 * LOAD_AVG_MAX - 1024 + sa->period_contrib
3488 * however, the PELT windows are not aligned between grq and gse.
3491 /* Set new sched_entity's utilization */
3492 se->avg.util_avg = gcfs_rq->avg.util_avg;
3493 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3495 /* Update parent cfs_rq utilization */
3496 add_positive(&cfs_rq->avg.util_avg, delta);
3497 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3500 static inline void
3501 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3503 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3504 unsigned long runnable_load_avg, load_avg;
3505 u64 runnable_load_sum, load_sum = 0;
3506 s64 delta_sum;
3508 if (!runnable_sum)
3509 return;
3511 gcfs_rq->prop_runnable_sum = 0;
3513 if (runnable_sum >= 0) {
3515 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3516 * the CPU is saturated running == runnable.
3518 runnable_sum += se->avg.load_sum;
3519 runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
3520 } else {
3522 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3523 * assuming all tasks are equally runnable.
3525 if (scale_load_down(gcfs_rq->load.weight)) {
3526 load_sum = div_s64(gcfs_rq->avg.load_sum,
3527 scale_load_down(gcfs_rq->load.weight));
3530 /* But make sure to not inflate se's runnable */
3531 runnable_sum = min(se->avg.load_sum, load_sum);
3535 * runnable_sum can't be lower than running_sum
3536 * As running sum is scale with cpu capacity wehreas the runnable sum
3537 * is not we rescale running_sum 1st
3539 running_sum = se->avg.util_sum /
3540 arch_scale_cpu_capacity(NULL, cpu_of(rq_of(cfs_rq)));
3541 runnable_sum = max(runnable_sum, running_sum);
3543 load_sum = (s64)se_weight(se) * runnable_sum;
3544 load_avg = div_s64(load_sum, LOAD_AVG_MAX);
3546 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3547 delta_avg = load_avg - se->avg.load_avg;
3549 se->avg.load_sum = runnable_sum;
3550 se->avg.load_avg = load_avg;
3551 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3552 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3554 runnable_load_sum = (s64)se_runnable(se) * runnable_sum;
3555 runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX);
3556 delta_sum = runnable_load_sum - se_weight(se) * se->avg.runnable_load_sum;
3557 delta_avg = runnable_load_avg - se->avg.runnable_load_avg;
3559 se->avg.runnable_load_sum = runnable_sum;
3560 se->avg.runnable_load_avg = runnable_load_avg;
3562 if (se->on_rq) {
3563 add_positive(&cfs_rq->avg.runnable_load_avg, delta_avg);
3564 add_positive(&cfs_rq->avg.runnable_load_sum, delta_sum);
3568 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3570 cfs_rq->propagate = 1;
3571 cfs_rq->prop_runnable_sum += runnable_sum;
3574 /* Update task and its cfs_rq load average */
3575 static inline int propagate_entity_load_avg(struct sched_entity *se)
3577 struct cfs_rq *cfs_rq, *gcfs_rq;
3579 if (entity_is_task(se))
3580 return 0;
3582 gcfs_rq = group_cfs_rq(se);
3583 if (!gcfs_rq->propagate)
3584 return 0;
3586 gcfs_rq->propagate = 0;
3588 cfs_rq = cfs_rq_of(se);
3590 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3592 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3593 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3595 return 1;
3599 * Check if we need to update the load and the utilization of a blocked
3600 * group_entity:
3602 static inline bool skip_blocked_update(struct sched_entity *se)
3604 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3607 * If sched_entity still have not zero load or utilization, we have to
3608 * decay it:
3610 if (se->avg.load_avg || se->avg.util_avg)
3611 return false;
3614 * If there is a pending propagation, we have to update the load and
3615 * the utilization of the sched_entity:
3617 if (gcfs_rq->propagate)
3618 return false;
3621 * Otherwise, the load and the utilization of the sched_entity is
3622 * already zero and there is no pending propagation, so it will be a
3623 * waste of time to try to decay it:
3625 return true;
3628 #else /* CONFIG_FAIR_GROUP_SCHED */
3630 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3632 static inline int propagate_entity_load_avg(struct sched_entity *se)
3634 return 0;
3637 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3639 #endif /* CONFIG_FAIR_GROUP_SCHED */
3642 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3643 * @now: current time, as per cfs_rq_clock_task()
3644 * @cfs_rq: cfs_rq to update
3646 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3647 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3648 * post_init_entity_util_avg().
3650 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3652 * Returns true if the load decayed or we removed load.
3654 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3655 * call update_tg_load_avg() when this function returns true.
3657 static inline int
3658 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3660 unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0;
3661 struct sched_avg *sa = &cfs_rq->avg;
3662 int decayed = 0;
3664 if (cfs_rq->removed.nr) {
3665 unsigned long r;
3666 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3668 raw_spin_lock(&cfs_rq->removed.lock);
3669 swap(cfs_rq->removed.util_avg, removed_util);
3670 swap(cfs_rq->removed.load_avg, removed_load);
3671 swap(cfs_rq->removed.runnable_sum, removed_runnable_sum);
3672 cfs_rq->removed.nr = 0;
3673 raw_spin_unlock(&cfs_rq->removed.lock);
3675 r = removed_load;
3676 sub_positive(&sa->load_avg, r);
3677 sub_positive(&sa->load_sum, r * divider);
3679 r = removed_util;
3680 sub_positive(&sa->util_avg, r);
3681 sub_positive(&sa->util_sum, r * divider);
3683 add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum);
3685 decayed = 1;
3688 decayed |= __update_load_avg_cfs_rq(now, cpu_of(rq_of(cfs_rq)), cfs_rq);
3690 #ifndef CONFIG_64BIT
3691 smp_wmb();
3692 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3693 #endif
3695 if (decayed)
3696 cfs_rq_util_change(cfs_rq);
3698 return decayed;
3702 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3703 * @cfs_rq: cfs_rq to attach to
3704 * @se: sched_entity to attach
3706 * Must call update_cfs_rq_load_avg() before this, since we rely on
3707 * cfs_rq->avg.last_update_time being current.
3709 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3711 u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3714 * When we attach the @se to the @cfs_rq, we must align the decay
3715 * window because without that, really weird and wonderful things can
3716 * happen.
3718 * XXX illustrate
3720 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3721 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3724 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3725 * period_contrib. This isn't strictly correct, but since we're
3726 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3727 * _sum a little.
3729 se->avg.util_sum = se->avg.util_avg * divider;
3731 se->avg.load_sum = divider;
3732 if (se_weight(se)) {
3733 se->avg.load_sum =
3734 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3737 se->avg.runnable_load_sum = se->avg.load_sum;
3739 enqueue_load_avg(cfs_rq, se);
3740 cfs_rq->avg.util_avg += se->avg.util_avg;
3741 cfs_rq->avg.util_sum += se->avg.util_sum;
3743 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3745 cfs_rq_util_change(cfs_rq);
3749 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3750 * @cfs_rq: cfs_rq to detach from
3751 * @se: sched_entity to detach
3753 * Must call update_cfs_rq_load_avg() before this, since we rely on
3754 * cfs_rq->avg.last_update_time being current.
3756 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3758 dequeue_load_avg(cfs_rq, se);
3759 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3760 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3762 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3764 cfs_rq_util_change(cfs_rq);
3768 * Optional action to be done while updating the load average
3770 #define UPDATE_TG 0x1
3771 #define SKIP_AGE_LOAD 0x2
3772 #define DO_ATTACH 0x4
3774 /* Update task and its cfs_rq load average */
3775 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3777 u64 now = cfs_rq_clock_task(cfs_rq);
3778 struct rq *rq = rq_of(cfs_rq);
3779 int cpu = cpu_of(rq);
3780 int decayed;
3783 * Track task load average for carrying it to new CPU after migrated, and
3784 * track group sched_entity load average for task_h_load calc in migration
3786 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3787 __update_load_avg_se(now, cpu, cfs_rq, se);
3789 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3790 decayed |= propagate_entity_load_avg(se);
3792 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3794 attach_entity_load_avg(cfs_rq, se);
3795 update_tg_load_avg(cfs_rq, 0);
3797 } else if (decayed && (flags & UPDATE_TG))
3798 update_tg_load_avg(cfs_rq, 0);
3801 #ifndef CONFIG_64BIT
3802 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3804 u64 last_update_time_copy;
3805 u64 last_update_time;
3807 do {
3808 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3809 smp_rmb();
3810 last_update_time = cfs_rq->avg.last_update_time;
3811 } while (last_update_time != last_update_time_copy);
3813 return last_update_time;
3815 #else
3816 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3818 return cfs_rq->avg.last_update_time;
3820 #endif
3823 * Synchronize entity load avg of dequeued entity without locking
3824 * the previous rq.
3826 void sync_entity_load_avg(struct sched_entity *se)
3828 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3829 u64 last_update_time;
3831 last_update_time = cfs_rq_last_update_time(cfs_rq);
3832 __update_load_avg_blocked_se(last_update_time, cpu_of(rq_of(cfs_rq)), se);
3836 * Task first catches up with cfs_rq, and then subtract
3837 * itself from the cfs_rq (task must be off the queue now).
3839 void remove_entity_load_avg(struct sched_entity *se)
3841 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3842 unsigned long flags;
3845 * tasks cannot exit without having gone through wake_up_new_task() ->
3846 * post_init_entity_util_avg() which will have added things to the
3847 * cfs_rq, so we can remove unconditionally.
3849 * Similarly for groups, they will have passed through
3850 * post_init_entity_util_avg() before unregister_sched_fair_group()
3851 * calls this.
3854 sync_entity_load_avg(se);
3856 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3857 ++cfs_rq->removed.nr;
3858 cfs_rq->removed.util_avg += se->avg.util_avg;
3859 cfs_rq->removed.load_avg += se->avg.load_avg;
3860 cfs_rq->removed.runnable_sum += se->avg.load_sum; /* == runnable_sum */
3861 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3864 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3866 return cfs_rq->avg.runnable_load_avg;
3869 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3871 return cfs_rq->avg.load_avg;
3874 static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3876 #else /* CONFIG_SMP */
3878 static inline int
3879 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3881 return 0;
3884 #define UPDATE_TG 0x0
3885 #define SKIP_AGE_LOAD 0x0
3886 #define DO_ATTACH 0x0
3888 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
3890 cfs_rq_util_change(cfs_rq);
3893 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3895 static inline void
3896 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3897 static inline void
3898 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3900 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3902 return 0;
3905 #endif /* CONFIG_SMP */
3907 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3909 #ifdef CONFIG_SCHED_DEBUG
3910 s64 d = se->vruntime - cfs_rq->min_vruntime;
3912 if (d < 0)
3913 d = -d;
3915 if (d > 3*sysctl_sched_latency)
3916 schedstat_inc(cfs_rq->nr_spread_over);
3917 #endif
3920 static void
3921 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3923 u64 vruntime = cfs_rq->min_vruntime;
3926 * The 'current' period is already promised to the current tasks,
3927 * however the extra weight of the new task will slow them down a
3928 * little, place the new task so that it fits in the slot that
3929 * stays open at the end.
3931 if (initial && sched_feat(START_DEBIT))
3932 vruntime += sched_vslice(cfs_rq, se);
3934 /* sleeps up to a single latency don't count. */
3935 if (!initial) {
3936 unsigned long thresh = sysctl_sched_latency;
3939 * Halve their sleep time's effect, to allow
3940 * for a gentler effect of sleepers:
3942 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3943 thresh >>= 1;
3945 vruntime -= thresh;
3948 /* ensure we never gain time by being placed backwards. */
3949 se->vruntime = max_vruntime(se->vruntime, vruntime);
3952 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3954 static inline void check_schedstat_required(void)
3956 #ifdef CONFIG_SCHEDSTATS
3957 if (schedstat_enabled())
3958 return;
3960 /* Force schedstat enabled if a dependent tracepoint is active */
3961 if (trace_sched_stat_wait_enabled() ||
3962 trace_sched_stat_sleep_enabled() ||
3963 trace_sched_stat_iowait_enabled() ||
3964 trace_sched_stat_blocked_enabled() ||
3965 trace_sched_stat_runtime_enabled()) {
3966 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3967 "stat_blocked and stat_runtime require the "
3968 "kernel parameter schedstats=enable or "
3969 "kernel.sched_schedstats=1\n");
3971 #endif
3976 * MIGRATION
3978 * dequeue
3979 * update_curr()
3980 * update_min_vruntime()
3981 * vruntime -= min_vruntime
3983 * enqueue
3984 * update_curr()
3985 * update_min_vruntime()
3986 * vruntime += min_vruntime
3988 * this way the vruntime transition between RQs is done when both
3989 * min_vruntime are up-to-date.
3991 * WAKEUP (remote)
3993 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3994 * vruntime -= min_vruntime
3996 * enqueue
3997 * update_curr()
3998 * update_min_vruntime()
3999 * vruntime += min_vruntime
4001 * this way we don't have the most up-to-date min_vruntime on the originating
4002 * CPU and an up-to-date min_vruntime on the destination CPU.
4005 static void
4006 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4008 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
4009 bool curr = cfs_rq->curr == se;
4012 * If we're the current task, we must renormalise before calling
4013 * update_curr().
4015 if (renorm && curr)
4016 se->vruntime += cfs_rq->min_vruntime;
4018 update_curr(cfs_rq);
4021 * Otherwise, renormalise after, such that we're placed at the current
4022 * moment in time, instead of some random moment in the past. Being
4023 * placed in the past could significantly boost this task to the
4024 * fairness detriment of existing tasks.
4026 if (renorm && !curr)
4027 se->vruntime += cfs_rq->min_vruntime;
4030 * When enqueuing a sched_entity, we must:
4031 * - Update loads to have both entity and cfs_rq synced with now.
4032 * - Add its load to cfs_rq->runnable_avg
4033 * - For group_entity, update its weight to reflect the new share of
4034 * its group cfs_rq
4035 * - Add its new weight to cfs_rq->load.weight
4037 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
4038 update_cfs_group(se);
4039 enqueue_runnable_load_avg(cfs_rq, se);
4040 account_entity_enqueue(cfs_rq, se);
4042 if (flags & ENQUEUE_WAKEUP)
4043 place_entity(cfs_rq, se, 0);
4045 check_schedstat_required();
4046 update_stats_enqueue(cfs_rq, se, flags);
4047 check_spread(cfs_rq, se);
4048 if (!curr)
4049 __enqueue_entity(cfs_rq, se);
4050 se->on_rq = 1;
4052 if (cfs_rq->nr_running == 1) {
4053 list_add_leaf_cfs_rq(cfs_rq);
4054 check_enqueue_throttle(cfs_rq);
4058 static void __clear_buddies_last(struct sched_entity *se)
4060 for_each_sched_entity(se) {
4061 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4062 if (cfs_rq->last != se)
4063 break;
4065 cfs_rq->last = NULL;
4069 static void __clear_buddies_next(struct sched_entity *se)
4071 for_each_sched_entity(se) {
4072 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4073 if (cfs_rq->next != se)
4074 break;
4076 cfs_rq->next = NULL;
4080 static void __clear_buddies_skip(struct sched_entity *se)
4082 for_each_sched_entity(se) {
4083 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4084 if (cfs_rq->skip != se)
4085 break;
4087 cfs_rq->skip = NULL;
4091 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4093 if (cfs_rq->last == se)
4094 __clear_buddies_last(se);
4096 if (cfs_rq->next == se)
4097 __clear_buddies_next(se);
4099 if (cfs_rq->skip == se)
4100 __clear_buddies_skip(se);
4103 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4105 static void
4106 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4109 * Update run-time statistics of the 'current'.
4111 update_curr(cfs_rq);
4114 * When dequeuing a sched_entity, we must:
4115 * - Update loads to have both entity and cfs_rq synced with now.
4116 * - Substract its load from the cfs_rq->runnable_avg.
4117 * - Substract its previous weight from cfs_rq->load.weight.
4118 * - For group entity, update its weight to reflect the new share
4119 * of its group cfs_rq.
4121 update_load_avg(cfs_rq, se, UPDATE_TG);
4122 dequeue_runnable_load_avg(cfs_rq, se);
4124 update_stats_dequeue(cfs_rq, se, flags);
4126 clear_buddies(cfs_rq, se);
4128 if (se != cfs_rq->curr)
4129 __dequeue_entity(cfs_rq, se);
4130 se->on_rq = 0;
4131 account_entity_dequeue(cfs_rq, se);
4134 * Normalize after update_curr(); which will also have moved
4135 * min_vruntime if @se is the one holding it back. But before doing
4136 * update_min_vruntime() again, which will discount @se's position and
4137 * can move min_vruntime forward still more.
4139 if (!(flags & DEQUEUE_SLEEP))
4140 se->vruntime -= cfs_rq->min_vruntime;
4142 /* return excess runtime on last dequeue */
4143 return_cfs_rq_runtime(cfs_rq);
4145 update_cfs_group(se);
4148 * Now advance min_vruntime if @se was the entity holding it back,
4149 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4150 * put back on, and if we advance min_vruntime, we'll be placed back
4151 * further than we started -- ie. we'll be penalized.
4153 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
4154 update_min_vruntime(cfs_rq);
4158 * Preempt the current task with a newly woken task if needed:
4160 static void
4161 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4163 unsigned long ideal_runtime, delta_exec;
4164 struct sched_entity *se;
4165 s64 delta;
4167 ideal_runtime = sched_slice(cfs_rq, curr);
4168 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4169 if (delta_exec > ideal_runtime) {
4170 resched_curr(rq_of(cfs_rq));
4172 * The current task ran long enough, ensure it doesn't get
4173 * re-elected due to buddy favours.
4175 clear_buddies(cfs_rq, curr);
4176 return;
4180 * Ensure that a task that missed wakeup preemption by a
4181 * narrow margin doesn't have to wait for a full slice.
4182 * This also mitigates buddy induced latencies under load.
4184 if (delta_exec < sysctl_sched_min_granularity)
4185 return;
4187 se = __pick_first_entity(cfs_rq);
4188 delta = curr->vruntime - se->vruntime;
4190 if (delta < 0)
4191 return;
4193 if (delta > ideal_runtime)
4194 resched_curr(rq_of(cfs_rq));
4197 static void
4198 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4200 /* 'current' is not kept within the tree. */
4201 if (se->on_rq) {
4203 * Any task has to be enqueued before it get to execute on
4204 * a CPU. So account for the time it spent waiting on the
4205 * runqueue.
4207 update_stats_wait_end(cfs_rq, se);
4208 __dequeue_entity(cfs_rq, se);
4209 update_load_avg(cfs_rq, se, UPDATE_TG);
4212 update_stats_curr_start(cfs_rq, se);
4213 cfs_rq->curr = se;
4216 * Track our maximum slice length, if the CPU's load is at
4217 * least twice that of our own weight (i.e. dont track it
4218 * when there are only lesser-weight tasks around):
4220 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
4221 schedstat_set(se->statistics.slice_max,
4222 max((u64)schedstat_val(se->statistics.slice_max),
4223 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4226 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4229 static int
4230 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4233 * Pick the next process, keeping these things in mind, in this order:
4234 * 1) keep things fair between processes/task groups
4235 * 2) pick the "next" process, since someone really wants that to run
4236 * 3) pick the "last" process, for cache locality
4237 * 4) do not run the "skip" process, if something else is available
4239 static struct sched_entity *
4240 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4242 struct sched_entity *left = __pick_first_entity(cfs_rq);
4243 struct sched_entity *se;
4246 * If curr is set we have to see if its left of the leftmost entity
4247 * still in the tree, provided there was anything in the tree at all.
4249 if (!left || (curr && entity_before(curr, left)))
4250 left = curr;
4252 se = left; /* ideally we run the leftmost entity */
4255 * Avoid running the skip buddy, if running something else can
4256 * be done without getting too unfair.
4258 if (cfs_rq->skip == se) {
4259 struct sched_entity *second;
4261 if (se == curr) {
4262 second = __pick_first_entity(cfs_rq);
4263 } else {
4264 second = __pick_next_entity(se);
4265 if (!second || (curr && entity_before(curr, second)))
4266 second = curr;
4269 if (second && wakeup_preempt_entity(second, left) < 1)
4270 se = second;
4274 * Prefer last buddy, try to return the CPU to a preempted task.
4276 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4277 se = cfs_rq->last;
4280 * Someone really wants this to run. If it's not unfair, run it.
4282 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4283 se = cfs_rq->next;
4285 clear_buddies(cfs_rq, se);
4287 return se;
4290 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4292 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4295 * If still on the runqueue then deactivate_task()
4296 * was not called and update_curr() has to be done:
4298 if (prev->on_rq)
4299 update_curr(cfs_rq);
4301 /* throttle cfs_rqs exceeding runtime */
4302 check_cfs_rq_runtime(cfs_rq);
4304 check_spread(cfs_rq, prev);
4306 if (prev->on_rq) {
4307 update_stats_wait_start(cfs_rq, prev);
4308 /* Put 'current' back into the tree. */
4309 __enqueue_entity(cfs_rq, prev);
4310 /* in !on_rq case, update occurred at dequeue */
4311 update_load_avg(cfs_rq, prev, 0);
4313 cfs_rq->curr = NULL;
4316 static void
4317 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4320 * Update run-time statistics of the 'current'.
