2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/cpuidle.h>
27 #include <linux/slab.h>
28 #include <linux/profile.h>
29 #include <linux/interrupt.h>
30 #include <linux/mempolicy.h>
31 #include <linux/migrate.h>
32 #include <linux/task_work.h>
33 #include <linux/module.h>
35 #include <trace/events/sched.h>
42 * Targeted preemption latency for CPU-bound tasks:
43 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
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 unsigned int sysctl_sched_latency = 6000000ULL;
54 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
56 unsigned int sysctl_sched_sync_hint_enable = 1;
57 unsigned int sysctl_sched_initial_task_util = 0;
58 unsigned int sysctl_sched_cstate_aware = 1;
60 #ifdef CONFIG_SCHED_WALT
61 unsigned int sysctl_sched_use_walt_cpu_util = 1;
62 unsigned int sysctl_sched_use_walt_task_util = 1;
63 __read_mostly unsigned int sysctl_sched_walt_cpu_high_irqload =
67 * The initial- and re-scaling of tunables is configurable
68 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
71 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
72 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
73 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
75 enum sched_tunable_scaling sysctl_sched_tunable_scaling
76 = SCHED_TUNABLESCALING_LOG;
79 * Minimal preemption granularity for CPU-bound tasks:
80 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
82 unsigned int sysctl_sched_min_granularity = 750000ULL;
83 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
86 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
88 static unsigned int sched_nr_latency = 8;
91 * After fork, child runs first. If set to 0 (default) then
92 * parent will (try to) run first.
94 unsigned int sysctl_sched_child_runs_first __read_mostly;
97 * SCHED_OTHER wake-up granularity.
98 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
100 * This option delays the preemption effects of decoupled workloads
101 * and reduces their over-scheduling. Synchronous workloads will still
102 * have immediate wakeup/sleep latencies.
104 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
105 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
107 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
110 * The exponential sliding window over which load is averaged for shares
114 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
116 #ifdef CONFIG_CFS_BANDWIDTH
118 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
119 * each time a cfs_rq requests quota.
121 * Note: in the case that the slice exceeds the runtime remaining (either due
122 * to consumption or the quota being specified to be smaller than the slice)
123 * we will always only issue the remaining available time.
125 * default: 5 msec, units: microseconds
127 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
131 * The margin used when comparing utilization with CPU capacity:
132 * util * margin < capacity * 1024
134 unsigned int capacity_margin = 1280; /* ~20% */
136 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
142 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
148 static inline void update_load_set(struct load_weight *lw, unsigned long w)
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
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);
168 switch (sysctl_sched_tunable_scaling) {
169 case SCHED_TUNABLESCALING_NONE:
172 case SCHED_TUNABLESCALING_LINEAR:
175 case SCHED_TUNABLESCALING_LOG:
177 factor = 1 + ilog2(cpus);
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);
196 void sched_init_granularity(void)
201 #define WMULT_CONST (~0U)
202 #define WMULT_SHIFT 32
204 static void __update_inv_weight(struct load_weight *lw)
208 if (likely(lw->inv_weight))
211 w = scale_load_down(lw->weight);
213 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
215 else if (unlikely(!w))
216 lw->inv_weight = WMULT_CONST;
218 lw->inv_weight = WMULT_CONST / w;
222 * delta_exec * weight / lw.weight
224 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
226 * Either weight := NICE_0_LOAD and lw \e 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)) {
247 /* hint to use a 32x32->64 mul */
248 fact = (u64)(u32)fact * lw->inv_weight;
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)
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 #ifdef CONFIG_SCHED_DEBUG
279 WARN_ON_ONCE(!entity_is_task(se));
281 return container_of(se, struct task_struct, se);
284 /* Walk up scheduling entities hierarchy */
285 #define for_each_sched_entity(se) \
286 for (; se; se = se->parent)
288 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
293 /* runqueue on which this entity is (to be) queued */
294 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
299 /* runqueue "owned" by this group */
300 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
305 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
307 if (!cfs_rq->on_list) {
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.
314 if (cfs_rq->tg->parent &&
315 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
316 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
317 &rq_of(cfs_rq)->leaf_cfs_rq_list);
319 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
320 &rq_of(cfs_rq)->leaf_cfs_rq_list);
327 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
329 if (cfs_rq->on_list) {
330 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
335 /* Iterate thr' all leaf cfs_rq's on a runqueue */
336 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
337 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
339 /* Do the two (enqueued) entities belong to the same group ? */
340 static inline struct cfs_rq *
341 is_same_group(struct sched_entity *se, struct sched_entity *pse)
343 if (se->cfs_rq == pse->cfs_rq)
349 static inline struct sched_entity *parent_entity(struct sched_entity *se)
355 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
357 int se_depth, pse_depth;
360 * preemption test can be made between sibling entities who are in the
361 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
362 * both tasks until we find their ancestors who are siblings of common
366 /* First walk up until both entities are at same depth */
367 se_depth = (*se)->depth;
368 pse_depth = (*pse)->depth;
370 while (se_depth > pse_depth) {
372 *se = parent_entity(*se);
375 while (pse_depth > se_depth) {
377 *pse = parent_entity(*pse);
380 while (!is_same_group(*se, *pse)) {
381 *se = parent_entity(*se);
382 *pse = parent_entity(*pse);
386 #else /* !CONFIG_FAIR_GROUP_SCHED */
388 static inline struct task_struct *task_of(struct sched_entity *se)
390 return container_of(se, struct task_struct, se);
393 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
395 return container_of(cfs_rq, struct rq, cfs);
398 #define entity_is_task(se) 1
400 #define for_each_sched_entity(se) \
401 for (; se; se = NULL)
403 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
405 return &task_rq(p)->cfs;
408 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
410 struct task_struct *p = task_of(se);
411 struct rq *rq = task_rq(p);
416 /* runqueue "owned" by this group */
417 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
422 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
426 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
430 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
431 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
433 static inline struct sched_entity *parent_entity(struct sched_entity *se)
439 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
443 #endif /* CONFIG_FAIR_GROUP_SCHED */
445 static __always_inline
446 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
448 /**************************************************************
449 * Scheduling class tree data structure manipulation methods:
452 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
454 s64 delta = (s64)(vruntime - max_vruntime);
456 max_vruntime = vruntime;
461 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
463 s64 delta = (s64)(vruntime - min_vruntime);
465 min_vruntime = vruntime;
470 static inline int entity_before(struct sched_entity *a,
471 struct sched_entity *b)
473 return (s64)(a->vruntime - b->vruntime) < 0;
476 static void update_min_vruntime(struct cfs_rq *cfs_rq)
478 u64 vruntime = cfs_rq->min_vruntime;
481 vruntime = cfs_rq->curr->vruntime;
483 if (cfs_rq->rb_leftmost) {
484 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
489 vruntime = se->vruntime;
491 vruntime = min_vruntime(vruntime, se->vruntime);
494 /* ensure we never gain time by being placed backwards. */
495 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
498 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
503 * Enqueue an entity into the rb-tree:
505 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
507 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
508 struct rb_node *parent = NULL;
509 struct sched_entity *entry;
513 * Find the right place in the rbtree:
517 entry = rb_entry(parent, struct sched_entity, run_node);
519 * We dont care about collisions. Nodes with
520 * the same key stay together.
522 if (entity_before(se, entry)) {
523 link = &parent->rb_left;
525 link = &parent->rb_right;
531 * Maintain a cache of leftmost tree entries (it is frequently
535 cfs_rq->rb_leftmost = &se->run_node;
537 rb_link_node(&se->run_node, parent, link);
538 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
541 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
543 if (cfs_rq->rb_leftmost == &se->run_node) {
544 struct rb_node *next_node;
546 next_node = rb_next(&se->run_node);
547 cfs_rq->rb_leftmost = next_node;
550 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
553 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
555 struct rb_node *left = cfs_rq->rb_leftmost;
560 return rb_entry(left, struct sched_entity, run_node);
563 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
565 struct rb_node *next = rb_next(&se->run_node);
570 return rb_entry(next, struct sched_entity, run_node);
573 #ifdef CONFIG_SCHED_DEBUG
574 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
576 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
581 return rb_entry(last, struct sched_entity, run_node);
584 /**************************************************************
585 * Scheduling class statistics methods:
588 int sched_proc_update_handler(struct ctl_table *table, int write,
589 void __user *buffer, size_t *lenp,
592 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
593 unsigned int factor = get_update_sysctl_factor();
598 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
599 sysctl_sched_min_granularity);
601 #define WRT_SYSCTL(name) \
602 (normalized_sysctl_##name = sysctl_##name / (factor))
603 WRT_SYSCTL(sched_min_granularity);
604 WRT_SYSCTL(sched_latency);
605 WRT_SYSCTL(sched_wakeup_granularity);
615 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
617 if (unlikely(se->load.weight != NICE_0_LOAD))
618 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
624 * The idea is to set a period in which each task runs once.
626 * When there are too many tasks (sched_nr_latency) we have to stretch
627 * this period because otherwise the slices get too small.
629 * p = (nr <= nl) ? l : l*nr/nl
631 static u64 __sched_period(unsigned long nr_running)
633 if (unlikely(nr_running > sched_nr_latency))
634 return nr_running * sysctl_sched_min_granularity;
636 return sysctl_sched_latency;
640 * We calculate the wall-time slice from the period by taking a part
641 * proportional to the weight.
645 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
647 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
649 for_each_sched_entity(se) {
650 struct load_weight *load;
651 struct load_weight lw;
653 cfs_rq = cfs_rq_of(se);
654 load = &cfs_rq->load;
656 if (unlikely(!se->on_rq)) {
659 update_load_add(&lw, se->load.weight);
662 slice = __calc_delta(slice, se->load.weight, load);
668 * We calculate the vruntime slice of a to-be-inserted task.
672 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
674 return calc_delta_fair(sched_slice(cfs_rq, se), se);
678 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
679 static unsigned long task_h_load(struct task_struct *p);
682 * We choose a half-life close to 1 scheduling period.
683 * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are
684 * dependent on this value.
686 #define LOAD_AVG_PERIOD 32
687 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
688 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */
690 /* Give new sched_entity start runnable values to heavy its load in infant time */
691 void init_entity_runnable_average(struct sched_entity *se)
693 struct sched_avg *sa = &se->avg;
695 sa->last_update_time = 0;
697 * sched_avg's period_contrib should be strictly less then 1024, so
698 * we give it 1023 to make sure it is almost a period (1024us), and
699 * will definitely be update (after enqueue).
701 sa->period_contrib = 1023;
702 sa->load_avg = scale_load_down(se->load.weight);
703 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
705 * In previous Android versions, we used to have:
706 * sa->util_avg = sched_freq() ?
707 * sysctl_sched_initial_task_util :
708 * scale_load_down(SCHED_LOAD_SCALE);
709 * sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
710 * However, that functionality has been moved to enqueue.
711 * It is unclear if we should restore this in enqueue.
714 * At this point, util_avg won't be used in select_task_rq_fair anyway
718 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
721 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
722 static int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq);
723 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force);
724 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se);
727 * With new tasks being created, their initial util_avgs are extrapolated
728 * based on the cfs_rq's current util_avg:
730 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
732 * However, in many cases, the above util_avg does not give a desired
733 * value. Moreover, the sum of the util_avgs may be divergent, such
734 * as when the series is a harmonic series.
736 * To solve this problem, we also cap the util_avg of successive tasks to
737 * only 1/2 of the left utilization budget:
739 * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
741 * where n denotes the nth task.
743 * For example, a simplest series from the beginning would be like:
745 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
746 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
748 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
749 * if util_avg > util_avg_cap.
751 void post_init_entity_util_avg(struct sched_entity *se)
753 struct cfs_rq *cfs_rq = cfs_rq_of(se);
754 struct sched_avg *sa = &se->avg;
755 long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
756 u64 now = cfs_rq_clock_task(cfs_rq);
760 if (cfs_rq->avg.util_avg != 0) {
761 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
762 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
764 if (sa->util_avg > cap)
770 * If we wish to restore tuning via setting initial util,
771 * this is where we should do it.
773 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
776 if (entity_is_task(se)) {
777 struct task_struct *p = task_of(se);
778 if (p->sched_class != &fair_sched_class) {
780 * For !fair tasks do:
782 update_cfs_rq_load_avg(now, cfs_rq, false);
783 attach_entity_load_avg(cfs_rq, se);
784 switched_from_fair(rq, p);
786 * such that the next switched_to_fair() has the
789 se->avg.last_update_time = now;
794 tg_update = update_cfs_rq_load_avg(now, cfs_rq, false);
795 attach_entity_load_avg(cfs_rq, se);
797 update_tg_load_avg(cfs_rq, false);
800 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq);
801 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq);
803 void init_entity_runnable_average(struct sched_entity *se)
806 void post_init_entity_util_avg(struct sched_entity *se)
809 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
812 #endif /* CONFIG_SMP */
815 * Update the current task's runtime statistics.
817 static void update_curr(struct cfs_rq *cfs_rq)
819 struct sched_entity *curr = cfs_rq->curr;
820 u64 now = rq_clock_task(rq_of(cfs_rq));
826 delta_exec = now - curr->exec_start;
827 if (unlikely((s64)delta_exec <= 0))
830 curr->exec_start = now;
832 schedstat_set(curr->statistics.exec_max,
833 max(delta_exec, curr->statistics.exec_max));
835 curr->sum_exec_runtime += delta_exec;
836 schedstat_add(cfs_rq, exec_clock, delta_exec);
838 curr->vruntime += calc_delta_fair(delta_exec, curr);
839 update_min_vruntime(cfs_rq);
841 if (entity_is_task(curr)) {
842 struct task_struct *curtask = task_of(curr);
844 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
845 cpuacct_charge(curtask, delta_exec);
846 account_group_exec_runtime(curtask, delta_exec);
849 account_cfs_rq_runtime(cfs_rq, delta_exec);
852 static void update_curr_fair(struct rq *rq)
854 update_curr(cfs_rq_of(&rq->curr->se));
858 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
860 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
864 * Task is being enqueued - update stats:
866 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
869 * Are we enqueueing a waiting task? (for current tasks
870 * a dequeue/enqueue event is a NOP)
872 if (se != cfs_rq->curr)
873 update_stats_wait_start(cfs_rq, se);
877 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
879 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
880 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
881 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
882 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
883 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
884 #ifdef CONFIG_SCHEDSTATS
885 if (entity_is_task(se)) {
886 trace_sched_stat_wait(task_of(se),
887 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
890 schedstat_set(se->statistics.wait_start, 0);
894 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
897 * Mark the end of the wait period if dequeueing a
900 if (se != cfs_rq->curr)
901 update_stats_wait_end(cfs_rq, se);
905 * We are picking a new current task - update its stats:
908 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
911 * We are starting a new run period:
913 se->exec_start = rq_clock_task(rq_of(cfs_rq));
916 /**************************************************
917 * Scheduling class queueing methods:
920 #ifdef CONFIG_NUMA_BALANCING
922 * Approximate time to scan a full NUMA task in ms. The task scan period is
923 * calculated based on the tasks virtual memory size and
924 * numa_balancing_scan_size.
926 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
927 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
929 /* Portion of address space to scan in MB */
930 unsigned int sysctl_numa_balancing_scan_size = 256;
932 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
933 unsigned int sysctl_numa_balancing_scan_delay = 1000;
935 static unsigned int task_nr_scan_windows(struct task_struct *p)
937 unsigned long rss = 0;
938 unsigned long nr_scan_pages;
941 * Calculations based on RSS as non-present and empty pages are skipped
942 * by the PTE scanner and NUMA hinting faults should be trapped based
945 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
946 rss = get_mm_rss(p->mm);
950 rss = round_up(rss, nr_scan_pages);
951 return rss / nr_scan_pages;
954 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
955 #define MAX_SCAN_WINDOW 2560
957 static unsigned int task_scan_min(struct task_struct *p)
959 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
960 unsigned int scan, floor;
961 unsigned int windows = 1;
963 if (scan_size < MAX_SCAN_WINDOW)
964 windows = MAX_SCAN_WINDOW / scan_size;
965 floor = 1000 / windows;
967 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
968 return max_t(unsigned int, floor, scan);
971 static unsigned int task_scan_max(struct task_struct *p)
973 unsigned int smin = task_scan_min(p);
976 /* Watch for min being lower than max due to floor calculations */
977 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
978 return max(smin, smax);
981 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
983 rq->nr_numa_running += (p->numa_preferred_nid != -1);
984 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
987 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
989 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
990 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
996 spinlock_t lock; /* nr_tasks, tasks */
1000 struct rcu_head rcu;
1001 nodemask_t active_nodes;
1002 unsigned long total_faults;
1004 * Faults_cpu is used to decide whether memory should move
1005 * towards the CPU. As a consequence, these stats are weighted
1006 * more by CPU use than by memory faults.
1008 unsigned long *faults_cpu;
1009 unsigned long faults[0];
1012 /* Shared or private faults. */
1013 #define NR_NUMA_HINT_FAULT_TYPES 2
1015 /* Memory and CPU locality */
1016 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1018 /* Averaged statistics, and temporary buffers. */
1019 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1021 pid_t task_numa_group_id(struct task_struct *p)
1023 return p->numa_group ? p->numa_group->gid : 0;
1027 * The averaged statistics, shared & private, memory & cpu,
1028 * occupy the first half of the array. The second half of the
1029 * array is for current counters, which are averaged into the
1030 * first set by task_numa_placement.
1032 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1034 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1037 static inline unsigned long task_faults(struct task_struct *p, int nid)
1039 if (!p->numa_faults)
1042 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1043 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1046 static inline unsigned long group_faults(struct task_struct *p, int nid)
1051 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1052 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1055 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1057 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1058 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1061 /* Handle placement on systems where not all nodes are directly connected. */
1062 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1063 int maxdist, bool task)
1065 unsigned long score = 0;
1069 * All nodes are directly connected, and the same distance
1070 * from each other. No need for fancy placement algorithms.
1072 if (sched_numa_topology_type == NUMA_DIRECT)
1076 * This code is called for each node, introducing N^2 complexity,
1077 * which should be ok given the number of nodes rarely exceeds 8.
1079 for_each_online_node(node) {
1080 unsigned long faults;
1081 int dist = node_distance(nid, node);
1084 * The furthest away nodes in the system are not interesting
1085 * for placement; nid was already counted.
1087 if (dist == sched_max_numa_distance || node == nid)
1091 * On systems with a backplane NUMA topology, compare groups
1092 * of nodes, and move tasks towards the group with the most
1093 * memory accesses. When comparing two nodes at distance
1094 * "hoplimit", only nodes closer by than "hoplimit" are part
1095 * of each group. Skip other nodes.
1097 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1101 /* Add up the faults from nearby nodes. */
1103 faults = task_faults(p, node);
1105 faults = group_faults(p, node);
1108 * On systems with a glueless mesh NUMA topology, there are
1109 * no fixed "groups of nodes". Instead, nodes that are not
1110 * directly connected bounce traffic through intermediate
1111 * nodes; a numa_group can occupy any set of nodes.
1112 * The further away a node is, the less the faults count.
1113 * This seems to result in good task placement.
1115 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1116 faults *= (sched_max_numa_distance - dist);
1117 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1127 * These return the fraction of accesses done by a particular task, or
1128 * task group, on a particular numa node. The group weight is given a
1129 * larger multiplier, in order to group tasks together that are almost
1130 * evenly spread out between numa nodes.
1132 static inline unsigned long task_weight(struct task_struct *p, int nid,
1135 unsigned long faults, total_faults;
1137 if (!p->numa_faults)
1140 total_faults = p->total_numa_faults;
1145 faults = task_faults(p, nid);
1146 faults += score_nearby_nodes(p, nid, dist, true);
1148 return 1000 * faults / total_faults;
1151 static inline unsigned long group_weight(struct task_struct *p, int nid,
1154 unsigned long faults, total_faults;
1159 total_faults = p->numa_group->total_faults;
1164 faults = group_faults(p, nid);
1165 faults += score_nearby_nodes(p, nid, dist, false);
1167 return 1000 * faults / total_faults;
1170 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1171 int src_nid, int dst_cpu)
1173 struct numa_group *ng = p->numa_group;
1174 int dst_nid = cpu_to_node(dst_cpu);
1175 int last_cpupid, this_cpupid;
1177 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1180 * Multi-stage node selection is used in conjunction with a periodic
1181 * migration fault to build a temporal task<->page relation. By using
1182 * a two-stage filter we remove short/unlikely relations.
1184 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1185 * a task's usage of a particular page (n_p) per total usage of this
1186 * page (n_t) (in a given time-span) to a probability.
1188 * Our periodic faults will sample this probability and getting the
1189 * same result twice in a row, given these samples are fully
1190 * independent, is then given by P(n)^2, provided our sample period
1191 * is sufficiently short compared to the usage pattern.
1193 * This quadric squishes small probabilities, making it less likely we
1194 * act on an unlikely task<->page relation.
1196 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1197 if (!cpupid_pid_unset(last_cpupid) &&
1198 cpupid_to_nid(last_cpupid) != dst_nid)
1201 /* Always allow migrate on private faults */
1202 if (cpupid_match_pid(p, last_cpupid))
1205 /* A shared fault, but p->numa_group has not been set up yet. */
1210 * Do not migrate if the destination is not a node that
1211 * is actively used by this numa group.
1213 if (!node_isset(dst_nid, ng->active_nodes))
1217 * Source is a node that is not actively used by this
1218 * numa group, while the destination is. Migrate.
1220 if (!node_isset(src_nid, ng->active_nodes))
1224 * Both source and destination are nodes in active
1225 * use by this numa group. Maximize memory bandwidth
1226 * by migrating from more heavily used groups, to less
1227 * heavily used ones, spreading the load around.
1228 * Use a 1/4 hysteresis to avoid spurious page movement.
1230 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1233 static unsigned long weighted_cpuload(const int cpu);
1234 static unsigned long source_load(int cpu, int type);
1235 static unsigned long target_load(int cpu, int type);
1236 static unsigned long capacity_of(int cpu);
1237 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1239 /* Cached statistics for all CPUs within a node */
1241 unsigned long nr_running;
1244 /* Total compute capacity of CPUs on a node */
1245 unsigned long compute_capacity;
1247 /* Approximate capacity in terms of runnable tasks on a node */
1248 unsigned long task_capacity;
1249 int has_free_capacity;
1253 * XXX borrowed from update_sg_lb_stats
1255 static void update_numa_stats(struct numa_stats *ns, int nid)
1257 int smt, cpu, cpus = 0;
1258 unsigned long capacity;
1260 memset(ns, 0, sizeof(*ns));
1261 for_each_cpu(cpu, cpumask_of_node(nid)) {
1262 struct rq *rq = cpu_rq(cpu);
1264 ns->nr_running += rq->nr_running;
1265 ns->load += weighted_cpuload(cpu);
1266 ns->compute_capacity += capacity_of(cpu);
1272 * If we raced with hotplug and there are no CPUs left in our mask
1273 * the @ns structure is NULL'ed and task_numa_compare() will
1274 * not find this node attractive.
1276 * We'll either bail at !has_free_capacity, or we'll detect a huge
1277 * imbalance and bail there.
1282 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1283 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1284 capacity = cpus / smt; /* cores */
1286 ns->task_capacity = min_t(unsigned, capacity,
1287 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1288 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1291 struct task_numa_env {
1292 struct task_struct *p;
1294 int src_cpu, src_nid;
1295 int dst_cpu, dst_nid;
1297 struct numa_stats src_stats, dst_stats;
1302 struct task_struct *best_task;
1307 static void task_numa_assign(struct task_numa_env *env,
1308 struct task_struct *p, long imp)
1311 put_task_struct(env->best_task);
1314 env->best_imp = imp;
1315 env->best_cpu = env->dst_cpu;
1318 static bool load_too_imbalanced(long src_load, long dst_load,
1319 struct task_numa_env *env)
1322 long orig_src_load, orig_dst_load;
1323 long src_capacity, dst_capacity;
1326 * The load is corrected for the CPU capacity available on each node.
1329 * ------------ vs ---------
1330 * src_capacity dst_capacity
1332 src_capacity = env->src_stats.compute_capacity;
1333 dst_capacity = env->dst_stats.compute_capacity;
1335 /* We care about the slope of the imbalance, not the direction. */
1336 if (dst_load < src_load)
1337 swap(dst_load, src_load);
1339 /* Is the difference below the threshold? */
1340 imb = dst_load * src_capacity * 100 -
1341 src_load * dst_capacity * env->imbalance_pct;
1346 * The imbalance is above the allowed threshold.
1347 * Compare it with the old imbalance.
1349 orig_src_load = env->src_stats.load;
1350 orig_dst_load = env->dst_stats.load;
1352 if (orig_dst_load < orig_src_load)
1353 swap(orig_dst_load, orig_src_load);
1355 old_imb = orig_dst_load * src_capacity * 100 -
1356 orig_src_load * dst_capacity * env->imbalance_pct;
1358 /* Would this change make things worse? */
1359 return (imb > old_imb);
1363 * This checks if the overall compute and NUMA accesses of the system would
1364 * be improved if the source tasks was migrated to the target dst_cpu taking
1365 * into account that it might be best if task running on the dst_cpu should
1366 * be exchanged with the source task
1368 static void task_numa_compare(struct task_numa_env *env,
1369 long taskimp, long groupimp)
1371 struct rq *src_rq = cpu_rq(env->src_cpu);
1372 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1373 struct task_struct *cur;
1374 long src_load, dst_load;
1376 long imp = env->p->numa_group ? groupimp : taskimp;
1378 int dist = env->dist;
1379 bool assigned = false;
1383 raw_spin_lock_irq(&dst_rq->lock);
1386 * No need to move the exiting task or idle task.