4322 update_curr(cfs_rq);
4325 * Ensure that runnable average is periodically updated.
4327 update_load_avg(cfs_rq, curr, UPDATE_TG);
4328 update_cfs_group(curr);
4330 #ifdef CONFIG_SCHED_HRTICK
4332 * queued ticks are scheduled to match the slice, so don't bother
4333 * validating it and just reschedule.
4335 if (queued) {
4336 resched_curr(rq_of(cfs_rq));
4337 return;
4340 * don't let the period tick interfere with the hrtick preemption
4342 if (!sched_feat(DOUBLE_TICK) &&
4343 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4344 return;
4345 #endif
4347 if (cfs_rq->nr_running > 1)
4348 check_preempt_tick(cfs_rq, curr);
4352 /**************************************************
4353 * CFS bandwidth control machinery
4356 #ifdef CONFIG_CFS_BANDWIDTH
4358 #ifdef HAVE_JUMP_LABEL
4359 static struct static_key __cfs_bandwidth_used;
4361 static inline bool cfs_bandwidth_used(void)
4363 return static_key_false(&__cfs_bandwidth_used);
4366 void cfs_bandwidth_usage_inc(void)
4368 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4371 void cfs_bandwidth_usage_dec(void)
4373 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4375 #else /* HAVE_JUMP_LABEL */
4376 static bool cfs_bandwidth_used(void)
4378 return true;
4381 void cfs_bandwidth_usage_inc(void) {}
4382 void cfs_bandwidth_usage_dec(void) {}
4383 #endif /* HAVE_JUMP_LABEL */
4386 * default period for cfs group bandwidth.
4387 * default: 0.1s, units: nanoseconds
4389 static inline u64 default_cfs_period(void)
4391 return 100000000ULL;
4394 static inline u64 sched_cfs_bandwidth_slice(void)
4396 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4400 * Replenish runtime according to assigned quota and update expiration time.
4401 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4402 * additional synchronization around rq->lock.
4404 * requires cfs_b->lock
4406 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4408 u64 now;
4410 if (cfs_b->quota == RUNTIME_INF)
4411 return;
4413 now = sched_clock_cpu(smp_processor_id());
4414 cfs_b->runtime = cfs_b->quota;
4415 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
4418 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4420 return &tg->cfs_bandwidth;
4423 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4424 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4426 if (unlikely(cfs_rq->throttle_count))
4427 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4429 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4432 /* returns 0 on failure to allocate runtime */
4433 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4435 struct task_group *tg = cfs_rq->tg;
4436 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4437 u64 amount = 0, min_amount, expires;
4439 /* note: this is a positive sum as runtime_remaining <= 0 */
4440 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4442 raw_spin_lock(&cfs_b->lock);
4443 if (cfs_b->quota == RUNTIME_INF)
4444 amount = min_amount;
4445 else {
4446 start_cfs_bandwidth(cfs_b);
4448 if (cfs_b->runtime > 0) {
4449 amount = min(cfs_b->runtime, min_amount);
4450 cfs_b->runtime -= amount;
4451 cfs_b->idle = 0;
4454 expires = cfs_b->runtime_expires;
4455 raw_spin_unlock(&cfs_b->lock);
4457 cfs_rq->runtime_remaining += amount;
4459 * we may have advanced our local expiration to account for allowed
4460 * spread between our sched_clock and the one on which runtime was
4461 * issued.
4463 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
4464 cfs_rq->runtime_expires = expires;
4466 return cfs_rq->runtime_remaining > 0;
4470 * Note: This depends on the synchronization provided by sched_clock and the
4471 * fact that rq->clock snapshots this value.
4473 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4475 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4477 /* if the deadline is ahead of our clock, nothing to do */
4478 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
4479 return;
4481 if (cfs_rq->runtime_remaining < 0)
4482 return;
4485 * If the local deadline has passed we have to consider the
4486 * possibility that our sched_clock is 'fast' and the global deadline
4487 * has not truly expired.
4489 * Fortunately we can check determine whether this the case by checking
4490 * whether the global deadline has advanced. It is valid to compare
4491 * cfs_b->runtime_expires without any locks since we only care about
4492 * exact equality, so a partial write will still work.
4495 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
4496 /* extend local deadline, drift is bounded above by 2 ticks */
4497 cfs_rq->runtime_expires += TICK_NSEC;
4498 } else {
4499 /* global deadline is ahead, expiration has passed */
4500 cfs_rq->runtime_remaining = 0;
4504 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4506 /* dock delta_exec before expiring quota (as it could span periods) */
4507 cfs_rq->runtime_remaining -= delta_exec;
4508 expire_cfs_rq_runtime(cfs_rq);
4510 if (likely(cfs_rq->runtime_remaining > 0))
4511 return;
4514 * if we're unable to extend our runtime we resched so that the active
4515 * hierarchy can be throttled
4517 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4518 resched_curr(rq_of(cfs_rq));
4521 static __always_inline
4522 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4524 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4525 return;
4527 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4530 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4532 return cfs_bandwidth_used() && cfs_rq->throttled;
4535 /* check whether cfs_rq, or any parent, is throttled */
4536 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4538 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4542 * Ensure that neither of the group entities corresponding to src_cpu or
4543 * dest_cpu are members of a throttled hierarchy when performing group
4544 * load-balance operations.
4546 static inline int throttled_lb_pair(struct task_group *tg,
4547 int src_cpu, int dest_cpu)
4549 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4551 src_cfs_rq = tg->cfs_rq[src_cpu];
4552 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4554 return throttled_hierarchy(src_cfs_rq) ||
4555 throttled_hierarchy(dest_cfs_rq);
4558 /* updated child weight may affect parent so we have to do this bottom up */
4559 static int tg_unthrottle_up(struct task_group *tg, void *data)
4561 struct rq *rq = data;
4562 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4564 cfs_rq->throttle_count--;
4565 if (!cfs_rq->throttle_count) {
4566 /* adjust cfs_rq_clock_task() */
4567 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4568 cfs_rq->throttled_clock_task;
4571 return 0;
4574 static int tg_throttle_down(struct task_group *tg, void *data)
4576 struct rq *rq = data;
4577 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4579 /* group is entering throttled state, stop time */
4580 if (!cfs_rq->throttle_count)
4581 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4582 cfs_rq->throttle_count++;
4584 return 0;
4587 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4589 struct rq *rq = rq_of(cfs_rq);
4590 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4591 struct sched_entity *se;
4592 long task_delta, dequeue = 1;
4593 bool empty;
4595 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4597 /* freeze hierarchy runnable averages while throttled */
4598 rcu_read_lock();
4599 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4600 rcu_read_unlock();
4602 task_delta = cfs_rq->h_nr_running;
4603 for_each_sched_entity(se) {
4604 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4605 /* throttled entity or throttle-on-deactivate */
4606 if (!se->on_rq)
4607 break;
4609 if (dequeue)
4610 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4611 qcfs_rq->h_nr_running -= task_delta;
4613 if (qcfs_rq->load.weight)
4614 dequeue = 0;
4617 if (!se)
4618 sub_nr_running(rq, task_delta);
4620 cfs_rq->throttled = 1;
4621 cfs_rq->throttled_clock = rq_clock(rq);
4622 raw_spin_lock(&cfs_b->lock);
4623 empty = list_empty(&cfs_b->throttled_cfs_rq);
4626 * Add to the _head_ of the list, so that an already-started
4627 * distribute_cfs_runtime will not see us
4629 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4632 * If we're the first throttled task, make sure the bandwidth
4633 * timer is running.
4635 if (empty)
4636 start_cfs_bandwidth(cfs_b);
4638 raw_spin_unlock(&cfs_b->lock);
4641 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4643 struct rq *rq = rq_of(cfs_rq);
4644 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4645 struct sched_entity *se;
4646 int enqueue = 1;
4647 long task_delta;
4649 se = cfs_rq->tg->se[cpu_of(rq)];
4651 cfs_rq->throttled = 0;
4653 update_rq_clock(rq);
4655 raw_spin_lock(&cfs_b->lock);
4656 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4657 list_del_rcu(&cfs_rq->throttled_list);
4658 raw_spin_unlock(&cfs_b->lock);
4660 /* update hierarchical throttle state */
4661 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4663 if (!cfs_rq->load.weight)
4664 return;
4666 task_delta = cfs_rq->h_nr_running;
4667 for_each_sched_entity(se) {
4668 if (se->on_rq)
4669 enqueue = 0;
4671 cfs_rq = cfs_rq_of(se);
4672 if (enqueue)
4673 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4674 cfs_rq->h_nr_running += task_delta;
4676 if (cfs_rq_throttled(cfs_rq))
4677 break;
4680 if (!se)
4681 add_nr_running(rq, task_delta);
4683 /* determine whether we need to wake up potentially idle cpu */
4684 if (rq->curr == rq->idle && rq->cfs.nr_running)
4685 resched_curr(rq);
4688 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4689 u64 remaining, u64 expires)
4691 struct cfs_rq *cfs_rq;
4692 u64 runtime;
4693 u64 starting_runtime = remaining;
4695 rcu_read_lock();
4696 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4697 throttled_list) {
4698 struct rq *rq = rq_of(cfs_rq);
4699 struct rq_flags rf;
4701 rq_lock(rq, &rf);
4702 if (!cfs_rq_throttled(cfs_rq))
4703 goto next;
4705 runtime = -cfs_rq->runtime_remaining + 1;
4706 if (runtime > remaining)
4707 runtime = remaining;
4708 remaining -= runtime;
4710 cfs_rq->runtime_remaining += runtime;
4711 cfs_rq->runtime_expires = expires;
4713 /* we check whether we're throttled above */
4714 if (cfs_rq->runtime_remaining > 0)
4715 unthrottle_cfs_rq(cfs_rq);
4717 next:
4718 rq_unlock(rq, &rf);
4720 if (!remaining)
4721 break;
4723 rcu_read_unlock();
4725 return starting_runtime - remaining;
4729 * Responsible for refilling a task_group's bandwidth and unthrottling its
4730 * cfs_rqs as appropriate. If there has been no activity within the last
4731 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4732 * used to track this state.
4734 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4736 u64 runtime, runtime_expires;
4737 int throttled;
4739 /* no need to continue the timer with no bandwidth constraint */
4740 if (cfs_b->quota == RUNTIME_INF)
4741 goto out_deactivate;
4743 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4744 cfs_b->nr_periods += overrun;
4747 * idle depends on !throttled (for the case of a large deficit), and if
4748 * we're going inactive then everything else can be deferred
4750 if (cfs_b->idle && !throttled)
4751 goto out_deactivate;
4753 __refill_cfs_bandwidth_runtime(cfs_b);
4755 if (!throttled) {
4756 /* mark as potentially idle for the upcoming period */
4757 cfs_b->idle = 1;
4758 return 0;
4761 /* account preceding periods in which throttling occurred */
4762 cfs_b->nr_throttled += overrun;
4764 runtime_expires = cfs_b->runtime_expires;
4767 * This check is repeated as we are holding onto the new bandwidth while
4768 * we unthrottle. This can potentially race with an unthrottled group
4769 * trying to acquire new bandwidth from the global pool. This can result
4770 * in us over-using our runtime if it is all used during this loop, but
4771 * only by limited amounts in that extreme case.
4773 while (throttled && cfs_b->runtime > 0) {
4774 runtime = cfs_b->runtime;
4775 raw_spin_unlock(&cfs_b->lock);
4776 /* we can't nest cfs_b->lock while distributing bandwidth */
4777 runtime = distribute_cfs_runtime(cfs_b, runtime,
4778 runtime_expires);
4779 raw_spin_lock(&cfs_b->lock);
4781 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4783 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4787 * While we are ensured activity in the period following an
4788 * unthrottle, this also covers the case in which the new bandwidth is
4789 * insufficient to cover the existing bandwidth deficit. (Forcing the
4790 * timer to remain active while there are any throttled entities.)
4792 cfs_b->idle = 0;
4794 return 0;
4796 out_deactivate:
4797 return 1;
4800 /* a cfs_rq won't donate quota below this amount */
4801 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4802 /* minimum remaining period time to redistribute slack quota */
4803 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4804 /* how long we wait to gather additional slack before distributing */
4805 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4808 * Are we near the end of the current quota period?
4810 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4811 * hrtimer base being cleared by hrtimer_start. In the case of
4812 * migrate_hrtimers, base is never cleared, so we are fine.
4814 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4816 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4817 u64 remaining;
4819 /* if the call-back is running a quota refresh is already occurring */
4820 if (hrtimer_callback_running(refresh_timer))
4821 return 1;
4823 /* is a quota refresh about to occur? */
4824 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4825 if (remaining < min_expire)
4826 return 1;
4828 return 0;
4831 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4833 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4835 /* if there's a quota refresh soon don't bother with slack */
4836 if (runtime_refresh_within(cfs_b, min_left))
4837 return;
4839 hrtimer_start(&cfs_b->slack_timer,
4840 ns_to_ktime(cfs_bandwidth_slack_period),
4841 HRTIMER_MODE_REL);
4844 /* we know any runtime found here is valid as update_curr() precedes return */
4845 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4847 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4848 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4850 if (slack_runtime <= 0)
4851 return;
4853 raw_spin_lock(&cfs_b->lock);
4854 if (cfs_b->quota != RUNTIME_INF &&
4855 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4856 cfs_b->runtime += slack_runtime;
4858 /* we are under rq->lock, defer unthrottling using a timer */
4859 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4860 !list_empty(&cfs_b->throttled_cfs_rq))
4861 start_cfs_slack_bandwidth(cfs_b);
4863 raw_spin_unlock(&cfs_b->lock);
4865 /* even if it's not valid for return we don't want to try again */
4866 cfs_rq->runtime_remaining -= slack_runtime;
4869 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4871 if (!cfs_bandwidth_used())
4872 return;
4874 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4875 return;
4877 __return_cfs_rq_runtime(cfs_rq);
4881 * This is done with a timer (instead of inline with bandwidth return) since
4882 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4884 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4886 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4887 u64 expires;
4889 /* confirm we're still not at a refresh boundary */
4890 raw_spin_lock(&cfs_b->lock);
4891 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4892 raw_spin_unlock(&cfs_b->lock);
4893 return;
4896 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4897 runtime = cfs_b->runtime;
4899 expires = cfs_b->runtime_expires;
4900 raw_spin_unlock(&cfs_b->lock);
4902 if (!runtime)
4903 return;
4905 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4907 raw_spin_lock(&cfs_b->lock);
4908 if (expires == cfs_b->runtime_expires)
4909 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4910 raw_spin_unlock(&cfs_b->lock);
4914 * When a group wakes up we want to make sure that its quota is not already
4915 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4916 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4918 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4920 if (!cfs_bandwidth_used())
4921 return;
4923 /* an active group must be handled by the update_curr()->put() path */
4924 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4925 return;
4927 /* ensure the group is not already throttled */
4928 if (cfs_rq_throttled(cfs_rq))
4929 return;
4931 /* update runtime allocation */
4932 account_cfs_rq_runtime(cfs_rq, 0);
4933 if (cfs_rq->runtime_remaining <= 0)
4934 throttle_cfs_rq(cfs_rq);
4937 static void sync_throttle(struct task_group *tg, int cpu)
4939 struct cfs_rq *pcfs_rq, *cfs_rq;
4941 if (!cfs_bandwidth_used())
4942 return;
4944 if (!tg->parent)
4945 return;
4947 cfs_rq = tg->cfs_rq[cpu];
4948 pcfs_rq = tg->parent->cfs_rq[cpu];
4950 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4951 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4954 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4955 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4957 if (!cfs_bandwidth_used())
4958 return false;
4960 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4961 return false;
4964 * it's possible for a throttled entity to be forced into a running
4965 * state (e.g. set_curr_task), in this case we're finished.
4967 if (cfs_rq_throttled(cfs_rq))
4968 return true;
4970 throttle_cfs_rq(cfs_rq);
4971 return true;
4974 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4976 struct cfs_bandwidth *cfs_b =
4977 container_of(timer, struct cfs_bandwidth, slack_timer);
4979 do_sched_cfs_slack_timer(cfs_b);
4981 return HRTIMER_NORESTART;
4984 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4986 struct cfs_bandwidth *cfs_b =
4987 container_of(timer, struct cfs_bandwidth, period_timer);
4988 int overrun;
4989 int idle = 0;
4991 raw_spin_lock(&cfs_b->lock);
4992 for (;;) {
4993 overrun = hrtimer_forward_now(timer, cfs_b->period);
4994 if (!overrun)
4995 break;
4997 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4999 if (idle)
5000 cfs_b->period_active = 0;
5001 raw_spin_unlock(&cfs_b->lock);
5003 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
5006 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5008 raw_spin_lock_init(&cfs_b->lock);
5009 cfs_b->runtime = 0;
5010 cfs_b->quota = RUNTIME_INF;
5011 cfs_b->period = ns_to_ktime(default_cfs_period());
5013 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
5014 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
5015 cfs_b->period_timer.function = sched_cfs_period_timer;
5016 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5017 cfs_b->slack_timer.function = sched_cfs_slack_timer;
5020 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5022 cfs_rq->runtime_enabled = 0;
5023 INIT_LIST_HEAD(&cfs_rq->throttled_list);
5026 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5028 lockdep_assert_held(&cfs_b->lock);
5030 if (!cfs_b->period_active) {
5031 cfs_b->period_active = 1;
5032 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
5033 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
5037 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5039 /* init_cfs_bandwidth() was not called */
5040 if (!cfs_b->throttled_cfs_rq.next)
5041 return;
5043 hrtimer_cancel(&cfs_b->period_timer);
5044 hrtimer_cancel(&cfs_b->slack_timer);
5048 * Both these cpu hotplug callbacks race against unregister_fair_sched_group()
5050 * The race is harmless, since modifying bandwidth settings of unhooked group
5051 * bits doesn't do much.
5054 /* cpu online calback */
5055 static void __maybe_unused update_runtime_enabled(struct rq *rq)
5057 struct task_group *tg;
5059 lockdep_assert_held(&rq->lock);
5061 rcu_read_lock();
5062 list_for_each_entry_rcu(tg, &task_groups, list) {
5063 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5064 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5066 raw_spin_lock(&cfs_b->lock);
5067 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5068 raw_spin_unlock(&cfs_b->lock);
5070 rcu_read_unlock();
5073 /* cpu offline callback */
5074 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5076 struct task_group *tg;
5078 lockdep_assert_held(&rq->lock);
5080 rcu_read_lock();
5081 list_for_each_entry_rcu(tg, &task_groups, list) {
5082 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5084 if (!cfs_rq->runtime_enabled)
5085 continue;
5088 * clock_task is not advancing so we just need to make sure
5089 * there's some valid quota amount
5091 cfs_rq->runtime_remaining = 1;
5093 * Offline rq is schedulable till cpu is completely disabled
5094 * in take_cpu_down(), so we prevent new cfs throttling here.
5096 cfs_rq->runtime_enabled = 0;
5098 if (cfs_rq_throttled(cfs_rq))
5099 unthrottle_cfs_rq(cfs_rq);
5101 rcu_read_unlock();
5104 #else /* CONFIG_CFS_BANDWIDTH */
5105 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
5107 return rq_clock_task(rq_of(cfs_rq));
5110 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5111 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5112 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5113 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5114 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5116 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5118 return 0;
5121 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5123 return 0;
5126 static inline int throttled_lb_pair(struct task_group *tg,
5127 int src_cpu, int dest_cpu)
5129 return 0;
5132 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5134 #ifdef CONFIG_FAIR_GROUP_SCHED
5135 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5136 #endif
5138 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5140 return NULL;
5142 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5143 static inline void update_runtime_enabled(struct rq *rq) {}
5144 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5146 #endif /* CONFIG_CFS_BANDWIDTH */
5148 /**************************************************
5149 * CFS operations on tasks:
5152 #ifdef CONFIG_SCHED_HRTICK
5153 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5155 struct sched_entity *se = &p->se;
5156 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5158 SCHED_WARN_ON(task_rq(p) != rq);
5160 if (rq->cfs.h_nr_running > 1) {
5161 u64 slice = sched_slice(cfs_rq, se);
5162 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5163 s64 delta = slice - ran;
5165 if (delta < 0) {
5166 if (rq->curr == p)
5167 resched_curr(rq);
5168 return;
5170 hrtick_start(rq, delta);
5175 * called from enqueue/dequeue and updates the hrtick when the
5176 * current task is from our class and nr_running is low enough
5177 * to matter.
5179 static void hrtick_update(struct rq *rq)
5181 struct task_struct *curr = rq->curr;
5183 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
5184 return;
5186 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5187 hrtick_start_fair(rq, curr);
5189 #else /* !CONFIG_SCHED_HRTICK */
5190 static inline void
5191 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5195 static inline void hrtick_update(struct rq *rq)
5198 #endif
5201 * The enqueue_task method is called before nr_running is
5202 * increased. Here we update the fair scheduling stats and
5203 * then put the task into the rbtree:
5205 static void
5206 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5208 struct cfs_rq *cfs_rq;
5209 struct sched_entity *se = &p->se;
5212 * If in_iowait is set, the code below may not trigger any cpufreq
5213 * utilization updates, so do it here explicitly with the IOWAIT flag
5214 * passed.
5216 if (p->in_iowait)
5217 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5219 for_each_sched_entity(se) {
5220 if (se->on_rq)
5221 break;
5222 cfs_rq = cfs_rq_of(se);
5223 enqueue_entity(cfs_rq, se, flags);
5226 * end evaluation on encountering a throttled cfs_rq
5228 * note: in the case of encountering a throttled cfs_rq we will
5229 * post the final h_nr_running increment below.
5231 if (cfs_rq_throttled(cfs_rq))
5232 break;
5233 cfs_rq->h_nr_running++;
5235 flags = ENQUEUE_WAKEUP;
5238 for_each_sched_entity(se) {
5239 cfs_rq = cfs_rq_of(se);
5240 cfs_rq->h_nr_running++;
5242 if (cfs_rq_throttled(cfs_rq))
5243 break;
5245 update_load_avg(cfs_rq, se, UPDATE_TG);
5246 update_cfs_group(se);
5249 if (!se)
5250 add_nr_running(rq, 1);
5252 hrtick_update(rq);
5255 static void set_next_buddy(struct sched_entity *se);
5258 * The dequeue_task method is called before nr_running is
5259 * decreased. We remove the task from the rbtree and
5260 * update the fair scheduling stats:
5262 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5264 struct cfs_rq *cfs_rq;
5265 struct sched_entity *se = &p->se;
5266 int task_sleep = flags & DEQUEUE_SLEEP;
5268 for_each_sched_entity(se) {
5269 cfs_rq = cfs_rq_of(se);
5270 dequeue_entity(cfs_rq, se, flags);
5273 * end evaluation on encountering a throttled cfs_rq
5275 * note: in the case of encountering a throttled cfs_rq we will
5276 * post the final h_nr_running decrement below.
5278 if (cfs_rq_throttled(cfs_rq))
5279 break;
5280 cfs_rq->h_nr_running--;
5282 /* Don't dequeue parent if it has other entities besides us */
5283 if (cfs_rq->load.weight) {
5284 /* Avoid re-evaluating load for this entity: */
5285 se = parent_entity(se);
5287 * Bias pick_next to pick a task from this cfs_rq, as
5288 * p is sleeping when it is within its sched_slice.
5290 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5291 set_next_buddy(se);
5292 break;
5294 flags |= DEQUEUE_SLEEP;
5297 for_each_sched_entity(se) {
5298 cfs_rq = cfs_rq_of(se);
5299 cfs_rq->h_nr_running--;
5301 if (cfs_rq_throttled(cfs_rq))
5302 break;
5304 update_load_avg(cfs_rq, se, UPDATE_TG);
5305 update_cfs_group(se);
5308 if (!se)
5309 sub_nr_running(rq, 1);
5311 hrtick_update(rq);
5314 #ifdef CONFIG_SMP
5316 /* Working cpumask for: load_balance, load_balance_newidle. */
5317 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5318 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5320 #ifdef CONFIG_NO_HZ_COMMON
5322 * per rq 'load' arrray crap; XXX kill this.
5326 * The exact cpuload calculated at every tick would be:
5328 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
5330 * If a cpu misses updates for n ticks (as it was idle) and update gets
5331 * called on the n+1-th tick when cpu may be busy, then we have:
5333 * load_n = (1 - 1/2^i)^n * load_0
5334 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
5336 * decay_load_missed() below does efficient calculation of
5338 * load' = (1 - 1/2^i)^n * load
5340 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
5341 * This allows us to precompute the above in said factors, thereby allowing the
5342 * reduction of an arbitrary n in O(log_2 n) steps. (See also
5343 * fixed_power_int())
5345 * The calculation is approximated on a 128 point scale.
5347 #define DEGRADE_SHIFT 7
5349 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
5350 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
5351 { 0, 0, 0, 0, 0, 0, 0, 0 },
5352 { 64, 32, 8, 0, 0, 0, 0, 0 },
5353 { 96, 72, 40, 12, 1, 0, 0, 0 },
5354 { 112, 98, 75, 43, 15, 1, 0, 0 },
5355 { 120, 112, 98, 76, 45, 16, 2, 0 }
5359 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
5360 * would be when CPU is idle and so we just decay the old load without
5361 * adding any new load.
5363 static unsigned long
5364 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
5366 int j = 0;
5368 if (!missed_updates)
5369 return load;
5371 if (missed_updates >= degrade_zero_ticks[idx])
5372 return 0;
5374 if (idx == 1)
5375 return load >> missed_updates;
5377 while (missed_updates) {
5378 if (missed_updates % 2)
5379 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
5381 missed_updates >>= 1;
5382 j++;
5384 return load;
5386 #endif /* CONFIG_NO_HZ_COMMON */
5389 * __cpu_load_update - update the rq->cpu_load[] statistics
5390 * @this_rq: The rq to update statistics for
5391 * @this_load: The current load
5392 * @pending_updates: The number of missed updates
5394 * Update rq->cpu_load[] statistics. This function is usually called every
5395 * scheduler tick (TICK_NSEC).
5397 * This function computes a decaying average:
5399 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
5401 * Because of NOHZ it might not get called on every tick which gives need for
5402 * the @pending_updates argument.
5404 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
5405 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
5406 * = A * (A * load[i]_n-2 + B) + B
5407 * = A * (A * (A * load[i]_n-3 + B) + B) + B
5408 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5409 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5410 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5411 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5413 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5414 * any change in load would have resulted in the tick being turned back on.
5416 * For regular NOHZ, this reduces to:
5418 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
5420 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5421 * term.
5423 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
5424 unsigned long pending_updates)
5426 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
5427 int i, scale;
5429 this_rq->nr_load_updates++;
5431 /* Update our load: */
5432 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
5433 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
5434 unsigned long old_load, new_load;
5436 /* scale is effectively 1 << i now, and >> i divides by scale */
5438 old_load = this_rq->cpu_load[i];
5439 #ifdef CONFIG_NO_HZ_COMMON
5440 old_load = decay_load_missed(old_load, pending_updates - 1, i);
5441 if (tickless_load) {
5442 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
5444 * old_load can never be a negative value because a
5445 * decayed tickless_load cannot be greater than the
5446 * original tickless_load.
5448 old_load += tickless_load;
5450 #endif
5451 new_load = this_load;
5453 * Round up the averaging division if load is increasing. This
5454 * prevents us from getting stuck on 9 if the load is 10, for
5455 * example.
5457 if (new_load > old_load)
5458 new_load += scale - 1;
5460 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
5463 sched_avg_update(this_rq);
5466 /* Used instead of source_load when we know the type == 0 */
5467 static unsigned long weighted_cpuload(struct rq *rq)
5469 return cfs_rq_runnable_load_avg(&rq->cfs);
5472 #ifdef CONFIG_NO_HZ_COMMON
5474 * There is no sane way to deal with nohz on smp when using jiffies because the
5475 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
5476 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5478 * Therefore we need to avoid the delta approach from the regular tick when
5479 * possible since that would seriously skew the load calculation. This is why we
5480 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5481 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5482 * loop exit, nohz_idle_balance, nohz full exit...)
5484 * This means we might still be one tick off for nohz periods.
5487 static void cpu_load_update_nohz(struct rq *this_rq,
5488 unsigned long curr_jiffies,
5489 unsigned long load)
5491 unsigned long pending_updates;
5493 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5494 if (pending_updates) {
5495 this_rq->last_load_update_tick = curr_jiffies;
5497 * In the regular NOHZ case, we were idle, this means load 0.
5498 * In the NOHZ_FULL case, we were non-idle, we should consider
5499 * its weighted load.
5501 cpu_load_update(this_rq, load, pending_updates);
5506 * Called from nohz_idle_balance() to update the load ratings before doing the
5507 * idle balance.
5509 static void cpu_load_update_idle(struct rq *this_rq)
5512 * bail if there's load or we're actually up-to-date.
5514 if (weighted_cpuload(this_rq))
5515 return;
5517 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
5521 * Record CPU load on nohz entry so we know the tickless load to account
5522 * on nohz exit. cpu_load[0] happens then to be updated more frequently
5523 * than other cpu_load[idx] but it should be fine as cpu_load readers
5524 * shouldn't rely into synchronized cpu_load[*] updates.
5526 void cpu_load_update_nohz_start(void)
5528 struct rq *this_rq = this_rq();
5531 * This is all lockless but should be fine. If weighted_cpuload changes
5532 * concurrently we'll exit nohz. And cpu_load write can race with
5533 * cpu_load_update_idle() but both updater would be writing the same.
5535 this_rq->cpu_load[0] = weighted_cpuload(this_rq);
5539 * Account the tickless load in the end of a nohz frame.
5541 void cpu_load_update_nohz_stop(void)
5543 unsigned long curr_jiffies = READ_ONCE(jiffies);
5544 struct rq *this_rq = this_rq();
5545 unsigned long load;
5546 struct rq_flags rf;
5548 if (curr_jiffies == this_rq->last_load_update_tick)
5549 return;
5551 load = weighted_cpuload(this_rq);
5552 rq_lock(this_rq, &rf);
5553 update_rq_clock(this_rq);
5554 cpu_load_update_nohz(this_rq, curr_jiffies, load);
5555 rq_unlock(this_rq, &rf);
5557 #else /* !CONFIG_NO_HZ_COMMON */
5558 static inline void cpu_load_update_nohz(struct rq *this_rq,
5559 unsigned long curr_jiffies,
5560 unsigned long load) { }
5561 #endif /* CONFIG_NO_HZ_COMMON */
5563 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
5565 #ifdef CONFIG_NO_HZ_COMMON
5566 /* See the mess around cpu_load_update_nohz(). */
5567 this_rq->last_load_update_tick = READ_ONCE(jiffies);
5568 #endif
5569 cpu_load_update(this_rq, load, 1);
5573 * Called from scheduler_tick()
5575 void cpu_load_update_active(struct rq *this_rq)
5577 unsigned long load = weighted_cpuload(this_rq);
5579 if (tick_nohz_tick_stopped())
5580 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
5581 else
5582 cpu_load_update_periodic(this_rq, load);
5586 * Return a low guess at the load of a migration-source cpu weighted
5587 * according to the scheduling class and "nice" value.
5589 * We want to under-estimate the load of migration sources, to
5590 * balance conservatively.
5592 static unsigned long source_load(int cpu, int type)
5594 struct rq *rq = cpu_rq(cpu);
5595 unsigned long total = weighted_cpuload(rq);
5597 if (type == 0 || !sched_feat(LB_BIAS))
5598 return total;
5600 return min(rq->cpu_load[type-1], total);
5604 * Return a high guess at the load of a migration-target cpu weighted
5605 * according to the scheduling class and "nice" value.
5607 static unsigned long target_load(int cpu, int type)
5609 struct rq *rq = cpu_rq(cpu);
5610 unsigned long total = weighted_cpuload(rq);
5612 if (type == 0 || !sched_feat(LB_BIAS))
5613 return total;
5615 return max(rq->cpu_load[type-1], total);
5618 static unsigned long capacity_of(int cpu)
5620 return cpu_rq(cpu)->cpu_capacity;
5623 static unsigned long capacity_orig_of(int cpu)
5625 return cpu_rq(cpu)->cpu_capacity_orig;
5628 static unsigned long cpu_avg_load_per_task(int cpu)
5630 struct rq *rq = cpu_rq(cpu);
5631 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5632 unsigned long load_avg = weighted_cpuload(rq);
5634 if (nr_running)
5635 return load_avg / nr_running;
5637 return 0;
5640 static void record_wakee(struct task_struct *p)
5643 * Only decay a single time; tasks that have less then 1 wakeup per
5644 * jiffy will not have built up many flips.
5646 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5647 current->wakee_flips >>= 1;
5648 current->wakee_flip_decay_ts = jiffies;
5651 if (current->last_wakee != p) {
5652 current->last_wakee = p;
5653 current->wakee_flips++;
5658 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5660 * A waker of many should wake a different task than the one last awakened
5661 * at a frequency roughly N times higher than one of its wakees.
5663 * In order to determine whether we should let the load spread vs consolidating
5664 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5665 * partner, and a factor of lls_size higher frequency in the other.
5667 * With both conditions met, we can be relatively sure that the relationship is
5668 * non-monogamous, with partner count exceeding socket size.
5670 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5671 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5672 * socket size.
5674 static int wake_wide(struct task_struct *p)
5676 unsigned int master = current->wakee_flips;
5677 unsigned int slave = p->wakee_flips;
5678 int factor = this_cpu_read(sd_llc_size);
5680 if (master < slave)
5681 swap(master, slave);
5682 if (slave < factor || master < slave * factor)
5683 return 0;
5684 return 1;
5688 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5689 * soonest. For the purpose of speed we only consider the waking and previous
5690 * CPU.
5692 * wake_affine_idle() - only considers 'now', it check if the waking CPU is (or
5693 * will be) idle.
5695 * wake_affine_weight() - considers the weight to reflect the average
5696 * scheduling latency of the CPUs. This seems to work
5697 * for the overloaded case.
5700 static bool
5701 wake_affine_idle(struct sched_domain *sd, struct task_struct *p,
5702 int this_cpu, int prev_cpu, int sync)
5704 if (idle_cpu(this_cpu))
5705 return true;
5707 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5708 return true;
5710 return false;
5713 static bool
5714 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5715 int this_cpu, int prev_cpu, int sync)
5717 s64 this_eff_load, prev_eff_load;
5718 unsigned long task_load;
5720 this_eff_load = target_load(this_cpu, sd->wake_idx);
5721 prev_eff_load = source_load(prev_cpu, sd->wake_idx);
5723 if (sync) {
5724 unsigned long current_load = task_h_load(current);
5726 if (current_load > this_eff_load)
5727 return true;
5729 this_eff_load -= current_load;
5732 task_load = task_h_load(p);
5734 this_eff_load += task_load;
5735 if (sched_feat(WA_BIAS))
5736 this_eff_load *= 100;
5737 this_eff_load *= capacity_of(prev_cpu);
5739 prev_eff_load -= task_load;
5740 if (sched_feat(WA_BIAS))
5741 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5742 prev_eff_load *= capacity_of(this_cpu);
5744 return this_eff_load <= prev_eff_load;
5747 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5748 int prev_cpu, int sync)
5750 int this_cpu = smp_processor_id();
5751 bool affine = false;
5753 if (sched_feat(WA_IDLE) && !affine)
5754 affine = wake_affine_idle(sd, p, this_cpu, prev_cpu, sync);
5756 if (sched_feat(WA_WEIGHT) && !affine)
5757 affine = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5759 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5760 if (affine) {
5761 schedstat_inc(sd->ttwu_move_affine);
5762 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5765 return affine;
5768 static inline int task_util(struct task_struct *p);
5769 static int cpu_util_wake(int cpu, struct task_struct *p);
5771 static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
5773 return capacity_orig_of(cpu) - cpu_util_wake(cpu, p);
5777 * find_idlest_group finds and returns the least busy CPU group within the
5778 * domain.
5780 * Assumes p is allowed on at least one CPU in sd.
5782 static struct sched_group *
5783 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5784 int this_cpu, int sd_flag)
5786 struct sched_group *idlest = NULL, *group = sd->groups;
5787 struct sched_group *most_spare_sg = NULL;
5788 unsigned long min_runnable_load = ULONG_MAX;
5789 unsigned long this_runnable_load = ULONG_MAX;
5790 unsigned long min_avg_load = ULONG_MAX, this_avg_load = ULONG_MAX;
5791 unsigned long most_spare = 0, this_spare = 0;
5792 int load_idx = sd->forkexec_idx;
5793 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5794 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5795 (sd->imbalance_pct-100) / 100;
5797 if (sd_flag & SD_BALANCE_WAKE)
5798 load_idx = sd->wake_idx;
5800 do {
5801 unsigned long load, avg_load, runnable_load;
5802 unsigned long spare_cap, max_spare_cap;
5803 int local_group;
5804 int i;
5806 /* Skip over this group if it has no CPUs allowed */
5807 if (!cpumask_intersects(sched_group_span(group),
5808 &p->cpus_allowed))
5809 continue;
5811 local_group = cpumask_test_cpu(this_cpu,
5812 sched_group_span(group));
5815 * Tally up the load of all CPUs in the group and find
5816 * the group containing the CPU with most spare capacity.
5818 avg_load = 0;
5819 runnable_load = 0;
5820 max_spare_cap = 0;
5822 for_each_cpu(i, sched_group_span(group)) {
5823 /* Bias balancing toward cpus of our domain */
5824 if (local_group)
5825 load = source_load(i, load_idx);
5826 else
5827 load = target_load(i, load_idx);
5829 runnable_load += load;
5831 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5833 spare_cap = capacity_spare_wake(i, p);
5835 if (spare_cap > max_spare_cap)
5836 max_spare_cap = spare_cap;
5839 /* Adjust by relative CPU capacity of the group */
5840 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5841 group->sgc->capacity;
5842 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5843 group->sgc->capacity;
5845 if (local_group) {
5846 this_runnable_load = runnable_load;
5847 this_avg_load = avg_load;
5848 this_spare = max_spare_cap;
5849 } else {
5850 if (min_runnable_load > (runnable_load + imbalance)) {
5852 * The runnable load is significantly smaller
5853 * so we can pick this new cpu
5855 min_runnable_load = runnable_load;
5856 min_avg_load = avg_load;
5857 idlest = group;
5858 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5859 (100*min_avg_load > imbalance_scale*avg_load)) {
5861 * The runnable loads are close so take the
5862 * blocked load into account through avg_load.
5864 min_avg_load = avg_load;
5865 idlest = group;
5868 if (most_spare < max_spare_cap) {
5869 most_spare = max_spare_cap;
5870 most_spare_sg = group;
5873 } while (group = group->next, group != sd->groups);
5876 * The cross-over point between using spare capacity or least load
5877 * is too conservative for high utilization tasks on partially
5878 * utilized systems if we require spare_capacity > task_util(p),
5879 * so we allow for some task stuffing by using
5880 * spare_capacity > task_util(p)/2.
5882 * Spare capacity can't be used for fork because the utilization has
5883 * not been set yet, we must first select a rq to compute the initial
5884 * utilization.
5886 if (sd_flag & SD_BALANCE_FORK)
5887 goto skip_spare;
5889 if (this_spare > task_util(p) / 2 &&
5890 imbalance_scale*this_spare > 100*most_spare)
5891 return NULL;
5893 if (most_spare > task_util(p) / 2)
5894 return most_spare_sg;
5896 skip_spare:
5897 if (!idlest)
5898 return NULL;
5900 if (min_runnable_load > (this_runnable_load + imbalance))
5901 return NULL;
5903 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5904 (100*this_avg_load < imbalance_scale*min_avg_load))
5905 return NULL;
5907 return idlest;
5911 * find_idlest_group_cpu - find the idlest cpu among the cpus in group.
5913 static int
5914 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5916 unsigned long load, min_load = ULONG_MAX;
5917 unsigned int min_exit_latency = UINT_MAX;
5918 u64 latest_idle_timestamp = 0;
5919 int least_loaded_cpu = this_cpu;
5920 int shallowest_idle_cpu = -1;
5921 int i;
5923 /* Check if we have any choice: */
5924 if (group->group_weight == 1)
5925 return cpumask_first(sched_group_span(group));
5927 /* Traverse only the allowed CPUs */
5928 for_each_cpu_and(i, sched_group_span(group), &p->cpus_allowed) {
5929 if (idle_cpu(i)) {
5930 struct rq *rq = cpu_rq(i);
5931 struct cpuidle_state *idle = idle_get_state(rq);
5932 if (idle && idle->exit_latency < min_exit_latency) {
5934 * We give priority to a CPU whose idle state
5935 * has the smallest exit latency irrespective
5936 * of any idle timestamp.
5938 min_exit_latency = idle->exit_latency;
5939 latest_idle_timestamp = rq->idle_stamp;
5940 shallowest_idle_cpu = i;
5941 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5942 rq->idle_stamp > latest_idle_timestamp) {
5944 * If equal or no active idle state, then
5945 * the most recently idled CPU might have
5946 * a warmer cache.