1388 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1392 * The task_struct must be protected here to protect the
1393 * p->numa_faults access in the task_weight since the
1394 * numa_faults could already be freed in the following path:
1395 * finish_task_switch()
1396 * --> put_task_struct()
1397 * --> __put_task_struct()
1398 * --> task_numa_free()
1400 get_task_struct(cur);
1403 raw_spin_unlock_irq(&dst_rq->lock);
1406 * Because we have preemption enabled we can get migrated around and
1407 * end try selecting ourselves (current == env->p) as a swap candidate.
1413 * "imp" is the fault differential for the source task between the
1414 * source and destination node. Calculate the total differential for
1415 * the source task and potential destination task. The more negative
1416 * the value is, the more rmeote accesses that would be expected to
1417 * be incurred if the tasks were swapped.
1420 /* Skip this swap candidate if cannot move to the source cpu */
1421 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1425 * If dst and source tasks are in the same NUMA group, or not
1426 * in any group then look only at task weights.
1428 if (cur->numa_group == env->p->numa_group) {
1429 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1430 task_weight(cur, env->dst_nid, dist);
1432 * Add some hysteresis to prevent swapping the
1433 * tasks within a group over tiny differences.
1435 if (cur->numa_group)
1439 * Compare the group weights. If a task is all by
1440 * itself (not part of a group), use the task weight
1443 if (cur->numa_group)
1444 imp += group_weight(cur, env->src_nid, dist) -
1445 group_weight(cur, env->dst_nid, dist);
1447 imp += task_weight(cur, env->src_nid, dist) -
1448 task_weight(cur, env->dst_nid, dist);
1452 if (imp <= env->best_imp && moveimp <= env->best_imp)
1456 /* Is there capacity at our destination? */
1457 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1458 !env->dst_stats.has_free_capacity)
1464 /* Balance doesn't matter much if we're running a task per cpu */
1465 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1466 dst_rq->nr_running == 1)
1470 * In the overloaded case, try and keep the load balanced.
1473 load = task_h_load(env->p);
1474 dst_load = env->dst_stats.load + load;
1475 src_load = env->src_stats.load - load;
1477 if (moveimp > imp && moveimp > env->best_imp) {
1479 * If the improvement from just moving env->p direction is
1480 * better than swapping tasks around, check if a move is
1481 * possible. Store a slightly smaller score than moveimp,
1482 * so an actually idle CPU will win.
1484 if (!load_too_imbalanced(src_load, dst_load, env)) {
1486 put_task_struct(cur);
1492 if (imp <= env->best_imp)
1496 load = task_h_load(cur);
1501 if (load_too_imbalanced(src_load, dst_load, env))
1505 * One idle CPU per node is evaluated for a task numa move.
1506 * Call select_idle_sibling to maybe find a better one.
1509 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1514 task_numa_assign(env, cur, imp);
1518 * The dst_rq->curr isn't assigned. The protection for task_struct is
1521 if (cur && !assigned)
1522 put_task_struct(cur);
1525 static void task_numa_find_cpu(struct task_numa_env *env,
1526 long taskimp, long groupimp)
1530 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1531 /* Skip this CPU if the source task cannot migrate */
1532 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1536 task_numa_compare(env, taskimp, groupimp);
1540 /* Only move tasks to a NUMA node less busy than the current node. */
1541 static bool numa_has_capacity(struct task_numa_env *env)
1543 struct numa_stats *src = &env->src_stats;
1544 struct numa_stats *dst = &env->dst_stats;
1546 if (src->has_free_capacity && !dst->has_free_capacity)
1550 * Only consider a task move if the source has a higher load
1551 * than the destination, corrected for CPU capacity on each node.
1553 * src->load dst->load
1554 * --------------------- vs ---------------------
1555 * src->compute_capacity dst->compute_capacity
1557 if (src->load * dst->compute_capacity * env->imbalance_pct >
1559 dst->load * src->compute_capacity * 100)
1565 static int task_numa_migrate(struct task_struct *p)
1567 struct task_numa_env env = {
1570 .src_cpu = task_cpu(p),
1571 .src_nid = task_node(p),
1573 .imbalance_pct = 112,
1579 struct sched_domain *sd;
1580 unsigned long taskweight, groupweight;
1582 long taskimp, groupimp;
1585 * Pick the lowest SD_NUMA domain, as that would have the smallest
1586 * imbalance and would be the first to start moving tasks about.
1588 * And we want to avoid any moving of tasks about, as that would create
1589 * random movement of tasks -- counter the numa conditions we're trying
1593 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1595 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1599 * Cpusets can break the scheduler domain tree into smaller
1600 * balance domains, some of which do not cross NUMA boundaries.
1601 * Tasks that are "trapped" in such domains cannot be migrated
1602 * elsewhere, so there is no point in (re)trying.
1604 if (unlikely(!sd)) {
1605 p->numa_preferred_nid = task_node(p);
1609 env.dst_nid = p->numa_preferred_nid;
1610 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1611 taskweight = task_weight(p, env.src_nid, dist);
1612 groupweight = group_weight(p, env.src_nid, dist);
1613 update_numa_stats(&env.src_stats, env.src_nid);
1614 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1615 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1616 update_numa_stats(&env.dst_stats, env.dst_nid);
1618 /* Try to find a spot on the preferred nid. */
1619 if (numa_has_capacity(&env))
1620 task_numa_find_cpu(&env, taskimp, groupimp);
1623 * Look at other nodes in these cases:
1624 * - there is no space available on the preferred_nid
1625 * - the task is part of a numa_group that is interleaved across
1626 * multiple NUMA nodes; in order to better consolidate the group,
1627 * we need to check other locations.
1629 if (env.best_cpu == -1 || (p->numa_group &&
1630 nodes_weight(p->numa_group->active_nodes) > 1)) {
1631 for_each_online_node(nid) {
1632 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1635 dist = node_distance(env.src_nid, env.dst_nid);
1636 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1638 taskweight = task_weight(p, env.src_nid, dist);
1639 groupweight = group_weight(p, env.src_nid, dist);
1642 /* Only consider nodes where both task and groups benefit */
1643 taskimp = task_weight(p, nid, dist) - taskweight;
1644 groupimp = group_weight(p, nid, dist) - groupweight;
1645 if (taskimp < 0 && groupimp < 0)
1650 update_numa_stats(&env.dst_stats, env.dst_nid);
1651 if (numa_has_capacity(&env))
1652 task_numa_find_cpu(&env, taskimp, groupimp);
1657 * If the task is part of a workload that spans multiple NUMA nodes,
1658 * and is migrating into one of the workload's active nodes, remember
1659 * this node as the task's preferred numa node, so the workload can
1661 * A task that migrated to a second choice node will be better off
1662 * trying for a better one later. Do not set the preferred node here.
1664 if (p->numa_group) {
1665 if (env.best_cpu == -1)
1670 if (node_isset(nid, p->numa_group->active_nodes))
1671 sched_setnuma(p, env.dst_nid);
1674 /* No better CPU than the current one was found. */
1675 if (env.best_cpu == -1)
1679 * Reset the scan period if the task is being rescheduled on an
1680 * alternative node to recheck if the tasks is now properly placed.
1682 p->numa_scan_period = task_scan_min(p);
1684 if (env.best_task == NULL) {
1685 ret = migrate_task_to(p, env.best_cpu);
1687 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1691 ret = migrate_swap(p, env.best_task);
1693 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1694 put_task_struct(env.best_task);
1698 /* Attempt to migrate a task to a CPU on the preferred node. */
1699 static void numa_migrate_preferred(struct task_struct *p)
1701 unsigned long interval = HZ;
1703 /* This task has no NUMA fault statistics yet */
1704 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1707 /* Periodically retry migrating the task to the preferred node */
1708 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1709 p->numa_migrate_retry = jiffies + interval;
1711 /* Success if task is already running on preferred CPU */
1712 if (task_node(p) == p->numa_preferred_nid)
1715 /* Otherwise, try migrate to a CPU on the preferred node */
1716 task_numa_migrate(p);
1720 * Find the nodes on which the workload is actively running. We do this by
1721 * tracking the nodes from which NUMA hinting faults are triggered. This can
1722 * be different from the set of nodes where the workload's memory is currently
1725 * The bitmask is used to make smarter decisions on when to do NUMA page
1726 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1727 * are added when they cause over 6/16 of the maximum number of faults, but
1728 * only removed when they drop below 3/16.
1730 static void update_numa_active_node_mask(struct numa_group *numa_group)
1732 unsigned long faults, max_faults = 0;
1735 for_each_online_node(nid) {
1736 faults = group_faults_cpu(numa_group, nid);
1737 if (faults > max_faults)
1738 max_faults = faults;
1741 for_each_online_node(nid) {
1742 faults = group_faults_cpu(numa_group, nid);
1743 if (!node_isset(nid, numa_group->active_nodes)) {
1744 if (faults > max_faults * 6 / 16)
1745 node_set(nid, numa_group->active_nodes);
1746 } else if (faults < max_faults * 3 / 16)
1747 node_clear(nid, numa_group->active_nodes);
1752 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1753 * increments. The more local the fault statistics are, the higher the scan
1754 * period will be for the next scan window. If local/(local+remote) ratio is
1755 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1756 * the scan period will decrease. Aim for 70% local accesses.
1758 #define NUMA_PERIOD_SLOTS 10
1759 #define NUMA_PERIOD_THRESHOLD 7
1762 * Increase the scan period (slow down scanning) if the majority of
1763 * our memory is already on our local node, or if the majority of
1764 * the page accesses are shared with other processes.
1765 * Otherwise, decrease the scan period.
1767 static void update_task_scan_period(struct task_struct *p,
1768 unsigned long shared, unsigned long private)
1770 unsigned int period_slot;
1774 unsigned long remote = p->numa_faults_locality[0];
1775 unsigned long local = p->numa_faults_locality[1];
1778 * If there were no record hinting faults then either the task is
1779 * completely idle or all activity is areas that are not of interest
1780 * to automatic numa balancing. Related to that, if there were failed
1781 * migration then it implies we are migrating too quickly or the local
1782 * node is overloaded. In either case, scan slower
1784 if (local + shared == 0 || p->numa_faults_locality[2]) {
1785 p->numa_scan_period = min(p->numa_scan_period_max,
1786 p->numa_scan_period << 1);
1788 p->mm->numa_next_scan = jiffies +
1789 msecs_to_jiffies(p->numa_scan_period);
1795 * Prepare to scale scan period relative to the current period.
1796 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1797 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1798 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1800 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1801 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1802 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1803 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1806 diff = slot * period_slot;
1808 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1811 * Scale scan rate increases based on sharing. There is an
1812 * inverse relationship between the degree of sharing and
1813 * the adjustment made to the scanning period. Broadly
1814 * speaking the intent is that there is little point
1815 * scanning faster if shared accesses dominate as it may
1816 * simply bounce migrations uselessly
1818 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1819 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1822 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1823 task_scan_min(p), task_scan_max(p));
1824 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1828 * Get the fraction of time the task has been running since the last
1829 * NUMA placement cycle. The scheduler keeps similar statistics, but
1830 * decays those on a 32ms period, which is orders of magnitude off
1831 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1832 * stats only if the task is so new there are no NUMA statistics yet.
1834 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1836 u64 runtime, delta, now;
1837 /* Use the start of this time slice to avoid calculations. */
1838 now = p->se.exec_start;
1839 runtime = p->se.sum_exec_runtime;
1841 if (p->last_task_numa_placement) {
1842 delta = runtime - p->last_sum_exec_runtime;
1843 *period = now - p->last_task_numa_placement;
1845 delta = p->se.avg.load_sum / p->se.load.weight;
1846 *period = LOAD_AVG_MAX;
1849 p->last_sum_exec_runtime = runtime;
1850 p->last_task_numa_placement = now;
1856 * Determine the preferred nid for a task in a numa_group. This needs to
1857 * be done in a way that produces consistent results with group_weight,
1858 * otherwise workloads might not converge.
1860 static int preferred_group_nid(struct task_struct *p, int nid)
1865 /* Direct connections between all NUMA nodes. */
1866 if (sched_numa_topology_type == NUMA_DIRECT)
1870 * On a system with glueless mesh NUMA topology, group_weight
1871 * scores nodes according to the number of NUMA hinting faults on
1872 * both the node itself, and on nearby nodes.
1874 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1875 unsigned long score, max_score = 0;
1876 int node, max_node = nid;
1878 dist = sched_max_numa_distance;
1880 for_each_online_node(node) {
1881 score = group_weight(p, node, dist);
1882 if (score > max_score) {
1891 * Finding the preferred nid in a system with NUMA backplane
1892 * interconnect topology is more involved. The goal is to locate
1893 * tasks from numa_groups near each other in the system, and
1894 * untangle workloads from different sides of the system. This requires
1895 * searching down the hierarchy of node groups, recursively searching
1896 * inside the highest scoring group of nodes. The nodemask tricks
1897 * keep the complexity of the search down.
1899 nodes = node_online_map;
1900 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1901 unsigned long max_faults = 0;
1902 nodemask_t max_group = NODE_MASK_NONE;
1905 /* Are there nodes at this distance from each other? */
1906 if (!find_numa_distance(dist))
1909 for_each_node_mask(a, nodes) {
1910 unsigned long faults = 0;
1911 nodemask_t this_group;
1912 nodes_clear(this_group);
1914 /* Sum group's NUMA faults; includes a==b case. */
1915 for_each_node_mask(b, nodes) {
1916 if (node_distance(a, b) < dist) {
1917 faults += group_faults(p, b);
1918 node_set(b, this_group);
1919 node_clear(b, nodes);
1923 /* Remember the top group. */
1924 if (faults > max_faults) {
1925 max_faults = faults;
1926 max_group = this_group;
1928 * subtle: at the smallest distance there is
1929 * just one node left in each "group", the
1930 * winner is the preferred nid.
1935 /* Next round, evaluate the nodes within max_group. */
1943 static void task_numa_placement(struct task_struct *p)
1945 int seq, nid, max_nid = -1, max_group_nid = -1;
1946 unsigned long max_faults = 0, max_group_faults = 0;
1947 unsigned long fault_types[2] = { 0, 0 };
1948 unsigned long total_faults;
1949 u64 runtime, period;
1950 spinlock_t *group_lock = NULL;
1953 * The p->mm->numa_scan_seq field gets updated without
1954 * exclusive access. Use READ_ONCE() here to ensure
1955 * that the field is read in a single access:
1957 seq = READ_ONCE(p->mm->numa_scan_seq);
1958 if (p->numa_scan_seq == seq)
1960 p->numa_scan_seq = seq;
1961 p->numa_scan_period_max = task_scan_max(p);
1963 total_faults = p->numa_faults_locality[0] +
1964 p->numa_faults_locality[1];
1965 runtime = numa_get_avg_runtime(p, &period);
1967 /* If the task is part of a group prevent parallel updates to group stats */
1968 if (p->numa_group) {
1969 group_lock = &p->numa_group->lock;
1970 spin_lock_irq(group_lock);
1973 /* Find the node with the highest number of faults */
1974 for_each_online_node(nid) {
1975 /* Keep track of the offsets in numa_faults array */
1976 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1977 unsigned long faults = 0, group_faults = 0;
1980 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1981 long diff, f_diff, f_weight;
1983 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1984 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1985 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1986 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1988 /* Decay existing window, copy faults since last scan */
1989 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1990 fault_types[priv] += p->numa_faults[membuf_idx];
1991 p->numa_faults[membuf_idx] = 0;
1994 * Normalize the faults_from, so all tasks in a group
1995 * count according to CPU use, instead of by the raw
1996 * number of faults. Tasks with little runtime have
1997 * little over-all impact on throughput, and thus their
1998 * faults are less important.
2000 f_weight = div64_u64(runtime << 16, period + 1);
2001 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2003 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2004 p->numa_faults[cpubuf_idx] = 0;
2006 p->numa_faults[mem_idx] += diff;
2007 p->numa_faults[cpu_idx] += f_diff;
2008 faults += p->numa_faults[mem_idx];
2009 p->total_numa_faults += diff;
2010 if (p->numa_group) {
2012 * safe because we can only change our own group
2014 * mem_idx represents the offset for a given
2015 * nid and priv in a specific region because it
2016 * is at the beginning of the numa_faults array.
2018 p->numa_group->faults[mem_idx] += diff;
2019 p->numa_group->faults_cpu[mem_idx] += f_diff;
2020 p->numa_group->total_faults += diff;
2021 group_faults += p->numa_group->faults[mem_idx];
2025 if (faults > max_faults) {
2026 max_faults = faults;
2030 if (group_faults > max_group_faults) {
2031 max_group_faults = group_faults;
2032 max_group_nid = nid;
2036 update_task_scan_period(p, fault_types[0], fault_types[1]);
2038 if (p->numa_group) {
2039 update_numa_active_node_mask(p->numa_group);
2040 spin_unlock_irq(group_lock);
2041 max_nid = preferred_group_nid(p, max_group_nid);
2045 /* Set the new preferred node */
2046 if (max_nid != p->numa_preferred_nid)
2047 sched_setnuma(p, max_nid);
2049 if (task_node(p) != p->numa_preferred_nid)
2050 numa_migrate_preferred(p);
2054 static inline int get_numa_group(struct numa_group *grp)
2056 return atomic_inc_not_zero(&grp->refcount);
2059 static inline void put_numa_group(struct numa_group *grp)
2061 if (atomic_dec_and_test(&grp->refcount))
2062 kfree_rcu(grp, rcu);
2065 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2068 struct numa_group *grp, *my_grp;
2069 struct task_struct *tsk;
2071 int cpu = cpupid_to_cpu(cpupid);
2074 if (unlikely(!p->numa_group)) {
2075 unsigned int size = sizeof(struct numa_group) +
2076 4*nr_node_ids*sizeof(unsigned long);
2078 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2082 atomic_set(&grp->refcount, 1);
2083 spin_lock_init(&grp->lock);
2085 /* Second half of the array tracks nids where faults happen */
2086 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2089 node_set(task_node(current), grp->active_nodes);
2091 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2092 grp->faults[i] = p->numa_faults[i];
2094 grp->total_faults = p->total_numa_faults;
2097 rcu_assign_pointer(p->numa_group, grp);
2101 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2103 if (!cpupid_match_pid(tsk, cpupid))
2106 grp = rcu_dereference(tsk->numa_group);
2110 my_grp = p->numa_group;
2115 * Only join the other group if its bigger; if we're the bigger group,
2116 * the other task will join us.
2118 if (my_grp->nr_tasks > grp->nr_tasks)
2122 * Tie-break on the grp address.
2124 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2127 /* Always join threads in the same process. */
2128 if (tsk->mm == current->mm)
2131 /* Simple filter to avoid false positives due to PID collisions */
2132 if (flags & TNF_SHARED)
2135 /* Update priv based on whether false sharing was detected */
2138 if (join && !get_numa_group(grp))
2146 BUG_ON(irqs_disabled());
2147 double_lock_irq(&my_grp->lock, &grp->lock);
2149 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2150 my_grp->faults[i] -= p->numa_faults[i];
2151 grp->faults[i] += p->numa_faults[i];
2153 my_grp->total_faults -= p->total_numa_faults;
2154 grp->total_faults += p->total_numa_faults;
2159 spin_unlock(&my_grp->lock);
2160 spin_unlock_irq(&grp->lock);
2162 rcu_assign_pointer(p->numa_group, grp);
2164 put_numa_group(my_grp);
2172 void task_numa_free(struct task_struct *p)
2174 struct numa_group *grp = p->numa_group;
2175 void *numa_faults = p->numa_faults;
2176 unsigned long flags;
2180 spin_lock_irqsave(&grp->lock, flags);
2181 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2182 grp->faults[i] -= p->numa_faults[i];
2183 grp->total_faults -= p->total_numa_faults;
2186 spin_unlock_irqrestore(&grp->lock, flags);
2187 RCU_INIT_POINTER(p->numa_group, NULL);
2188 put_numa_group(grp);
2191 p->numa_faults = NULL;
2196 * Got a PROT_NONE fault for a page on @node.
2198 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2200 struct task_struct *p = current;
2201 bool migrated = flags & TNF_MIGRATED;
2202 int cpu_node = task_node(current);
2203 int local = !!(flags & TNF_FAULT_LOCAL);
2206 if (!static_branch_likely(&sched_numa_balancing))
2209 /* for example, ksmd faulting in a user's mm */
2213 /* Allocate buffer to track faults on a per-node basis */
2214 if (unlikely(!p->numa_faults)) {
2215 int size = sizeof(*p->numa_faults) *
2216 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2218 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2219 if (!p->numa_faults)
2222 p->total_numa_faults = 0;
2223 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2227 * First accesses are treated as private, otherwise consider accesses
2228 * to be private if the accessing pid has not changed
2230 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2233 priv = cpupid_match_pid(p, last_cpupid);
2234 if (!priv && !(flags & TNF_NO_GROUP))
2235 task_numa_group(p, last_cpupid, flags, &priv);
2239 * If a workload spans multiple NUMA nodes, a shared fault that
2240 * occurs wholly within the set of nodes that the workload is
2241 * actively using should be counted as local. This allows the
2242 * scan rate to slow down when a workload has settled down.
2244 if (!priv && !local && p->numa_group &&
2245 node_isset(cpu_node, p->numa_group->active_nodes) &&
2246 node_isset(mem_node, p->numa_group->active_nodes))
2249 task_numa_placement(p);
2252 * Retry task to preferred node migration periodically, in case it
2253 * case it previously failed, or the scheduler moved us.
2255 if (time_after(jiffies, p->numa_migrate_retry))
2256 numa_migrate_preferred(p);
2259 p->numa_pages_migrated += pages;
2260 if (flags & TNF_MIGRATE_FAIL)
2261 p->numa_faults_locality[2] += pages;
2263 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2264 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2265 p->numa_faults_locality[local] += pages;
2268 static void reset_ptenuma_scan(struct task_struct *p)
2271 * We only did a read acquisition of the mmap sem, so
2272 * p->mm->numa_scan_seq is written to without exclusive access
2273 * and the update is not guaranteed to be atomic. That's not
2274 * much of an issue though, since this is just used for
2275 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2276 * expensive, to avoid any form of compiler optimizations:
2278 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2279 p->mm->numa_scan_offset = 0;
2283 * The expensive part of numa migration is done from task_work context.
2284 * Triggered from task_tick_numa().
2286 void task_numa_work(struct callback_head *work)
2288 unsigned long migrate, next_scan, now = jiffies;
2289 struct task_struct *p = current;
2290 struct mm_struct *mm = p->mm;
2291 struct vm_area_struct *vma;
2292 unsigned long start, end;
2293 unsigned long nr_pte_updates = 0;
2294 long pages, virtpages;
2296 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2298 work->next = work; /* protect against double add */
2300 * Who cares about NUMA placement when they're dying.
2302 * NOTE: make sure not to dereference p->mm before this check,
2303 * exit_task_work() happens _after_ exit_mm() so we could be called
2304 * without p->mm even though we still had it when we enqueued this
2307 if (p->flags & PF_EXITING)
2310 if (!mm->numa_next_scan) {
2311 mm->numa_next_scan = now +
2312 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2316 * Enforce maximal scan/migration frequency..
2318 migrate = mm->numa_next_scan;
2319 if (time_before(now, migrate))
2322 if (p->numa_scan_period == 0) {
2323 p->numa_scan_period_max = task_scan_max(p);
2324 p->numa_scan_period = task_scan_min(p);
2327 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2328 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2332 * Delay this task enough that another task of this mm will likely win
2333 * the next time around.
2335 p->node_stamp += 2 * TICK_NSEC;
2337 start = mm->numa_scan_offset;
2338 pages = sysctl_numa_balancing_scan_size;
2339 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2340 virtpages = pages * 8; /* Scan up to this much virtual space */
2345 down_read(&mm->mmap_sem);
2346 vma = find_vma(mm, start);
2348 reset_ptenuma_scan(p);
2352 for (; vma; vma = vma->vm_next) {
2353 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2354 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2359 * Shared library pages mapped by multiple processes are not
2360 * migrated as it is expected they are cache replicated. Avoid
2361 * hinting faults in read-only file-backed mappings or the vdso
2362 * as migrating the pages will be of marginal benefit.
2365 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2369 * Skip inaccessible VMAs to avoid any confusion between
2370 * PROT_NONE and NUMA hinting ptes
2372 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2376 start = max(start, vma->vm_start);
2377 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2378 end = min(end, vma->vm_end);
2379 nr_pte_updates = change_prot_numa(vma, start, end);
2382 * Try to scan sysctl_numa_balancing_size worth of
2383 * hpages that have at least one present PTE that
2384 * is not already pte-numa. If the VMA contains
2385 * areas that are unused or already full of prot_numa
2386 * PTEs, scan up to virtpages, to skip through those
2390 pages -= (end - start) >> PAGE_SHIFT;
2391 virtpages -= (end - start) >> PAGE_SHIFT;
2394 if (pages <= 0 || virtpages <= 0)
2398 } while (end != vma->vm_end);
2403 * It is possible to reach the end of the VMA list but the last few
2404 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2405 * would find the !migratable VMA on the next scan but not reset the
2406 * scanner to the start so check it now.
2409 mm->numa_scan_offset = start;
2411 reset_ptenuma_scan(p);
2412 up_read(&mm->mmap_sem);
2416 * Drive the periodic memory faults..