5948 latest_idle_timestamp = rq->idle_stamp;
5949 shallowest_idle_cpu = i;
5951 } else if (shallowest_idle_cpu == -1) {
5952 load = weighted_cpuload(cpu_rq(i));
5953 if (load < min_load || (load == min_load && i == this_cpu)) {
5954 min_load = load;
5955 least_loaded_cpu = i;
5960 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5963 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
5964 int cpu, int prev_cpu, int sd_flag)
5966 int new_cpu = cpu;
5968 if (!cpumask_intersects(sched_domain_span(sd), &p->cpus_allowed))
5969 return prev_cpu;
5971 while (sd) {
5972 struct sched_group *group;
5973 struct sched_domain *tmp;
5974 int weight;
5976 if (!(sd->flags & sd_flag)) {
5977 sd = sd->child;
5978 continue;
5981 group = find_idlest_group(sd, p, cpu, sd_flag);
5982 if (!group) {
5983 sd = sd->child;
5984 continue;
5987 new_cpu = find_idlest_group_cpu(group, p, cpu);
5988 if (new_cpu == cpu) {
5989 /* Now try balancing at a lower domain level of cpu */
5990 sd = sd->child;
5991 continue;
5994 /* Now try balancing at a lower domain level of new_cpu */
5995 cpu = new_cpu;
5996 weight = sd->span_weight;
5997 sd = NULL;
5998 for_each_domain(cpu, tmp) {
5999 if (weight <= tmp->span_weight)
6000 break;
6001 if (tmp->flags & sd_flag)
6002 sd = tmp;
6004 /* while loop will break here if sd == NULL */
6007 return new_cpu;
6010 #ifdef CONFIG_SCHED_SMT
6012 static inline void set_idle_cores(int cpu, int val)
6014 struct sched_domain_shared *sds;
6016 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6017 if (sds)
6018 WRITE_ONCE(sds->has_idle_cores, val);
6021 static inline bool test_idle_cores(int cpu, bool def)
6023 struct sched_domain_shared *sds;
6025 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6026 if (sds)
6027 return READ_ONCE(sds->has_idle_cores);
6029 return def;
6033 * Scans the local SMT mask to see if the entire core is idle, and records this
6034 * information in sd_llc_shared->has_idle_cores.
6036 * Since SMT siblings share all cache levels, inspecting this limited remote
6037 * state should be fairly cheap.
6039 void __update_idle_core(struct rq *rq)
6041 int core = cpu_of(rq);
6042 int cpu;
6044 rcu_read_lock();
6045 if (test_idle_cores(core, true))
6046 goto unlock;
6048 for_each_cpu(cpu, cpu_smt_mask(core)) {
6049 if (cpu == core)
6050 continue;
6052 if (!idle_cpu(cpu))
6053 goto unlock;
6056 set_idle_cores(core, 1);
6057 unlock:
6058 rcu_read_unlock();
6062 * Scan the entire LLC domain for idle cores; this dynamically switches off if
6063 * there are no idle cores left in the system; tracked through
6064 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6066 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6068 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6069 int core, cpu;
6071 if (!static_branch_likely(&sched_smt_present))
6072 return -1;
6074 if (!test_idle_cores(target, false))
6075 return -1;
6077 cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
6079 for_each_cpu_wrap(core, cpus, target) {
6080 bool idle = true;
6082 for_each_cpu(cpu, cpu_smt_mask(core)) {
6083 cpumask_clear_cpu(cpu, cpus);
6084 if (!idle_cpu(cpu))
6085 idle = false;
6088 if (idle)
6089 return core;
6093 * Failed to find an idle core; stop looking for one.
6095 set_idle_cores(target, 0);
6097 return -1;
6101 * Scan the local SMT mask for idle CPUs.
6103 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6105 int cpu;
6107 if (!static_branch_likely(&sched_smt_present))
6108 return -1;
6110 for_each_cpu(cpu, cpu_smt_mask(target)) {
6111 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6112 continue;
6113 if (idle_cpu(cpu))
6114 return cpu;
6117 return -1;
6120 #else /* CONFIG_SCHED_SMT */
6122 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6124 return -1;
6127 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6129 return -1;
6132 #endif /* CONFIG_SCHED_SMT */
6135 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6136 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6137 * average idle time for this rq (as found in rq->avg_idle).
6139 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
6141 struct sched_domain *this_sd;
6142 u64 avg_cost, avg_idle;
6143 u64 time, cost;
6144 s64 delta;
6145 int cpu, nr = INT_MAX;
6147 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6148 if (!this_sd)
6149 return -1;
6152 * Due to large variance we need a large fuzz factor; hackbench in
6153 * particularly is sensitive here.
6155 avg_idle = this_rq()->avg_idle / 512;
6156 avg_cost = this_sd->avg_scan_cost + 1;
6158 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
6159 return -1;
6161 if (sched_feat(SIS_PROP)) {
6162 u64 span_avg = sd->span_weight * avg_idle;
6163 if (span_avg > 4*avg_cost)
6164 nr = div_u64(span_avg, avg_cost);
6165 else
6166 nr = 4;
6169 time = local_clock();
6171 for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
6172 if (!--nr)
6173 return -1;
6174 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6175 continue;
6176 if (idle_cpu(cpu))
6177 break;
6180 time = local_clock() - time;
6181 cost = this_sd->avg_scan_cost;
6182 delta = (s64)(time - cost) / 8;
6183 this_sd->avg_scan_cost += delta;
6185 return cpu;
6189 * Try and locate an idle core/thread in the LLC cache domain.
6191 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6193 struct sched_domain *sd;
6194 int i;
6196 if (idle_cpu(target))
6197 return target;
6200 * If the previous cpu is cache affine and idle, don't be stupid.
6202 if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev))
6203 return prev;
6205 sd = rcu_dereference(per_cpu(sd_llc, target));
6206 if (!sd)
6207 return target;
6209 i = select_idle_core(p, sd, target);
6210 if ((unsigned)i < nr_cpumask_bits)
6211 return i;
6213 i = select_idle_cpu(p, sd, target);
6214 if ((unsigned)i < nr_cpumask_bits)
6215 return i;
6217 i = select_idle_smt(p, sd, target);
6218 if ((unsigned)i < nr_cpumask_bits)
6219 return i;
6221 return target;
6225 * cpu_util returns the amount of capacity of a CPU that is used by CFS
6226 * tasks. The unit of the return value must be the one of capacity so we can
6227 * compare the utilization with the capacity of the CPU that is available for
6228 * CFS task (ie cpu_capacity).
6230 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6231 * recent utilization of currently non-runnable tasks on a CPU. It represents
6232 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6233 * capacity_orig is the cpu_capacity available at the highest frequency
6234 * (arch_scale_freq_capacity()).
6235 * The utilization of a CPU converges towards a sum equal to or less than the
6236 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6237 * the running time on this CPU scaled by capacity_curr.
6239 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6240 * higher than capacity_orig because of unfortunate rounding in
6241 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6242 * the average stabilizes with the new running time. We need to check that the
6243 * utilization stays within the range of [0..capacity_orig] and cap it if
6244 * necessary. Without utilization capping, a group could be seen as overloaded
6245 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6246 * available capacity. We allow utilization to overshoot capacity_curr (but not
6247 * capacity_orig) as it useful for predicting the capacity required after task
6248 * migrations (scheduler-driven DVFS).
6250 static int cpu_util(int cpu)
6252 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
6253 unsigned long capacity = capacity_orig_of(cpu);
6255 return (util >= capacity) ? capacity : util;
6258 static inline int task_util(struct task_struct *p)
6260 return p->se.avg.util_avg;
6264 * cpu_util_wake: Compute cpu utilization with any contributions from
6265 * the waking task p removed.
6267 static int cpu_util_wake(int cpu, struct task_struct *p)
6269 unsigned long util, capacity;
6271 /* Task has no contribution or is new */
6272 if (cpu != task_cpu(p) || !p->se.avg.last_update_time)
6273 return cpu_util(cpu);
6275 capacity = capacity_orig_of(cpu);
6276 util = max_t(long, cpu_rq(cpu)->cfs.avg.util_avg - task_util(p), 0);
6278 return (util >= capacity) ? capacity : util;
6282 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6283 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6285 * In that case WAKE_AFFINE doesn't make sense and we'll let
6286 * BALANCE_WAKE sort things out.
6288 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6290 long min_cap, max_cap;
6292 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6293 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
6295 /* Minimum capacity is close to max, no need to abort wake_affine */
6296 if (max_cap - min_cap < max_cap >> 3)
6297 return 0;
6299 /* Bring task utilization in sync with prev_cpu */
6300 sync_entity_load_avg(&p->se);
6302 return min_cap * 1024 < task_util(p) * capacity_margin;
6306 * select_task_rq_fair: Select target runqueue for the waking task in domains
6307 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6308 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6310 * Balances load by selecting the idlest cpu in the idlest group, or under
6311 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
6313 * Returns the target cpu number.
6315 * preempt must be disabled.
6317 static int
6318 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6320 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
6321 int cpu = smp_processor_id();
6322 int new_cpu = prev_cpu;
6323 int want_affine = 0;
6324 int sync = wake_flags & WF_SYNC;
6326 if (sd_flag & SD_BALANCE_WAKE) {
6327 record_wakee(p);
6328 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
6329 && cpumask_test_cpu(cpu, &p->cpus_allowed);
6332 rcu_read_lock();
6333 for_each_domain(cpu, tmp) {
6334 if (!(tmp->flags & SD_LOAD_BALANCE))
6335 break;
6338 * If both cpu and prev_cpu are part of this domain,
6339 * cpu is a valid SD_WAKE_AFFINE target.
6341 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6342 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6343 affine_sd = tmp;
6344 break;
6347 if (tmp->flags & sd_flag)
6348 sd = tmp;
6349 else if (!want_affine)
6350 break;
6353 if (affine_sd) {
6354 sd = NULL; /* Prefer wake_affine over balance flags */
6355 if (cpu == prev_cpu)
6356 goto pick_cpu;
6358 if (wake_affine(affine_sd, p, prev_cpu, sync))
6359 new_cpu = cpu;
6362 if (sd && !(sd_flag & SD_BALANCE_FORK)) {
6364 * We're going to need the task's util for capacity_spare_wake
6365 * in find_idlest_group. Sync it up to prev_cpu's
6366 * last_update_time.
6368 sync_entity_load_avg(&p->se);
6371 if (!sd) {
6372 pick_cpu:
6373 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
6374 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6376 } else {
6377 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6379 rcu_read_unlock();
6381 return new_cpu;
6384 static void detach_entity_cfs_rq(struct sched_entity *se);
6387 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
6388 * cfs_rq_of(p) references at time of call are still valid and identify the
6389 * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6391 static void migrate_task_rq_fair(struct task_struct *p)
6394 * As blocked tasks retain absolute vruntime the migration needs to
6395 * deal with this by subtracting the old and adding the new
6396 * min_vruntime -- the latter is done by enqueue_entity() when placing
6397 * the task on the new runqueue.
6399 if (p->state == TASK_WAKING) {
6400 struct sched_entity *se = &p->se;
6401 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6402 u64 min_vruntime;
6404 #ifndef CONFIG_64BIT
6405 u64 min_vruntime_copy;
6407 do {
6408 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6409 smp_rmb();
6410 min_vruntime = cfs_rq->min_vruntime;
6411 } while (min_vruntime != min_vruntime_copy);
6412 #else
6413 min_vruntime = cfs_rq->min_vruntime;
6414 #endif
6416 se->vruntime -= min_vruntime;
6419 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6421 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6422 * rq->lock and can modify state directly.
6424 lockdep_assert_held(&task_rq(p)->lock);
6425 detach_entity_cfs_rq(&p->se);
6427 } else {
6429 * We are supposed to update the task to "current" time, then
6430 * its up to date and ready to go to new CPU/cfs_rq. But we
6431 * have difficulty in getting what current time is, so simply
6432 * throw away the out-of-date time. This will result in the
6433 * wakee task is less decayed, but giving the wakee more load
6434 * sounds not bad.
6436 remove_entity_load_avg(&p->se);
6439 /* Tell new CPU we are migrated */
6440 p->se.avg.last_update_time = 0;
6442 /* We have migrated, no longer consider this task hot */
6443 p->se.exec_start = 0;
6446 static void task_dead_fair(struct task_struct *p)
6448 remove_entity_load_avg(&p->se);
6450 #endif /* CONFIG_SMP */
6452 static unsigned long
6453 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
6455 unsigned long gran = sysctl_sched_wakeup_granularity;
6458 * Since its curr running now, convert the gran from real-time
6459 * to virtual-time in his units.
6461 * By using 'se' instead of 'curr' we penalize light tasks, so
6462 * they get preempted easier. That is, if 'se' < 'curr' then
6463 * the resulting gran will be larger, therefore penalizing the
6464 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6465 * be smaller, again penalizing the lighter task.
6467 * This is especially important for buddies when the leftmost
6468 * task is higher priority than the buddy.
6470 return calc_delta_fair(gran, se);
6474 * Should 'se' preempt 'curr'.
6476 * |s1
6477 * |s2
6478 * |s3
6480 * |<--->|c
6482 * w(c, s1) = -1
6483 * w(c, s2) = 0
6484 * w(c, s3) = 1
6487 static int
6488 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6490 s64 gran, vdiff = curr->vruntime - se->vruntime;
6492 if (vdiff <= 0)
6493 return -1;
6495 gran = wakeup_gran(curr, se);
6496 if (vdiff > gran)
6497 return 1;
6499 return 0;
6502 static void set_last_buddy(struct sched_entity *se)
6504 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6505 return;
6507 for_each_sched_entity(se) {
6508 if (SCHED_WARN_ON(!se->on_rq))
6509 return;
6510 cfs_rq_of(se)->last = se;
6514 static void set_next_buddy(struct sched_entity *se)
6516 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6517 return;
6519 for_each_sched_entity(se) {
6520 if (SCHED_WARN_ON(!se->on_rq))
6521 return;
6522 cfs_rq_of(se)->next = se;
6526 static void set_skip_buddy(struct sched_entity *se)
6528 for_each_sched_entity(se)
6529 cfs_rq_of(se)->skip = se;
6533 * Preempt the current task with a newly woken task if needed:
6535 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6537 struct task_struct *curr = rq->curr;
6538 struct sched_entity *se = &curr->se, *pse = &p->se;
6539 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6540 int scale = cfs_rq->nr_running >= sched_nr_latency;
6541 int next_buddy_marked = 0;
6543 if (unlikely(se == pse))
6544 return;
6547 * This is possible from callers such as attach_tasks(), in which we
6548 * unconditionally check_prempt_curr() after an enqueue (which may have
6549 * lead to a throttle). This both saves work and prevents false
6550 * next-buddy nomination below.
6552 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6553 return;
6555 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6556 set_next_buddy(pse);
6557 next_buddy_marked = 1;
6561 * We can come here with TIF_NEED_RESCHED already set from new task
6562 * wake up path.
6564 * Note: this also catches the edge-case of curr being in a throttled
6565 * group (e.g. via set_curr_task), since update_curr() (in the
6566 * enqueue of curr) will have resulted in resched being set. This
6567 * prevents us from potentially nominating it as a false LAST_BUDDY
6568 * below.
6570 if (test_tsk_need_resched(curr))
6571 return;
6573 /* Idle tasks are by definition preempted by non-idle tasks. */
6574 if (unlikely(curr->policy == SCHED_IDLE) &&
6575 likely(p->policy != SCHED_IDLE))
6576 goto preempt;
6579 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6580 * is driven by the tick):
6582 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6583 return;
6585 find_matching_se(&se, &pse);
6586 update_curr(cfs_rq_of(se));
6587 BUG_ON(!pse);
6588 if (wakeup_preempt_entity(se, pse) == 1) {
6590 * Bias pick_next to pick the sched entity that is
6591 * triggering this preemption.
6593 if (!next_buddy_marked)
6594 set_next_buddy(pse);
6595 goto preempt;
6598 return;
6600 preempt:
6601 resched_curr(rq);
6603 * Only set the backward buddy when the current task is still
6604 * on the rq. This can happen when a wakeup gets interleaved
6605 * with schedule on the ->pre_schedule() or idle_balance()
6606 * point, either of which can * drop the rq lock.
6608 * Also, during early boot the idle thread is in the fair class,
6609 * for obvious reasons its a bad idea to schedule back to it.
6611 if (unlikely(!se->on_rq || curr == rq->idle))
6612 return;
6614 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6615 set_last_buddy(se);
6618 static struct task_struct *
6619 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6621 struct cfs_rq *cfs_rq = &rq->cfs;
6622 struct sched_entity *se;
6623 struct task_struct *p;
6624 int new_tasks;
6626 again:
6627 if (!cfs_rq->nr_running)
6628 goto idle;
6630 #ifdef CONFIG_FAIR_GROUP_SCHED
6631 if (prev->sched_class != &fair_sched_class)
6632 goto simple;
6635 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6636 * likely that a next task is from the same cgroup as the current.
6638 * Therefore attempt to avoid putting and setting the entire cgroup
6639 * hierarchy, only change the part that actually changes.
6642 do {
6643 struct sched_entity *curr = cfs_rq->curr;
6646 * Since we got here without doing put_prev_entity() we also
6647 * have to consider cfs_rq->curr. If it is still a runnable
6648 * entity, update_curr() will update its vruntime, otherwise
6649 * forget we've ever seen it.
6651 if (curr) {
6652 if (curr->on_rq)
6653 update_curr(cfs_rq);
6654 else
6655 curr = NULL;
6658 * This call to check_cfs_rq_runtime() will do the
6659 * throttle and dequeue its entity in the parent(s).
6660 * Therefore the nr_running test will indeed
6661 * be correct.
6663 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6664 cfs_rq = &rq->cfs;
6666 if (!cfs_rq->nr_running)
6667 goto idle;
6669 goto simple;
6673 se = pick_next_entity(cfs_rq, curr);
6674 cfs_rq = group_cfs_rq(se);
6675 } while (cfs_rq);
6677 p = task_of(se);
6680 * Since we haven't yet done put_prev_entity and if the selected task
6681 * is a different task than we started out with, try and touch the
6682 * least amount of cfs_rqs.
6684 if (prev != p) {
6685 struct sched_entity *pse = &prev->se;
6687 while (!(cfs_rq = is_same_group(se, pse))) {
6688 int se_depth = se->depth;
6689 int pse_depth = pse->depth;
6691 if (se_depth <= pse_depth) {
6692 put_prev_entity(cfs_rq_of(pse), pse);
6693 pse = parent_entity(pse);
6695 if (se_depth >= pse_depth) {
6696 set_next_entity(cfs_rq_of(se), se);
6697 se = parent_entity(se);
6701 put_prev_entity(cfs_rq, pse);
6702 set_next_entity(cfs_rq, se);
6705 goto done;
6706 simple:
6707 #endif
6709 put_prev_task(rq, prev);
6711 do {
6712 se = pick_next_entity(cfs_rq, NULL);
6713 set_next_entity(cfs_rq, se);
6714 cfs_rq = group_cfs_rq(se);
6715 } while (cfs_rq);
6717 p = task_of(se);
6719 done: __maybe_unused
6720 #ifdef CONFIG_SMP
6722 * Move the next running task to the front of
6723 * the list, so our cfs_tasks list becomes MRU
6724 * one.
6726 list_move(&p->se.group_node, &rq->cfs_tasks);
6727 #endif
6729 if (hrtick_enabled(rq))
6730 hrtick_start_fair(rq, p);
6732 return p;
6734 idle:
6735 new_tasks = idle_balance(rq, rf);
6738 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6739 * possible for any higher priority task to appear. In that case we
6740 * must re-start the pick_next_entity() loop.
6742 if (new_tasks < 0)
6743 return RETRY_TASK;
6745 if (new_tasks > 0)
6746 goto again;
6748 return NULL;
6752 * Account for a descheduled task:
6754 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6756 struct sched_entity *se = &prev->se;
6757 struct cfs_rq *cfs_rq;
6759 for_each_sched_entity(se) {
6760 cfs_rq = cfs_rq_of(se);
6761 put_prev_entity(cfs_rq, se);
6766 * sched_yield() is very simple
6768 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6770 static void yield_task_fair(struct rq *rq)
6772 struct task_struct *curr = rq->curr;
6773 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6774 struct sched_entity *se = &curr->se;
6777 * Are we the only task in the tree?
6779 if (unlikely(rq->nr_running == 1))
6780 return;
6782 clear_buddies(cfs_rq, se);
6784 if (curr->policy != SCHED_BATCH) {
6785 update_rq_clock(rq);
6787 * Update run-time statistics of the 'current'.
6789 update_curr(cfs_rq);
6791 * Tell update_rq_clock() that we've just updated,
6792 * so we don't do microscopic update in schedule()
6793 * and double the fastpath cost.
6795 rq_clock_skip_update(rq, true);
6798 set_skip_buddy(se);
6801 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6803 struct sched_entity *se = &p->se;
6805 /* throttled hierarchies are not runnable */
6806 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6807 return false;
6809 /* Tell the scheduler that we'd really like pse to run next. */
6810 set_next_buddy(se);
6812 yield_task_fair(rq);
6814 return true;
6817 #ifdef CONFIG_SMP
6818 /**************************************************
6819 * Fair scheduling class load-balancing methods.
6821 * BASICS
6823 * The purpose of load-balancing is to achieve the same basic fairness the
6824 * per-cpu scheduler provides, namely provide a proportional amount of compute
6825 * time to each task. This is expressed in the following equation:
6827 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6829 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6830 * W_i,0 is defined as:
6832 * W_i,0 = \Sum_j w_i,j (2)
6834 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6835 * is derived from the nice value as per sched_prio_to_weight[].
6837 * The weight average is an exponential decay average of the instantaneous
6838 * weight:
6840 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6842 * C_i is the compute capacity of cpu i, typically it is the
6843 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6844 * can also include other factors [XXX].