2418 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2420 struct callback_head *work = &curr->numa_work;
2424 * We don't care about NUMA placement if we don't have memory.
2426 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2430 * Using runtime rather than walltime has the dual advantage that
2431 * we (mostly) drive the selection from busy threads and that the
2432 * task needs to have done some actual work before we bother with
2435 now = curr->se.sum_exec_runtime;
2436 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2438 if (now > curr->node_stamp + period) {
2439 if (!curr->node_stamp)
2440 curr->numa_scan_period = task_scan_min(curr);
2441 curr->node_stamp += period;
2443 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2444 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2445 task_work_add(curr, work, true);
2450 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2454 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2458 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2461 #endif /* CONFIG_NUMA_BALANCING */
2464 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2466 update_load_add(&cfs_rq->load, se->load.weight);
2467 if (!parent_entity(se))
2468 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2470 if (entity_is_task(se)) {
2471 struct rq *rq = rq_of(cfs_rq);
2473 account_numa_enqueue(rq, task_of(se));
2474 list_add(&se->group_node, &rq->cfs_tasks);
2477 cfs_rq->nr_running++;
2481 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2483 update_load_sub(&cfs_rq->load, se->load.weight);
2484 if (!parent_entity(se))
2485 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2486 if (entity_is_task(se)) {
2487 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2488 list_del_init(&se->group_node);
2490 cfs_rq->nr_running--;
2493 #ifdef CONFIG_FAIR_GROUP_SCHED
2495 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2500 * Use this CPU's real-time load instead of the last load contribution
2501 * as the updating of the contribution is delayed, and we will use the
2502 * the real-time load to calc the share. See update_tg_load_avg().
2504 tg_weight = atomic_long_read(&tg->load_avg);
2505 tg_weight -= cfs_rq->tg_load_avg_contrib;
2506 tg_weight += cfs_rq->load.weight;
2511 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2513 long tg_weight, load, shares;
2515 tg_weight = calc_tg_weight(tg, cfs_rq);
2516 load = cfs_rq->load.weight;
2518 shares = (tg->shares * load);
2520 shares /= tg_weight;
2522 if (shares < MIN_SHARES)
2523 shares = MIN_SHARES;
2524 if (shares > tg->shares)
2525 shares = tg->shares;
2529 # else /* CONFIG_SMP */
2530 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2534 # endif /* CONFIG_SMP */
2535 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2536 unsigned long weight)
2539 /* commit outstanding execution time */
2540 if (cfs_rq->curr == se)
2541 update_curr(cfs_rq);
2542 account_entity_dequeue(cfs_rq, se);
2545 update_load_set(&se->load, weight);
2548 account_entity_enqueue(cfs_rq, se);
2551 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2553 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2555 struct task_group *tg;
2556 struct sched_entity *se;
2560 se = tg->se[cpu_of(rq_of(cfs_rq))];
2561 if (!se || throttled_hierarchy(cfs_rq))
2564 if (likely(se->load.weight == tg->shares))
2567 shares = calc_cfs_shares(cfs_rq, tg);
2569 reweight_entity(cfs_rq_of(se), se, shares);
2571 #else /* CONFIG_FAIR_GROUP_SCHED */
2572 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2575 #endif /* CONFIG_FAIR_GROUP_SCHED */
2578 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2579 static const u32 runnable_avg_yN_inv[] = {
2580 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2581 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2582 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2583 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2584 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2585 0x85aac367, 0x82cd8698,
2589 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2590 * over-estimates when re-combining.
2592 static const u32 runnable_avg_yN_sum[] = {
2593 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2594 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2595 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2600 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2602 static __always_inline u64 decay_load(u64 val, u64 n)
2604 unsigned int local_n;
2608 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2611 /* after bounds checking we can collapse to 32-bit */
2615 * As y^PERIOD = 1/2, we can combine
2616 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2617 * With a look-up table which covers y^n (n<PERIOD)
2619 * To achieve constant time decay_load.
2621 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2622 val >>= local_n / LOAD_AVG_PERIOD;
2623 local_n %= LOAD_AVG_PERIOD;
2626 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2631 * For updates fully spanning n periods, the contribution to runnable
2632 * average will be: \Sum 1024*y^n
2634 * We can compute this reasonably efficiently by combining:
2635 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2637 static u32 __compute_runnable_contrib(u64 n)
2641 if (likely(n <= LOAD_AVG_PERIOD))
2642 return runnable_avg_yN_sum[n];
2643 else if (unlikely(n >= LOAD_AVG_MAX_N))
2644 return LOAD_AVG_MAX;
2646 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2648 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2649 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2651 n -= LOAD_AVG_PERIOD;
2652 } while (n > LOAD_AVG_PERIOD);
2654 contrib = decay_load(contrib, n);
2655 return contrib + runnable_avg_yN_sum[n];
2658 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2659 #error "load tracking assumes 2^10 as unit"
2662 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2665 * We can represent the historical contribution to runnable average as the
2666 * coefficients of a geometric series. To do this we sub-divide our runnable
2667 * history into segments of approximately 1ms (1024us); label the segment that
2668 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2670 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2672 * (now) (~1ms ago) (~2ms ago)
2674 * Let u_i denote the fraction of p_i that the entity was runnable.
2676 * We then designate the fractions u_i as our co-efficients, yielding the
2677 * following representation of historical load:
2678 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2680 * We choose y based on the with of a reasonably scheduling period, fixing:
2683 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2684 * approximately half as much as the contribution to load within the last ms
2687 * When a period "rolls over" and we have new u_0`, multiplying the previous
2688 * sum again by y is sufficient to update:
2689 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2690 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2692 static __always_inline int
2693 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2694 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2696 u64 delta, scaled_delta, periods;
2698 unsigned int delta_w, scaled_delta_w, decayed = 0;
2699 unsigned long scale_freq, scale_cpu;
2701 delta = now - sa->last_update_time;
2703 * This should only happen when time goes backwards, which it
2704 * unfortunately does during sched clock init when we swap over to TSC.
2706 if ((s64)delta < 0) {
2707 sa->last_update_time = now;
2712 * Use 1024ns as the unit of measurement since it's a reasonable
2713 * approximation of 1us and fast to compute.
2718 sa->last_update_time = now;
2720 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2721 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2722 trace_sched_contrib_scale_f(cpu, scale_freq, scale_cpu);
2724 /* delta_w is the amount already accumulated against our next period */
2725 delta_w = sa->period_contrib;
2726 if (delta + delta_w >= 1024) {
2729 /* how much left for next period will start over, we don't know yet */
2730 sa->period_contrib = 0;
2733 * Now that we know we're crossing a period boundary, figure
2734 * out how much from delta we need to complete the current
2735 * period and accrue it.
2737 delta_w = 1024 - delta_w;
2738 scaled_delta_w = cap_scale(delta_w, scale_freq);
2740 sa->load_sum += weight * scaled_delta_w;
2742 cfs_rq->runnable_load_sum +=
2743 weight * scaled_delta_w;
2747 sa->util_sum += scaled_delta_w * scale_cpu;
2751 /* Figure out how many additional periods this update spans */
2752 periods = delta / 1024;
2755 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2757 cfs_rq->runnable_load_sum =
2758 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2760 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2762 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2763 contrib = __compute_runnable_contrib(periods);
2764 contrib = cap_scale(contrib, scale_freq);
2766 sa->load_sum += weight * contrib;
2768 cfs_rq->runnable_load_sum += weight * contrib;
2771 sa->util_sum += contrib * scale_cpu;
2774 /* Remainder of delta accrued against u_0` */
2775 scaled_delta = cap_scale(delta, scale_freq);
2777 sa->load_sum += weight * scaled_delta;
2779 cfs_rq->runnable_load_sum += weight * scaled_delta;
2782 sa->util_sum += scaled_delta * scale_cpu;
2784 sa->period_contrib += delta;
2787 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2789 cfs_rq->runnable_load_avg =
2790 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2792 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2798 #ifdef CONFIG_FAIR_GROUP_SCHED
2800 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2801 * and effective_load (which is not done because it is too costly).
2803 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2805 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2807 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2808 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2809 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2813 #else /* CONFIG_FAIR_GROUP_SCHED */
2814 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2815 #endif /* CONFIG_FAIR_GROUP_SCHED */
2817 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
2819 if (&this_rq()->cfs == cfs_rq) {
2821 * There are a few boundary cases this might miss but it should
2822 * get called often enough that that should (hopefully) not be
2823 * a real problem -- added to that it only calls on the local
2824 * CPU, so if we enqueue remotely we'll miss an update, but
2825 * the next tick/schedule should update.
2827 * It will not get called when we go idle, because the idle
2828 * thread is a different class (!fair), nor will the utilization
2829 * number include things like RT tasks.
2831 * As is, the util number is not freq-invariant (we'd have to
2832 * implement arch_scale_freq_capacity() for that).
2836 cpufreq_update_util(rq_of(cfs_rq), 0);
2840 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2843 * Unsigned subtract and clamp on underflow.
2845 * Explicitly do a load-store to ensure the intermediate value never hits
2846 * memory. This allows lockless observations without ever seeing the negative
2849 #define sub_positive(_ptr, _val) do { \
2850 typeof(_ptr) ptr = (_ptr); \
2851 typeof(*ptr) val = (_val); \
2852 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2856 WRITE_ONCE(*ptr, res); \
2860 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
2861 * @now: current time, as per cfs_rq_clock_task()
2862 * @cfs_rq: cfs_rq to update
2863 * @update_freq: should we call cfs_rq_util_change() or will the call do so
2865 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
2866 * avg. The immediate corollary is that all (fair) tasks must be attached, see
2867 * post_init_entity_util_avg().
2869 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
2871 * Returns true if the load decayed or we removed utilization. It is expected
2872 * that one calls update_tg_load_avg() on this condition, but after you've
2873 * modified the cfs_rq avg (attach/detach), such that we propagate the new
2877 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
2879 struct sched_avg *sa = &cfs_rq->avg;
2880 int decayed, removed = 0, removed_util = 0;
2882 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2883 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2884 sub_positive(&sa->load_avg, r);
2885 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
2889 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2890 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2891 sub_positive(&sa->util_avg, r);
2892 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
2896 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2897 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2899 #ifndef CONFIG_64BIT
2901 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2904 /* Trace CPU load, unless cfs_rq belongs to a non-root task_group */
2905 if (cfs_rq == &rq_of(cfs_rq)->cfs)
2906 trace_sched_load_avg_cpu(cpu_of(rq_of(cfs_rq)), cfs_rq);
2908 if (update_freq && (decayed || removed_util))
2909 cfs_rq_util_change(cfs_rq);
2911 return decayed || removed;
2914 /* Update task and its cfs_rq load average */
2915 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2917 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2918 u64 now = cfs_rq_clock_task(cfs_rq);
2919 int cpu = cpu_of(rq_of(cfs_rq));
2922 * Track task load average for carrying it to new CPU after migrated, and
2923 * track group sched_entity load average for task_h_load calc in migration
2925 __update_load_avg(now, cpu, &se->avg,
2926 se->on_rq * scale_load_down(se->load.weight),
2927 cfs_rq->curr == se, NULL);
2929 if (update_cfs_rq_load_avg(now, cfs_rq, true) && update_tg)
2930 update_tg_load_avg(cfs_rq, 0);
2932 if (entity_is_task(se))
2933 trace_sched_load_avg_task(task_of(se), &se->avg);
2937 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
2938 * @cfs_rq: cfs_rq to attach to
2939 * @se: sched_entity to attach
2941 * Must call update_cfs_rq_load_avg() before this, since we rely on
2942 * cfs_rq->avg.last_update_time being current.
2944 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2946 if (!sched_feat(ATTACH_AGE_LOAD))
2950 * If we got migrated (either between CPUs or between cgroups) we'll
2951 * have aged the average right before clearing @last_update_time.
2953 if (se->avg.last_update_time) {
2954 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2955 &se->avg, 0, 0, NULL);
2958 * XXX: we could have just aged the entire load away if we've been
2959 * absent from the fair class for too long.
2964 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2965 cfs_rq->avg.load_avg += se->avg.load_avg;
2966 cfs_rq->avg.load_sum += se->avg.load_sum;
2967 cfs_rq->avg.util_avg += se->avg.util_avg;
2968 cfs_rq->avg.util_sum += se->avg.util_sum;
2970 cfs_rq_util_change(cfs_rq);
2974 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
2975 * @cfs_rq: cfs_rq to detach from
2976 * @se: sched_entity to detach
2978 * Must call update_cfs_rq_load_avg() before this, since we rely on
2979 * cfs_rq->avg.last_update_time being current.
2981 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2983 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2984 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2985 cfs_rq->curr == se, NULL);
2987 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2988 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
2989 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
2990 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
2992 cfs_rq_util_change(cfs_rq);
2995 /* Add the load generated by se into cfs_rq's load average */
2997 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2999 struct sched_avg *sa = &se->avg;
3000 u64 now = cfs_rq_clock_task(cfs_rq);
3001 int migrated, decayed;
3003 migrated = !sa->last_update_time;
3005 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
3006 se->on_rq * scale_load_down(se->load.weight),
3007 cfs_rq->curr == se, NULL);
3010 decayed = update_cfs_rq_load_avg(now, cfs_rq, !migrated);
3012 cfs_rq->runnable_load_avg += sa->load_avg;
3013 cfs_rq->runnable_load_sum += sa->load_sum;
3016 attach_entity_load_avg(cfs_rq, se);
3018 if (decayed || migrated)
3019 update_tg_load_avg(cfs_rq, 0);
3022 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3024 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3026 update_load_avg(se, 1);
3028 cfs_rq->runnable_load_avg =
3029 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
3030 cfs_rq->runnable_load_sum =
3031 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
3034 #ifndef CONFIG_64BIT
3035 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3037 u64 last_update_time_copy;
3038 u64 last_update_time;
3041 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3043 last_update_time = cfs_rq->avg.last_update_time;
3044 } while (last_update_time != last_update_time_copy);
3046 return last_update_time;
3049 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3051 return cfs_rq->avg.last_update_time;
3056 * Synchronize entity load avg of dequeued entity without locking
3059 void sync_entity_load_avg(struct sched_entity *se)
3061 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3062 u64 last_update_time;
3064 last_update_time = cfs_rq_last_update_time(cfs_rq);
3065 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
3069 * Task first catches up with cfs_rq, and then subtract
3070 * itself from the cfs_rq (task must be off the queue now).
3072 void remove_entity_load_avg(struct sched_entity *se)
3074 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3077 * Newly created task or never used group entity should not be removed
3078 * from its (source) cfs_rq
3080 if (se->avg.last_update_time == 0)
3083 sync_entity_load_avg(se);
3084 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3085 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3089 * Update the rq's load with the elapsed running time before entering
3090 * idle. if the last scheduled task is not a CFS task, idle_enter will
3091 * be the only way to update the runnable statistic.
3093 void idle_enter_fair(struct rq *this_rq)
3098 * Update the rq's load with the elapsed idle time before a task is
3099 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
3100 * be the only way to update the runnable statistic.
3102 void idle_exit_fair(struct rq *this_rq)
3106 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3108 return cfs_rq->runnable_load_avg;
3111 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3113 return cfs_rq->avg.load_avg;
3116 static int idle_balance(struct rq *this_rq);
3118 #else /* CONFIG_SMP */
3120 static inline void update_load_avg(struct sched_entity *se, int update_tg)
3122 cpufreq_update_util(rq_of(cfs_rq_of(se)), 0);
3126 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3128 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3129 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3132 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3134 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3136 static inline int idle_balance(struct rq *rq)
3141 #endif /* CONFIG_SMP */
3143 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
3145 #ifdef CONFIG_SCHEDSTATS
3146 struct task_struct *tsk = NULL;
3148 if (entity_is_task(se))
3151 if (se->statistics.sleep_start) {
3152 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
3157 if (unlikely(delta > se->statistics.sleep_max))
3158 se->statistics.sleep_max = delta;
3160 se->statistics.sleep_start = 0;
3161 se->statistics.sum_sleep_runtime += delta;
3164 account_scheduler_latency(tsk, delta >> 10, 1);
3165 trace_sched_stat_sleep(tsk, delta);
3168 if (se->statistics.block_start) {
3169 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
3174 if (unlikely(delta > se->statistics.block_max))
3175 se->statistics.block_max = delta;
3177 se->statistics.block_start = 0;
3178 se->statistics.sum_sleep_runtime += delta;
3181 if (tsk->in_iowait) {
3182 se->statistics.iowait_sum += delta;
3183 se->statistics.iowait_count++;
3184 trace_sched_stat_iowait(tsk, delta);
3187 trace_sched_stat_blocked(tsk, delta);
3188 trace_sched_blocked_reason(tsk);
3191 * Blocking time is in units of nanosecs, so shift by
3192 * 20 to get a milliseconds-range estimation of the
3193 * amount of time that the task spent sleeping:
3195 if (unlikely(prof_on == SLEEP_PROFILING)) {
3196 profile_hits(SLEEP_PROFILING,
3197 (void *)get_wchan(tsk),
3200 account_scheduler_latency(tsk, delta >> 10, 0);
3206 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3208 #ifdef CONFIG_SCHED_DEBUG
3209 s64 d = se->vruntime - cfs_rq->min_vruntime;
3214 if (d > 3*sysctl_sched_latency)
3215 schedstat_inc(cfs_rq, nr_spread_over);
3220 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3222 u64 vruntime = cfs_rq->min_vruntime;
3225 * The 'current' period is already promised to the current tasks,
3226 * however the extra weight of the new task will slow them down a
3227 * little, place the new task so that it fits in the slot that
3228 * stays open at the end.
3230 if (initial && sched_feat(START_DEBIT))
3231 vruntime += sched_vslice(cfs_rq, se);
3233 /* sleeps up to a single latency don't count. */
3235 unsigned long thresh = sysctl_sched_latency;
3238 * Halve their sleep time's effect, to allow
3239 * for a gentler effect of sleepers:
3241 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3247 /* ensure we never gain time by being placed backwards. */
3248 se->vruntime = max_vruntime(se->vruntime, vruntime);
3251 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3254 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3257 * Update the normalized vruntime before updating min_vruntime
3258 * through calling update_curr().
3260 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3261 se->vruntime += cfs_rq->min_vruntime;
3264 * Update run-time statistics of the 'current'.
3266 update_curr(cfs_rq);
3267 enqueue_entity_load_avg(cfs_rq, se);
3268 account_entity_enqueue(cfs_rq, se);
3269 update_cfs_shares(cfs_rq);
3271 if (flags & ENQUEUE_WAKEUP) {
3272 place_entity(cfs_rq, se, 0);
3273 enqueue_sleeper(cfs_rq, se);
3276 update_stats_enqueue(cfs_rq, se);
3277 check_spread(cfs_rq, se);
3278 if (se != cfs_rq->curr)
3279 __enqueue_entity(cfs_rq, se);
3282 if (cfs_rq->nr_running == 1) {
3283 list_add_leaf_cfs_rq(cfs_rq);
3284 check_enqueue_throttle(cfs_rq);
3288 static void __clear_buddies_last(struct sched_entity *se)
3290 for_each_sched_entity(se) {
3291 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3292 if (cfs_rq->last != se)
3295 cfs_rq->last = NULL;
3299 static void __clear_buddies_next(struct sched_entity *se)
3301 for_each_sched_entity(se) {
3302 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3303 if (cfs_rq->next != se)
3306 cfs_rq->next = NULL;
3310 static void __clear_buddies_skip(struct sched_entity *se)
3312 for_each_sched_entity(se) {
3313 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3314 if (cfs_rq->skip != se)
3317 cfs_rq->skip = NULL;
3321 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3323 if (cfs_rq->last == se)
3324 __clear_buddies_last(se);
3326 if (cfs_rq->next == se)
3327 __clear_buddies_next(se);
3329 if (cfs_rq->skip == se)
3330 __clear_buddies_skip(se);
3333 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3336 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3339 * Update run-time statistics of the 'current'.
3341 update_curr(cfs_rq);
3342 dequeue_entity_load_avg(cfs_rq, se);
3344 update_stats_dequeue(cfs_rq, se);
3345 if (flags & DEQUEUE_SLEEP) {
3346 #ifdef CONFIG_SCHEDSTATS
3347 if (entity_is_task(se)) {
3348 struct task_struct *tsk = task_of(se);
3350 if (tsk->state & TASK_INTERRUPTIBLE)
3351 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3352 if (tsk->state & TASK_UNINTERRUPTIBLE)
3353 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3358 clear_buddies(cfs_rq, se);
3360 if (se != cfs_rq->curr)
3361 __dequeue_entity(cfs_rq, se);
3363 account_entity_dequeue(cfs_rq, se);
3366 * Normalize the entity after updating the min_vruntime because the
3367 * update can refer to the ->curr item and we need to reflect this
3368 * movement in our normalized position.
3370 if (!(flags & DEQUEUE_SLEEP))
3371 se->vruntime -= cfs_rq->min_vruntime;
3373 /* return excess runtime on last dequeue */
3374 return_cfs_rq_runtime(cfs_rq);
3376 update_min_vruntime(cfs_rq);
3377 update_cfs_shares(cfs_rq);
3381 * Preempt the current task with a newly woken task if needed:
3384 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3386 unsigned long ideal_runtime, delta_exec;
3387 struct sched_entity *se;
3390 ideal_runtime = sched_slice(cfs_rq, curr);
3391 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3392 if (delta_exec > ideal_runtime) {
3393 resched_curr(rq_of(cfs_rq));
3395 * The current task ran long enough, ensure it doesn't get
3396 * re-elected due to buddy favours.
3398 clear_buddies(cfs_rq, curr);
3403 * Ensure that a task that missed wakeup preemption by a
3404 * narrow margin doesn't have to wait for a full slice.
3405 * This also mitigates buddy induced latencies under load.
3407 if (delta_exec < sysctl_sched_min_granularity)
3410 se = __pick_first_entity(cfs_rq);
3411 delta = curr->vruntime - se->vruntime;
3416 if (delta > ideal_runtime)
3417 resched_curr(rq_of(cfs_rq));
3421 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3423 /* 'current' is not kept within the tree. */
3426 * Any task has to be enqueued before it get to execute on
3427 * a CPU. So account for the time it spent waiting on the
3430 update_stats_wait_end(cfs_rq, se);
3431 __dequeue_entity(cfs_rq, se);
3432 update_load_avg(se, 1);
3435 update_stats_curr_start(cfs_rq, se);
3437 #ifdef CONFIG_SCHEDSTATS
3439 * Track our maximum slice length, if the CPU's load is at
3440 * least twice that of our own weight (i.e. dont track it
3441 * when there are only lesser-weight tasks around):
3443 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3444 se->statistics.slice_max = max(se->statistics.slice_max,
3445 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3448 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3452 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3455 * Pick the next process, keeping these things in mind, in this order:
3456 * 1) keep things fair between processes/task groups
3457 * 2) pick the "next" process, since someone really wants that to run
3458 * 3) pick the "last" process, for cache locality
3459 * 4) do not run the "skip" process, if something else is available
3461 static struct sched_entity *
3462 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3464 struct sched_entity *left = __pick_first_entity(cfs_rq);
3465 struct sched_entity *se;
3468 * If curr is set we have to see if its left of the leftmost entity
3469 * still in the tree, provided there was anything in the tree at all.
3471 if (!left || (curr && entity_before(curr, left)))
3474 se = left; /* ideally we run the leftmost entity */
3477 * Avoid running the skip buddy, if running something else can
3478 * be done without getting too unfair.
3480 if (cfs_rq->skip == se) {
3481 struct sched_entity *second;
3484 second = __pick_first_entity(cfs_rq);
3486 second = __pick_next_entity(se);
3487 if (!second || (curr && entity_before(curr, second)))
3491 if (second && wakeup_preempt_entity(second, left) < 1)
3496 * Prefer last buddy, try to return the CPU to a preempted task.
3498 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3502 * Someone really wants this to run. If it's not unfair, run it.
3504 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3507 clear_buddies(cfs_rq, se);
3512 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3514 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3517 * If still on the runqueue then deactivate_task()
3518 * was not called and update_curr() has to be done:
3521 update_curr(cfs_rq);
3523 /* throttle cfs_rqs exceeding runtime */
3524 check_cfs_rq_runtime(cfs_rq);
3526 check_spread(cfs_rq, prev);
3528 update_stats_wait_start(cfs_rq, prev);
3529 /* Put 'current' back into the tree. */
3530 __enqueue_entity(cfs_rq, prev);
3531 /* in !on_rq case, update occurred at dequeue */
3532 update_load_avg(prev, 0);
3534 cfs_rq->curr = NULL;
3538 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3541 * Update run-time statistics of the 'current'.
3543 update_curr(cfs_rq);
3546 * Ensure that runnable average is periodically updated.
3548 update_load_avg(curr, 1);
3549 update_cfs_shares(cfs_rq);
3551 #ifdef CONFIG_SCHED_HRTICK
3553 * queued ticks are scheduled to match the slice, so don't bother
3554 * validating it and just reschedule.
3557 resched_curr(rq_of(cfs_rq));
3561 * don't let the period tick interfere with the hrtick preemption
3563 if (!sched_feat(DOUBLE_TICK) &&
3564 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3568 if (cfs_rq->nr_running > 1)
3569 check_preempt_tick(cfs_rq, curr);
3573 /**************************************************
3574 * CFS bandwidth control machinery
3577 #ifdef CONFIG_CFS_BANDWIDTH
3579 #ifdef HAVE_JUMP_LABEL
3580 static struct static_key __cfs_bandwidth_used;
3582 static inline bool cfs_bandwidth_used(void)
3584 return static_key_false(&__cfs_bandwidth_used);
3587 void cfs_bandwidth_usage_inc(void)
3589 static_key_slow_inc(&__cfs_bandwidth_used);
3592 void cfs_bandwidth_usage_dec(void)
3594 static_key_slow_dec(&__cfs_bandwidth_used);
3596 #else /* HAVE_JUMP_LABEL */
3597 static bool cfs_bandwidth_used(void)
3602 void cfs_bandwidth_usage_inc(void) {}
3603 void cfs_bandwidth_usage_dec(void) {}
3604 #endif /* HAVE_JUMP_LABEL */
3607 * default period for cfs group bandwidth.
3608 * default: 0.1s, units: nanoseconds
3610 static inline u64 default_cfs_period(void)
3612 return 100000000ULL;
3615 static inline u64 sched_cfs_bandwidth_slice(void)
3617 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3621 * Replenish runtime according to assigned quota and update expiration time.
3622 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3623 * additional synchronization around rq->lock.
3625 * requires cfs_b->lock
3627 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3631 if (cfs_b->quota == RUNTIME_INF)
3634 now = sched_clock_cpu(smp_processor_id());
3635 cfs_b->runtime = cfs_b->quota;
3636 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3639 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3641 return &tg->cfs_bandwidth;
3644 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3645 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3647 if (unlikely(cfs_rq->throttle_count))
3648 return cfs_rq->throttled_clock_task;
3650 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3653 /* returns 0 on failure to allocate runtime */
3654 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3656 struct task_group *tg = cfs_rq->tg;
3657 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3658 u64 amount = 0, min_amount, expires;
3660 /* note: this is a positive sum as runtime_remaining <= 0 */
3661 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3663 raw_spin_lock(&cfs_b->lock);
3664 if (cfs_b->quota == RUNTIME_INF)
3665 amount = min_amount;
3667 start_cfs_bandwidth(cfs_b);
3669 if (cfs_b->runtime > 0) {
3670 amount = min(cfs_b->runtime, min_amount);
3671 cfs_b->runtime -= amount;
3675 expires = cfs_b->runtime_expires;
3676 raw_spin_unlock(&cfs_b->lock);
3678 cfs_rq->runtime_remaining += amount;
3680 * we may have advanced our local expiration to account for allowed
3681 * spread between our sched_clock and the one on which runtime was
3684 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3685 cfs_rq->runtime_expires = expires;
3687 return cfs_rq->runtime_remaining > 0;
3691 * Note: This depends on the synchronization provided by sched_clock and the
3692 * fact that rq->clock snapshots this value.
3694 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3696 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3698 /* if the deadline is ahead of our clock, nothing to do */
3699 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3702 if (cfs_rq->runtime_remaining < 0)
3706 * If the local deadline has passed we have to consider the
3707 * possibility that our sched_clock is 'fast' and the global deadline
3708 * has not truly expired.
3710 * Fortunately we can check determine whether this the case by checking
3711 * whether the global deadline has advanced. It is valid to compare
3712 * cfs_b->runtime_expires without any locks since we only care about
3713 * exact equality, so a partial write will still work.
3716 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3717 /* extend local deadline, drift is bounded above by 2 ticks */
3718 cfs_rq->runtime_expires += TICK_NSEC;
3720 /* global deadline is ahead, expiration has passed */
3721 cfs_rq->runtime_remaining = 0;
3725 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3727 /* dock delta_exec before expiring quota (as it could span periods) */
3728 cfs_rq->runtime_remaining -= delta_exec;
3729 expire_cfs_rq_runtime(cfs_rq);
3731 if (likely(cfs_rq->runtime_remaining > 0))
3735 * if we're unable to extend our runtime we resched so that the active
3736 * hierarchy can be throttled
3738 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3739 resched_curr(rq_of(cfs_rq));
3742 static __always_inline
3743 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3745 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3748 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3751 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3753 return cfs_bandwidth_used() && cfs_rq->throttled;
3756 /* check whether cfs_rq, or any parent, is throttled */
3757 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3759 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3763 * Ensure that neither of the group entities corresponding to src_cpu or
3764 * dest_cpu are members of a throttled hierarchy when performing group
3765 * load-balance operations.
3767 static inline int throttled_lb_pair(struct task_group *tg,
3768 int src_cpu, int dest_cpu)
3770 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3772 src_cfs_rq = tg->cfs_rq[src_cpu];
3773 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3775 return throttled_hierarchy(src_cfs_rq) ||
3776 throttled_hierarchy(dest_cfs_rq);
3779 /* updated child weight may affect parent so we have to do this bottom up */
3780 static int tg_unthrottle_up(struct task_group *tg, void *data)
3782 struct rq *rq = data;
3783 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3785 cfs_rq->throttle_count--;
3787 if (!cfs_rq->throttle_count) {
3788 /* adjust cfs_rq_clock_task() */
3789 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3790 cfs_rq->throttled_clock_task;
3797 static int tg_throttle_down(struct task_group *tg, void *data)
3799 struct rq *rq = data;
3800 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3802 /* group is entering throttled state, stop time */
3803 if (!cfs_rq->throttle_count)
3804 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3805 cfs_rq->throttle_count++;
3810 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3812 struct rq *rq = rq_of(cfs_rq);
3813 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3814 struct sched_entity *se;
3815 long task_delta, dequeue = 1;
3818 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3820 /* freeze hierarchy runnable averages while throttled */
3822 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3825 task_delta = cfs_rq->h_nr_running;
3826 for_each_sched_entity(se) {
3827 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3828 /* throttled entity or throttle-on-deactivate */
3833 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3834 qcfs_rq->h_nr_running -= task_delta;
3836 if (qcfs_rq->load.weight)
3841 sub_nr_running(rq, task_delta);
3843 cfs_rq->throttled = 1;
3844 cfs_rq->throttled_clock = rq_clock(rq);
3845 raw_spin_lock(&cfs_b->lock);
3846 empty = list_empty(&cfs_b->throttled_cfs_rq);
3849 * Add to the _head_ of the list, so that an already-started
3850 * distribute_cfs_runtime will not see us
3852 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3855 * If we're the first throttled task, make sure the bandwidth
3859 start_cfs_bandwidth(cfs_b);
3861 raw_spin_unlock(&cfs_b->lock);
3864 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3866 struct rq *rq = rq_of(cfs_rq);
3867 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3868 struct sched_entity *se;
3872 se = cfs_rq->tg->se[cpu_of(rq)];
3874 cfs_rq->throttled = 0;
3876 update_rq_clock(rq);
3878 raw_spin_lock(&cfs_b->lock);
3879 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3880 list_del_rcu(&cfs_rq->throttled_list);
3881 raw_spin_unlock(&cfs_b->lock);
3883 /* update hierarchical throttle state */
3884 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3886 if (!cfs_rq->load.weight)
3889 task_delta = cfs_rq->h_nr_running;
3890 for_each_sched_entity(se) {
3894 cfs_rq = cfs_rq_of(se);
3896 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3897 cfs_rq->h_nr_running += task_delta;
3899 if (cfs_rq_throttled(cfs_rq))
3904 add_nr_running(rq, task_delta);
3906 /* determine whether we need to wake up potentially idle cpu */
3907 if (rq->curr == rq->idle && rq->cfs.nr_running)
3911 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3912 u64 remaining, u64 expires)
3914 struct cfs_rq *cfs_rq;
3916 u64 starting_runtime = remaining;
3919 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3921 struct rq *rq = rq_of(cfs_rq);
3923 raw_spin_lock(&rq->lock);
3924 if (!cfs_rq_throttled(cfs_rq))
3927 runtime = -cfs_rq->runtime_remaining + 1;
3928 if (runtime > remaining)
3929 runtime = remaining;
3930 remaining -= runtime;
3932 cfs_rq->runtime_remaining += runtime;
3933 cfs_rq->runtime_expires = expires;
3935 /* we check whether we're throttled above */
3936 if (cfs_rq->runtime_remaining > 0)
3937 unthrottle_cfs_rq(cfs_rq);
3940 raw_spin_unlock(&rq->lock);
3947 return starting_runtime - remaining;
3951 * Responsible for refilling a task_group's bandwidth and unthrottling its
3952 * cfs_rqs as appropriate. If there has been no activity within the last
3953 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3954 * used to track this state.
3956 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3958 u64 runtime, runtime_expires;
3961 /* no need to continue the timer with no bandwidth constraint */
3962 if (cfs_b->quota == RUNTIME_INF)
3963 goto out_deactivate;
3965 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3966 cfs_b->nr_periods += overrun;
3969 * idle depends on !throttled (for the case of a large deficit), and if
3970 * we're going inactive then everything else can be deferred
3972 if (cfs_b->idle && !throttled)
3973 goto out_deactivate;
3975 __refill_cfs_bandwidth_runtime(cfs_b);
3978 /* mark as potentially idle for the upcoming period */
3983 /* account preceding periods in which throttling occurred */
3984 cfs_b->nr_throttled += overrun;
3986 runtime_expires = cfs_b->runtime_expires;
3989 * This check is repeated as we are holding onto the new bandwidth while
3990 * we unthrottle. This can potentially race with an unthrottled group
3991 * trying to acquire new bandwidth from the global pool. This can result
3992 * in us over-using our runtime if it is all used during this loop, but
3993 * only by limited amounts in that extreme case.
3995 while (throttled && cfs_b->runtime > 0) {
3996 runtime = cfs_b->runtime;
3997 raw_spin_unlock(&cfs_b->lock);
3998 /* we can't nest cfs_b->lock while distributing bandwidth */
3999 runtime = distribute_cfs_runtime(cfs_b, runtime,
4001 raw_spin_lock(&cfs_b->lock);
4003 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4005 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4009 * While we are ensured activity in the period following an
4010 * unthrottle, this also covers the case in which the new bandwidth is
4011 * insufficient to cover the existing bandwidth deficit. (Forcing the
4012 * timer to remain active while there are any throttled entities.)
4022 /* a cfs_rq won't donate quota below this amount */
4023 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4024 /* minimum remaining period time to redistribute slack quota */
4025 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4026 /* how long we wait to gather additional slack before distributing */
4027 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4030 * Are we near the end of the current quota period?
4032 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4033 * hrtimer base being cleared by hrtimer_start. In the case of
4034 * migrate_hrtimers, base is never cleared, so we are fine.
4036 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4038 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4041 /* if the call-back is running a quota refresh is already occurring */
4042 if (hrtimer_callback_running(refresh_timer))
4045 /* is a quota refresh about to occur? */
4046 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4047 if (remaining < min_expire)
4053 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4055 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4057 /* if there's a quota refresh soon don't bother with slack */
4058 if (runtime_refresh_within(cfs_b, min_left))
4061 hrtimer_start(&cfs_b->slack_timer,
4062 ns_to_ktime(cfs_bandwidth_slack_period),
4066 /* we know any runtime found here is valid as update_curr() precedes return */
4067 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4069 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4070 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4072 if (slack_runtime <= 0)
4075 raw_spin_lock(&cfs_b->lock);
4076 if (cfs_b->quota != RUNTIME_INF &&
4077 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4078 cfs_b->runtime += slack_runtime;
4080 /* we are under rq->lock, defer unthrottling using a timer */
4081 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4082 !list_empty(&cfs_b->throttled_cfs_rq))
4083 start_cfs_slack_bandwidth(cfs_b);
4085 raw_spin_unlock(&cfs_b->lock);
4087 /* even if it's not valid for return we don't want to try again */
4088 cfs_rq->runtime_remaining -= slack_runtime;
4091 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4093 if (!cfs_bandwidth_used())
4096 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4099 __return_cfs_rq_runtime(cfs_rq);
4103 * This is done with a timer (instead of inline with bandwidth return) since
4104 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4106 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4108 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4111 /* confirm we're still not at a refresh boundary */
4112 raw_spin_lock(&cfs_b->lock);
4113 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4114 raw_spin_unlock(&cfs_b->lock);
4118 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4119 runtime = cfs_b->runtime;
4121 expires = cfs_b->runtime_expires;
4122 raw_spin_unlock(&cfs_b->lock);
4127 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4129 raw_spin_lock(&cfs_b->lock);
4130 if (expires == cfs_b->runtime_expires)
4131 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4132 raw_spin_unlock(&cfs_b->lock);
4136 * When a group wakes up we want to make sure that its quota is not already
4137 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4138 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4140 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4142 if (!cfs_bandwidth_used())
4145 /* Synchronize hierarchical throttle counter: */
4146 if (unlikely(!cfs_rq->throttle_uptodate)) {
4147 struct rq *rq = rq_of(cfs_rq);
4148 struct cfs_rq *pcfs_rq;
4149 struct task_group *tg;
4151 cfs_rq->throttle_uptodate = 1;
4153 /* Get closest up-to-date node, because leaves go first: */
4154 for (tg = cfs_rq->tg->parent; tg; tg = tg->parent) {
4155 pcfs_rq = tg->cfs_rq[cpu_of(rq)];
4156 if (pcfs_rq->throttle_uptodate)
4160 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4161 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4165 /* an active group must be handled by the update_curr()->put() path */
4166 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4169 /* ensure the group is not already throttled */
4170 if (cfs_rq_throttled(cfs_rq))
4173 /* update runtime allocation */
4174 account_cfs_rq_runtime(cfs_rq, 0);
4175 if (cfs_rq->runtime_remaining <= 0)
4176 throttle_cfs_rq(cfs_rq);
4179 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4180 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4182 if (!cfs_bandwidth_used())
4185 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4189 * it's possible for a throttled entity to be forced into a running
4190 * state (e.g. set_curr_task), in this case we're finished.
4192 if (cfs_rq_throttled(cfs_rq))
4195 throttle_cfs_rq(cfs_rq);
4199 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4201 struct cfs_bandwidth *cfs_b =
4202 container_of(timer, struct cfs_bandwidth, slack_timer);
4204 do_sched_cfs_slack_timer(cfs_b);
4206 return HRTIMER_NORESTART;
4209 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4211 struct cfs_bandwidth *cfs_b =
4212 container_of(timer, struct cfs_bandwidth, period_timer);
4216 raw_spin_lock(&cfs_b->lock);
4218 overrun = hrtimer_forward_now(timer, cfs_b->period);
4222 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4225 cfs_b->period_active = 0;
4226 raw_spin_unlock(&cfs_b->lock);
4228 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4231 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4233 raw_spin_lock_init(&cfs_b->lock);
4235 cfs_b->quota = RUNTIME_INF;
4236 cfs_b->period = ns_to_ktime(default_cfs_period());
4238 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4239 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4240 cfs_b->period_timer.function = sched_cfs_period_timer;
4241 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4242 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4245 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4247 cfs_rq->runtime_enabled = 0;
4248 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4251 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4253 lockdep_assert_held(&cfs_b->lock);
4255 if (!cfs_b->period_active) {
4256 cfs_b->period_active = 1;
4257 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4258 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4262 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4264 /* init_cfs_bandwidth() was not called */
4265 if (!cfs_b->throttled_cfs_rq.next)
4268 hrtimer_cancel(&cfs_b->period_timer);
4269 hrtimer_cancel(&cfs_b->slack_timer);
4272 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4274 struct cfs_rq *cfs_rq;
4276 for_each_leaf_cfs_rq(rq, cfs_rq) {
4277 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4279 raw_spin_lock(&cfs_b->lock);
4280 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4281 raw_spin_unlock(&cfs_b->lock);
4285 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4287 struct cfs_rq *cfs_rq;
4289 for_each_leaf_cfs_rq(rq, cfs_rq) {
4290 if (!cfs_rq->runtime_enabled)
4294 * clock_task is not advancing so we just need to make sure
4295 * there's some valid quota amount
4297 cfs_rq->runtime_remaining = 1;
4299 * Offline rq is schedulable till cpu is completely disabled
4300 * in take_cpu_down(), so we prevent new cfs throttling here.
4302 cfs_rq->runtime_enabled = 0;
4304 if (cfs_rq_throttled(cfs_rq))
4305 unthrottle_cfs_rq(cfs_rq);
4309 #else /* CONFIG_CFS_BANDWIDTH */
4310 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4312 return rq_clock_task(rq_of(cfs_rq));
4315 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4316 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4317 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4318 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4320 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4325 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4330 static inline int throttled_lb_pair(struct task_group *tg,
4331 int src_cpu, int dest_cpu)
4336 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4338 #ifdef CONFIG_FAIR_GROUP_SCHED
4339 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4342 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4346 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4347 static inline void update_runtime_enabled(struct rq *rq) {}
4348 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4350 #endif /* CONFIG_CFS_BANDWIDTH */
4352 /**************************************************
4353 * CFS operations on tasks:
4356 #ifdef CONFIG_SCHED_HRTICK
4357 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4359 struct sched_entity *se = &p->se;
4360 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4362 WARN_ON(task_rq(p) != rq);
4364 if (cfs_rq->nr_running > 1) {
4365 u64 slice = sched_slice(cfs_rq, se);
4366 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4367 s64 delta = slice - ran;
4374 hrtick_start(rq, delta);
4379 * called from enqueue/dequeue and updates the hrtick when the
4380 * current task is from our class and nr_running is low enough
4383 static void hrtick_update(struct rq *rq)
4385 struct task_struct *curr = rq->curr;
4387 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4390 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4391 hrtick_start_fair(rq, curr);
4393 #else /* !CONFIG_SCHED_HRTICK */
4395 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4399 static inline void hrtick_update(struct rq *rq)
4405 static bool cpu_overutilized(int cpu);
4406 unsigned long boosted_cpu_util(int cpu);
4408 #define boosted_cpu_util(cpu) cpu_util(cpu)
4412 static void update_capacity_of(int cpu)
4414 unsigned long req_cap;
4419 /* Convert scale-invariant capacity to cpu. */
4420 req_cap = boosted_cpu_util(cpu);
4421 req_cap = req_cap * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
4422 set_cfs_cpu_capacity(cpu, true, req_cap);
4427 * The enqueue_task method is called before nr_running is
4428 * increased. Here we update the fair scheduling stats and
4429 * then put the task into the rbtree:
4432 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4434 struct cfs_rq *cfs_rq;
4435 struct sched_entity *se = &p->se;
4437 int task_new = flags & ENQUEUE_WAKEUP_NEW;
4438 int task_wakeup = flags & ENQUEUE_WAKEUP;
4442 * If in_iowait is set, the code below may not trigger any cpufreq
4443 * utilization updates, so do it here explicitly with the IOWAIT flag
4447 cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_IOWAIT);
4449 for_each_sched_entity(se) {
4452 cfs_rq = cfs_rq_of(se);
4453 enqueue_entity(cfs_rq, se, flags);
4456 * end evaluation on encountering a throttled cfs_rq
4458 * note: in the case of encountering a throttled cfs_rq we will
4459 * post the final h_nr_running increment below.
4461 if (cfs_rq_throttled(cfs_rq))
4463 cfs_rq->h_nr_running++;
4464 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4466 flags = ENQUEUE_WAKEUP;
4469 for_each_sched_entity(se) {
4470 cfs_rq = cfs_rq_of(se);
4471 cfs_rq->h_nr_running++;
4472 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4474 if (cfs_rq_throttled(cfs_rq))
4477 update_load_avg(se, 1);
4478 update_cfs_shares(cfs_rq);
4482 add_nr_running(rq, 1);
4487 * Update SchedTune accounting.
4489 * We do it before updating the CPU capacity to ensure the
4490 * boost value of the current task is accounted for in the
4491 * selection of the OPP.
4493 * We do it also in the case where we enqueue a throttled task;
4494 * we could argue that a throttled task should not boost a CPU,
4496 * a) properly implementing CPU boosting considering throttled
4497 * tasks will increase a lot the complexity of the solution
4498 * b) it's not easy to quantify the benefits introduced by
4499 * such a more complex solution.
4500 * Thus, for the time being we go for the simple solution and boost
4501 * also for throttled RQs.
4503 schedtune_enqueue_task(p, cpu_of(rq));
4506 walt_inc_cumulative_runnable_avg(rq, p);
4507 if (!task_new && !rq->rd->overutilized &&
4508 cpu_overutilized(rq->cpu)) {
4509 rq->rd->overutilized = true;
4510 trace_sched_overutilized(true);
4514 * We want to potentially trigger a freq switch
4515 * request only for tasks that are waking up; this is
4516 * because we get here also during load balancing, but
4517 * in these cases it seems wise to trigger as single
4518 * request after load balancing is done.
4520 if (task_new || task_wakeup)
4521 update_capacity_of(cpu_of(rq));
4524 #endif /* CONFIG_SMP */
4528 static void set_next_buddy(struct sched_entity *se);
4531 * The dequeue_task method is called before nr_running is
4532 * decreased. We remove the task from the rbtree and
4533 * update the fair scheduling stats:
4535 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4537 struct cfs_rq *cfs_rq;
4538 struct sched_entity *se = &p->se;
4539 int task_sleep = flags & DEQUEUE_SLEEP;
4541 for_each_sched_entity(se) {
4542 cfs_rq = cfs_rq_of(se);
4543 dequeue_entity(cfs_rq, se, flags);
4546 * end evaluation on encountering a throttled cfs_rq
4548 * note: in the case of encountering a throttled cfs_rq we will
4549 * post the final h_nr_running decrement below.
4551 if (cfs_rq_throttled(cfs_rq))
4553 cfs_rq->h_nr_running--;
4554 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4556 /* Don't dequeue parent if it has other entities besides us */
4557 if (cfs_rq->load.weight) {
4558 /* Avoid re-evaluating load for this entity: */
4559 se = parent_entity(se);
4561 * Bias pick_next to pick a task from this cfs_rq, as
4562 * p is sleeping when it is within its sched_slice.
4564 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4568 flags |= DEQUEUE_SLEEP;
4571 for_each_sched_entity(se) {
4572 cfs_rq = cfs_rq_of(se);
4573 cfs_rq->h_nr_running--;
4574 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4576 if (cfs_rq_throttled(cfs_rq))
4579 update_load_avg(se, 1);
4580 update_cfs_shares(cfs_rq);
4584 sub_nr_running(rq, 1);
4589 * Update SchedTune accounting
4591 * We do it before updating the CPU capacity to ensure the
4592 * boost value of the current task is accounted for in the
4593 * selection of the OPP.
4595 schedtune_dequeue_task(p, cpu_of(rq));
4598 walt_dec_cumulative_runnable_avg(rq, p);
4601 * We want to potentially trigger a freq switch
4602 * request only for tasks that are going to sleep;
4603 * this is because we get here also during load
4604 * balancing, but in these cases it seems wise to
4605 * trigger as single request after load balancing is
4609 if (rq->cfs.nr_running)
4610 update_capacity_of(cpu_of(rq));
4611 else if (sched_freq())
4612 set_cfs_cpu_capacity(cpu_of(rq), false, 0);
4616 #endif /* CONFIG_SMP */
4624 * per rq 'load' arrray crap; XXX kill this.
4628 * The exact cpuload at various idx values, calculated at every tick would be
4629 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4631 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4632 * on nth tick when cpu may be busy, then we have:
4633 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4634 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4636 * decay_load_missed() below does efficient calculation of
4637 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4638 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4640 * The calculation is approximated on a 128 point scale.
4641 * degrade_zero_ticks is the number of ticks after which load at any
4642 * particular idx is approximated to be zero.
4643 * degrade_factor is a precomputed table, a row for each load idx.
4644 * Each column corresponds to degradation factor for a power of two ticks,
4645 * based on 128 point scale.
4647 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4648 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4650 * With this power of 2 load factors, we can degrade the load n times
4651 * by looking at 1 bits in n and doing as many mult/shift instead of
4652 * n mult/shifts needed by the exact degradation.
4654 #define DEGRADE_SHIFT 7
4655 static const unsigned char
4656 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4657 static const unsigned char
4658 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4659 {0, 0, 0, 0, 0, 0, 0, 0},
4660 {64, 32, 8, 0, 0, 0, 0, 0},
4661 {96, 72, 40, 12, 1, 0, 0},
4662 {112, 98, 75, 43, 15, 1, 0},
4663 {120, 112, 98, 76, 45, 16, 2} };
4666 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4667 * would be when CPU is idle and so we just decay the old load without
4668 * adding any new load.
4670 static unsigned long
4671 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4675 if (!missed_updates)
4678 if (missed_updates >= degrade_zero_ticks[idx])
4682 return load >> missed_updates;
4684 while (missed_updates) {
4685 if (missed_updates % 2)
4686 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4688 missed_updates >>= 1;
4695 * Update rq->cpu_load[] statistics. This function is usually called every
4696 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4697 * every tick. We fix it up based on jiffies.
4699 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4700 unsigned long pending_updates)
4704 this_rq->nr_load_updates++;
4706 /* Update our load: */
4707 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4708 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4709 unsigned long old_load, new_load;
4711 /* scale is effectively 1 << i now, and >> i divides by scale */
4713 old_load = this_rq->cpu_load[i];
4714 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4715 new_load = this_load;
4717 * Round up the averaging division if load is increasing. This
4718 * prevents us from getting stuck on 9 if the load is 10, for
4721 if (new_load > old_load)
4722 new_load += scale - 1;
4724 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4727 sched_avg_update(this_rq);
4730 /* Used instead of source_load when we know the type == 0 */
4731 static unsigned long weighted_cpuload(const int cpu)
4733 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4736 #ifdef CONFIG_NO_HZ_COMMON
4738 * There is no sane way to deal with nohz on smp when using jiffies because the
4739 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4740 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4742 * Therefore we cannot use the delta approach from the regular tick since that
4743 * would seriously skew the load calculation. However we'll make do for those
4744 * updates happening while idle (nohz_idle_balance) or coming out of idle
4745 * (tick_nohz_idle_exit).