6846 * To achieve this balance we define a measure of imbalance which follows
6847 * directly from (1):
6849 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6851 * We them move tasks around to minimize the imbalance. In the continuous
6852 * function space it is obvious this converges, in the discrete case we get
6853 * a few fun cases generally called infeasible weight scenarios.
6855 * [XXX expand on:
6856 * - infeasible weights;
6857 * - local vs global optima in the discrete case. ]
6860 * SCHED DOMAINS
6862 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6863 * for all i,j solution, we create a tree of cpus that follows the hardware
6864 * topology where each level pairs two lower groups (or better). This results
6865 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6866 * tree to only the first of the previous level and we decrease the frequency
6867 * of load-balance at each level inv. proportional to the number of cpus in
6868 * the groups.
6870 * This yields:
6872 * log_2 n 1 n
6873 * \Sum { --- * --- * 2^i } = O(n) (5)
6874 * i = 0 2^i 2^i
6875 * `- size of each group
6876 * | | `- number of cpus doing load-balance
6877 * | `- freq
6878 * `- sum over all levels
6880 * Coupled with a limit on how many tasks we can migrate every balance pass,
6881 * this makes (5) the runtime complexity of the balancer.
6883 * An important property here is that each CPU is still (indirectly) connected
6884 * to every other cpu in at most O(log n) steps:
6886 * The adjacency matrix of the resulting graph is given by:
6888 * log_2 n
6889 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6890 * k = 0
6892 * And you'll find that:
6894 * A^(log_2 n)_i,j != 0 for all i,j (7)
6896 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6897 * The task movement gives a factor of O(m), giving a convergence complexity
6898 * of:
6900 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6903 * WORK CONSERVING
6905 * In order to avoid CPUs going idle while there's still work to do, new idle
6906 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6907 * tree itself instead of relying on other CPUs to bring it work.
6909 * This adds some complexity to both (5) and (8) but it reduces the total idle
6910 * time.
6912 * [XXX more?]
6915 * CGROUPS
6917 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6919 * s_k,i
6920 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6921 * S_k
6923 * Where
6925 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6927 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6929 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6930 * property.
6932 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6933 * rewrite all of this once again.]
6936 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6938 enum fbq_type { regular, remote, all };
6940 #define LBF_ALL_PINNED 0x01
6941 #define LBF_NEED_BREAK 0x02
6942 #define LBF_DST_PINNED 0x04
6943 #define LBF_SOME_PINNED 0x08
6945 struct lb_env {
6946 struct sched_domain *sd;
6948 struct rq *src_rq;
6949 int src_cpu;
6951 int dst_cpu;
6952 struct rq *dst_rq;
6954 struct cpumask *dst_grpmask;
6955 int new_dst_cpu;
6956 enum cpu_idle_type idle;
6957 long imbalance;
6958 /* The set of CPUs under consideration for load-balancing */
6959 struct cpumask *cpus;
6961 unsigned int flags;
6963 unsigned int loop;
6964 unsigned int loop_break;
6965 unsigned int loop_max;
6967 enum fbq_type fbq_type;
6968 struct list_head tasks;
6972 * Is this task likely cache-hot:
6974 static int task_hot(struct task_struct *p, struct lb_env *env)
6976 s64 delta;
6978 lockdep_assert_held(&env->src_rq->lock);
6980 if (p->sched_class != &fair_sched_class)
6981 return 0;
6983 if (unlikely(p->policy == SCHED_IDLE))
6984 return 0;
6987 * Buddy candidates are cache hot:
6989 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6990 (&p->se == cfs_rq_of(&p->se)->next ||
6991 &p->se == cfs_rq_of(&p->se)->last))
6992 return 1;
6994 if (sysctl_sched_migration_cost == -1)
6995 return 1;
6996 if (sysctl_sched_migration_cost == 0)
6997 return 0;
6999 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7001 return delta < (s64)sysctl_sched_migration_cost;
7004 #ifdef CONFIG_NUMA_BALANCING
7006 * Returns 1, if task migration degrades locality
7007 * Returns 0, if task migration improves locality i.e migration preferred.
7008 * Returns -1, if task migration is not affected by locality.
7010 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7012 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7013 unsigned long src_faults, dst_faults;
7014 int src_nid, dst_nid;
7016 if (!static_branch_likely(&sched_numa_balancing))
7017 return -1;
7019 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7020 return -1;
7022 src_nid = cpu_to_node(env->src_cpu);
7023 dst_nid = cpu_to_node(env->dst_cpu);
7025 if (src_nid == dst_nid)
7026 return -1;
7028 /* Migrating away from the preferred node is always bad. */
7029 if (src_nid == p->numa_preferred_nid) {
7030 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7031 return 1;
7032 else
7033 return -1;
7036 /* Encourage migration to the preferred node. */
7037 if (dst_nid == p->numa_preferred_nid)
7038 return 0;
7040 /* Leaving a core idle is often worse than degrading locality. */
7041 if (env->idle != CPU_NOT_IDLE)
7042 return -1;
7044 if (numa_group) {
7045 src_faults = group_faults(p, src_nid);
7046 dst_faults = group_faults(p, dst_nid);
7047 } else {
7048 src_faults = task_faults(p, src_nid);
7049 dst_faults = task_faults(p, dst_nid);
7052 return dst_faults < src_faults;
7055 #else
7056 static inline int migrate_degrades_locality(struct task_struct *p,
7057 struct lb_env *env)
7059 return -1;
7061 #endif
7064 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7066 static
7067 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7069 int tsk_cache_hot;
7071 lockdep_assert_held(&env->src_rq->lock);
7074 * We do not migrate tasks that are:
7075 * 1) throttled_lb_pair, or
7076 * 2) cannot be migrated to this CPU due to cpus_allowed, or
7077 * 3) running (obviously), or
7078 * 4) are cache-hot on their current CPU.
7080 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7081 return 0;
7083 if (!cpumask_test_cpu(env->dst_cpu, &p->cpus_allowed)) {
7084 int cpu;
7086 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7088 env->flags |= LBF_SOME_PINNED;
7091 * Remember if this task can be migrated to any other cpu in
7092 * our sched_group. We may want to revisit it if we couldn't
7093 * meet load balance goals by pulling other tasks on src_cpu.
7095 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7096 * already computed one in current iteration.
7098 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
7099 return 0;
7101 /* Prevent to re-select dst_cpu via env's cpus */
7102 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7103 if (cpumask_test_cpu(cpu, &p->cpus_allowed)) {
7104 env->flags |= LBF_DST_PINNED;
7105 env->new_dst_cpu = cpu;
7106 break;
7110 return 0;
7113 /* Record that we found atleast one task that could run on dst_cpu */
7114 env->flags &= ~LBF_ALL_PINNED;
7116 if (task_running(env->src_rq, p)) {
7117 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7118 return 0;
7122 * Aggressive migration if:
7123 * 1) destination numa is preferred
7124 * 2) task is cache cold, or
7125 * 3) too many balance attempts have failed.
7127 tsk_cache_hot = migrate_degrades_locality(p, env);
7128 if (tsk_cache_hot == -1)
7129 tsk_cache_hot = task_hot(p, env);
7131 if (tsk_cache_hot <= 0 ||
7132 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7133 if (tsk_cache_hot == 1) {
7134 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7135 schedstat_inc(p->se.statistics.nr_forced_migrations);
7137 return 1;
7140 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7141 return 0;
7145 * detach_task() -- detach the task for the migration specified in env
7147 static void detach_task(struct task_struct *p, struct lb_env *env)
7149 lockdep_assert_held(&env->src_rq->lock);
7151 p->on_rq = TASK_ON_RQ_MIGRATING;
7152 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7153 set_task_cpu(p, env->dst_cpu);
7157 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7158 * part of active balancing operations within "domain".
7160 * Returns a task if successful and NULL otherwise.
7162 static struct task_struct *detach_one_task(struct lb_env *env)
7164 struct task_struct *p;
7166 lockdep_assert_held(&env->src_rq->lock);
7168 list_for_each_entry_reverse(p,
7169 &env->src_rq->cfs_tasks, se.group_node) {
7170 if (!can_migrate_task(p, env))
7171 continue;
7173 detach_task(p, env);
7176 * Right now, this is only the second place where
7177 * lb_gained[env->idle] is updated (other is detach_tasks)
7178 * so we can safely collect stats here rather than
7179 * inside detach_tasks().
7181 schedstat_inc(env->sd->lb_gained[env->idle]);
7182 return p;
7184 return NULL;
7187 static const unsigned int sched_nr_migrate_break = 32;
7190 * detach_tasks() -- tries to detach up to imbalance weighted load from
7191 * busiest_rq, as part of a balancing operation within domain "sd".
7193 * Returns number of detached tasks if successful and 0 otherwise.
7195 static int detach_tasks(struct lb_env *env)
7197 struct list_head *tasks = &env->src_rq->cfs_tasks;
7198 struct task_struct *p;
7199 unsigned long load;
7200 int detached = 0;
7202 lockdep_assert_held(&env->src_rq->lock);
7204 if (env->imbalance <= 0)
7205 return 0;
7207 while (!list_empty(tasks)) {
7209 * We don't want to steal all, otherwise we may be treated likewise,
7210 * which could at worst lead to a livelock crash.
7212 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7213 break;
7215 p = list_last_entry(tasks, struct task_struct, se.group_node);
7217 env->loop++;
7218 /* We've more or less seen every task there is, call it quits */
7219 if (env->loop > env->loop_max)
7220 break;
7222 /* take a breather every nr_migrate tasks */
7223 if (env->loop > env->loop_break) {
7224 env->loop_break += sched_nr_migrate_break;
7225 env->flags |= LBF_NEED_BREAK;
7226 break;
7229 if (!can_migrate_task(p, env))
7230 goto next;
7232 load = task_h_load(p);
7234 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7235 goto next;
7237 if ((load / 2) > env->imbalance)
7238 goto next;
7240 detach_task(p, env);
7241 list_add(&p->se.group_node, &env->tasks);
7243 detached++;
7244 env->imbalance -= load;
7246 #ifdef CONFIG_PREEMPT
7248 * NEWIDLE balancing is a source of latency, so preemptible
7249 * kernels will stop after the first task is detached to minimize
7250 * the critical section.
7252 if (env->idle == CPU_NEWLY_IDLE)
7253 break;
7254 #endif
7257 * We only want to steal up to the prescribed amount of
7258 * weighted load.
7260 if (env->imbalance <= 0)
7261 break;
7263 continue;
7264 next:
7265 list_move(&p->se.group_node, tasks);
7269 * Right now, this is one of only two places we collect this stat
7270 * so we can safely collect detach_one_task() stats here rather
7271 * than inside detach_one_task().
7273 schedstat_add(env->sd->lb_gained[env->idle], detached);
7275 return detached;
7279 * attach_task() -- attach the task detached by detach_task() to its new rq.
7281 static void attach_task(struct rq *rq, struct task_struct *p)
7283 lockdep_assert_held(&rq->lock);
7285 BUG_ON(task_rq(p) != rq);
7286 activate_task(rq, p, ENQUEUE_NOCLOCK);
7287 p->on_rq = TASK_ON_RQ_QUEUED;
7288 check_preempt_curr(rq, p, 0);
7292 * attach_one_task() -- attaches the task returned from detach_one_task() to
7293 * its new rq.
7295 static void attach_one_task(struct rq *rq, struct task_struct *p)
7297 struct rq_flags rf;
7299 rq_lock(rq, &rf);
7300 update_rq_clock(rq);
7301 attach_task(rq, p);
7302 rq_unlock(rq, &rf);
7306 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7307 * new rq.
7309 static void attach_tasks(struct lb_env *env)
7311 struct list_head *tasks = &env->tasks;
7312 struct task_struct *p;
7313 struct rq_flags rf;
7315 rq_lock(env->dst_rq, &rf);
7316 update_rq_clock(env->dst_rq);
7318 while (!list_empty(tasks)) {
7319 p = list_first_entry(tasks, struct task_struct, se.group_node);
7320 list_del_init(&p->se.group_node);
7322 attach_task(env->dst_rq, p);
7325 rq_unlock(env->dst_rq, &rf);
7328 #ifdef CONFIG_FAIR_GROUP_SCHED
7330 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7332 if (cfs_rq->load.weight)
7333 return false;
7335 if (cfs_rq->avg.load_sum)
7336 return false;
7338 if (cfs_rq->avg.util_sum)
7339 return false;
7341 if (cfs_rq->avg.runnable_load_sum)
7342 return false;
7344 return true;
7347 static void update_blocked_averages(int cpu)
7349 struct rq *rq = cpu_rq(cpu);
7350 struct cfs_rq *cfs_rq, *pos;
7351 struct rq_flags rf;
7353 rq_lock_irqsave(rq, &rf);
7354 update_rq_clock(rq);
7357 * Iterates the task_group tree in a bottom up fashion, see
7358 * list_add_leaf_cfs_rq() for details.
7360 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7361 struct sched_entity *se;
7363 /* throttled entities do not contribute to load */
7364 if (throttled_hierarchy(cfs_rq))
7365 continue;
7367 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
7368 update_tg_load_avg(cfs_rq, 0);
7370 /* Propagate pending load changes to the parent, if any: */
7371 se = cfs_rq->tg->se[cpu];
7372 if (se && !skip_blocked_update(se))
7373 update_load_avg(cfs_rq_of(se), se, 0);
7376 * There can be a lot of idle CPU cgroups. Don't let fully
7377 * decayed cfs_rqs linger on the list.
7379 if (cfs_rq_is_decayed(cfs_rq))
7380 list_del_leaf_cfs_rq(cfs_rq);
7382 rq_unlock_irqrestore(rq, &rf);
7386 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7387 * This needs to be done in a top-down fashion because the load of a child
7388 * group is a fraction of its parents load.
7390 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7392 struct rq *rq = rq_of(cfs_rq);
7393 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7394 unsigned long now = jiffies;
7395 unsigned long load;
7397 if (cfs_rq->last_h_load_update == now)
7398 return;
7400 cfs_rq->h_load_next = NULL;
7401 for_each_sched_entity(se) {
7402 cfs_rq = cfs_rq_of(se);
7403 cfs_rq->h_load_next = se;
7404 if (cfs_rq->last_h_load_update == now)
7405 break;
7408 if (!se) {
7409 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7410 cfs_rq->last_h_load_update = now;
7413 while ((se = cfs_rq->h_load_next) != NULL) {
7414 load = cfs_rq->h_load;
7415 load = div64_ul(load * se->avg.load_avg,
7416 cfs_rq_load_avg(cfs_rq) + 1);
7417 cfs_rq = group_cfs_rq(se);
7418 cfs_rq->h_load = load;
7419 cfs_rq->last_h_load_update = now;
7423 static unsigned long task_h_load(struct task_struct *p)
7425 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7427 update_cfs_rq_h_load(cfs_rq);
7428 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7429 cfs_rq_load_avg(cfs_rq) + 1);
7431 #else
7432 static inline void update_blocked_averages(int cpu)
7434 struct rq *rq = cpu_rq(cpu);
7435 struct cfs_rq *cfs_rq = &rq->cfs;
7436 struct rq_flags rf;
7438 rq_lock_irqsave(rq, &rf);
7439 update_rq_clock(rq);
7440 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
7441 rq_unlock_irqrestore(rq, &rf);
7444 static unsigned long task_h_load(struct task_struct *p)
7446 return p->se.avg.load_avg;
7448 #endif
7450 /********** Helpers for find_busiest_group ************************/
7452 enum group_type {
7453 group_other = 0,
7454 group_imbalanced,
7455 group_overloaded,
7459 * sg_lb_stats - stats of a sched_group required for load_balancing
7461 struct sg_lb_stats {
7462 unsigned long avg_load; /*Avg load across the CPUs of the group */
7463 unsigned long group_load; /* Total load over the CPUs of the group */
7464 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7465 unsigned long load_per_task;
7466 unsigned long group_capacity;
7467 unsigned long group_util; /* Total utilization of the group */
7468 unsigned int sum_nr_running; /* Nr tasks running in the group */
7469 unsigned int idle_cpus;
7470 unsigned int group_weight;
7471 enum group_type group_type;
7472 int group_no_capacity;
7473 #ifdef CONFIG_NUMA_BALANCING
7474 unsigned int nr_numa_running;
7475 unsigned int nr_preferred_running;
7476 #endif
7480 * sd_lb_stats - Structure to store the statistics of a sched_domain
7481 * during load balancing.
7483 struct sd_lb_stats {
7484 struct sched_group *busiest; /* Busiest group in this sd */
7485 struct sched_group *local; /* Local group in this sd */
7486 unsigned long total_running;
7487 unsigned long total_load; /* Total load of all groups in sd */
7488 unsigned long total_capacity; /* Total capacity of all groups in sd */
7489 unsigned long avg_load; /* Average load across all groups in sd */
7491 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7492 struct sg_lb_stats local_stat; /* Statistics of the local group */
7495 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7498 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7499 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7500 * We must however clear busiest_stat::avg_load because
7501 * update_sd_pick_busiest() reads this before assignment.
7503 *sds = (struct sd_lb_stats){
7504 .busiest = NULL,
7505 .local = NULL,
7506 .total_running = 0UL,
7507 .total_load = 0UL,
7508 .total_capacity = 0UL,
7509 .busiest_stat = {
7510 .avg_load = 0UL,
7511 .sum_nr_running = 0,
7512 .group_type = group_other,
7518 * get_sd_load_idx - Obtain the load index for a given sched domain.
7519 * @sd: The sched_domain whose load_idx is to be obtained.
7520 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7522 * Return: The load index.
7524 static inline int get_sd_load_idx(struct sched_domain *sd,
7525 enum cpu_idle_type idle)
7527 int load_idx;
7529 switch (idle) {
7530 case CPU_NOT_IDLE:
7531 load_idx = sd->busy_idx;
7532 break;
7534 case CPU_NEWLY_IDLE:
7535 load_idx = sd->newidle_idx;
7536 break;
7537 default:
7538 load_idx = sd->idle_idx;
7539 break;
7542 return load_idx;
7545 static unsigned long scale_rt_capacity(int cpu)
7547 struct rq *rq = cpu_rq(cpu);
7548 u64 total, used, age_stamp, avg;
7549 s64 delta;
7552 * Since we're reading these variables without serialization make sure
7553 * we read them once before doing sanity checks on them.
7555 age_stamp = READ_ONCE(rq->age_stamp);
7556 avg = READ_ONCE(rq->rt_avg);
7557 delta = __rq_clock_broken(rq) - age_stamp;
7559 if (unlikely(delta < 0))
7560 delta = 0;
7562 total = sched_avg_period() + delta;
7564 used = div_u64(avg, total);
7566 if (likely(used < SCHED_CAPACITY_SCALE))
7567 return SCHED_CAPACITY_SCALE - used;
7569 return 1;
7572 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7574 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7575 struct sched_group *sdg = sd->groups;
7577 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7579 capacity *= scale_rt_capacity(cpu);
7580 capacity >>= SCHED_CAPACITY_SHIFT;
7582 if (!capacity)
7583 capacity = 1;
7585 cpu_rq(cpu)->cpu_capacity = capacity;
7586 sdg->sgc->capacity = capacity;
7587 sdg->sgc->min_capacity = capacity;
7590 void update_group_capacity(struct sched_domain *sd, int cpu)
7592 struct sched_domain *child = sd->child;
7593 struct sched_group *group, *sdg = sd->groups;
7594 unsigned long capacity, min_capacity;
7595 unsigned long interval;
7597 interval = msecs_to_jiffies(sd->balance_interval);
7598 interval = clamp(interval, 1UL, max_load_balance_interval);
7599 sdg->sgc->next_update = jiffies + interval;
7601 if (!child) {
7602 update_cpu_capacity(sd, cpu);
7603 return;
7606 capacity = 0;
7607 min_capacity = ULONG_MAX;
7609 if (child->flags & SD_OVERLAP) {
7611 * SD_OVERLAP domains cannot assume that child groups
7612 * span the current group.
7615 for_each_cpu(cpu, sched_group_span(sdg)) {
7616 struct sched_group_capacity *sgc;
7617 struct rq *rq = cpu_rq(cpu);
7620 * build_sched_domains() -> init_sched_groups_capacity()
7621 * gets here before we've attached the domains to the
7622 * runqueues.
7624 * Use capacity_of(), which is set irrespective of domains
7625 * in update_cpu_capacity().
7627 * This avoids capacity from being 0 and
7628 * causing divide-by-zero issues on boot.
7630 if (unlikely(!rq->sd)) {
7631 capacity += capacity_of(cpu);
7632 } else {
7633 sgc = rq->sd->groups->sgc;
7634 capacity += sgc->capacity;
7637 min_capacity = min(capacity, min_capacity);
7639 } else {
7641 * !SD_OVERLAP domains can assume that child groups
7642 * span the current group.
7645 group = child->groups;
7646 do {
7647 struct sched_group_capacity *sgc = group->sgc;
7649 capacity += sgc->capacity;
7650 min_capacity = min(sgc->min_capacity, min_capacity);
7651 group = group->next;
7652 } while (group != child->groups);
7655 sdg->sgc->capacity = capacity;
7656 sdg->sgc->min_capacity = min_capacity;
7660 * Check whether the capacity of the rq has been noticeably reduced by side
7661 * activity. The imbalance_pct is used for the threshold.