4747 * This means we might still be one tick off for nohz periods.
4751 * Called from nohz_idle_balance() to update the load ratings before doing the
4754 static void update_idle_cpu_load(struct rq *this_rq)
4756 unsigned long curr_jiffies = READ_ONCE(jiffies);
4757 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4758 unsigned long pending_updates;
4761 * bail if there's load or we're actually up-to-date.
4763 if (load || curr_jiffies == this_rq->last_load_update_tick)
4766 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4767 this_rq->last_load_update_tick = curr_jiffies;
4769 __update_cpu_load(this_rq, load, pending_updates);
4773 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4775 void update_cpu_load_nohz(void)
4777 struct rq *this_rq = this_rq();
4778 unsigned long curr_jiffies = READ_ONCE(jiffies);
4779 unsigned long pending_updates;
4781 if (curr_jiffies == this_rq->last_load_update_tick)
4784 raw_spin_lock(&this_rq->lock);
4785 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4786 if (pending_updates) {
4787 this_rq->last_load_update_tick = curr_jiffies;
4789 * We were idle, this means load 0, the current load might be
4790 * !0 due to remote wakeups and the sort.
4792 __update_cpu_load(this_rq, 0, pending_updates);
4794 raw_spin_unlock(&this_rq->lock);
4796 #endif /* CONFIG_NO_HZ */
4799 * Called from scheduler_tick()
4801 void update_cpu_load_active(struct rq *this_rq)
4803 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4805 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4807 this_rq->last_load_update_tick = jiffies;
4808 __update_cpu_load(this_rq, load, 1);
4812 * Return a low guess at the load of a migration-source cpu weighted
4813 * according to the scheduling class and "nice" value.
4815 * We want to under-estimate the load of migration sources, to
4816 * balance conservatively.
4818 static unsigned long source_load(int cpu, int type)
4820 struct rq *rq = cpu_rq(cpu);
4821 unsigned long total = weighted_cpuload(cpu);
4823 if (type == 0 || !sched_feat(LB_BIAS))
4826 return min(rq->cpu_load[type-1], total);
4830 * Return a high guess at the load of a migration-target cpu weighted
4831 * according to the scheduling class and "nice" value.
4833 static unsigned long target_load(int cpu, int type)
4835 struct rq *rq = cpu_rq(cpu);
4836 unsigned long total = weighted_cpuload(cpu);
4838 if (type == 0 || !sched_feat(LB_BIAS))
4841 return max(rq->cpu_load[type-1], total);
4845 static unsigned long cpu_avg_load_per_task(int cpu)
4847 struct rq *rq = cpu_rq(cpu);
4848 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4849 unsigned long load_avg = weighted_cpuload(cpu);
4852 return load_avg / nr_running;
4857 static void record_wakee(struct task_struct *p)
4860 * Rough decay (wiping) for cost saving, don't worry
4861 * about the boundary, really active task won't care
4864 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4865 current->wakee_flips >>= 1;
4866 current->wakee_flip_decay_ts = jiffies;
4869 if (current->last_wakee != p) {
4870 current->last_wakee = p;
4871 current->wakee_flips++;
4875 static void task_waking_fair(struct task_struct *p)
4877 struct sched_entity *se = &p->se;
4878 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4881 #ifndef CONFIG_64BIT
4882 u64 min_vruntime_copy;
4885 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4887 min_vruntime = cfs_rq->min_vruntime;
4888 } while (min_vruntime != min_vruntime_copy);
4890 min_vruntime = cfs_rq->min_vruntime;
4893 se->vruntime -= min_vruntime;
4897 #ifdef CONFIG_FAIR_GROUP_SCHED
4899 * effective_load() calculates the load change as seen from the root_task_group
4901 * Adding load to a group doesn't make a group heavier, but can cause movement
4902 * of group shares between cpus. Assuming the shares were perfectly aligned one
4903 * can calculate the shift in shares.
4905 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4906 * on this @cpu and results in a total addition (subtraction) of @wg to the
4907 * total group weight.
4909 * Given a runqueue weight distribution (rw_i) we can compute a shares
4910 * distribution (s_i) using:
4912 * s_i = rw_i / \Sum rw_j (1)
4914 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4915 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4916 * shares distribution (s_i):
4918 * rw_i = { 2, 4, 1, 0 }
4919 * s_i = { 2/7, 4/7, 1/7, 0 }
4921 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4922 * task used to run on and the CPU the waker is running on), we need to
4923 * compute the effect of waking a task on either CPU and, in case of a sync
4924 * wakeup, compute the effect of the current task going to sleep.
4926 * So for a change of @wl to the local @cpu with an overall group weight change
4927 * of @wl we can compute the new shares distribution (s'_i) using:
4929 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4931 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4932 * differences in waking a task to CPU 0. The additional task changes the
4933 * weight and shares distributions like:
4935 * rw'_i = { 3, 4, 1, 0 }
4936 * s'_i = { 3/8, 4/8, 1/8, 0 }
4938 * We can then compute the difference in effective weight by using:
4940 * dw_i = S * (s'_i - s_i) (3)
4942 * Where 'S' is the group weight as seen by its parent.
4944 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4945 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4946 * 4/7) times the weight of the group.
4948 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4950 struct sched_entity *se = tg->se[cpu];
4952 if (!tg->parent) /* the trivial, non-cgroup case */
4955 for_each_sched_entity(se) {
4956 struct cfs_rq *cfs_rq = se->my_q;
4957 long W, w = cfs_rq_load_avg(cfs_rq);
4962 * W = @wg + \Sum rw_j
4964 W = wg + atomic_long_read(&tg->load_avg);
4966 /* Ensure \Sum rw_j >= rw_i */
4967 W -= cfs_rq->tg_load_avg_contrib;
4976 * wl = S * s'_i; see (2)
4979 wl = (w * (long)tg->shares) / W;
4984 * Per the above, wl is the new se->load.weight value; since
4985 * those are clipped to [MIN_SHARES, ...) do so now. See
4986 * calc_cfs_shares().
4988 if (wl < MIN_SHARES)
4992 * wl = dw_i = S * (s'_i - s_i); see (3)
4994 wl -= se->avg.load_avg;
4997 * Recursively apply this logic to all parent groups to compute
4998 * the final effective load change on the root group. Since
4999 * only the @tg group gets extra weight, all parent groups can
5000 * only redistribute existing shares. @wl is the shift in shares
5001 * resulting from this level per the above.
5010 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
5018 * Returns the current capacity of cpu after applying both
5019 * cpu and freq scaling.
5021 unsigned long capacity_curr_of(int cpu)
5023 return cpu_rq(cpu)->cpu_capacity_orig *
5024 arch_scale_freq_capacity(NULL, cpu)
5025 >> SCHED_CAPACITY_SHIFT;
5028 static inline bool energy_aware(void)
5030 return sched_feat(ENERGY_AWARE);
5034 struct sched_group *sg_top;
5035 struct sched_group *sg_cap;
5042 struct task_struct *task;
5057 * __cpu_norm_util() returns the cpu util relative to a specific capacity,
5058 * i.e. it's busy ratio, in the range [0..SCHED_LOAD_SCALE] which is useful for
5059 * energy calculations. Using the scale-invariant util returned by
5060 * cpu_util() and approximating scale-invariant util by:
5062 * util ~ (curr_freq/max_freq)*1024 * capacity_orig/1024 * running_time/time
5064 * the normalized util can be found using the specific capacity.
5066 * capacity = capacity_orig * curr_freq/max_freq
5068 * norm_util = running_time/time ~ util/capacity
5070 static unsigned long __cpu_norm_util(int cpu, unsigned long capacity, int delta)
5072 int util = __cpu_util(cpu, delta);
5074 if (util >= capacity)
5075 return SCHED_CAPACITY_SCALE;
5077 return (util << SCHED_CAPACITY_SHIFT)/capacity;
5080 static int calc_util_delta(struct energy_env *eenv, int cpu)
5082 if (cpu == eenv->src_cpu)
5083 return -eenv->util_delta;
5084 if (cpu == eenv->dst_cpu)
5085 return eenv->util_delta;
5090 unsigned long group_max_util(struct energy_env *eenv)
5093 unsigned long max_util = 0;
5095 for_each_cpu(i, sched_group_cpus(eenv->sg_cap)) {
5096 delta = calc_util_delta(eenv, i);
5097 max_util = max(max_util, __cpu_util(i, delta));
5104 * group_norm_util() returns the approximated group util relative to it's
5105 * current capacity (busy ratio) in the range [0..SCHED_LOAD_SCALE] for use in
5106 * energy calculations. Since task executions may or may not overlap in time in
5107 * the group the true normalized util is between max(cpu_norm_util(i)) and
5108 * sum(cpu_norm_util(i)) when iterating over all cpus in the group, i. The
5109 * latter is used as the estimate as it leads to a more pessimistic energy
5110 * estimate (more busy).
5113 long group_norm_util(struct energy_env *eenv, struct sched_group *sg)
5116 unsigned long util_sum = 0;
5117 unsigned long capacity = sg->sge->cap_states[eenv->cap_idx].cap;
5119 for_each_cpu(i, sched_group_cpus(sg)) {
5120 delta = calc_util_delta(eenv, i);
5121 util_sum += __cpu_norm_util(i, capacity, delta);
5124 if (util_sum > SCHED_CAPACITY_SCALE)
5125 return SCHED_CAPACITY_SCALE;
5129 static int find_new_capacity(struct energy_env *eenv,
5130 const struct sched_group_energy * const sge)
5133 unsigned long util = group_max_util(eenv);
5135 for (idx = 0; idx < sge->nr_cap_states; idx++) {
5136 if (sge->cap_states[idx].cap >= util)
5140 eenv->cap_idx = idx;
5145 static int group_idle_state(struct sched_group *sg)
5147 int i, state = INT_MAX;
5149 /* Find the shallowest idle state in the sched group. */
5150 for_each_cpu(i, sched_group_cpus(sg))
5151 state = min(state, idle_get_state_idx(cpu_rq(i)));
5153 /* Take non-cpuidle idling into account (active idle/arch_cpu_idle()) */
5160 * sched_group_energy(): Computes the absolute energy consumption of cpus
5161 * belonging to the sched_group including shared resources shared only by
5162 * members of the group. Iterates over all cpus in the hierarchy below the
5163 * sched_group starting from the bottom working it's way up before going to
5164 * the next cpu until all cpus are covered at all levels. The current
5165 * implementation is likely to gather the same util statistics multiple times.
5166 * This can probably be done in a faster but more complex way.
5167 * Note: sched_group_energy() may fail when racing with sched_domain updates.
5169 static int sched_group_energy(struct energy_env *eenv)
5171 struct sched_domain *sd;
5172 int cpu, total_energy = 0;
5173 struct cpumask visit_cpus;
5174 struct sched_group *sg;
5176 WARN_ON(!eenv->sg_top->sge);
5178 cpumask_copy(&visit_cpus, sched_group_cpus(eenv->sg_top));
5180 while (!cpumask_empty(&visit_cpus)) {
5181 struct sched_group *sg_shared_cap = NULL;
5183 cpu = cpumask_first(&visit_cpus);
5186 * Is the group utilization affected by cpus outside this
5189 sd = rcu_dereference(per_cpu(sd_scs, cpu));
5193 * We most probably raced with hotplug; returning a
5194 * wrong energy estimation is better than entering an
5200 sg_shared_cap = sd->parent->groups;
5202 for_each_domain(cpu, sd) {
5205 /* Has this sched_domain already been visited? */
5206 if (sd->child && group_first_cpu(sg) != cpu)
5210 unsigned long group_util;
5211 int sg_busy_energy, sg_idle_energy;
5212 int cap_idx, idle_idx;
5214 if (sg_shared_cap && sg_shared_cap->group_weight >= sg->group_weight)
5215 eenv->sg_cap = sg_shared_cap;
5219 cap_idx = find_new_capacity(eenv, sg->sge);
5221 if (sg->group_weight == 1) {
5222 /* Remove capacity of src CPU (before task move) */
5223 if (eenv->util_delta == 0 &&
5224 cpumask_test_cpu(eenv->src_cpu, sched_group_cpus(sg))) {
5225 eenv->cap.before = sg->sge->cap_states[cap_idx].cap;
5226 eenv->cap.delta -= eenv->cap.before;
5228 /* Add capacity of dst CPU (after task move) */
5229 if (eenv->util_delta != 0 &&
5230 cpumask_test_cpu(eenv->dst_cpu, sched_group_cpus(sg))) {
5231 eenv->cap.after = sg->sge->cap_states[cap_idx].cap;
5232 eenv->cap.delta += eenv->cap.after;
5236 idle_idx = group_idle_state(sg);
5237 group_util = group_norm_util(eenv, sg);
5238 sg_busy_energy = (group_util * sg->sge->cap_states[cap_idx].power)
5239 >> SCHED_CAPACITY_SHIFT;
5240 sg_idle_energy = ((SCHED_LOAD_SCALE-group_util)
5241 * sg->sge->idle_states[idle_idx].power)
5242 >> SCHED_CAPACITY_SHIFT;
5244 total_energy += sg_busy_energy + sg_idle_energy;
5247 cpumask_xor(&visit_cpus, &visit_cpus, sched_group_cpus(sg));
5249 if (cpumask_equal(sched_group_cpus(sg), sched_group_cpus(eenv->sg_top)))
5252 } while (sg = sg->next, sg != sd->groups);
5255 cpumask_clear_cpu(cpu, &visit_cpus);
5259 eenv->energy = total_energy;
5263 static inline bool cpu_in_sg(struct sched_group *sg, int cpu)
5265 return cpu != -1 && cpumask_test_cpu(cpu, sched_group_cpus(sg));
5269 * energy_diff(): Estimate the energy impact of changing the utilization
5270 * distribution. eenv specifies the change: utilisation amount, source, and
5271 * destination cpu. Source or destination cpu may be -1 in which case the
5272 * utilization is removed from or added to the system (e.g. task wake-up). If
5273 * both are specified, the utilization is migrated.
5275 static inline int __energy_diff(struct energy_env *eenv)
5277 struct sched_domain *sd;
5278 struct sched_group *sg;
5279 int sd_cpu = -1, energy_before = 0, energy_after = 0;
5282 struct energy_env eenv_before = {
5284 .src_cpu = eenv->src_cpu,
5285 .dst_cpu = eenv->dst_cpu,
5286 .nrg = { 0, 0, 0, 0},
5290 if (eenv->src_cpu == eenv->dst_cpu)
5293 sd_cpu = (eenv->src_cpu != -1) ? eenv->src_cpu : eenv->dst_cpu;
5294 sd = rcu_dereference(per_cpu(sd_ea, sd_cpu));
5297 return 0; /* Error */
5302 if (cpu_in_sg(sg, eenv->src_cpu) || cpu_in_sg(sg, eenv->dst_cpu)) {
5303 eenv_before.sg_top = eenv->sg_top = sg;
5305 if (sched_group_energy(&eenv_before))
5306 return 0; /* Invalid result abort */
5307 energy_before += eenv_before.energy;
5309 /* Keep track of SRC cpu (before) capacity */
5310 eenv->cap.before = eenv_before.cap.before;
5311 eenv->cap.delta = eenv_before.cap.delta;
5313 if (sched_group_energy(eenv))
5314 return 0; /* Invalid result abort */
5315 energy_after += eenv->energy;
5317 } while (sg = sg->next, sg != sd->groups);
5319 eenv->nrg.before = energy_before;
5320 eenv->nrg.after = energy_after;
5321 eenv->nrg.diff = eenv->nrg.after - eenv->nrg.before;
5324 trace_sched_energy_diff(eenv->task,
5325 eenv->src_cpu, eenv->dst_cpu, eenv->util_delta,
5326 eenv->nrg.before, eenv->nrg.after, eenv->nrg.diff,
5327 eenv->cap.before, eenv->cap.after, eenv->cap.delta,
5328 eenv->nrg.delta, eenv->payoff);
5331 * Dead-zone margin preventing too many migrations.
5334 margin = eenv->nrg.before >> 6; /* ~1.56% */
5336 diff = eenv->nrg.after - eenv->nrg.before;
5338 eenv->nrg.diff = (abs(diff) < margin) ? 0 : eenv->nrg.diff;
5340 return eenv->nrg.diff;
5343 #ifdef CONFIG_SCHED_TUNE
5345 struct target_nrg schedtune_target_nrg;
5348 * System energy normalization
5349 * Returns the normalized value, in the range [0..SCHED_LOAD_SCALE],
5350 * corresponding to the specified energy variation.
5353 normalize_energy(int energy_diff)
5356 #ifdef CONFIG_SCHED_DEBUG
5359 /* Check for boundaries */
5360 max_delta = schedtune_target_nrg.max_power;
5361 max_delta -= schedtune_target_nrg.min_power;
5362 WARN_ON(abs(energy_diff) >= max_delta);
5365 /* Do scaling using positive numbers to increase the range */
5366 normalized_nrg = (energy_diff < 0) ? -energy_diff : energy_diff;
5368 /* Scale by energy magnitude */
5369 normalized_nrg <<= SCHED_LOAD_SHIFT;
5371 /* Normalize on max energy for target platform */
5372 normalized_nrg = reciprocal_divide(
5373 normalized_nrg, schedtune_target_nrg.rdiv);
5375 return (energy_diff < 0) ? -normalized_nrg : normalized_nrg;
5379 energy_diff(struct energy_env *eenv)
5381 int boost = schedtune_task_boost(eenv->task);
5384 /* Conpute "absolute" energy diff */
5385 __energy_diff(eenv);
5387 /* Return energy diff when boost margin is 0 */
5389 return eenv->nrg.diff;
5391 /* Compute normalized energy diff */
5392 nrg_delta = normalize_energy(eenv->nrg.diff);
5393 eenv->nrg.delta = nrg_delta;
5395 eenv->payoff = schedtune_accept_deltas(
5401 * When SchedTune is enabled, the energy_diff() function will return
5402 * the computed energy payoff value. Since the energy_diff() return
5403 * value is expected to be negative by its callers, this evaluation
5404 * function return a negative value each time the evaluation return a
5405 * positive payoff, which is the condition for the acceptance of
5406 * a scheduling decision
5408 return -eenv->payoff;
5410 #else /* CONFIG_SCHED_TUNE */
5411 #define energy_diff(eenv) __energy_diff(eenv)
5415 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5416 * A waker of many should wake a different task than the one last awakened
5417 * at a frequency roughly N times higher than one of its wakees. In order
5418 * to determine whether we should let the load spread vs consolodating to
5419 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
5420 * partner, and a factor of lls_size higher frequency in the other. With
5421 * both conditions met, we can be relatively sure that the relationship is
5422 * non-monogamous, with partner count exceeding socket size. Waker/wakee
5423 * being client/server, worker/dispatcher, interrupt source or whatever is
5424 * irrelevant, spread criteria is apparent partner count exceeds socket size.
5426 static int wake_wide(struct task_struct *p)
5428 unsigned int master = current->wakee_flips;
5429 unsigned int slave = p->wakee_flips;
5430 int factor = this_cpu_read(sd_llc_size);
5433 swap(master, slave);
5434 if (slave < factor || master < slave * factor)
5439 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5440 int prev_cpu, int sync)
5442 s64 this_load, load;
5443 s64 this_eff_load, prev_eff_load;
5445 struct task_group *tg;
5446 unsigned long weight;
5450 this_cpu = smp_processor_id();
5451 load = source_load(prev_cpu, idx);
5452 this_load = target_load(this_cpu, idx);
5455 * If sync wakeup then subtract the (maximum possible)
5456 * effect of the currently running task from the load
5457 * of the current CPU:
5460 tg = task_group(current);
5461 weight = current->se.avg.load_avg;
5463 this_load += effective_load(tg, this_cpu, -weight, -weight);
5464 load += effective_load(tg, prev_cpu, 0, -weight);
5468 weight = p->se.avg.load_avg;
5471 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5472 * due to the sync cause above having dropped this_load to 0, we'll
5473 * always have an imbalance, but there's really nothing you can do
5474 * about that, so that's good too.
5476 * Otherwise check if either cpus are near enough in load to allow this
5477 * task to be woken on this_cpu.
5479 this_eff_load = 100;
5480 this_eff_load *= capacity_of(prev_cpu);
5482 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5483 prev_eff_load *= capacity_of(this_cpu);
5485 if (this_load > 0) {
5486 this_eff_load *= this_load +
5487 effective_load(tg, this_cpu, weight, weight);
5489 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5492 balanced = this_eff_load <= prev_eff_load;
5494 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5499 schedstat_inc(sd, ttwu_move_affine);
5500 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5505 static inline unsigned long task_util(struct task_struct *p)
5507 #ifdef CONFIG_SCHED_WALT
5508 if (!walt_disabled && sysctl_sched_use_walt_task_util) {
5509 unsigned long demand = p->ravg.demand;
5510 return (demand << 10) / walt_ravg_window;
5513 return p->se.avg.util_avg;
5516 static inline unsigned long boosted_task_util(struct task_struct *task);
5518 static inline bool __task_fits(struct task_struct *p, int cpu, int util)
5520 unsigned long capacity = capacity_of(cpu);
5522 util += boosted_task_util(p);
5524 return (capacity * 1024) > (util * capacity_margin);
5527 static inline bool task_fits_max(struct task_struct *p, int cpu)
5529 unsigned long capacity = capacity_of(cpu);
5530 unsigned long max_capacity = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5532 if (capacity == max_capacity)
5535 if (capacity * capacity_margin > max_capacity * 1024)
5538 return __task_fits(p, cpu, 0);
5541 static bool cpu_overutilized(int cpu)
5543 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5546 #ifdef CONFIG_SCHED_TUNE
5549 schedtune_margin(unsigned long signal, long boost)
5551 long long margin = 0;
5554 * Signal proportional compensation (SPC)
5556 * The Boost (B) value is used to compute a Margin (M) which is
5557 * proportional to the complement of the original Signal (S):
5558 * M = B * (SCHED_LOAD_SCALE - S), if B is positive
5559 * M = B * S, if B is negative
5560 * The obtained M could be used by the caller to "boost" S.
5563 margin = SCHED_LOAD_SCALE - signal;
5566 margin = -signal * boost;
5568 * Fast integer division by constant:
5569 * Constant : (C) = 100
5570 * Precision : 0.1% (P) = 0.1
5571 * Reference : C * 100 / P (R) = 100000
5574 * Shift bits : ceil(log(R,2)) (S) = 17
5575 * Mult const : round(2^S/C) (M) = 1311
5588 schedtune_cpu_margin(unsigned long util, int cpu)
5590 int boost = schedtune_cpu_boost(cpu);
5595 return schedtune_margin(util, boost);
5599 schedtune_task_margin(struct task_struct *task)
5601 int boost = schedtune_task_boost(task);
5608 util = task_util(task);
5609 margin = schedtune_margin(util, boost);
5614 #else /* CONFIG_SCHED_TUNE */
5617 schedtune_cpu_margin(unsigned long util, int cpu)
5623 schedtune_task_margin(struct task_struct *task)
5628 #endif /* CONFIG_SCHED_TUNE */
5631 boosted_cpu_util(int cpu)
5633 unsigned long util = cpu_util(cpu);
5634 long margin = schedtune_cpu_margin(util, cpu);
5636 trace_sched_boost_cpu(cpu, util, margin);
5638 return util + margin;
5641 static inline unsigned long
5642 boosted_task_util(struct task_struct *task)
5644 unsigned long util = task_util(task);
5645 long margin = schedtune_task_margin(task);
5647 trace_sched_boost_task(task, util, margin);
5649 return util + margin;
5652 static int cpu_util_wake(int cpu, struct task_struct *p);
5654 static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
5656 return capacity_orig_of(cpu) - cpu_util_wake(cpu, p);
5660 * find_idlest_group finds and returns the least busy CPU group within the
5663 static struct sched_group *
5664 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5665 int this_cpu, int sd_flag)
5667 struct sched_group *idlest = NULL, *group = sd->groups;
5668 struct sched_group *most_spare_sg = NULL;
5669 unsigned long min_load = ULONG_MAX, this_load = 0;
5670 unsigned long most_spare = 0, this_spare = 0;
5671 int load_idx = sd->forkexec_idx;
5672 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5674 if (sd_flag & SD_BALANCE_WAKE)
5675 load_idx = sd->wake_idx;
5678 unsigned long load, avg_load, spare_cap, max_spare_cap;
5682 /* Skip over this group if it has no CPUs allowed */
5683 if (!cpumask_intersects(sched_group_cpus(group),
5684 tsk_cpus_allowed(p)))
5687 local_group = cpumask_test_cpu(this_cpu,
5688 sched_group_cpus(group));
5691 * Tally up the load of all CPUs in the group and find
5692 * the group containing the CPU with most spare capacity.
5697 for_each_cpu(i, sched_group_cpus(group)) {
5698 /* Bias balancing toward cpus of our domain */
5700 load = source_load(i, load_idx);
5702 load = target_load(i, load_idx);
5706 spare_cap = capacity_spare_wake(i, p);
5708 if (spare_cap > max_spare_cap)
5709 max_spare_cap = spare_cap;
5712 /* Adjust by relative CPU capacity of the group */
5713 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5716 this_load = avg_load;
5717 this_spare = max_spare_cap;
5719 if (avg_load < min_load) {
5720 min_load = avg_load;
5724 if (most_spare < max_spare_cap) {
5725 most_spare = max_spare_cap;
5726 most_spare_sg = group;
5729 } while (group = group->next, group != sd->groups);
5732 * The cross-over point between using spare capacity or least load
5733 * is too conservative for high utilization tasks on partially
5734 * utilized systems if we require spare_capacity > task_util(p),
5735 * so we allow for some task stuffing by using
5736 * spare_capacity > task_util(p)/2.