7662 * Return true is the capacity is reduced
7664 static inline int
7665 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7667 return ((rq->cpu_capacity * sd->imbalance_pct) <
7668 (rq->cpu_capacity_orig * 100));
7672 * Group imbalance indicates (and tries to solve) the problem where balancing
7673 * groups is inadequate due to ->cpus_allowed constraints.
7675 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7676 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7677 * Something like:
7679 * { 0 1 2 3 } { 4 5 6 7 }
7680 * * * * *
7682 * If we were to balance group-wise we'd place two tasks in the first group and
7683 * two tasks in the second group. Clearly this is undesired as it will overload
7684 * cpu 3 and leave one of the cpus in the second group unused.
7686 * The current solution to this issue is detecting the skew in the first group
7687 * by noticing the lower domain failed to reach balance and had difficulty
7688 * moving tasks due to affinity constraints.
7690 * When this is so detected; this group becomes a candidate for busiest; see
7691 * update_sd_pick_busiest(). And calculate_imbalance() and
7692 * find_busiest_group() avoid some of the usual balance conditions to allow it
7693 * to create an effective group imbalance.
7695 * This is a somewhat tricky proposition since the next run might not find the
7696 * group imbalance and decide the groups need to be balanced again. A most
7697 * subtle and fragile situation.
7700 static inline int sg_imbalanced(struct sched_group *group)
7702 return group->sgc->imbalance;
7706 * group_has_capacity returns true if the group has spare capacity that could
7707 * be used by some tasks.
7708 * We consider that a group has spare capacity if the * number of task is
7709 * smaller than the number of CPUs or if the utilization is lower than the
7710 * available capacity for CFS tasks.
7711 * For the latter, we use a threshold to stabilize the state, to take into
7712 * account the variance of the tasks' load and to return true if the available
7713 * capacity in meaningful for the load balancer.
7714 * As an example, an available capacity of 1% can appear but it doesn't make
7715 * any benefit for the load balance.
7717 static inline bool
7718 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7720 if (sgs->sum_nr_running < sgs->group_weight)
7721 return true;
7723 if ((sgs->group_capacity * 100) >
7724 (sgs->group_util * env->sd->imbalance_pct))
7725 return true;
7727 return false;
7731 * group_is_overloaded returns true if the group has more tasks than it can
7732 * handle.
7733 * group_is_overloaded is not equals to !group_has_capacity because a group
7734 * with the exact right number of tasks, has no more spare capacity but is not
7735 * overloaded so both group_has_capacity and group_is_overloaded return
7736 * false.
7738 static inline bool
7739 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7741 if (sgs->sum_nr_running <= sgs->group_weight)
7742 return false;
7744 if ((sgs->group_capacity * 100) <
7745 (sgs->group_util * env->sd->imbalance_pct))
7746 return true;
7748 return false;
7752 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7753 * per-CPU capacity than sched_group ref.
7755 static inline bool
7756 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7758 return sg->sgc->min_capacity * capacity_margin <
7759 ref->sgc->min_capacity * 1024;
7762 static inline enum
7763 group_type group_classify(struct sched_group *group,
7764 struct sg_lb_stats *sgs)
7766 if (sgs->group_no_capacity)
7767 return group_overloaded;
7769 if (sg_imbalanced(group))
7770 return group_imbalanced;
7772 return group_other;
7776 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7777 * @env: The load balancing environment.
7778 * @group: sched_group whose statistics are to be updated.
7779 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7780 * @local_group: Does group contain this_cpu.
7781 * @sgs: variable to hold the statistics for this group.
7782 * @overload: Indicate more than one runnable task for any CPU.
7784 static inline void update_sg_lb_stats(struct lb_env *env,
7785 struct sched_group *group, int load_idx,
7786 int local_group, struct sg_lb_stats *sgs,
7787 bool *overload)
7789 unsigned long load;
7790 int i, nr_running;
7792 memset(sgs, 0, sizeof(*sgs));
7794 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
7795 struct rq *rq = cpu_rq(i);
7797 /* Bias balancing toward cpus of our domain */
7798 if (local_group)
7799 load = target_load(i, load_idx);
7800 else
7801 load = source_load(i, load_idx);
7803 sgs->group_load += load;
7804 sgs->group_util += cpu_util(i);
7805 sgs->sum_nr_running += rq->cfs.h_nr_running;
7807 nr_running = rq->nr_running;
7808 if (nr_running > 1)
7809 *overload = true;
7811 #ifdef CONFIG_NUMA_BALANCING
7812 sgs->nr_numa_running += rq->nr_numa_running;
7813 sgs->nr_preferred_running += rq->nr_preferred_running;
7814 #endif
7815 sgs->sum_weighted_load += weighted_cpuload(rq);
7817 * No need to call idle_cpu() if nr_running is not 0
7819 if (!nr_running && idle_cpu(i))
7820 sgs->idle_cpus++;
7823 /* Adjust by relative CPU capacity of the group */
7824 sgs->group_capacity = group->sgc->capacity;
7825 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7827 if (sgs->sum_nr_running)
7828 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7830 sgs->group_weight = group->group_weight;
7832 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7833 sgs->group_type = group_classify(group, sgs);
7837 * update_sd_pick_busiest - return 1 on busiest group
7838 * @env: The load balancing environment.
7839 * @sds: sched_domain statistics
7840 * @sg: sched_group candidate to be checked for being the busiest
7841 * @sgs: sched_group statistics
7843 * Determine if @sg is a busier group than the previously selected
7844 * busiest group.
7846 * Return: %true if @sg is a busier group than the previously selected
7847 * busiest group. %false otherwise.
7849 static bool update_sd_pick_busiest(struct lb_env *env,
7850 struct sd_lb_stats *sds,
7851 struct sched_group *sg,
7852 struct sg_lb_stats *sgs)
7854 struct sg_lb_stats *busiest = &sds->busiest_stat;
7856 if (sgs->group_type > busiest->group_type)
7857 return true;
7859 if (sgs->group_type < busiest->group_type)
7860 return false;
7862 if (sgs->avg_load <= busiest->avg_load)
7863 return false;
7865 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
7866 goto asym_packing;
7869 * Candidate sg has no more than one task per CPU and
7870 * has higher per-CPU capacity. Migrating tasks to less
7871 * capable CPUs may harm throughput. Maximize throughput,
7872 * power/energy consequences are not considered.
7874 if (sgs->sum_nr_running <= sgs->group_weight &&
7875 group_smaller_cpu_capacity(sds->local, sg))
7876 return false;
7878 asym_packing:
7879 /* This is the busiest node in its class. */
7880 if (!(env->sd->flags & SD_ASYM_PACKING))
7881 return true;
7883 /* No ASYM_PACKING if target cpu is already busy */
7884 if (env->idle == CPU_NOT_IDLE)
7885 return true;
7887 * ASYM_PACKING needs to move all the work to the highest
7888 * prority CPUs in the group, therefore mark all groups
7889 * of lower priority than ourself as busy.
7891 if (sgs->sum_nr_running &&
7892 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
7893 if (!sds->busiest)
7894 return true;
7896 /* Prefer to move from lowest priority cpu's work */
7897 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
7898 sg->asym_prefer_cpu))
7899 return true;
7902 return false;
7905 #ifdef CONFIG_NUMA_BALANCING
7906 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7908 if (sgs->sum_nr_running > sgs->nr_numa_running)
7909 return regular;
7910 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7911 return remote;
7912 return all;
7915 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7917 if (rq->nr_running > rq->nr_numa_running)
7918 return regular;
7919 if (rq->nr_running > rq->nr_preferred_running)
7920 return remote;
7921 return all;
7923 #else
7924 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7926 return all;
7929 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7931 return regular;
7933 #endif /* CONFIG_NUMA_BALANCING */
7936 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7937 * @env: The load balancing environment.
7938 * @sds: variable to hold the statistics for this sched_domain.
7940 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7942 struct sched_domain *child = env->sd->child;
7943 struct sched_group *sg = env->sd->groups;
7944 struct sg_lb_stats *local = &sds->local_stat;
7945 struct sg_lb_stats tmp_sgs;
7946 int load_idx, prefer_sibling = 0;
7947 bool overload = false;
7949 if (child && child->flags & SD_PREFER_SIBLING)
7950 prefer_sibling = 1;
7952 load_idx = get_sd_load_idx(env->sd, env->idle);
7954 do {
7955 struct sg_lb_stats *sgs = &tmp_sgs;
7956 int local_group;
7958 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
7959 if (local_group) {
7960 sds->local = sg;
7961 sgs = local;
7963 if (env->idle != CPU_NEWLY_IDLE ||
7964 time_after_eq(jiffies, sg->sgc->next_update))
7965 update_group_capacity(env->sd, env->dst_cpu);
7968 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7969 &overload);
7971 if (local_group)
7972 goto next_group;
7975 * In case the child domain prefers tasks go to siblings
7976 * first, lower the sg capacity so that we'll try
7977 * and move all the excess tasks away. We lower the capacity
7978 * of a group only if the local group has the capacity to fit
7979 * these excess tasks. The extra check prevents the case where
7980 * you always pull from the heaviest group when it is already
7981 * under-utilized (possible with a large weight task outweighs
7982 * the tasks on the system).
7984 if (prefer_sibling && sds->local &&
7985 group_has_capacity(env, local) &&
7986 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
7987 sgs->group_no_capacity = 1;
7988 sgs->group_type = group_classify(sg, sgs);
7991 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7992 sds->busiest = sg;
7993 sds->busiest_stat = *sgs;
7996 next_group:
7997 /* Now, start updating sd_lb_stats */
7998 sds->total_running += sgs->sum_nr_running;
7999 sds->total_load += sgs->group_load;
8000 sds->total_capacity += sgs->group_capacity;
8002 sg = sg->next;
8003 } while (sg != env->sd->groups);
8005 if (env->sd->flags & SD_NUMA)
8006 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8008 if (!env->sd->parent) {
8009 /* update overload indicator if we are at root domain */
8010 if (env->dst_rq->rd->overload != overload)
8011 env->dst_rq->rd->overload = overload;
8016 * check_asym_packing - Check to see if the group is packed into the
8017 * sched domain.
8019 * This is primarily intended to used at the sibling level. Some
8020 * cores like POWER7 prefer to use lower numbered SMT threads. In the
8021 * case of POWER7, it can move to lower SMT modes only when higher
8022 * threads are idle. When in lower SMT modes, the threads will
8023 * perform better since they share less core resources. Hence when we
8024 * have idle threads, we want them to be the higher ones.
8026 * This packing function is run on idle threads. It checks to see if
8027 * the busiest CPU in this domain (core in the P7 case) has a higher
8028 * CPU number than the packing function is being run on. Here we are
8029 * assuming lower CPU number will be equivalent to lower a SMT thread
8030 * number.
8032 * Return: 1 when packing is required and a task should be moved to
8033 * this CPU. The amount of the imbalance is returned in env->imbalance.
8035 * @env: The load balancing environment.
8036 * @sds: Statistics of the sched_domain which is to be packed
8038 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
8040 int busiest_cpu;
8042 if (!(env->sd->flags & SD_ASYM_PACKING))
8043 return 0;
8045 if (env->idle == CPU_NOT_IDLE)
8046 return 0;
8048 if (!sds->busiest)
8049 return 0;
8051 busiest_cpu = sds->busiest->asym_prefer_cpu;
8052 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
8053 return 0;
8055 env->imbalance = DIV_ROUND_CLOSEST(
8056 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
8057 SCHED_CAPACITY_SCALE);
8059 return 1;
8063 * fix_small_imbalance - Calculate the minor imbalance that exists
8064 * amongst the groups of a sched_domain, during
8065 * load balancing.
8066 * @env: The load balancing environment.
8067 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
8069 static inline
8070 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8072 unsigned long tmp, capa_now = 0, capa_move = 0;
8073 unsigned int imbn = 2;
8074 unsigned long scaled_busy_load_per_task;
8075 struct sg_lb_stats *local, *busiest;
8077 local = &sds->local_stat;
8078 busiest = &sds->busiest_stat;
8080 if (!local->sum_nr_running)
8081 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
8082 else if (busiest->load_per_task > local->load_per_task)
8083 imbn = 1;
8085 scaled_busy_load_per_task =
8086 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8087 busiest->group_capacity;
8089 if (busiest->avg_load + scaled_busy_load_per_task >=
8090 local->avg_load + (scaled_busy_load_per_task * imbn)) {
8091 env->imbalance = busiest->load_per_task;
8092 return;
8096 * OK, we don't have enough imbalance to justify moving tasks,
8097 * however we may be able to increase total CPU capacity used by
8098 * moving them.
8101 capa_now += busiest->group_capacity *
8102 min(busiest->load_per_task, busiest->avg_load);
8103 capa_now += local->group_capacity *
8104 min(local->load_per_task, local->avg_load);
8105 capa_now /= SCHED_CAPACITY_SCALE;
8107 /* Amount of load we'd subtract */
8108 if (busiest->avg_load > scaled_busy_load_per_task) {
8109 capa_move += busiest->group_capacity *
8110 min(busiest->load_per_task,
8111 busiest->avg_load - scaled_busy_load_per_task);
8114 /* Amount of load we'd add */
8115 if (busiest->avg_load * busiest->group_capacity <
8116 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
8117 tmp = (busiest->avg_load * busiest->group_capacity) /
8118 local->group_capacity;
8119 } else {
8120 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8121 local->group_capacity;
8123 capa_move += local->group_capacity *
8124 min(local->load_per_task, local->avg_load + tmp);
8125 capa_move /= SCHED_CAPACITY_SCALE;
8127 /* Move if we gain throughput */
8128 if (capa_move > capa_now)
8129 env->imbalance = busiest->load_per_task;
8133 * calculate_imbalance - Calculate the amount of imbalance present within the
8134 * groups of a given sched_domain during load balance.
8135 * @env: load balance environment
8136 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8138 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8140 unsigned long max_pull, load_above_capacity = ~0UL;
8141 struct sg_lb_stats *local, *busiest;
8143 local = &sds->local_stat;
8144 busiest = &sds->busiest_stat;
8146 if (busiest->group_type == group_imbalanced) {
8148 * In the group_imb case we cannot rely on group-wide averages
8149 * to ensure cpu-load equilibrium, look at wider averages. XXX
8151 busiest->load_per_task =
8152 min(busiest->load_per_task, sds->avg_load);
8156 * Avg load of busiest sg can be less and avg load of local sg can
8157 * be greater than avg load across all sgs of sd because avg load
8158 * factors in sg capacity and sgs with smaller group_type are
8159 * skipped when updating the busiest sg:
8161 if (busiest->avg_load <= sds->avg_load ||
8162 local->avg_load >= sds->avg_load) {
8163 env->imbalance = 0;
8164 return fix_small_imbalance(env, sds);
8168 * If there aren't any idle cpus, avoid creating some.
8170 if (busiest->group_type == group_overloaded &&
8171 local->group_type == group_overloaded) {
8172 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
8173 if (load_above_capacity > busiest->group_capacity) {
8174 load_above_capacity -= busiest->group_capacity;
8175 load_above_capacity *= scale_load_down(NICE_0_LOAD);
8176 load_above_capacity /= busiest->group_capacity;
8177 } else
8178 load_above_capacity = ~0UL;
8182 * We're trying to get all the cpus to the average_load, so we don't
8183 * want to push ourselves above the average load, nor do we wish to
8184 * reduce the max loaded cpu below the average load. At the same time,
8185 * we also don't want to reduce the group load below the group
8186 * capacity. Thus we look for the minimum possible imbalance.
8188 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8190 /* How much load to actually move to equalise the imbalance */
8191 env->imbalance = min(
8192 max_pull * busiest->group_capacity,
8193 (sds->avg_load - local->avg_load) * local->group_capacity
8194 ) / SCHED_CAPACITY_SCALE;
8197 * if *imbalance is less than the average load per runnable task
8198 * there is no guarantee that any tasks will be moved so we'll have
8199 * a think about bumping its value to force at least one task to be
8200 * moved
8202 if (env->imbalance < busiest->load_per_task)
8203 return fix_small_imbalance(env, sds);
8206 /******* find_busiest_group() helpers end here *********************/
8209 * find_busiest_group - Returns the busiest group within the sched_domain
8210 * if there is an imbalance.
8212 * Also calculates the amount of weighted load which should be moved
8213 * to restore balance.
8215 * @env: The load balancing environment.
8217 * Return: - The busiest group if imbalance exists.
8219 static struct sched_group *find_busiest_group(struct lb_env *env)
8221 struct sg_lb_stats *local, *busiest;
8222 struct sd_lb_stats sds;
8224 init_sd_lb_stats(&sds);
8227 * Compute the various statistics relavent for load balancing at
8228 * this level.
8230 update_sd_lb_stats(env, &sds);
8231 local = &sds.local_stat;
8232 busiest = &sds.busiest_stat;
8234 /* ASYM feature bypasses nice load balance check */
8235 if (check_asym_packing(env, &sds))
8236 return sds.busiest;
8238 /* There is no busy sibling group to pull tasks from */
8239 if (!sds.busiest || busiest->sum_nr_running == 0)
8240 goto out_balanced;
8242 /* XXX broken for overlapping NUMA groups */
8243 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8244 / sds.total_capacity;
8247 * If the busiest group is imbalanced the below checks don't
8248 * work because they assume all things are equal, which typically
8249 * isn't true due to cpus_allowed constraints and the like.
8251 if (busiest->group_type == group_imbalanced)
8252 goto force_balance;
8255 * When dst_cpu is idle, prevent SMP nice and/or asymmetric group
8256 * capacities from resulting in underutilization due to avg_load.
8258 if (env->idle != CPU_NOT_IDLE && group_has_capacity(env, local) &&
8259 busiest->group_no_capacity)
8260 goto force_balance;
8263 * If the local group is busier than the selected busiest group
8264 * don't try and pull any tasks.
8266 if (local->avg_load >= busiest->avg_load)
8267 goto out_balanced;
8270 * Don't pull any tasks if this group is already above the domain
8271 * average load.
8273 if (local->avg_load >= sds.avg_load)
8274 goto out_balanced;
8276 if (env->idle == CPU_IDLE) {
8278 * This cpu is idle. If the busiest group is not overloaded
8279 * and there is no imbalance between this and busiest group
8280 * wrt idle cpus, it is balanced. The imbalance becomes
8281 * significant if the diff is greater than 1 otherwise we
8282 * might end up to just move the imbalance on another group
8284 if ((busiest->group_type != group_overloaded) &&
8285 (local->idle_cpus <= (busiest->idle_cpus + 1)))
8286 goto out_balanced;
8287 } else {
8289 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8290 * imbalance_pct to be conservative.
8292 if (100 * busiest->avg_load <=
8293 env->sd->imbalance_pct * local->avg_load)
8294 goto out_balanced;
8297 force_balance:
8298 /* Looks like there is an imbalance. Compute it */
8299 calculate_imbalance(env, &sds);
8300 return sds.busiest;
8302 out_balanced:
8303 env->imbalance = 0;
8304 return NULL;
8308 * find_busiest_queue - find the busiest runqueue among the cpus in group.
8310 static struct rq *find_busiest_queue(struct lb_env *env,
8311 struct sched_group *group)
8313 struct rq *busiest = NULL, *rq;
8314 unsigned long busiest_load = 0, busiest_capacity = 1;
8315 int i;
8317 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8318 unsigned long capacity, wl;
8319 enum fbq_type rt;
8321 rq = cpu_rq(i);
8322 rt = fbq_classify_rq(rq);
8325 * We classify groups/runqueues into three groups:
8326 * - regular: there are !numa tasks
8327 * - remote: there are numa tasks that run on the 'wrong' node
8328 * - all: there is no distinction
8330 * In order to avoid migrating ideally placed numa tasks,
8331 * ignore those when there's better options.
8333 * If we ignore the actual busiest queue to migrate another
8334 * task, the next balance pass can still reduce the busiest
8335 * queue by moving tasks around inside the node.
8337 * If we cannot move enough load due to this classification
8338 * the next pass will adjust the group classification and
8339 * allow migration of more tasks.
8341 * Both cases only affect the total convergence complexity.
8343 if (rt > env->fbq_type)
8344 continue;
8346 capacity = capacity_of(i);
8348 wl = weighted_cpuload(rq);
8351 * When comparing with imbalance, use weighted_cpuload()
8352 * which is not scaled with the cpu capacity.
8355 if (rq->nr_running == 1 && wl > env->imbalance &&
8356 !check_cpu_capacity(rq, env->sd))
8357 continue;
8360 * For the load comparisons with the other cpu's, consider
8361 * the weighted_cpuload() scaled with the cpu capacity, so
8362 * that the load can be moved away from the cpu that is
8363 * potentially running at a lower capacity.
8365 * Thus we're looking for max(wl_i / capacity_i), crosswise
8366 * multiplication to rid ourselves of the division works out
8367 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8368 * our previous maximum.
8370 if (wl * busiest_capacity > busiest_load * capacity) {
8371 busiest_load = wl;
8372 busiest_capacity = capacity;
8373 busiest = rq;
8377 return busiest;
8381 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8382 * so long as it is large enough.
8384 #define MAX_PINNED_INTERVAL 512
8386 static int need_active_balance(struct lb_env *env)
8388 struct sched_domain *sd = env->sd;
8390 if (env->idle == CPU_NEWLY_IDLE) {
8393 * ASYM_PACKING needs to force migrate tasks from busy but
8394 * lower priority CPUs in order to pack all tasks in the
8395 * highest priority CPUs.
8397 if ((sd->flags & SD_ASYM_PACKING) &&
8398 sched_asym_prefer(env->dst_cpu, env->src_cpu))
8399 return 1;
8403 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8404 * It's worth migrating the task if the src_cpu's capacity is reduced
8405 * because of other sched_class or IRQs if more capacity stays
8406 * available on dst_cpu.
8408 if ((env->idle != CPU_NOT_IDLE) &&
8409 (env->src_rq->cfs.h_nr_running == 1)) {
8410 if ((check_cpu_capacity(env->src_rq, sd)) &&
8411 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8412 return 1;
8415 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8418 static int active_load_balance_cpu_stop(void *data);
8420 static int should_we_balance(struct lb_env *env)
8422 struct sched_group *sg = env->sd->groups;
8423 int cpu, balance_cpu = -1;
8426 * Ensure the balancing environment is consistent; can happen
8427 * when the softirq triggers 'during' hotplug.
8429 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
8430 return 0;
8433 * In the newly idle case, we will allow all the cpu's
8434 * to do the newly idle load balance.
8436 if (env->idle == CPU_NEWLY_IDLE)
8437 return 1;
8439 /* Try to find first idle cpu */
8440 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8441 if (!idle_cpu(cpu))
8442 continue;
8444 balance_cpu = cpu;
8445 break;
8448 if (balance_cpu == -1)
8449 balance_cpu = group_balance_cpu(sg);
8452 * First idle cpu or the first cpu(busiest) in this sched group
8453 * is eligible for doing load balancing at this and above domains.
8455 return balance_cpu == env->dst_cpu;
8459 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8460 * tasks if there is an imbalance.
8462 static int load_balance(int this_cpu, struct rq *this_rq,
8463 struct sched_domain *sd, enum cpu_idle_type idle,
8464 int *continue_balancing)
8466 int ld_moved, cur_ld_moved, active_balance = 0;
8467 struct sched_domain *sd_parent = sd->parent;
8468 struct sched_group *group;
8469 struct rq *busiest;
8470 struct rq_flags rf;
8471 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8473 struct lb_env env = {
8474 .sd = sd,
8475 .dst_cpu = this_cpu,
8476 .dst_rq = this_rq,
8477 .dst_grpmask = sched_group_span(sd->groups),
8478 .idle = idle,
8479 .loop_break = sched_nr_migrate_break,
8480 .cpus = cpus,
8481 .fbq_type = all,
8482 .tasks = LIST_HEAD_INIT(env.tasks),
8485 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8487 schedstat_inc(sd->lb_count[idle]);
8489 redo:
8490 if (!should_we_balance(&env)) {
8491 *continue_balancing = 0;
8492 goto out_balanced;
8495 group = find_busiest_group(&env);
8496 if (!group) {
8497 schedstat_inc(sd->lb_nobusyg[idle]);
8498 goto out_balanced;
8501 busiest = find_busiest_queue(&env, group);
8502 if (!busiest) {
8503 schedstat_inc(sd->lb_nobusyq[idle]);
8504 goto out_balanced;
8507 BUG_ON(busiest == env.dst_rq);
8509 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8511 env.src_cpu = busiest->cpu;
8512 env.src_rq = busiest;
8514 ld_moved = 0;
8515 if (busiest->nr_running > 1) {
8517 * Attempt to move tasks. If find_busiest_group has found
8518 * an imbalance but busiest->nr_running <= 1, the group is
8519 * still unbalanced. ld_moved simply stays zero, so it is
8520 * correctly treated as an imbalance.
8522 env.flags |= LBF_ALL_PINNED;
8523 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8525 more_balance:
8526 rq_lock_irqsave(busiest, &rf);
8527 update_rq_clock(busiest);
8530 * cur_ld_moved - load moved in current iteration
8531 * ld_moved - cumulative load moved across iterations
8533 cur_ld_moved = detach_tasks(&env);
8536 * We've detached some tasks from busiest_rq. Every
8537 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8538 * unlock busiest->lock, and we are able to be sure
8539 * that nobody can manipulate the tasks in parallel.
8540 * See task_rq_lock() family for the details.
8543 rq_unlock(busiest, &rf);
8545 if (cur_ld_moved) {
8546 attach_tasks(&env);
8547 ld_moved += cur_ld_moved;
8550 local_irq_restore(rf.flags);
8552 if (env.flags & LBF_NEED_BREAK) {
8553 env.flags &= ~LBF_NEED_BREAK;
8554 goto more_balance;
8558 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8559 * us and move them to an alternate dst_cpu in our sched_group
8560 * where they can run. The upper limit on how many times we
8561 * iterate on same src_cpu is dependent on number of cpus in our
8562 * sched_group.
8564 * This changes load balance semantics a bit on who can move
8565 * load to a given_cpu. In addition to the given_cpu itself
8566 * (or a ilb_cpu acting on its behalf where given_cpu is
8567 * nohz-idle), we now have balance_cpu in a position to move
8568 * load to given_cpu. In rare situations, this may cause
8569 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8570 * _independently_ and at _same_ time to move some load to
8571 * given_cpu) causing exceess load to be moved to given_cpu.
8572 * This however should not happen so much in practice and
8573 * moreover subsequent load balance cycles should correct the
8574 * excess load moved.
8576 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8578 /* Prevent to re-select dst_cpu via env's cpus */
8579 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8581 env.dst_rq = cpu_rq(env.new_dst_cpu);
8582 env.dst_cpu = env.new_dst_cpu;
8583 env.flags &= ~LBF_DST_PINNED;
8584 env.loop = 0;
8585 env.loop_break = sched_nr_migrate_break;
8588 * Go back to "more_balance" rather than "redo" since we
8589 * need to continue with same src_cpu.
8591 goto more_balance;
8595 * We failed to reach balance because of affinity.
8597 if (sd_parent) {
8598 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8600 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8601 *group_imbalance = 1;
8604 /* All tasks on this runqueue were pinned by CPU affinity */
8605 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8606 cpumask_clear_cpu(cpu_of(busiest), cpus);
8608 * Attempting to continue load balancing at the current
8609 * sched_domain level only makes sense if there are
8610 * active CPUs remaining as possible busiest CPUs to
8611 * pull load from which are not contained within the
8612 * destination group that is receiving any migrated
8613 * load.
8615 if (!cpumask_subset(cpus, env.dst_grpmask)) {
8616 env.loop = 0;
8617 env.loop_break = sched_nr_migrate_break;
8618 goto redo;
8620 goto out_all_pinned;
8624 if (!ld_moved) {
8625 schedstat_inc(sd->lb_failed[idle]);
8627 * Increment the failure counter only on periodic balance.
8628 * We do not want newidle balance, which can be very
8629 * frequent, pollute the failure counter causing
8630 * excessive cache_hot migrations and active balances.
8632 if (idle != CPU_NEWLY_IDLE)
8633 sd->nr_balance_failed++;
8635 if (need_active_balance(&env)) {
8636 unsigned long flags;
8638 raw_spin_lock_irqsave(&busiest->lock, flags);
8640 /* don't kick the active_load_balance_cpu_stop,
8641 * if the curr task on busiest cpu can't be
8642 * moved to this_cpu
8644 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
8645 raw_spin_unlock_irqrestore(&busiest->lock,
8646 flags);
8647 env.flags |= LBF_ALL_PINNED;
8648 goto out_one_pinned;
8652 * ->active_balance synchronizes accesses to
8653 * ->active_balance_work. Once set, it's cleared
8654 * only after active load balance is finished.
8656 if (!busiest->active_balance) {
8657 busiest->active_balance = 1;
8658 busiest->push_cpu = this_cpu;
8659 active_balance = 1;
8661 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8663 if (active_balance) {
8664 stop_one_cpu_nowait(cpu_of(busiest),
8665 active_load_balance_cpu_stop, busiest,
8666 &busiest->active_balance_work);
8669 /* We've kicked active balancing, force task migration. */
8670 sd->nr_balance_failed = sd->cache_nice_tries+1;
8672 } else
8673 sd->nr_balance_failed = 0;
8675 if (likely(!active_balance)) {
8676 /* We were unbalanced, so reset the balancing interval */
8677 sd->balance_interval = sd->min_interval;
8678 } else {
8680 * If we've begun active balancing, start to back off. This
8681 * case may not be covered by the all_pinned logic if there
8682 * is only 1 task on the busy runqueue (because we don't call
8683 * detach_tasks).
8685 if (sd->balance_interval < sd->max_interval)
8686 sd->balance_interval *= 2;
8689 goto out;
8691 out_balanced:
8693 * We reach balance although we may have faced some affinity
8694 * constraints. Clear the imbalance flag if it was set.
8696 if (sd_parent) {
8697 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8699 if (*group_imbalance)
8700 *group_imbalance = 0;
8703 out_all_pinned:
8705 * We reach balance because all tasks are pinned at this level so
8706 * we can't migrate them. Let the imbalance flag set so parent level
8707 * can try to migrate them.
8709 schedstat_inc(sd->lb_balanced[idle]);
8711 sd->nr_balance_failed = 0;
8713 out_one_pinned:
8714 /* tune up the balancing interval */
8715 if (((env.flags & LBF_ALL_PINNED) &&
8716 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8717 (sd->balance_interval < sd->max_interval))
8718 sd->balance_interval *= 2;
8720 ld_moved = 0;
8721 out:
8722 return ld_moved;
8725 static inline unsigned long
8726 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8728 unsigned long interval = sd->balance_interval;
8730 if (cpu_busy)
8731 interval *= sd->busy_factor;
8733 /* scale ms to jiffies */
8734 interval = msecs_to_jiffies(interval);
8735 interval = clamp(interval, 1UL, max_load_balance_interval);
8737 return interval;
8740 static inline void
8741 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
8743 unsigned long interval, next;
8745 /* used by idle balance, so cpu_busy = 0 */
8746 interval = get_sd_balance_interval(sd, 0);
8747 next = sd->last_balance + interval;
8749 if (time_after(*next_balance, next))
8750 *next_balance = next;
8754 * idle_balance is called by schedule() if this_cpu is about to become
8755 * idle. Attempts to pull tasks from other CPUs.
8757 static int idle_balance(struct rq *this_rq, struct rq_flags *rf)
8759 unsigned long next_balance = jiffies + HZ;
8760 int this_cpu = this_rq->cpu;
8761 struct sched_domain *sd;
8762 int pulled_task = 0;
8763 u64 curr_cost = 0;
8766 * We must set idle_stamp _before_ calling idle_balance(), such that we
8767 * measure the duration of idle_balance() as idle time.
8769 this_rq->idle_stamp = rq_clock(this_rq);
8772 * Do not pull tasks towards !active CPUs...
8774 if (!cpu_active(this_cpu))
8775 return 0;
8778 * This is OK, because current is on_cpu, which avoids it being picked
8779 * for load-balance and preemption/IRQs are still disabled avoiding
8780 * further scheduler activity on it and we're being very careful to
8781 * re-start the picking loop.
8783 rq_unpin_lock(this_rq, rf);
8785 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
8786 !this_rq->rd->overload) {
8787 rcu_read_lock();
8788 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8789 if (sd)
8790 update_next_balance(sd, &next_balance);
8791 rcu_read_unlock();
8793 goto out;
8796 raw_spin_unlock(&this_rq->lock);
8798 update_blocked_averages(this_cpu);
8799 rcu_read_lock();
8800 for_each_domain(this_cpu, sd) {
8801 int continue_balancing = 1;
8802 u64 t0, domain_cost;
8804 if (!(sd->flags & SD_LOAD_BALANCE))
8805 continue;
8807 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8808 update_next_balance(sd, &next_balance);
8809 break;
8812 if (sd->flags & SD_BALANCE_NEWIDLE) {
8813 t0 = sched_clock_cpu(this_cpu);
8815 pulled_task = load_balance(this_cpu, this_rq,
8816 sd, CPU_NEWLY_IDLE,
8817 &continue_balancing);
8819 domain_cost = sched_clock_cpu(this_cpu) - t0;
8820 if (domain_cost > sd->max_newidle_lb_cost)
8821 sd->max_newidle_lb_cost = domain_cost;
8823 curr_cost += domain_cost;
8826 update_next_balance(sd, &next_balance);
8829 * Stop searching for tasks to pull if there are
8830 * now runnable tasks on this rq.
8832 if (pulled_task || this_rq->nr_running > 0)
8833 break;
8835 rcu_read_unlock();
8837 raw_spin_lock(&this_rq->lock);
8839 if (curr_cost > this_rq->max_idle_balance_cost)
8840 this_rq->max_idle_balance_cost = curr_cost;
8843 * While browsing the domains, we released the rq lock, a task could
8844 * have been enqueued in the meantime. Since we're not going idle,
8845 * pretend we pulled a task.
8847 if (this_rq->cfs.h_nr_running && !pulled_task)
8848 pulled_task = 1;
8850 out:
8851 /* Move the next balance forward */
8852 if (time_after(this_rq->next_balance, next_balance))
8853 this_rq->next_balance = next_balance;
8855 /* Is there a task of a high priority class? */
8856 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8857 pulled_task = -1;
8859 if (pulled_task)
8860 this_rq->idle_stamp = 0;
8862 rq_repin_lock(this_rq, rf);
8864 return pulled_task;
8868 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8869 * running tasks off the busiest CPU onto idle CPUs. It requires at
8870 * least 1 task to be running on each physical CPU where possible, and
8871 * avoids physical / logical imbalances.
8873 static int active_load_balance_cpu_stop(void *data)
8875 struct rq *busiest_rq = data;
8876 int busiest_cpu = cpu_of(busiest_rq);
8877 int target_cpu = busiest_rq->push_cpu;
8878 struct rq *target_rq = cpu_rq(target_cpu);
8879 struct sched_domain *sd;
8880 struct task_struct *p = NULL;
8881 struct rq_flags rf;
8883 rq_lock_irq(busiest_rq, &rf);
8885 * Between queueing the stop-work and running it is a hole in which
8886 * CPUs can become inactive. We should not move tasks from or to
8887 * inactive CPUs.
8889 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
8890 goto out_unlock;
8892 /* make sure the requested cpu hasn't gone down in the meantime */
8893 if (unlikely(busiest_cpu != smp_processor_id() ||
8894 !busiest_rq->active_balance))
8895 goto out_unlock;
8897 /* Is there any task to move? */
8898 if (busiest_rq->nr_running <= 1)
8899 goto out_unlock;
8902 * This condition is "impossible", if it occurs
8903 * we need to fix it. Originally reported by
8904 * Bjorn Helgaas on a 128-cpu setup.
8906 BUG_ON(busiest_rq == target_rq);
8908 /* Search for an sd spanning us and the target CPU. */
8909 rcu_read_lock();
8910 for_each_domain(target_cpu, sd) {
8911 if ((sd->flags & SD_LOAD_BALANCE) &&
8912 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8913 break;
8916 if (likely(sd)) {
8917 struct lb_env env = {
8918 .sd = sd,
8919 .dst_cpu = target_cpu,
8920 .dst_rq = target_rq,
8921 .src_cpu = busiest_rq->cpu,
8922 .src_rq = busiest_rq,
8923 .idle = CPU_IDLE,
8925 * can_migrate_task() doesn't need to compute new_dst_cpu
8926 * for active balancing. Since we have CPU_IDLE, but no
8927 * @dst_grpmask we need to make that test go away with lying
8928 * about DST_PINNED.
8930 .flags = LBF_DST_PINNED,
8933 schedstat_inc(sd->alb_count);
8934 update_rq_clock(busiest_rq);
8936 p = detach_one_task(&env);
8937 if (p) {
8938 schedstat_inc(sd->alb_pushed);
8939 /* Active balancing done, reset the failure counter. */
8940 sd->nr_balance_failed = 0;
8941 } else {
8942 schedstat_inc(sd->alb_failed);
8945 rcu_read_unlock();
8946 out_unlock:
8947 busiest_rq->active_balance = 0;
8948 rq_unlock(busiest_rq, &rf);
8950 if (p)
8951 attach_one_task(target_rq, p);
8953 local_irq_enable();
8955 return 0;
8958 static inline int on_null_domain(struct rq *rq)
8960 return unlikely(!rcu_dereference_sched(rq->sd));
8963 #ifdef CONFIG_NO_HZ_COMMON
8965 * idle load balancing details
8966 * - When one of the busy CPUs notice that there may be an idle rebalancing
8967 * needed, they will kick the idle load balancer, which then does idle
8968 * load balancing for all the idle CPUs.
8970 static struct {
8971 cpumask_var_t idle_cpus_mask;
8972 atomic_t nr_cpus;
8973 unsigned long next_balance; /* in jiffy units */
8974 } nohz ____cacheline_aligned;
8976 static inline int find_new_ilb(void)
8978 int ilb = cpumask_first(nohz.idle_cpus_mask);
8980 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8981 return ilb;
8983 return nr_cpu_ids;
8987 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8988 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8989 * CPU (if there is one).
8991 static void nohz_balancer_kick(void)
8993 int ilb_cpu;
8995 nohz.next_balance++;
8997 ilb_cpu = find_new_ilb();
8999 if (ilb_cpu >= nr_cpu_ids)
9000 return;
9002 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
9003 return;
9005 * Use smp_send_reschedule() instead of resched_cpu().
9006 * This way we generate a sched IPI on the target cpu which
9007 * is idle. And the softirq performing nohz idle load balance
9008 * will be run before returning from the IPI.
9010 smp_send_reschedule(ilb_cpu);
9011 return;
9014 void nohz_balance_exit_idle(unsigned int cpu)
9016 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
9018 * Completely isolated CPUs don't ever set, so we must test.
9020 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
9021 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
9022 atomic_dec(&nohz.nr_cpus);
9024 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
9028 static inline void set_cpu_sd_state_busy(void)
9030 struct sched_domain *sd;
9031 int cpu = smp_processor_id();
9033 rcu_read_lock();
9034 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9036 if (!sd || !sd->nohz_idle)
9037 goto unlock;
9038 sd->nohz_idle = 0;
9040 atomic_inc(&sd->shared->nr_busy_cpus);
9041 unlock:
9042 rcu_read_unlock();
9045 void set_cpu_sd_state_idle(void)
9047 struct sched_domain *sd;
9048 int cpu = smp_processor_id();
9050 rcu_read_lock();
9051 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9053 if (!sd || sd->nohz_idle)
9054 goto unlock;
9055 sd->nohz_idle = 1;
9057 atomic_dec(&sd->shared->nr_busy_cpus);
9058 unlock:
9059 rcu_read_unlock();
9063 * This routine will record that the cpu is going idle with tick stopped.
9064 * This info will be used in performing idle load balancing in the future.
9066 void nohz_balance_enter_idle(int cpu)
9069 * If this cpu is going down, then nothing needs to be done.
9071 if (!cpu_active(cpu))
9072 return;
9074 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
9075 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
9076 return;
9078 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
9079 return;
9082 * If we're a completely isolated CPU, we don't play.
9084 if (on_null_domain(cpu_rq(cpu)))
9085 return;
9087 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9088 atomic_inc(&nohz.nr_cpus);
9089 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
9091 #endif
9093 static DEFINE_SPINLOCK(balancing);
9096 * Scale the max load_balance interval with the number of CPUs in the system.
9097 * This trades load-balance latency on larger machines for less cross talk.
9099 void update_max_interval(void)
9101 max_load_balance_interval = HZ*num_online_cpus()/10;
9105 * It checks each scheduling domain to see if it is due to be balanced,
9106 * and initiates a balancing operation if so.
9108 * Balancing parameters are set up in init_sched_domains.
9110 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
9112 int continue_balancing = 1;
9113 int cpu = rq->cpu;
9114 unsigned long interval;
9115 struct sched_domain *sd;
9116 /* Earliest time when we have to do rebalance again */
9117 unsigned long next_balance = jiffies + 60*HZ;
9118 int update_next_balance = 0;
9119 int need_serialize, need_decay = 0;
9120 u64 max_cost = 0;
9122 update_blocked_averages(cpu);
9124 rcu_read_lock();
9125 for_each_domain(cpu, sd) {
9127 * Decay the newidle max times here because this is a regular
9128 * visit to all the domains. Decay ~1% per second.
9130 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
9131 sd->max_newidle_lb_cost =
9132 (sd->max_newidle_lb_cost * 253) / 256;
9133 sd->next_decay_max_lb_cost = jiffies + HZ;
9134 need_decay = 1;
9136 max_cost += sd->max_newidle_lb_cost;
9138 if (!(sd->flags & SD_LOAD_BALANCE))
9139 continue;
9142 * Stop the load balance at this level. There is another
9143 * CPU in our sched group which is doing load balancing more
9144 * actively.