5738 if (this_spare > task_util(p) / 2 &&
5739 imbalance*this_spare > 100*most_spare)
5741 else if (most_spare > task_util(p) / 2)
5742 return most_spare_sg;
5744 if (!idlest || 100*this_load < imbalance*min_load)
5750 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5753 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5755 unsigned long load, min_load = ULONG_MAX;
5756 unsigned int min_exit_latency = UINT_MAX;
5757 u64 latest_idle_timestamp = 0;
5758 int least_loaded_cpu = this_cpu;
5759 int shallowest_idle_cpu = -1;
5762 /* Check if we have any choice: */
5763 if (group->group_weight == 1)
5764 return cpumask_first(sched_group_cpus(group));
5766 /* Traverse only the allowed CPUs */
5767 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5769 struct rq *rq = cpu_rq(i);
5770 struct cpuidle_state *idle = idle_get_state(rq);
5771 if (idle && idle->exit_latency < min_exit_latency) {
5773 * We give priority to a CPU whose idle state
5774 * has the smallest exit latency irrespective
5775 * of any idle timestamp.
5777 min_exit_latency = idle->exit_latency;
5778 latest_idle_timestamp = rq->idle_stamp;
5779 shallowest_idle_cpu = i;
5780 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5781 rq->idle_stamp > latest_idle_timestamp) {
5783 * If equal or no active idle state, then
5784 * the most recently idled CPU might have
5787 latest_idle_timestamp = rq->idle_stamp;
5788 shallowest_idle_cpu = i;
5790 } else if (shallowest_idle_cpu == -1) {
5791 load = weighted_cpuload(i);
5792 if (load < min_load || (load == min_load && i == this_cpu)) {
5794 least_loaded_cpu = i;
5799 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5803 * Try and locate an idle CPU in the sched_domain.
5805 static int select_idle_sibling(struct task_struct *p, int prev, int target)
5807 struct sched_domain *sd;
5808 struct sched_group *sg;
5809 int best_idle_cpu = -1;
5810 int best_idle_cstate = INT_MAX;
5811 unsigned long best_idle_capacity = ULONG_MAX;
5813 if (!sysctl_sched_cstate_aware) {
5814 if (idle_cpu(target))
5818 * If the prevous cpu is cache affine and idle, don't be stupid.
5820 if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev))
5825 * Otherwise, iterate the domains and find an elegible idle cpu.
5827 sd = rcu_dereference(per_cpu(sd_llc, target));
5828 for_each_lower_domain(sd) {
5832 if (!cpumask_intersects(sched_group_cpus(sg),
5833 tsk_cpus_allowed(p)))
5836 if (sysctl_sched_cstate_aware) {
5837 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg)) {
5838 int idle_idx = idle_get_state_idx(cpu_rq(i));
5839 unsigned long new_usage = boosted_task_util(p);
5840 unsigned long capacity_orig = capacity_orig_of(i);
5842 if (new_usage > capacity_orig || !idle_cpu(i))
5845 if (i == target && new_usage <= capacity_curr_of(target))
5848 if (idle_idx < best_idle_cstate &&
5849 capacity_orig <= best_idle_capacity) {
5851 best_idle_cstate = idle_idx;
5852 best_idle_capacity = capacity_orig;
5856 for_each_cpu(i, sched_group_cpus(sg)) {
5857 if (i == target || !idle_cpu(i))
5861 target = cpumask_first_and(sched_group_cpus(sg),
5862 tsk_cpus_allowed(p));
5867 } while (sg != sd->groups);
5870 if (best_idle_cpu >= 0)
5871 target = best_idle_cpu;
5877 static int start_cpu(bool boosted)
5879 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
5881 RCU_LOCKDEP_WARN(rcu_read_lock_sched_held(),
5882 "sched RCU must be held");
5884 return boosted ? rd->max_cap_orig_cpu : rd->min_cap_orig_cpu;
5887 static inline int find_best_target(struct task_struct *p, bool boosted, bool prefer_idle)
5889 int target_cpu = -1;
5890 unsigned long target_util = prefer_idle ? ULONG_MAX : 0;
5891 unsigned long backup_capacity = ULONG_MAX;
5892 int best_idle_cpu = -1;
5893 int best_idle_cstate = INT_MAX;
5894 int backup_cpu = -1;
5895 unsigned long min_util = boosted_task_util(p);
5896 struct sched_domain *sd;
5897 struct sched_group *sg;
5898 int cpu = start_cpu(boosted);
5903 sd = rcu_dereference(per_cpu(sd_ea, cpu));
5913 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg)) {
5914 unsigned long cur_capacity, new_util;
5920 * p's blocked utilization is still accounted for on prev_cpu
5921 * so prev_cpu will receive a negative bias due to the double
5922 * accounting. However, the blocked utilization may be zero.
5924 new_util = cpu_util(i) + task_util(p);
5927 * Ensure minimum capacity to grant the required boost.
5928 * The target CPU can be already at a capacity level higher
5929 * than the one required to boost the task.
5931 new_util = max(min_util, new_util);
5933 if (new_util > capacity_orig_of(i))
5936 #ifdef CONFIG_SCHED_WALT
5937 if (walt_cpu_high_irqload(i))
5942 * Unconditionally favoring tasks that prefer idle cpus to
5945 if (idle_cpu(i) && prefer_idle)
5948 cur_capacity = capacity_curr_of(i);
5950 if (new_util < cur_capacity) {
5951 if (cpu_rq(i)->nr_running) {
5953 * Find a target cpu with the lowest/highest
5954 * utilization if prefer_idle/!prefer_idle.
5956 if ((prefer_idle && target_util > new_util) ||
5957 (!prefer_idle && target_util < new_util)) {
5958 target_util = new_util;
5961 } else if (!prefer_idle) {
5962 int idle_idx = idle_get_state_idx(cpu_rq(i));
5964 if (best_idle_cpu < 0 ||
5965 (sysctl_sched_cstate_aware &&
5966 best_idle_cstate > idle_idx)) {
5967 best_idle_cstate = idle_idx;
5971 } else if (backup_capacity > cur_capacity) {
5972 /* Find a backup cpu with least capacity. */
5973 backup_capacity = cur_capacity;
5977 } while (sg = sg->next, sg != sd->groups);
5980 target_cpu = best_idle_cpu >= 0 ? best_idle_cpu : backup_cpu;
5986 * cpu_util_wake: Compute cpu utilization with any contributions from
5987 * the waking task p removed.
5989 static int cpu_util_wake(int cpu, struct task_struct *p)
5991 unsigned long util, capacity;
5993 /* Task has no contribution or is new */
5994 if (cpu != task_cpu(p) || !p->se.avg.last_update_time)
5995 return cpu_util(cpu);
5997 capacity = capacity_orig_of(cpu);
5998 util = max_t(long, cpu_rq(cpu)->cfs.avg.util_avg - task_util(p), 0);
6000 return (util >= capacity) ? capacity : util;
6004 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6005 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6007 * In that case WAKE_AFFINE doesn't make sense and we'll let
6008 * BALANCE_WAKE sort things out.
6010 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6012 long min_cap, max_cap;
6014 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6015 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity.val;
6017 /* Minimum capacity is close to max, no need to abort wake_affine */
6018 if (max_cap - min_cap < max_cap >> 3)
6021 /* Bring task utilization in sync with prev_cpu */
6022 sync_entity_load_avg(&p->se);
6024 return min_cap * 1024 < task_util(p) * capacity_margin;
6027 static int select_energy_cpu_brute(struct task_struct *p, int prev_cpu, int sync)
6029 struct sched_domain *sd;
6030 int target_cpu = prev_cpu, tmp_target;
6031 bool boosted, prefer_idle;
6033 if (sysctl_sched_sync_hint_enable && sync) {
6034 int cpu = smp_processor_id();
6036 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
6041 #ifdef CONFIG_CGROUP_SCHEDTUNE
6042 boosted = schedtune_task_boost(p) > 0;
6043 prefer_idle = schedtune_prefer_idle(p) > 0;
6045 boosted = get_sysctl_sched_cfs_boost() > 0;
6049 sd = rcu_dereference(per_cpu(sd_ea, prev_cpu));
6050 /* Find a cpu with sufficient capacity */
6051 tmp_target = find_best_target(p, boosted, prefer_idle);
6055 if (tmp_target >= 0) {
6056 target_cpu = tmp_target;
6057 if ((boosted || prefer_idle) && idle_cpu(target_cpu))
6061 if (target_cpu != prev_cpu) {
6062 struct energy_env eenv = {
6063 .util_delta = task_util(p),
6064 .src_cpu = prev_cpu,
6065 .dst_cpu = target_cpu,
6069 /* Not enough spare capacity on previous cpu */
6070 if (cpu_overutilized(prev_cpu))
6073 if (energy_diff(&eenv) >= 0)
6074 target_cpu = prev_cpu;
6083 * select_task_rq_fair: Select target runqueue for the waking task in domains
6084 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6085 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6087 * Balances load by selecting the idlest cpu in the idlest group, or under
6088 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
6090 * Returns the target cpu number.
6092 * preempt must be disabled.
6095 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6097 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
6098 int cpu = smp_processor_id();
6099 int new_cpu = prev_cpu;
6100 int want_affine = 0;
6101 int sync = wake_flags & WF_SYNC;
6103 if (sd_flag & SD_BALANCE_WAKE)
6104 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
6105 && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
6107 if (energy_aware() && !(cpu_rq(prev_cpu)->rd->overutilized))
6108 return select_energy_cpu_brute(p, prev_cpu, sync);
6111 for_each_domain(cpu, tmp) {
6112 if (!(tmp->flags & SD_LOAD_BALANCE))
6116 * If both cpu and prev_cpu are part of this domain,
6117 * cpu is a valid SD_WAKE_AFFINE target.
6119 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6120 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6125 if (tmp->flags & sd_flag)
6127 else if (!want_affine)
6132 sd = NULL; /* Prefer wake_affine over balance flags */
6133 if (cpu != prev_cpu && wake_affine(affine_sd, p, prev_cpu, sync))
6138 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
6139 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6142 struct sched_group *group;
6145 if (!(sd->flags & sd_flag)) {
6150 group = find_idlest_group(sd, p, cpu, sd_flag);
6156 new_cpu = find_idlest_cpu(group, p, cpu);
6157 if (new_cpu == -1 || new_cpu == cpu) {
6158 /* Now try balancing at a lower domain level of cpu */
6163 /* Now try balancing at a lower domain level of new_cpu */
6165 weight = sd->span_weight;
6167 for_each_domain(cpu, tmp) {
6168 if (weight <= tmp->span_weight)
6170 if (tmp->flags & sd_flag)
6173 /* while loop will break here if sd == NULL */
6181 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
6182 * cfs_rq_of(p) references at time of call are still valid and identify the
6183 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
6184 * other assumptions, including the state of rq->lock, should be made.
6186 static void migrate_task_rq_fair(struct task_struct *p)
6189 * We are supposed to update the task to "current" time, then its up to date
6190 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
6191 * what current time is, so simply throw away the out-of-date time. This
6192 * will result in the wakee task is less decayed, but giving the wakee more
6193 * load sounds not bad.
6195 remove_entity_load_avg(&p->se);
6197 /* Tell new CPU we are migrated */
6198 p->se.avg.last_update_time = 0;
6200 /* We have migrated, no longer consider this task hot */
6201 p->se.exec_start = 0;
6204 static void task_dead_fair(struct task_struct *p)
6206 remove_entity_load_avg(&p->se);
6209 #define task_fits_max(p, cpu) true
6210 #endif /* CONFIG_SMP */
6212 static unsigned long
6213 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
6215 unsigned long gran = sysctl_sched_wakeup_granularity;
6218 * Since its curr running now, convert the gran from real-time
6219 * to virtual-time in his units.
6221 * By using 'se' instead of 'curr' we penalize light tasks, so
6222 * they get preempted easier. That is, if 'se' < 'curr' then
6223 * the resulting gran will be larger, therefore penalizing the
6224 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6225 * be smaller, again penalizing the lighter task.
6227 * This is especially important for buddies when the leftmost
6228 * task is higher priority than the buddy.
6230 return calc_delta_fair(gran, se);
6234 * Should 'se' preempt 'curr'.
6248 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6250 s64 gran, vdiff = curr->vruntime - se->vruntime;
6255 gran = wakeup_gran(curr, se);
6262 static void set_last_buddy(struct sched_entity *se)
6264 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6267 for_each_sched_entity(se)
6268 cfs_rq_of(se)->last = se;
6271 static void set_next_buddy(struct sched_entity *se)
6273 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6276 for_each_sched_entity(se)
6277 cfs_rq_of(se)->next = se;
6280 static void set_skip_buddy(struct sched_entity *se)
6282 for_each_sched_entity(se)
6283 cfs_rq_of(se)->skip = se;
6287 * Preempt the current task with a newly woken task if needed:
6289 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6291 struct task_struct *curr = rq->curr;
6292 struct sched_entity *se = &curr->se, *pse = &p->se;
6293 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6294 int scale = cfs_rq->nr_running >= sched_nr_latency;
6295 int next_buddy_marked = 0;
6297 if (unlikely(se == pse))
6301 * This is possible from callers such as attach_tasks(), in which we
6302 * unconditionally check_prempt_curr() after an enqueue (which may have
6303 * lead to a throttle). This both saves work and prevents false
6304 * next-buddy nomination below.
6306 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6309 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6310 set_next_buddy(pse);
6311 next_buddy_marked = 1;
6315 * We can come here with TIF_NEED_RESCHED already set from new task
6318 * Note: this also catches the edge-case of curr being in a throttled
6319 * group (e.g. via set_curr_task), since update_curr() (in the
6320 * enqueue of curr) will have resulted in resched being set. This
6321 * prevents us from potentially nominating it as a false LAST_BUDDY
6324 if (test_tsk_need_resched(curr))
6327 /* Idle tasks are by definition preempted by non-idle tasks. */
6328 if (unlikely(curr->policy == SCHED_IDLE) &&
6329 likely(p->policy != SCHED_IDLE))
6333 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6334 * is driven by the tick):
6336 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6339 find_matching_se(&se, &pse);
6340 update_curr(cfs_rq_of(se));
6342 if (wakeup_preempt_entity(se, pse) == 1) {
6344 * Bias pick_next to pick the sched entity that is
6345 * triggering this preemption.
6347 if (!next_buddy_marked)
6348 set_next_buddy(pse);
6357 * Only set the backward buddy when the current task is still
6358 * on the rq. This can happen when a wakeup gets interleaved
6359 * with schedule on the ->pre_schedule() or idle_balance()
6360 * point, either of which can * drop the rq lock.
6362 * Also, during early boot the idle thread is in the fair class,
6363 * for obvious reasons its a bad idea to schedule back to it.
6365 if (unlikely(!se->on_rq || curr == rq->idle))
6368 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6372 static struct task_struct *
6373 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
6375 struct cfs_rq *cfs_rq = &rq->cfs;
6376 struct sched_entity *se;
6377 struct task_struct *p;
6381 #ifdef CONFIG_FAIR_GROUP_SCHED
6382 if (!cfs_rq->nr_running)
6385 if (prev->sched_class != &fair_sched_class)
6389 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6390 * likely that a next task is from the same cgroup as the current.
6392 * Therefore attempt to avoid putting and setting the entire cgroup
6393 * hierarchy, only change the part that actually changes.
6397 struct sched_entity *curr = cfs_rq->curr;
6400 * Since we got here without doing put_prev_entity() we also
6401 * have to consider cfs_rq->curr. If it is still a runnable
6402 * entity, update_curr() will update its vruntime, otherwise
6403 * forget we've ever seen it.
6407 update_curr(cfs_rq);
6412 * This call to check_cfs_rq_runtime() will do the
6413 * throttle and dequeue its entity in the parent(s).
6414 * Therefore the 'simple' nr_running test will indeed
6417 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
6421 se = pick_next_entity(cfs_rq, curr);
6422 cfs_rq = group_cfs_rq(se);
6428 * Since we haven't yet done put_prev_entity and if the selected task
6429 * is a different task than we started out with, try and touch the
6430 * least amount of cfs_rqs.
6433 struct sched_entity *pse = &prev->se;
6435 while (!(cfs_rq = is_same_group(se, pse))) {
6436 int se_depth = se->depth;
6437 int pse_depth = pse->depth;
6439 if (se_depth <= pse_depth) {
6440 put_prev_entity(cfs_rq_of(pse), pse);
6441 pse = parent_entity(pse);
6443 if (se_depth >= pse_depth) {
6444 set_next_entity(cfs_rq_of(se), se);
6445 se = parent_entity(se);
6449 put_prev_entity(cfs_rq, pse);
6450 set_next_entity(cfs_rq, se);
6453 if (hrtick_enabled(rq))
6454 hrtick_start_fair(rq, p);
6456 rq->misfit_task = !task_fits_max(p, rq->cpu);
6463 if (!cfs_rq->nr_running)
6466 put_prev_task(rq, prev);
6469 se = pick_next_entity(cfs_rq, NULL);
6470 set_next_entity(cfs_rq, se);
6471 cfs_rq = group_cfs_rq(se);
6476 if (hrtick_enabled(rq))
6477 hrtick_start_fair(rq, p);
6479 rq->misfit_task = !task_fits_max(p, rq->cpu);
6484 rq->misfit_task = 0;
6486 * This is OK, because current is on_cpu, which avoids it being picked
6487 * for load-balance and preemption/IRQs are still disabled avoiding
6488 * further scheduler activity on it and we're being very careful to
6489 * re-start the picking loop.
6491 lockdep_unpin_lock(&rq->lock);
6492 new_tasks = idle_balance(rq);
6493 lockdep_pin_lock(&rq->lock);
6495 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6496 * possible for any higher priority task to appear. In that case we
6497 * must re-start the pick_next_entity() loop.
6509 * Account for a descheduled task:
6511 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6513 struct sched_entity *se = &prev->se;
6514 struct cfs_rq *cfs_rq;
6516 for_each_sched_entity(se) {
6517 cfs_rq = cfs_rq_of(se);
6518 put_prev_entity(cfs_rq, se);
6523 * sched_yield() is very simple
6525 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6527 static void yield_task_fair(struct rq *rq)
6529 struct task_struct *curr = rq->curr;
6530 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6531 struct sched_entity *se = &curr->se;
6534 * Are we the only task in the tree?
6536 if (unlikely(rq->nr_running == 1))
6539 clear_buddies(cfs_rq, se);
6541 if (curr->policy != SCHED_BATCH) {
6542 update_rq_clock(rq);
6544 * Update run-time statistics of the 'current'.
6546 update_curr(cfs_rq);
6548 * Tell update_rq_clock() that we've just updated,
6549 * so we don't do microscopic update in schedule()
6550 * and double the fastpath cost.
6552 rq_clock_skip_update(rq, true);
6558 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6560 struct sched_entity *se = &p->se;
6562 /* throttled hierarchies are not runnable */
6563 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6566 /* Tell the scheduler that we'd really like pse to run next. */
6569 yield_task_fair(rq);
6575 /**************************************************
6576 * Fair scheduling class load-balancing methods.
6580 * The purpose of load-balancing is to achieve the same basic fairness the
6581 * per-cpu scheduler provides, namely provide a proportional amount of compute
6582 * time to each task. This is expressed in the following equation:
6584 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6586 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6587 * W_i,0 is defined as:
6589 * W_i,0 = \Sum_j w_i,j (2)
6591 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6592 * is derived from the nice value as per prio_to_weight[].
6594 * The weight average is an exponential decay average of the instantaneous
6597 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6599 * C_i is the compute capacity of cpu i, typically it is the
6600 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6601 * can also include other factors [XXX].
6603 * To achieve this balance we define a measure of imbalance which follows
6604 * directly from (1):
6606 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6608 * We them move tasks around to minimize the imbalance. In the continuous
6609 * function space it is obvious this converges, in the discrete case we get
6610 * a few fun cases generally called infeasible weight scenarios.
6613 * - infeasible weights;
6614 * - local vs global optima in the discrete case. ]
6619 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6620 * for all i,j solution, we create a tree of cpus that follows the hardware
6621 * topology where each level pairs two lower groups (or better). This results
6622 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6623 * tree to only the first of the previous level and we decrease the frequency
6624 * of load-balance at each level inv. proportional to the number of cpus in
6630 * \Sum { --- * --- * 2^i } = O(n) (5)
6632 * `- size of each group
6633 * | | `- number of cpus doing load-balance
6635 * `- sum over all levels
6637 * Coupled with a limit on how many tasks we can migrate every balance pass,
6638 * this makes (5) the runtime complexity of the balancer.
6640 * An important property here is that each CPU is still (indirectly) connected
6641 * to every other cpu in at most O(log n) steps:
6643 * The adjacency matrix of the resulting graph is given by:
6646 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6649 * And you'll find that:
6651 * A^(log_2 n)_i,j != 0 for all i,j (7)
6653 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6654 * The task movement gives a factor of O(m), giving a convergence complexity
6657 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6662 * In order to avoid CPUs going idle while there's still work to do, new idle
6663 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6664 * tree itself instead of relying on other CPUs to bring it work.
6666 * This adds some complexity to both (5) and (8) but it reduces the total idle
6674 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6677 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6682 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6684 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6686 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6689 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6690 * rewrite all of this once again.]
6693 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6695 enum fbq_type { regular, remote, all };
6704 #define LBF_ALL_PINNED 0x01
6705 #define LBF_NEED_BREAK 0x02
6706 #define LBF_DST_PINNED 0x04
6707 #define LBF_SOME_PINNED 0x08
6710 struct sched_domain *sd;
6718 struct cpumask *dst_grpmask;
6720 enum cpu_idle_type idle;
6722 unsigned int src_grp_nr_running;
6723 /* The set of CPUs under consideration for load-balancing */
6724 struct cpumask *cpus;
6729 unsigned int loop_break;
6730 unsigned int loop_max;
6732 enum fbq_type fbq_type;
6733 enum group_type busiest_group_type;
6734 struct list_head tasks;
6738 * Is this task likely cache-hot:
6740 static int task_hot(struct task_struct *p, struct lb_env *env)
6744 lockdep_assert_held(&env->src_rq->lock);
6746 if (p->sched_class != &fair_sched_class)
6749 if (unlikely(p->policy == SCHED_IDLE))
6753 * Buddy candidates are cache hot:
6755 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6756 (&p->se == cfs_rq_of(&p->se)->next ||
6757 &p->se == cfs_rq_of(&p->se)->last))
6760 if (sysctl_sched_migration_cost == -1)
6762 if (sysctl_sched_migration_cost == 0)
6765 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6767 return delta < (s64)sysctl_sched_migration_cost;
6770 #ifdef CONFIG_NUMA_BALANCING
6772 * Returns 1, if task migration degrades locality
6773 * Returns 0, if task migration improves locality i.e migration preferred.
6774 * Returns -1, if task migration is not affected by locality.
6776 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6778 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6779 unsigned long src_faults, dst_faults;
6780 int src_nid, dst_nid;
6782 if (!static_branch_likely(&sched_numa_balancing))
6785 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6788 src_nid = cpu_to_node(env->src_cpu);
6789 dst_nid = cpu_to_node(env->dst_cpu);
6791 if (src_nid == dst_nid)
6794 /* Migrating away from the preferred node is always bad. */
6795 if (src_nid == p->numa_preferred_nid) {
6796 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6802 /* Encourage migration to the preferred node. */
6803 if (dst_nid == p->numa_preferred_nid)
6807 src_faults = group_faults(p, src_nid);
6808 dst_faults = group_faults(p, dst_nid);
6810 src_faults = task_faults(p, src_nid);
6811 dst_faults = task_faults(p, dst_nid);
6814 return dst_faults < src_faults;
6818 static inline int migrate_degrades_locality(struct task_struct *p,
6826 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6829 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6833 lockdep_assert_held(&env->src_rq->lock);
6836 * We do not migrate tasks that are:
6837 * 1) throttled_lb_pair, or
6838 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6839 * 3) running (obviously), or
6840 * 4) are cache-hot on their current CPU.
6842 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6845 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6848 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6850 env->flags |= LBF_SOME_PINNED;
6853 * Remember if this task can be migrated to any other cpu in
6854 * our sched_group. We may want to revisit it if we couldn't
6855 * meet load balance goals by pulling other tasks on src_cpu.
6857 * Also avoid computing new_dst_cpu if we have already computed
6858 * one in current iteration.
6860 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6863 /* Prevent to re-select dst_cpu via env's cpus */
6864 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6865 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6866 env->flags |= LBF_DST_PINNED;
6867 env->new_dst_cpu = cpu;
6875 /* Record that we found atleast one task that could run on dst_cpu */
6876 env->flags &= ~LBF_ALL_PINNED;
6878 if (task_running(env->src_rq, p)) {
6879 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6884 * Aggressive migration if:
6885 * 1) destination numa is preferred
6886 * 2) task is cache cold, or
6887 * 3) too many balance attempts have failed.