9146 if (!continue_balancing) {
9147 if (need_decay)
9148 continue;
9149 break;
9152 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9154 need_serialize = sd->flags & SD_SERIALIZE;
9155 if (need_serialize) {
9156 if (!spin_trylock(&balancing))
9157 goto out;
9160 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9161 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9163 * The LBF_DST_PINNED logic could have changed
9164 * env->dst_cpu, so we can't know our idle
9165 * state even if we migrated tasks. Update it.
9167 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9169 sd->last_balance = jiffies;
9170 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9172 if (need_serialize)
9173 spin_unlock(&balancing);
9174 out:
9175 if (time_after(next_balance, sd->last_balance + interval)) {
9176 next_balance = sd->last_balance + interval;
9177 update_next_balance = 1;
9180 if (need_decay) {
9182 * Ensure the rq-wide value also decays but keep it at a
9183 * reasonable floor to avoid funnies with rq->avg_idle.
9185 rq->max_idle_balance_cost =
9186 max((u64)sysctl_sched_migration_cost, max_cost);
9188 rcu_read_unlock();
9191 * next_balance will be updated only when there is a need.
9192 * When the cpu is attached to null domain for ex, it will not be
9193 * updated.
9195 if (likely(update_next_balance)) {
9196 rq->next_balance = next_balance;
9198 #ifdef CONFIG_NO_HZ_COMMON
9200 * If this CPU has been elected to perform the nohz idle
9201 * balance. Other idle CPUs have already rebalanced with
9202 * nohz_idle_balance() and nohz.next_balance has been
9203 * updated accordingly. This CPU is now running the idle load
9204 * balance for itself and we need to update the
9205 * nohz.next_balance accordingly.
9207 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9208 nohz.next_balance = rq->next_balance;
9209 #endif
9213 #ifdef CONFIG_NO_HZ_COMMON
9215 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9216 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9218 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9220 int this_cpu = this_rq->cpu;
9221 struct rq *rq;
9222 int balance_cpu;
9223 /* Earliest time when we have to do rebalance again */
9224 unsigned long next_balance = jiffies + 60*HZ;
9225 int update_next_balance = 0;
9227 if (idle != CPU_IDLE ||
9228 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
9229 goto end;
9231 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9232 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9233 continue;
9236 * If this cpu gets work to do, stop the load balancing
9237 * work being done for other cpus. Next load
9238 * balancing owner will pick it up.
9240 if (need_resched())
9241 break;
9243 rq = cpu_rq(balance_cpu);
9246 * If time for next balance is due,
9247 * do the balance.
9249 if (time_after_eq(jiffies, rq->next_balance)) {
9250 struct rq_flags rf;
9252 rq_lock_irq(rq, &rf);
9253 update_rq_clock(rq);
9254 cpu_load_update_idle(rq);
9255 rq_unlock_irq(rq, &rf);
9257 rebalance_domains(rq, CPU_IDLE);
9260 if (time_after(next_balance, rq->next_balance)) {
9261 next_balance = rq->next_balance;
9262 update_next_balance = 1;
9267 * next_balance will be updated only when there is a need.
9268 * When the CPU is attached to null domain for ex, it will not be
9269 * updated.
9271 if (likely(update_next_balance))
9272 nohz.next_balance = next_balance;
9273 end:
9274 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
9278 * Current heuristic for kicking the idle load balancer in the presence
9279 * of an idle cpu in the system.
9280 * - This rq has more than one task.
9281 * - This rq has at least one CFS task and the capacity of the CPU is
9282 * significantly reduced because of RT tasks or IRQs.
9283 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
9284 * multiple busy cpu.
9285 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
9286 * domain span are idle.
9288 static inline bool nohz_kick_needed(struct rq *rq)
9290 unsigned long now = jiffies;
9291 struct sched_domain_shared *sds;
9292 struct sched_domain *sd;
9293 int nr_busy, i, cpu = rq->cpu;
9294 bool kick = false;
9296 if (unlikely(rq->idle_balance))
9297 return false;
9300 * We may be recently in ticked or tickless idle mode. At the first
9301 * busy tick after returning from idle, we will update the busy stats.
9303 set_cpu_sd_state_busy();
9304 nohz_balance_exit_idle(cpu);
9307 * None are in tickless mode and hence no need for NOHZ idle load
9308 * balancing.
9310 if (likely(!atomic_read(&nohz.nr_cpus)))
9311 return false;
9313 if (time_before(now, nohz.next_balance))
9314 return false;
9316 if (rq->nr_running >= 2)
9317 return true;
9319 rcu_read_lock();
9320 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
9321 if (sds) {
9323 * XXX: write a coherent comment on why we do this.
9324 * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com
9326 nr_busy = atomic_read(&sds->nr_busy_cpus);
9327 if (nr_busy > 1) {
9328 kick = true;
9329 goto unlock;
9334 sd = rcu_dereference(rq->sd);
9335 if (sd) {
9336 if ((rq->cfs.h_nr_running >= 1) &&
9337 check_cpu_capacity(rq, sd)) {
9338 kick = true;
9339 goto unlock;
9343 sd = rcu_dereference(per_cpu(sd_asym, cpu));
9344 if (sd) {
9345 for_each_cpu(i, sched_domain_span(sd)) {
9346 if (i == cpu ||
9347 !cpumask_test_cpu(i, nohz.idle_cpus_mask))
9348 continue;
9350 if (sched_asym_prefer(i, cpu)) {
9351 kick = true;
9352 goto unlock;
9356 unlock:
9357 rcu_read_unlock();
9358 return kick;
9360 #else
9361 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
9362 #endif
9365 * run_rebalance_domains is triggered when needed from the scheduler tick.
9366 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9368 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
9370 struct rq *this_rq = this_rq();
9371 enum cpu_idle_type idle = this_rq->idle_balance ?
9372 CPU_IDLE : CPU_NOT_IDLE;
9375 * If this cpu has a pending nohz_balance_kick, then do the
9376 * balancing on behalf of the other idle cpus whose ticks are
9377 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9378 * give the idle cpus a chance to load balance. Else we may
9379 * load balance only within the local sched_domain hierarchy
9380 * and abort nohz_idle_balance altogether if we pull some load.
9382 nohz_idle_balance(this_rq, idle);
9383 rebalance_domains(this_rq, idle);
9387 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9389 void trigger_load_balance(struct rq *rq)
9391 /* Don't need to rebalance while attached to NULL domain */
9392 if (unlikely(on_null_domain(rq)))
9393 return;
9395 if (time_after_eq(jiffies, rq->next_balance))
9396 raise_softirq(SCHED_SOFTIRQ);
9397 #ifdef CONFIG_NO_HZ_COMMON
9398 if (nohz_kick_needed(rq))
9399 nohz_balancer_kick();
9400 #endif
9403 static void rq_online_fair(struct rq *rq)
9405 update_sysctl();
9407 update_runtime_enabled(rq);
9410 static void rq_offline_fair(struct rq *rq)
9412 update_sysctl();
9414 /* Ensure any throttled groups are reachable by pick_next_task */
9415 unthrottle_offline_cfs_rqs(rq);
9418 #endif /* CONFIG_SMP */
9421 * scheduler tick hitting a task of our scheduling class:
9423 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9425 struct cfs_rq *cfs_rq;
9426 struct sched_entity *se = &curr->se;
9428 for_each_sched_entity(se) {
9429 cfs_rq = cfs_rq_of(se);
9430 entity_tick(cfs_rq, se, queued);
9433 if (static_branch_unlikely(&sched_numa_balancing))
9434 task_tick_numa(rq, curr);
9438 * called on fork with the child task as argument from the parent's context
9439 * - child not yet on the tasklist
9440 * - preemption disabled
9442 static void task_fork_fair(struct task_struct *p)
9444 struct cfs_rq *cfs_rq;
9445 struct sched_entity *se = &p->se, *curr;
9446 struct rq *rq = this_rq();
9447 struct rq_flags rf;
9449 rq_lock(rq, &rf);
9450 update_rq_clock(rq);
9452 cfs_rq = task_cfs_rq(current);
9453 curr = cfs_rq->curr;
9454 if (curr) {
9455 update_curr(cfs_rq);
9456 se->vruntime = curr->vruntime;
9458 place_entity(cfs_rq, se, 1);
9460 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9462 * Upon rescheduling, sched_class::put_prev_task() will place
9463 * 'current' within the tree based on its new key value.
9465 swap(curr->vruntime, se->vruntime);
9466 resched_curr(rq);
9469 se->vruntime -= cfs_rq->min_vruntime;
9470 rq_unlock(rq, &rf);
9474 * Priority of the task has changed. Check to see if we preempt
9475 * the current task.
9477 static void
9478 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9480 if (!task_on_rq_queued(p))
9481 return;
9484 * Reschedule if we are currently running on this runqueue and
9485 * our priority decreased, or if we are not currently running on
9486 * this runqueue and our priority is higher than the current's
9488 if (rq->curr == p) {
9489 if (p->prio > oldprio)
9490 resched_curr(rq);
9491 } else
9492 check_preempt_curr(rq, p, 0);
9495 static inline bool vruntime_normalized(struct task_struct *p)
9497 struct sched_entity *se = &p->se;
9500 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9501 * the dequeue_entity(.flags=0) will already have normalized the
9502 * vruntime.
9504 if (p->on_rq)
9505 return true;
9508 * When !on_rq, vruntime of the task has usually NOT been normalized.
9509 * But there are some cases where it has already been normalized:
9511 * - A forked child which is waiting for being woken up by
9512 * wake_up_new_task().
9513 * - A task which has been woken up by try_to_wake_up() and
9514 * waiting for actually being woken up by sched_ttwu_pending().
9516 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
9517 return true;
9519 return false;
9522 #ifdef CONFIG_FAIR_GROUP_SCHED
9524 * Propagate the changes of the sched_entity across the tg tree to make it
9525 * visible to the root
9527 static void propagate_entity_cfs_rq(struct sched_entity *se)
9529 struct cfs_rq *cfs_rq;
9531 /* Start to propagate at parent */
9532 se = se->parent;
9534 for_each_sched_entity(se) {
9535 cfs_rq = cfs_rq_of(se);
9537 if (cfs_rq_throttled(cfs_rq))
9538 break;
9540 update_load_avg(cfs_rq, se, UPDATE_TG);
9543 #else
9544 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
9545 #endif
9547 static void detach_entity_cfs_rq(struct sched_entity *se)
9549 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9551 /* Catch up with the cfs_rq and remove our load when we leave */
9552 update_load_avg(cfs_rq, se, 0);
9553 detach_entity_load_avg(cfs_rq, se);
9554 update_tg_load_avg(cfs_rq, false);
9555 propagate_entity_cfs_rq(se);
9558 static void attach_entity_cfs_rq(struct sched_entity *se)
9560 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9562 #ifdef CONFIG_FAIR_GROUP_SCHED
9564 * Since the real-depth could have been changed (only FAIR
9565 * class maintain depth value), reset depth properly.
9567 se->depth = se->parent ? se->parent->depth + 1 : 0;
9568 #endif
9570 /* Synchronize entity with its cfs_rq */
9571 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
9572 attach_entity_load_avg(cfs_rq, se);
9573 update_tg_load_avg(cfs_rq, false);
9574 propagate_entity_cfs_rq(se);
9577 static void detach_task_cfs_rq(struct task_struct *p)
9579 struct sched_entity *se = &p->se;
9580 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9582 if (!vruntime_normalized(p)) {
9584 * Fix up our vruntime so that the current sleep doesn't
9585 * cause 'unlimited' sleep bonus.
9587 place_entity(cfs_rq, se, 0);
9588 se->vruntime -= cfs_rq->min_vruntime;
9591 detach_entity_cfs_rq(se);
9594 static void attach_task_cfs_rq(struct task_struct *p)
9596 struct sched_entity *se = &p->se;
9597 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9599 attach_entity_cfs_rq(se);
9601 if (!vruntime_normalized(p))
9602 se->vruntime += cfs_rq->min_vruntime;
9605 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9607 detach_task_cfs_rq(p);
9610 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9612 attach_task_cfs_rq(p);
9614 if (task_on_rq_queued(p)) {
9616 * We were most likely switched from sched_rt, so
9617 * kick off the schedule if running, otherwise just see
9618 * if we can still preempt the current task.
9620 if (rq->curr == p)
9621 resched_curr(rq);
9622 else
9623 check_preempt_curr(rq, p, 0);
9627 /* Account for a task changing its policy or group.
9629 * This routine is mostly called to set cfs_rq->curr field when a task
9630 * migrates between groups/classes.
9632 static void set_curr_task_fair(struct rq *rq)
9634 struct sched_entity *se = &rq->curr->se;
9636 for_each_sched_entity(se) {
9637 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9639 set_next_entity(cfs_rq, se);
9640 /* ensure bandwidth has been allocated on our new cfs_rq */
9641 account_cfs_rq_runtime(cfs_rq, 0);
9645 void init_cfs_rq(struct cfs_rq *cfs_rq)
9647 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
9648 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9649 #ifndef CONFIG_64BIT
9650 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9651 #endif
9652 #ifdef CONFIG_SMP
9653 raw_spin_lock_init(&cfs_rq->removed.lock);
9654 #endif
9657 #ifdef CONFIG_FAIR_GROUP_SCHED
9658 static void task_set_group_fair(struct task_struct *p)
9660 struct sched_entity *se = &p->se;
9662 set_task_rq(p, task_cpu(p));
9663 se->depth = se->parent ? se->parent->depth + 1 : 0;
9666 static void task_move_group_fair(struct task_struct *p)
9668 detach_task_cfs_rq(p);
9669 set_task_rq(p, task_cpu(p));
9671 #ifdef CONFIG_SMP
9672 /* Tell se's cfs_rq has been changed -- migrated */
9673 p->se.avg.last_update_time = 0;
9674 #endif
9675 attach_task_cfs_rq(p);
9678 static void task_change_group_fair(struct task_struct *p, int type)
9680 switch (type) {
9681 case TASK_SET_GROUP:
9682 task_set_group_fair(p);
9683 break;
9685 case TASK_MOVE_GROUP:
9686 task_move_group_fair(p);
9687 break;
9691 void free_fair_sched_group(struct task_group *tg)
9693 int i;
9695 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9697 for_each_possible_cpu(i) {
9698 if (tg->cfs_rq)
9699 kfree(tg->cfs_rq[i]);
9700 if (tg->se)
9701 kfree(tg->se[i]);
9704 kfree(tg->cfs_rq);
9705 kfree(tg->se);
9708 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9710 struct sched_entity *se;
9711 struct cfs_rq *cfs_rq;
9712 int i;
9714 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9715 if (!tg->cfs_rq)
9716 goto err;
9717 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9718 if (!tg->se)
9719 goto err;
9721 tg->shares = NICE_0_LOAD;
9723 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9725 for_each_possible_cpu(i) {
9726 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9727 GFP_KERNEL, cpu_to_node(i));
9728 if (!cfs_rq)
9729 goto err;
9731 se = kzalloc_node(sizeof(struct sched_entity),
9732 GFP_KERNEL, cpu_to_node(i));
9733 if (!se)
9734 goto err_free_rq;
9736 init_cfs_rq(cfs_rq);
9737 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9738 init_entity_runnable_average(se);
9741 return 1;
9743 err_free_rq:
9744 kfree(cfs_rq);
9745 err:
9746 return 0;
9749 void online_fair_sched_group(struct task_group *tg)
9751 struct sched_entity *se;
9752 struct rq *rq;
9753 int i;
9755 for_each_possible_cpu(i) {
9756 rq = cpu_rq(i);
9757 se = tg->se[i];
9759 raw_spin_lock_irq(&rq->lock);
9760 update_rq_clock(rq);
9761 attach_entity_cfs_rq(se);
9762 sync_throttle(tg, i);
9763 raw_spin_unlock_irq(&rq->lock);
9767 void unregister_fair_sched_group(struct task_group *tg)
9769 unsigned long flags;
9770 struct rq *rq;
9771 int cpu;
9773 for_each_possible_cpu(cpu) {
9774 if (tg->se[cpu])
9775 remove_entity_load_avg(tg->se[cpu]);
9778 * Only empty task groups can be destroyed; so we can speculatively
9779 * check on_list without danger of it being re-added.
9781 if (!tg->cfs_rq[cpu]->on_list)
9782 continue;
9784 rq = cpu_rq(cpu);
9786 raw_spin_lock_irqsave(&rq->lock, flags);
9787 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9788 raw_spin_unlock_irqrestore(&rq->lock, flags);
9792 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9793 struct sched_entity *se, int cpu,
9794 struct sched_entity *parent)
9796 struct rq *rq = cpu_rq(cpu);
9798 cfs_rq->tg = tg;
9799 cfs_rq->rq = rq;
9800 init_cfs_rq_runtime(cfs_rq);
9802 tg->cfs_rq[cpu] = cfs_rq;
9803 tg->se[cpu] = se;
9805 /* se could be NULL for root_task_group */
9806 if (!se)
9807 return;
9809 if (!parent) {
9810 se->cfs_rq = &rq->cfs;
9811 se->depth = 0;
9812 } else {
9813 se->cfs_rq = parent->my_q;
9814 se->depth = parent->depth + 1;
9817 se->my_q = cfs_rq;
9818 /* guarantee group entities always have weight */
9819 update_load_set(&se->load, NICE_0_LOAD);
9820 se->parent = parent;
9823 static DEFINE_MUTEX(shares_mutex);
9825 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9827 int i;
9830 * We can't change the weight of the root cgroup.
9832 if (!tg->se[0])
9833 return -EINVAL;
9835 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9837 mutex_lock(&shares_mutex);
9838 if (tg->shares == shares)
9839 goto done;
9841 tg->shares = shares;
9842 for_each_possible_cpu(i) {
9843 struct rq *rq = cpu_rq(i);
9844 struct sched_entity *se = tg->se[i];
9845 struct rq_flags rf;
9847 /* Propagate contribution to hierarchy */
9848 rq_lock_irqsave(rq, &rf);
9849 update_rq_clock(rq);
9850 for_each_sched_entity(se) {
9851 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
9852 update_cfs_group(se);
9854 rq_unlock_irqrestore(rq, &rf);
9857 done:
9858 mutex_unlock(&shares_mutex);
9859 return 0;
9861 #else /* CONFIG_FAIR_GROUP_SCHED */
9863 void free_fair_sched_group(struct task_group *tg) { }
9865 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9867 return 1;
9870 void online_fair_sched_group(struct task_group *tg) { }
9872 void unregister_fair_sched_group(struct task_group *tg) { }
9874 #endif /* CONFIG_FAIR_GROUP_SCHED */
9877 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9879 struct sched_entity *se = &task->se;
9880 unsigned int rr_interval = 0;
9883 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9884 * idle runqueue:
9886 if (rq->cfs.load.weight)
9887 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9889 return rr_interval;
9893 * All the scheduling class methods:
9895 const struct sched_class fair_sched_class = {
9896 .next = &idle_sched_class,
9897 .enqueue_task = enqueue_task_fair,
9898 .dequeue_task = dequeue_task_fair,
9899 .yield_task = yield_task_fair,
9900 .yield_to_task = yield_to_task_fair,
9902 .check_preempt_curr = check_preempt_wakeup,
9904 .pick_next_task = pick_next_task_fair,
9905 .put_prev_task = put_prev_task_fair,
9907 #ifdef CONFIG_SMP
9908 .select_task_rq = select_task_rq_fair,
9909 .migrate_task_rq = migrate_task_rq_fair,
9911 .rq_online = rq_online_fair,
9912 .rq_offline = rq_offline_fair,
9914 .task_dead = task_dead_fair,
9915 .set_cpus_allowed = set_cpus_allowed_common,
9916 #endif
9918 .set_curr_task = set_curr_task_fair,
9919 .task_tick = task_tick_fair,
9920 .task_fork = task_fork_fair,
9922 .prio_changed = prio_changed_fair,
9923 .switched_from = switched_from_fair,
9924 .switched_to = switched_to_fair,
9926 .get_rr_interval = get_rr_interval_fair,
9928 .update_curr = update_curr_fair,
9930 #ifdef CONFIG_FAIR_GROUP_SCHED
9931 .task_change_group = task_change_group_fair,
9932 #endif
9935 #ifdef CONFIG_SCHED_DEBUG
9936 void print_cfs_stats(struct seq_file *m, int cpu)
9938 struct cfs_rq *cfs_rq, *pos;
9940 rcu_read_lock();
9941 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
9942 print_cfs_rq(m, cpu, cfs_rq);
9943 rcu_read_unlock();
9946 #ifdef CONFIG_NUMA_BALANCING
9947 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9949 int node;
9950 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9952 for_each_online_node(node) {
9953 if (p->numa_faults) {
9954 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9955 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9957 if (p->numa_group) {
9958 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9959 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9961 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9964 #endif /* CONFIG_NUMA_BALANCING */
9965 #endif /* CONFIG_SCHED_DEBUG */
9967 __init void init_sched_fair_class(void)
9969 #ifdef CONFIG_SMP
9970 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9972 #ifdef CONFIG_NO_HZ_COMMON
9973 nohz.next_balance = jiffies;
9974 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
9975 #endif
9976 #endif /* SMP */