6889 tsk_cache_hot = migrate_degrades_locality(p, env);
6890 if (tsk_cache_hot == -1)
6891 tsk_cache_hot = task_hot(p, env);
6893 if (tsk_cache_hot <= 0 ||
6894 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6895 if (tsk_cache_hot == 1) {
6896 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6897 schedstat_inc(p, se.statistics.nr_forced_migrations);
6902 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6907 * detach_task() -- detach the task for the migration specified in env
6909 static void detach_task(struct task_struct *p, struct lb_env *env)
6911 lockdep_assert_held(&env->src_rq->lock);
6913 deactivate_task(env->src_rq, p, 0);
6914 p->on_rq = TASK_ON_RQ_MIGRATING;
6915 double_lock_balance(env->src_rq, env->dst_rq);
6916 set_task_cpu(p, env->dst_cpu);
6917 double_unlock_balance(env->src_rq, env->dst_rq);
6921 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6922 * part of active balancing operations within "domain".
6924 * Returns a task if successful and NULL otherwise.
6926 static struct task_struct *detach_one_task(struct lb_env *env)
6928 struct task_struct *p, *n;
6930 lockdep_assert_held(&env->src_rq->lock);
6932 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6933 if (!can_migrate_task(p, env))
6936 detach_task(p, env);
6939 * Right now, this is only the second place where
6940 * lb_gained[env->idle] is updated (other is detach_tasks)
6941 * so we can safely collect stats here rather than
6942 * inside detach_tasks().
6944 schedstat_inc(env->sd, lb_gained[env->idle]);
6950 static const unsigned int sched_nr_migrate_break = 32;
6953 * detach_tasks() -- tries to detach up to imbalance weighted load from
6954 * busiest_rq, as part of a balancing operation within domain "sd".
6956 * Returns number of detached tasks if successful and 0 otherwise.
6958 static int detach_tasks(struct lb_env *env)
6960 struct list_head *tasks = &env->src_rq->cfs_tasks;
6961 struct task_struct *p;
6965 lockdep_assert_held(&env->src_rq->lock);
6967 if (env->imbalance <= 0)
6970 while (!list_empty(tasks)) {
6972 * We don't want to steal all, otherwise we may be treated likewise,
6973 * which could at worst lead to a livelock crash.
6975 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6978 p = list_first_entry(tasks, struct task_struct, se.group_node);
6981 /* We've more or less seen every task there is, call it quits */
6982 if (env->loop > env->loop_max)
6985 /* take a breather every nr_migrate tasks */
6986 if (env->loop > env->loop_break) {
6987 env->loop_break += sched_nr_migrate_break;
6988 env->flags |= LBF_NEED_BREAK;
6992 if (!can_migrate_task(p, env))
6995 load = task_h_load(p);
6997 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7000 if ((load / 2) > env->imbalance)
7003 detach_task(p, env);
7004 list_add(&p->se.group_node, &env->tasks);
7007 env->imbalance -= load;
7009 #ifdef CONFIG_PREEMPT
7011 * NEWIDLE balancing is a source of latency, so preemptible
7012 * kernels will stop after the first task is detached to minimize
7013 * the critical section.
7015 if (env->idle == CPU_NEWLY_IDLE)
7020 * We only want to steal up to the prescribed amount of
7023 if (env->imbalance <= 0)
7028 list_move_tail(&p->se.group_node, tasks);
7032 * Right now, this is one of only two places we collect this stat
7033 * so we can safely collect detach_one_task() stats here rather
7034 * than inside detach_one_task().
7036 schedstat_add(env->sd, lb_gained[env->idle], detached);
7042 * attach_task() -- attach the task detached by detach_task() to its new rq.
7044 static void attach_task(struct rq *rq, struct task_struct *p)
7046 lockdep_assert_held(&rq->lock);
7048 BUG_ON(task_rq(p) != rq);
7049 p->on_rq = TASK_ON_RQ_QUEUED;
7050 activate_task(rq, p, 0);
7051 check_preempt_curr(rq, p, 0);
7055 * attach_one_task() -- attaches the task returned from detach_one_task() to
7058 static void attach_one_task(struct rq *rq, struct task_struct *p)
7060 raw_spin_lock(&rq->lock);
7063 * We want to potentially raise target_cpu's OPP.
7065 update_capacity_of(cpu_of(rq));
7066 raw_spin_unlock(&rq->lock);
7070 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7073 static void attach_tasks(struct lb_env *env)
7075 struct list_head *tasks = &env->tasks;
7076 struct task_struct *p;
7078 raw_spin_lock(&env->dst_rq->lock);
7080 while (!list_empty(tasks)) {
7081 p = list_first_entry(tasks, struct task_struct, se.group_node);
7082 list_del_init(&p->se.group_node);
7084 attach_task(env->dst_rq, p);
7088 * We want to potentially raise env.dst_cpu's OPP.
7090 update_capacity_of(env->dst_cpu);
7092 raw_spin_unlock(&env->dst_rq->lock);
7095 #ifdef CONFIG_FAIR_GROUP_SCHED
7096 static void update_blocked_averages(int cpu)
7098 struct rq *rq = cpu_rq(cpu);
7099 struct cfs_rq *cfs_rq;
7100 unsigned long flags;
7102 raw_spin_lock_irqsave(&rq->lock, flags);
7103 update_rq_clock(rq);
7106 * Iterates the task_group tree in a bottom up fashion, see
7107 * list_add_leaf_cfs_rq() for details.
7109 for_each_leaf_cfs_rq(rq, cfs_rq) {
7110 /* throttled entities do not contribute to load */
7111 if (throttled_hierarchy(cfs_rq))
7114 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq,
7116 update_tg_load_avg(cfs_rq, 0);
7118 raw_spin_unlock_irqrestore(&rq->lock, flags);
7122 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7123 * This needs to be done in a top-down fashion because the load of a child
7124 * group is a fraction of its parents load.
7126 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7128 struct rq *rq = rq_of(cfs_rq);
7129 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7130 unsigned long now = jiffies;
7133 if (cfs_rq->last_h_load_update == now)
7136 cfs_rq->h_load_next = NULL;
7137 for_each_sched_entity(se) {
7138 cfs_rq = cfs_rq_of(se);
7139 cfs_rq->h_load_next = se;
7140 if (cfs_rq->last_h_load_update == now)
7145 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7146 cfs_rq->last_h_load_update = now;
7149 while ((se = cfs_rq->h_load_next) != NULL) {
7150 load = cfs_rq->h_load;
7151 load = div64_ul(load * se->avg.load_avg,
7152 cfs_rq_load_avg(cfs_rq) + 1);
7153 cfs_rq = group_cfs_rq(se);
7154 cfs_rq->h_load = load;
7155 cfs_rq->last_h_load_update = now;
7159 static unsigned long task_h_load(struct task_struct *p)
7161 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7163 update_cfs_rq_h_load(cfs_rq);
7164 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7165 cfs_rq_load_avg(cfs_rq) + 1);
7168 static inline void update_blocked_averages(int cpu)
7170 struct rq *rq = cpu_rq(cpu);
7171 struct cfs_rq *cfs_rq = &rq->cfs;
7172 unsigned long flags;
7174 raw_spin_lock_irqsave(&rq->lock, flags);
7175 update_rq_clock(rq);
7176 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true);
7177 raw_spin_unlock_irqrestore(&rq->lock, flags);
7180 static unsigned long task_h_load(struct task_struct *p)
7182 return p->se.avg.load_avg;
7186 /********** Helpers for find_busiest_group ************************/
7189 * sg_lb_stats - stats of a sched_group required for load_balancing
7191 struct sg_lb_stats {
7192 unsigned long avg_load; /*Avg load across the CPUs of the group */
7193 unsigned long group_load; /* Total load over the CPUs of the group */
7194 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7195 unsigned long load_per_task;
7196 unsigned long group_capacity;
7197 unsigned long group_util; /* Total utilization of the group */
7198 unsigned int sum_nr_running; /* Nr tasks running in the group */
7199 unsigned int idle_cpus;
7200 unsigned int group_weight;
7201 enum group_type group_type;
7202 int group_no_capacity;
7203 int group_misfit_task; /* A cpu has a task too big for its capacity */
7204 #ifdef CONFIG_NUMA_BALANCING
7205 unsigned int nr_numa_running;
7206 unsigned int nr_preferred_running;
7211 * sd_lb_stats - Structure to store the statistics of a sched_domain
7212 * during load balancing.
7214 struct sd_lb_stats {
7215 struct sched_group *busiest; /* Busiest group in this sd */
7216 struct sched_group *local; /* Local group in this sd */
7217 unsigned long total_load; /* Total load of all groups in sd */
7218 unsigned long total_capacity; /* Total capacity of all groups in sd */
7219 unsigned long avg_load; /* Average load across all groups in sd */
7221 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7222 struct sg_lb_stats local_stat; /* Statistics of the local group */
7225 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7228 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7229 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7230 * We must however clear busiest_stat::avg_load because
7231 * update_sd_pick_busiest() reads this before assignment.
7233 *sds = (struct sd_lb_stats){
7237 .total_capacity = 0UL,
7240 .sum_nr_running = 0,
7241 .group_type = group_other,
7247 * get_sd_load_idx - Obtain the load index for a given sched domain.
7248 * @sd: The sched_domain whose load_idx is to be obtained.
7249 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7251 * Return: The load index.
7253 static inline int get_sd_load_idx(struct sched_domain *sd,
7254 enum cpu_idle_type idle)
7260 load_idx = sd->busy_idx;
7263 case CPU_NEWLY_IDLE:
7264 load_idx = sd->newidle_idx;
7267 load_idx = sd->idle_idx;
7274 static unsigned long scale_rt_capacity(int cpu)
7276 struct rq *rq = cpu_rq(cpu);
7277 u64 total, used, age_stamp, avg;
7281 * Since we're reading these variables without serialization make sure
7282 * we read them once before doing sanity checks on them.
7284 age_stamp = READ_ONCE(rq->age_stamp);
7285 avg = READ_ONCE(rq->rt_avg);
7286 delta = __rq_clock_broken(rq) - age_stamp;
7288 if (unlikely(delta < 0))
7291 total = sched_avg_period() + delta;
7293 used = div_u64(avg, total);
7296 * deadline bandwidth is defined at system level so we must
7297 * weight this bandwidth with the max capacity of the system.
7298 * As a reminder, avg_bw is 20bits width and
7299 * scale_cpu_capacity is 10 bits width
7301 used += div_u64(rq->dl.avg_bw, arch_scale_cpu_capacity(NULL, cpu));
7303 if (likely(used < SCHED_CAPACITY_SCALE))
7304 return SCHED_CAPACITY_SCALE - used;
7309 void init_max_cpu_capacity(struct max_cpu_capacity *mcc)
7311 raw_spin_lock_init(&mcc->lock);
7316 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7318 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7319 struct sched_group *sdg = sd->groups;
7320 struct max_cpu_capacity *mcc;
7321 unsigned long max_capacity;
7323 unsigned long flags;
7325 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7327 mcc = &cpu_rq(cpu)->rd->max_cpu_capacity;
7329 raw_spin_lock_irqsave(&mcc->lock, flags);
7330 max_capacity = mcc->val;
7331 max_cap_cpu = mcc->cpu;
7333 if ((max_capacity > capacity && max_cap_cpu == cpu) ||
7334 (max_capacity < capacity)) {
7335 mcc->val = capacity;
7337 #ifdef CONFIG_SCHED_DEBUG
7338 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7339 printk_deferred(KERN_INFO "CPU%d: update max cpu_capacity %lu\n",
7344 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7346 skip_unlock: __attribute__ ((unused));
7347 capacity *= scale_rt_capacity(cpu);
7348 capacity >>= SCHED_CAPACITY_SHIFT;
7353 cpu_rq(cpu)->cpu_capacity = capacity;
7354 sdg->sgc->capacity = capacity;
7355 sdg->sgc->max_capacity = capacity;
7356 sdg->sgc->min_capacity = capacity;
7359 void update_group_capacity(struct sched_domain *sd, int cpu)
7361 struct sched_domain *child = sd->child;
7362 struct sched_group *group, *sdg = sd->groups;
7363 unsigned long capacity, max_capacity, min_capacity;
7364 unsigned long interval;
7366 interval = msecs_to_jiffies(sd->balance_interval);
7367 interval = clamp(interval, 1UL, max_load_balance_interval);
7368 sdg->sgc->next_update = jiffies + interval;
7371 update_cpu_capacity(sd, cpu);
7377 min_capacity = ULONG_MAX;
7379 if (child->flags & SD_OVERLAP) {
7381 * SD_OVERLAP domains cannot assume that child groups
7382 * span the current group.
7385 for_each_cpu(cpu, sched_group_cpus(sdg)) {
7386 struct sched_group_capacity *sgc;
7387 struct rq *rq = cpu_rq(cpu);
7390 * build_sched_domains() -> init_sched_groups_capacity()
7391 * gets here before we've attached the domains to the
7394 * Use capacity_of(), which is set irrespective of domains
7395 * in update_cpu_capacity().
7397 * This avoids capacity from being 0 and
7398 * causing divide-by-zero issues on boot.
7400 if (unlikely(!rq->sd)) {
7401 capacity += capacity_of(cpu);
7403 sgc = rq->sd->groups->sgc;
7404 capacity += sgc->capacity;
7407 max_capacity = max(capacity, max_capacity);
7408 min_capacity = min(capacity, min_capacity);
7412 * !SD_OVERLAP domains can assume that child groups
7413 * span the current group.
7416 group = child->groups;
7418 struct sched_group_capacity *sgc = group->sgc;
7420 capacity += sgc->capacity;
7421 max_capacity = max(sgc->max_capacity, max_capacity);
7422 min_capacity = min(sgc->min_capacity, min_capacity);
7423 group = group->next;
7424 } while (group != child->groups);
7427 sdg->sgc->capacity = capacity;
7428 sdg->sgc->max_capacity = max_capacity;
7429 sdg->sgc->min_capacity = min_capacity;
7433 * Check whether the capacity of the rq has been noticeably reduced by side
7434 * activity. The imbalance_pct is used for the threshold.
7435 * Return true is the capacity is reduced
7438 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7440 return ((rq->cpu_capacity * sd->imbalance_pct) <
7441 (rq->cpu_capacity_orig * 100));
7445 * Group imbalance indicates (and tries to solve) the problem where balancing
7446 * groups is inadequate due to tsk_cpus_allowed() constraints.
7448 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7449 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7452 * { 0 1 2 3 } { 4 5 6 7 }
7455 * If we were to balance group-wise we'd place two tasks in the first group and
7456 * two tasks in the second group. Clearly this is undesired as it will overload
7457 * cpu 3 and leave one of the cpus in the second group unused.
7459 * The current solution to this issue is detecting the skew in the first group
7460 * by noticing the lower domain failed to reach balance and had difficulty
7461 * moving tasks due to affinity constraints.
7463 * When this is so detected; this group becomes a candidate for busiest; see
7464 * update_sd_pick_busiest(). And calculate_imbalance() and
7465 * find_busiest_group() avoid some of the usual balance conditions to allow it
7466 * to create an effective group imbalance.
7468 * This is a somewhat tricky proposition since the next run might not find the
7469 * group imbalance and decide the groups need to be balanced again. A most
7470 * subtle and fragile situation.
7473 static inline int sg_imbalanced(struct sched_group *group)
7475 return group->sgc->imbalance;
7479 * group_has_capacity returns true if the group has spare capacity that could
7480 * be used by some tasks.
7481 * We consider that a group has spare capacity if the * number of task is
7482 * smaller than the number of CPUs or if the utilization is lower than the
7483 * available capacity for CFS tasks.
7484 * For the latter, we use a threshold to stabilize the state, to take into
7485 * account the variance of the tasks' load and to return true if the available
7486 * capacity in meaningful for the load balancer.
7487 * As an example, an available capacity of 1% can appear but it doesn't make
7488 * any benefit for the load balance.
7491 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7493 if (sgs->sum_nr_running < sgs->group_weight)
7496 if ((sgs->group_capacity * 100) >
7497 (sgs->group_util * env->sd->imbalance_pct))
7504 * group_is_overloaded returns true if the group has more tasks than it can
7506 * group_is_overloaded is not equals to !group_has_capacity because a group
7507 * with the exact right number of tasks, has no more spare capacity but is not
7508 * overloaded so both group_has_capacity and group_is_overloaded return
7512 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7514 if (sgs->sum_nr_running <= sgs->group_weight)
7517 if ((sgs->group_capacity * 100) <
7518 (sgs->group_util * env->sd->imbalance_pct))
7526 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7527 * per-cpu capacity than sched_group ref.
7530 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7532 return sg->sgc->max_capacity + capacity_margin - SCHED_LOAD_SCALE <
7533 ref->sgc->max_capacity;
7537 group_type group_classify(struct sched_group *group,
7538 struct sg_lb_stats *sgs)
7540 if (sgs->group_no_capacity)
7541 return group_overloaded;
7543 if (sg_imbalanced(group))
7544 return group_imbalanced;
7546 if (sgs->group_misfit_task)
7547 return group_misfit_task;
7553 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7554 * @env: The load balancing environment.
7555 * @group: sched_group whose statistics are to be updated.
7556 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7557 * @local_group: Does group contain this_cpu.
7558 * @sgs: variable to hold the statistics for this group.
7559 * @overload: Indicate more than one runnable task for any CPU.
7560 * @overutilized: Indicate overutilization for any CPU.
7562 static inline void update_sg_lb_stats(struct lb_env *env,
7563 struct sched_group *group, int load_idx,
7564 int local_group, struct sg_lb_stats *sgs,
7565 bool *overload, bool *overutilized)
7570 memset(sgs, 0, sizeof(*sgs));
7572 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7573 struct rq *rq = cpu_rq(i);
7575 /* Bias balancing toward cpus of our domain */
7577 load = target_load(i, load_idx);
7579 load = source_load(i, load_idx);
7581 sgs->group_load += load;
7582 sgs->group_util += cpu_util(i);
7583 sgs->sum_nr_running += rq->cfs.h_nr_running;
7585 nr_running = rq->nr_running;
7589 #ifdef CONFIG_NUMA_BALANCING
7590 sgs->nr_numa_running += rq->nr_numa_running;
7591 sgs->nr_preferred_running += rq->nr_preferred_running;
7593 sgs->sum_weighted_load += weighted_cpuload(i);
7595 * No need to call idle_cpu() if nr_running is not 0
7597 if (!nr_running && idle_cpu(i))
7600 if (cpu_overutilized(i)) {
7601 *overutilized = true;
7602 if (!sgs->group_misfit_task && rq->misfit_task)
7603 sgs->group_misfit_task = capacity_of(i);
7607 /* Adjust by relative CPU capacity of the group */
7608 sgs->group_capacity = group->sgc->capacity;
7609 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7611 if (sgs->sum_nr_running)
7612 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7614 sgs->group_weight = group->group_weight;
7616 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7617 sgs->group_type = group_classify(group, sgs);
7621 * update_sd_pick_busiest - return 1 on busiest group
7622 * @env: The load balancing environment.
7623 * @sds: sched_domain statistics
7624 * @sg: sched_group candidate to be checked for being the busiest
7625 * @sgs: sched_group statistics
7627 * Determine if @sg is a busier group than the previously selected
7630 * Return: %true if @sg is a busier group than the previously selected
7631 * busiest group. %false otherwise.
7633 static bool update_sd_pick_busiest(struct lb_env *env,
7634 struct sd_lb_stats *sds,
7635 struct sched_group *sg,
7636 struct sg_lb_stats *sgs)
7638 struct sg_lb_stats *busiest = &sds->busiest_stat;
7640 if (sgs->group_type > busiest->group_type)
7643 if (sgs->group_type < busiest->group_type)
7647 * Candidate sg doesn't face any serious load-balance problems
7648 * so don't pick it if the local sg is already filled up.
7650 if (sgs->group_type == group_other &&
7651 !group_has_capacity(env, &sds->local_stat))
7654 if (sgs->avg_load <= busiest->avg_load)
7657 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
7661 * Candidate sg has no more than one task per CPU and
7662 * has higher per-CPU capacity. Migrating tasks to less
7663 * capable CPUs may harm throughput. Maximize throughput,
7664 * power/energy consequences are not considered.
7666 if (sgs->sum_nr_running <= sgs->group_weight &&
7667 group_smaller_cpu_capacity(sds->local, sg))
7671 /* This is the busiest node in its class. */
7672 if (!(env->sd->flags & SD_ASYM_PACKING))
7676 * ASYM_PACKING needs to move all the work to the lowest
7677 * numbered CPUs in the group, therefore mark all groups
7678 * higher than ourself as busy.
7680 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
7684 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
7691 #ifdef CONFIG_NUMA_BALANCING
7692 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7694 if (sgs->sum_nr_running > sgs->nr_numa_running)
7696 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7701 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7703 if (rq->nr_running > rq->nr_numa_running)
7705 if (rq->nr_running > rq->nr_preferred_running)
7710 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7715 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7719 #endif /* CONFIG_NUMA_BALANCING */
7722 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7723 * @env: The load balancing environment.
7724 * @sds: variable to hold the statistics for this sched_domain.
7726 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7728 struct sched_domain *child = env->sd->child;
7729 struct sched_group *sg = env->sd->groups;
7730 struct sg_lb_stats tmp_sgs;
7731 int load_idx, prefer_sibling = 0;
7732 bool overload = false, overutilized = false;
7734 if (child && child->flags & SD_PREFER_SIBLING)
7737 load_idx = get_sd_load_idx(env->sd, env->idle);
7740 struct sg_lb_stats *sgs = &tmp_sgs;
7743 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
7746 sgs = &sds->local_stat;
7748 if (env->idle != CPU_NEWLY_IDLE ||
7749 time_after_eq(jiffies, sg->sgc->next_update))
7750 update_group_capacity(env->sd, env->dst_cpu);
7753 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7754 &overload, &overutilized);
7760 * In case the child domain prefers tasks go to siblings
7761 * first, lower the sg capacity so that we'll try
7762 * and move all the excess tasks away. We lower the capacity
7763 * of a group only if the local group has the capacity to fit
7764 * these excess tasks. The extra check prevents the case where
7765 * you always pull from the heaviest group when it is already
7766 * under-utilized (possible with a large weight task outweighs
7767 * the tasks on the system).
7769 if (prefer_sibling && sds->local &&
7770 group_has_capacity(env, &sds->local_stat) &&
7771 (sgs->sum_nr_running > 1)) {
7772 sgs->group_no_capacity = 1;
7773 sgs->group_type = group_classify(sg, sgs);
7777 * Ignore task groups with misfit tasks if local group has no
7778 * capacity or if per-cpu capacity isn't higher.
7780 if (sgs->group_type == group_misfit_task &&
7781 (!group_has_capacity(env, &sds->local_stat) ||
7782 !group_smaller_cpu_capacity(sg, sds->local)))
7783 sgs->group_type = group_other;
7785 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7787 sds->busiest_stat = *sgs;
7791 /* Now, start updating sd_lb_stats */
7792 sds->total_load += sgs->group_load;
7793 sds->total_capacity += sgs->group_capacity;
7796 } while (sg != env->sd->groups);
7798 if (env->sd->flags & SD_NUMA)
7799 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7801 env->src_grp_nr_running = sds->busiest_stat.sum_nr_running;
7803 if (!env->sd->parent) {
7804 /* update overload indicator if we are at root domain */
7805 if (env->dst_rq->rd->overload != overload)
7806 env->dst_rq->rd->overload = overload;
7808 /* Update over-utilization (tipping point, U >= 0) indicator */
7809 if (env->dst_rq->rd->overutilized != overutilized) {
7810 env->dst_rq->rd->overutilized = overutilized;
7811 trace_sched_overutilized(overutilized);
7814 if (!env->dst_rq->rd->overutilized && overutilized) {
7815 env->dst_rq->rd->overutilized = true;
7816 trace_sched_overutilized(true);
7823 * check_asym_packing - Check to see if the group is packed into the
7826 * This is primarily intended to used at the sibling level. Some
7827 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7828 * case of POWER7, it can move to lower SMT modes only when higher
7829 * threads are idle. When in lower SMT modes, the threads will
7830 * perform better since they share less core resources. Hence when we
7831 * have idle threads, we want them to be the higher ones.
7833 * This packing function is run on idle threads. It checks to see if
7834 * the busiest CPU in this domain (core in the P7 case) has a higher
7835 * CPU number than the packing function is being run on. Here we are
7836 * assuming lower CPU number will be equivalent to lower a SMT thread
7839 * Return: 1 when packing is required and a task should be moved to
7840 * this CPU. The amount of the imbalance is returned in *imbalance.
7842 * @env: The load balancing environment.
7843 * @sds: Statistics of the sched_domain which is to be packed
7845 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7849 if (!(env->sd->flags & SD_ASYM_PACKING))
7855 busiest_cpu = group_first_cpu(sds->busiest);
7856 if (env->dst_cpu > busiest_cpu)
7859 env->imbalance = DIV_ROUND_CLOSEST(
7860 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7861 SCHED_CAPACITY_SCALE);
7867 * fix_small_imbalance - Calculate the minor imbalance that exists
7868 * amongst the groups of a sched_domain, during
7870 * @env: The load balancing environment.
7871 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7874 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7876 unsigned long tmp, capa_now = 0, capa_move = 0;
7877 unsigned int imbn = 2;
7878 unsigned long scaled_busy_load_per_task;
7879 struct sg_lb_stats *local, *busiest;
7881 local = &sds->local_stat;
7882 busiest = &sds->busiest_stat;
7884 if (!local->sum_nr_running)
7885 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7886 else if (busiest->load_per_task > local->load_per_task)
7889 scaled_busy_load_per_task =
7890 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7891 busiest->group_capacity;
7893 if (busiest->avg_load + scaled_busy_load_per_task >=
7894 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7895 env->imbalance = busiest->load_per_task;
7900 * OK, we don't have enough imbalance to justify moving tasks,
7901 * however we may be able to increase total CPU capacity used by
7905 capa_now += busiest->group_capacity *
7906 min(busiest->load_per_task, busiest->avg_load);
7907 capa_now += local->group_capacity *
7908 min(local->load_per_task, local->avg_load);
7909 capa_now /= SCHED_CAPACITY_SCALE;
7911 /* Amount of load we'd subtract */
7912 if (busiest->avg_load > scaled_busy_load_per_task) {
7913 capa_move += busiest->group_capacity *
7914 min(busiest->load_per_task,
7915 busiest->avg_load - scaled_busy_load_per_task);
7918 /* Amount of load we'd add */
7919 if (busiest->avg_load * busiest->group_capacity <
7920 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7921 tmp = (busiest->avg_load * busiest->group_capacity) /
7922 local->group_capacity;
7924 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7925 local->group_capacity;
7927 capa_move += local->group_capacity *
7928 min(local->load_per_task, local->avg_load + tmp);
7929 capa_move /= SCHED_CAPACITY_SCALE;
7931 /* Move if we gain throughput */
7932 if (capa_move > capa_now)
7933 env->imbalance = busiest->load_per_task;
7937 * calculate_imbalance - Calculate the amount of imbalance present within the
7938 * groups of a given sched_domain during load balance.
7939 * @env: load balance environment
7940 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7942 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7944 unsigned long max_pull, load_above_capacity = ~0UL;
7945 struct sg_lb_stats *local, *busiest;
7947 local = &sds->local_stat;
7948 busiest = &sds->busiest_stat;
7950 if (busiest->group_type == group_imbalanced) {
7952 * In the group_imb case we cannot rely on group-wide averages
7953 * to ensure cpu-load equilibrium, look at wider averages. XXX
7955 busiest->load_per_task =
7956 min(busiest->load_per_task, sds->avg_load);
7960 * In the presence of smp nice balancing, certain scenarios can have
7961 * max load less than avg load(as we skip the groups at or below
7962 * its cpu_capacity, while calculating max_load..)
7964 if (busiest->avg_load <= sds->avg_load ||
7965 local->avg_load >= sds->avg_load) {
7966 /* Misfitting tasks should be migrated in any case */
7967 if (busiest->group_type == group_misfit_task) {
7968 env->imbalance = busiest->group_misfit_task;
7973 * Busiest group is overloaded, local is not, use the spare
7974 * cycles to maximize throughput
7976 if (busiest->group_type == group_overloaded &&
7977 local->group_type <= group_misfit_task) {
7978 env->imbalance = busiest->load_per_task;
7983 return fix_small_imbalance(env, sds);
7987 * If there aren't any idle cpus, avoid creating some.
7989 if (busiest->group_type == group_overloaded &&
7990 local->group_type == group_overloaded) {
7991 load_above_capacity = busiest->sum_nr_running *
7993 if (load_above_capacity > busiest->group_capacity)
7994 load_above_capacity -= busiest->group_capacity;
7996 load_above_capacity = ~0UL;
8000 * We're trying to get all the cpus to the average_load, so we don't
8001 * want to push ourselves above the average load, nor do we wish to
8002 * reduce the max loaded cpu below the average load. At the same time,
8003 * we also don't want to reduce the group load below the group capacity
8004 * (so that we can implement power-savings policies etc). Thus we look
8005 * for the minimum possible imbalance.
8007 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8009 /* How much load to actually move to equalise the imbalance */
8010 env->imbalance = min(
8011 max_pull * busiest->group_capacity,
8012 (sds->avg_load - local->avg_load) * local->group_capacity
8013 ) / SCHED_CAPACITY_SCALE;
8015 /* Boost imbalance to allow misfit task to be balanced. */
8016 if (busiest->group_type == group_misfit_task)
8017 env->imbalance = max_t(long, env->imbalance,
8018 busiest->group_misfit_task);
8021 * if *imbalance is less than the average load per runnable task
8022 * there is no guarantee that any tasks will be moved so we'll have
8023 * a think about bumping its value to force at least one task to be
8026 if (env->imbalance < busiest->load_per_task)
8027 return fix_small_imbalance(env, sds);
8030 /******* find_busiest_group() helpers end here *********************/
8033 * find_busiest_group - Returns the busiest group within the sched_domain
8034 * if there is an imbalance. If there isn't an imbalance, and
8035 * the user has opted for power-savings, it returns a group whose
8036 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
8037 * such a group exists.
8039 * Also calculates the amount of weighted load which should be moved
8040 * to restore balance.
8042 * @env: The load balancing environment.
8044 * Return: - The busiest group if imbalance exists.
8045 * - If no imbalance and user has opted for power-savings balance,
8046 * return the least loaded group whose CPUs can be
8047 * put to idle by rebalancing its tasks onto our group.
8049 static struct sched_group *find_busiest_group(struct lb_env *env)
8051 struct sg_lb_stats *local, *busiest;
8052 struct sd_lb_stats sds;
8054 init_sd_lb_stats(&sds);
8057 * Compute the various statistics relavent for load balancing at
8060 update_sd_lb_stats(env, &sds);
8062 if (energy_aware() && !env->dst_rq->rd->overutilized)
8065 local = &sds.local_stat;
8066 busiest = &sds.busiest_stat;
8068 /* ASYM feature bypasses nice load balance check */
8069 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
8070 check_asym_packing(env, &sds))
8073 /* There is no busy sibling group to pull tasks from */
8074 if (!sds.busiest || busiest->sum_nr_running == 0)
8077 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8078 / sds.total_capacity;
8081 * If the busiest group is imbalanced the below checks don't
8082 * work because they assume all things are equal, which typically
8083 * isn't true due to cpus_allowed constraints and the like.
8085 if (busiest->group_type == group_imbalanced)
8088 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
8089 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
8090 busiest->group_no_capacity)
8093 /* Misfitting tasks should be dealt with regardless of the avg load */
8094 if (busiest->group_type == group_misfit_task) {
8099 * If the local group is busier than the selected busiest group
8100 * don't try and pull any tasks.
8102 if (local->avg_load >= busiest->avg_load)
8106 * Don't pull any tasks if this group is already above the domain
8109 if (local->avg_load >= sds.avg_load)
8112 if (env->idle == CPU_IDLE) {
8114 * This cpu is idle. If the busiest group is not overloaded
8115 * and there is no imbalance between this and busiest group
8116 * wrt idle cpus, it is balanced. The imbalance becomes
8117 * significant if the diff is greater than 1 otherwise we
8118 * might end up to just move the imbalance on another group
8120 if ((busiest->group_type != group_overloaded) &&
8121 (local->idle_cpus <= (busiest->idle_cpus + 1)) &&
8122 !group_smaller_cpu_capacity(sds.busiest, sds.local))
8126 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8127 * imbalance_pct to be conservative.
8129 if (100 * busiest->avg_load <=
8130 env->sd->imbalance_pct * local->avg_load)
8135 env->busiest_group_type = busiest->group_type;
8136 /* Looks like there is an imbalance. Compute it */
8137 calculate_imbalance(env, &sds);
8146 * find_busiest_queue - find the busiest runqueue among the cpus in group.
8148 static struct rq *find_busiest_queue(struct lb_env *env,
8149 struct sched_group *group)
8151 struct rq *busiest = NULL, *rq;
8152 unsigned long busiest_load = 0, busiest_capacity = 1;
8155 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
8156 unsigned long capacity, wl;
8160 rt = fbq_classify_rq(rq);
8163 * We classify groups/runqueues into three groups:
8164 * - regular: there are !numa tasks
8165 * - remote: there are numa tasks that run on the 'wrong' node
8166 * - all: there is no distinction
8168 * In order to avoid migrating ideally placed numa tasks,
8169 * ignore those when there's better options.
8171 * If we ignore the actual busiest queue to migrate another
8172 * task, the next balance pass can still reduce the busiest
8173 * queue by moving tasks around inside the node.
8175 * If we cannot move enough load due to this classification
8176 * the next pass will adjust the group classification and
8177 * allow migration of more tasks.
8179 * Both cases only affect the total convergence complexity.
8181 if (rt > env->fbq_type)
8184 capacity = capacity_of(i);
8186 wl = weighted_cpuload(i);
8189 * When comparing with imbalance, use weighted_cpuload()
8190 * which is not scaled with the cpu capacity.
8193 if (rq->nr_running == 1 && wl > env->imbalance &&
8194 !check_cpu_capacity(rq, env->sd) &&
8195 env->busiest_group_type != group_misfit_task)
8199 * For the load comparisons with the other cpu's, consider
8200 * the weighted_cpuload() scaled with the cpu capacity, so
8201 * that the load can be moved away from the cpu that is
8202 * potentially running at a lower capacity.
8204 * Thus we're looking for max(wl_i / capacity_i), crosswise
8205 * multiplication to rid ourselves of the division works out
8206 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8207 * our previous maximum.
8209 if (wl * busiest_capacity > busiest_load * capacity) {
8211 busiest_capacity = capacity;
8220 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8221 * so long as it is large enough.
8223 #define MAX_PINNED_INTERVAL 512
8225 /* Working cpumask for load_balance and load_balance_newidle. */
8226 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
8228 static int need_active_balance(struct lb_env *env)
8230 struct sched_domain *sd = env->sd;
8232 if (env->idle == CPU_NEWLY_IDLE) {
8235 * ASYM_PACKING needs to force migrate tasks from busy but
8236 * higher numbered CPUs in order to pack all tasks in the
8237 * lowest numbered CPUs.
8239 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
8244 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8245 * It's worth migrating the task if the src_cpu's capacity is reduced
8246 * because of other sched_class or IRQs if more capacity stays
8247 * available on dst_cpu.
8249 if ((env->idle != CPU_NOT_IDLE) &&
8250 (env->src_rq->cfs.h_nr_running == 1)) {
8251 if ((check_cpu_capacity(env->src_rq, sd)) &&
8252 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8256 if ((capacity_of(env->src_cpu) < capacity_of(env->dst_cpu)) &&
8257 env->src_rq->cfs.h_nr_running == 1 &&
8258 cpu_overutilized(env->src_cpu) &&
8259 !cpu_overutilized(env->dst_cpu)) {
8263 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8266 static int active_load_balance_cpu_stop(void *data);
8268 static int should_we_balance(struct lb_env *env)
8270 struct sched_group *sg = env->sd->groups;
8271 struct cpumask *sg_cpus, *sg_mask;
8272 int cpu, balance_cpu = -1;
8275 * In the newly idle case, we will allow all the cpu's
8276 * to do the newly idle load balance.
8278 if (env->idle == CPU_NEWLY_IDLE)
8281 sg_cpus = sched_group_cpus(sg);
8282 sg_mask = sched_group_mask(sg);
8283 /* Try to find first idle cpu */
8284 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
8285 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
8292 if (balance_cpu == -1)
8293 balance_cpu = group_balance_cpu(sg);
8296 * First idle cpu or the first cpu(busiest) in this sched group
8297 * is eligible for doing load balancing at this and above domains.
8299 return balance_cpu == env->dst_cpu;
8303 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8304 * tasks if there is an imbalance.
8306 static int load_balance(int this_cpu, struct rq *this_rq,
8307 struct sched_domain *sd, enum cpu_idle_type idle,
8308 int *continue_balancing)
8310 int ld_moved, cur_ld_moved, active_balance = 0;
8311 struct sched_domain *sd_parent = sd->parent;
8312 struct sched_group *group;
8314 unsigned long flags;
8315 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8317 struct lb_env env = {
8319 .dst_cpu = this_cpu,
8321 .dst_grpmask = sched_group_cpus(sd->groups),
8323 .loop_break = sched_nr_migrate_break,
8326 .tasks = LIST_HEAD_INIT(env.tasks),
8330 * For NEWLY_IDLE load_balancing, we don't need to consider
8331 * other cpus in our group
8333 if (idle == CPU_NEWLY_IDLE)
8334 env.dst_grpmask = NULL;
8336 cpumask_copy(cpus, cpu_active_mask);
8338 schedstat_inc(sd, lb_count[idle]);
8341 if (!should_we_balance(&env)) {
8342 *continue_balancing = 0;
8346 group = find_busiest_group(&env);
8348 schedstat_inc(sd, lb_nobusyg[idle]);
8352 busiest = find_busiest_queue(&env, group);
8354 schedstat_inc(sd, lb_nobusyq[idle]);
8358 BUG_ON(busiest == env.dst_rq);
8360 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
8362 env.src_cpu = busiest->cpu;
8363 env.src_rq = busiest;
8366 if (busiest->nr_running > 1) {
8368 * Attempt to move tasks. If find_busiest_group has found
8369 * an imbalance but busiest->nr_running <= 1, the group is
8370 * still unbalanced. ld_moved simply stays zero, so it is
8371 * correctly treated as an imbalance.
8373 env.flags |= LBF_ALL_PINNED;
8374 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8377 raw_spin_lock_irqsave(&busiest->lock, flags);
8380 * cur_ld_moved - load moved in current iteration
8381 * ld_moved - cumulative load moved across iterations
8383 cur_ld_moved = detach_tasks(&env);
8385 * We want to potentially lower env.src_cpu's OPP.
8388 update_capacity_of(env.src_cpu);
8391 * We've detached some tasks from busiest_rq. Every
8392 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8393 * unlock busiest->lock, and we are able to be sure
8394 * that nobody can manipulate the tasks in parallel.
8395 * See task_rq_lock() family for the details.
8398 raw_spin_unlock(&busiest->lock);
8402 ld_moved += cur_ld_moved;
8405 local_irq_restore(flags);
8407 if (env.flags & LBF_NEED_BREAK) {
8408 env.flags &= ~LBF_NEED_BREAK;
8413 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8414 * us and move them to an alternate dst_cpu in our sched_group
8415 * where they can run. The upper limit on how many times we
8416 * iterate on same src_cpu is dependent on number of cpus in our
8419 * This changes load balance semantics a bit on who can move
8420 * load to a given_cpu. In addition to the given_cpu itself
8421 * (or a ilb_cpu acting on its behalf where given_cpu is
8422 * nohz-idle), we now have balance_cpu in a position to move
8423 * load to given_cpu. In rare situations, this may cause
8424 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8425 * _independently_ and at _same_ time to move some load to
8426 * given_cpu) causing exceess load to be moved to given_cpu.
8427 * This however should not happen so much in practice and
8428 * moreover subsequent load balance cycles should correct the
8429 * excess load moved.
8431 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8433 /* Prevent to re-select dst_cpu via env's cpus */
8434 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8436 env.dst_rq = cpu_rq(env.new_dst_cpu);
8437 env.dst_cpu = env.new_dst_cpu;
8438 env.flags &= ~LBF_DST_PINNED;
8440 env.loop_break = sched_nr_migrate_break;
8443 * Go back to "more_balance" rather than "redo" since we
8444 * need to continue with same src_cpu.
8450 * We failed to reach balance because of affinity.
8453 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8455 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8456 *group_imbalance = 1;
8459 /* All tasks on this runqueue were pinned by CPU affinity */
8460 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8461 cpumask_clear_cpu(cpu_of(busiest), cpus);
8462 if (!cpumask_empty(cpus)) {
8464 env.loop_break = sched_nr_migrate_break;
8467 goto out_all_pinned;
8472 schedstat_inc(sd, lb_failed[idle]);
8474 * Increment the failure counter only on periodic balance.
8475 * We do not want newidle balance, which can be very
8476 * frequent, pollute the failure counter causing
8477 * excessive cache_hot migrations and active balances.
8479 if (idle != CPU_NEWLY_IDLE)
8480 if (env.src_grp_nr_running > 1)
8481 sd->nr_balance_failed++;
8483 if (need_active_balance(&env)) {
8484 raw_spin_lock_irqsave(&busiest->lock, flags);
8486 /* don't kick the active_load_balance_cpu_stop,
8487 * if the curr task on busiest cpu can't be
8490 if (!cpumask_test_cpu(this_cpu,
8491 tsk_cpus_allowed(busiest->curr))) {
8492 raw_spin_unlock_irqrestore(&busiest->lock,
8494 env.flags |= LBF_ALL_PINNED;
8495 goto out_one_pinned;
8499 * ->active_balance synchronizes accesses to
8500 * ->active_balance_work. Once set, it's cleared
8501 * only after active load balance is finished.
8503 if (!busiest->active_balance) {
8504 busiest->active_balance = 1;
8505 busiest->push_cpu = this_cpu;
8508 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8510 if (active_balance) {
8511 stop_one_cpu_nowait(cpu_of(busiest),
8512 active_load_balance_cpu_stop, busiest,
8513 &busiest->active_balance_work);
8517 * We've kicked active balancing, reset the failure
8520 sd->nr_balance_failed = sd->cache_nice_tries+1;
8523 sd->nr_balance_failed = 0;
8525 if (likely(!active_balance)) {
8526 /* We were unbalanced, so reset the balancing interval */
8527 sd->balance_interval = sd->min_interval;
8530 * If we've begun active balancing, start to back off. This
8531 * case may not be covered by the all_pinned logic if there
8532 * is only 1 task on the busy runqueue (because we don't call
8535 if (sd->balance_interval < sd->max_interval)
8536 sd->balance_interval *= 2;
8543 * We reach balance although we may have faced some affinity
8544 * constraints. Clear the imbalance flag if it was set.
8547 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8549 if (*group_imbalance)
8550 *group_imbalance = 0;
8555 * We reach balance because all tasks are pinned at this level so
8556 * we can't migrate them. Let the imbalance flag set so parent level
8557 * can try to migrate them.
8559 schedstat_inc(sd, lb_balanced[idle]);
8561 sd->nr_balance_failed = 0;
8564 /* tune up the balancing interval */
8565 if (((env.flags & LBF_ALL_PINNED) &&
8566 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8567 (sd->balance_interval < sd->max_interval))
8568 sd->balance_interval *= 2;
8575 static inline unsigned long
8576 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8578 unsigned long interval = sd->balance_interval;
8581 interval *= sd->busy_factor;
8583 /* scale ms to jiffies */
8584 interval = msecs_to_jiffies(interval);
8585 interval = clamp(interval, 1UL, max_load_balance_interval);
8591 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
8593 unsigned long interval, next;
8595 interval = get_sd_balance_interval(sd, cpu_busy);
8596 next = sd->last_balance + interval;
8598 if (time_after(*next_balance, next))
8599 *next_balance = next;
8603 * idle_balance is called by schedule() if this_cpu is about to become
8604 * idle. Attempts to pull tasks from other CPUs.
8606 static int idle_balance(struct rq *this_rq)
8608 unsigned long next_balance = jiffies + HZ;
8609 int this_cpu = this_rq->cpu;
8610 struct sched_domain *sd;
8611 int pulled_task = 0;
8613 long removed_util=0;
8615 idle_enter_fair(this_rq);
8618 * We must set idle_stamp _before_ calling idle_balance(), such that we
8619 * measure the duration of idle_balance() as idle time.
8621 this_rq->idle_stamp = rq_clock(this_rq);
8623 if (!energy_aware() &&
8624 (this_rq->avg_idle < sysctl_sched_migration_cost ||
8625 !this_rq->rd->overload)) {
8627 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8629 update_next_balance(sd, 0, &next_balance);
8635 raw_spin_unlock(&this_rq->lock);
8638 * If removed_util_avg is !0 we most probably migrated some task away
8639 * from this_cpu. In this case we might be willing to trigger an OPP
8640 * update, but we want to do so if we don't find anybody else to pull
8641 * here (we will trigger an OPP update with the pulled task's enqueue
8644 * Record removed_util before calling update_blocked_averages, and use
8645 * it below (before returning) to see if an OPP update is required.
8647 removed_util = atomic_long_read(&(this_rq->cfs).removed_util_avg);
8648 update_blocked_averages(this_cpu);
8650 for_each_domain(this_cpu, sd) {
8651 int continue_balancing = 1;
8652 u64 t0, domain_cost;
8654 if (!(sd->flags & SD_LOAD_BALANCE))
8657 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8658 update_next_balance(sd, 0, &next_balance);
8662 if (sd->flags & SD_BALANCE_NEWIDLE) {
8663 t0 = sched_clock_cpu(this_cpu);
8665 pulled_task = load_balance(this_cpu, this_rq,
8667 &continue_balancing);
8669 domain_cost = sched_clock_cpu(this_cpu) - t0;
8670 if (domain_cost > sd->max_newidle_lb_cost)
8671 sd->max_newidle_lb_cost = domain_cost;
8673 curr_cost += domain_cost;
8676 update_next_balance(sd, 0, &next_balance);
8679 * Stop searching for tasks to pull if there are
8680 * now runnable tasks on this rq.
8682 if (pulled_task || this_rq->nr_running > 0)
8687 raw_spin_lock(&this_rq->lock);
8689 if (curr_cost > this_rq->max_idle_balance_cost)
8690 this_rq->max_idle_balance_cost = curr_cost;
8693 * While browsing the domains, we released the rq lock, a task could
8694 * have been enqueued in the meantime. Since we're not going idle,
8695 * pretend we pulled a task.
8697 if (this_rq->cfs.h_nr_running && !pulled_task)
8701 /* Move the next balance forward */
8702 if (time_after(this_rq->next_balance, next_balance))
8703 this_rq->next_balance = next_balance;
8705 /* Is there a task of a high priority class? */
8706 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8710 idle_exit_fair(this_rq);
8711 this_rq->idle_stamp = 0;
8712 } else if (removed_util) {
8714 * No task pulled and someone has been migrated away.
8715 * Good case to trigger an OPP update.
8717 update_capacity_of(this_cpu);
8724 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8725 * running tasks off the busiest CPU onto idle CPUs. It requires at
8726 * least 1 task to be running on each physical CPU where possible, and
8727 * avoids physical / logical imbalances.
8729 static int active_load_balance_cpu_stop(void *data)
8731 struct rq *busiest_rq = data;
8732 int busiest_cpu = cpu_of(busiest_rq);
8733 int target_cpu = busiest_rq->push_cpu;
8734 struct rq *target_rq = cpu_rq(target_cpu);
8735 struct sched_domain *sd;
8736 struct task_struct *p = NULL;
8738 raw_spin_lock_irq(&busiest_rq->lock);
8740 /* make sure the requested cpu hasn't gone down in the meantime */
8741 if (unlikely(busiest_cpu != smp_processor_id() ||
8742 !busiest_rq->active_balance))
8745 /* Is there any task to move? */
8746 if (busiest_rq->nr_running <= 1)
8750 * This condition is "impossible", if it occurs
8751 * we need to fix it. Originally reported by
8752 * Bjorn Helgaas on a 128-cpu setup.
8754 BUG_ON(busiest_rq == target_rq);
8756 /* Search for an sd spanning us and the target CPU. */
8758 for_each_domain(target_cpu, sd) {
8759 if ((sd->flags & SD_LOAD_BALANCE) &&
8760 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8765 struct lb_env env = {
8767 .dst_cpu = target_cpu,
8768 .dst_rq = target_rq,
8769 .src_cpu = busiest_rq->cpu,
8770 .src_rq = busiest_rq,
8774 schedstat_inc(sd, alb_count);
8776 p = detach_one_task(&env);
8778 schedstat_inc(sd, alb_pushed);
8780 * We want to potentially lower env.src_cpu's OPP.
8782 update_capacity_of(env.src_cpu);
8785 schedstat_inc(sd, alb_failed);
8789 busiest_rq->active_balance = 0;
8790 raw_spin_unlock(&busiest_rq->lock);
8793 attach_one_task(target_rq, p);
8800 static inline int on_null_domain(struct rq *rq)
8802 return unlikely(!rcu_dereference_sched(rq->sd));
8805 #ifdef CONFIG_NO_HZ_COMMON
8807 * idle load balancing details
8808 * - When one of the busy CPUs notice that there may be an idle rebalancing
8809 * needed, they will kick the idle load balancer, which then does idle
8810 * load balancing for all the idle CPUs.
8813 cpumask_var_t idle_cpus_mask;
8815 unsigned long next_balance; /* in jiffy units */
8816 } nohz ____cacheline_aligned;
8818 static inline int find_new_ilb(void)
8820 int ilb = cpumask_first(nohz.idle_cpus_mask);
8822 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8829 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8830 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8831 * CPU (if there is one).
8833 static void nohz_balancer_kick(void)
8837 nohz.next_balance++;
8839 ilb_cpu = find_new_ilb();
8841 if (ilb_cpu >= nr_cpu_ids)
8844 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8847 * Use smp_send_reschedule() instead of resched_cpu().
8848 * This way we generate a sched IPI on the target cpu which
8849 * is idle. And the softirq performing nohz idle load balance
8850 * will be run before returning from the IPI.
8852 smp_send_reschedule(ilb_cpu);
8856 static inline void nohz_balance_exit_idle(int cpu)
8858 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8860 * Completely isolated CPUs don't ever set, so we must test.
8862 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8863 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8864 atomic_dec(&nohz.nr_cpus);
8866 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8870 static inline void set_cpu_sd_state_busy(void)
8872 struct sched_domain *sd;
8873 int cpu = smp_processor_id();
8876 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8878 if (!sd || !sd->nohz_idle)
8882 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
8887 void set_cpu_sd_state_idle(void)
8889 struct sched_domain *sd;
8890 int cpu = smp_processor_id();
8893 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8895 if (!sd || sd->nohz_idle)
8899 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
8905 * This routine will record that the cpu is going idle with tick stopped.
8906 * This info will be used in performing idle load balancing in the future.
8908 void nohz_balance_enter_idle(int cpu)
8911 * If this cpu is going down, then nothing needs to be done.
8913 if (!cpu_active(cpu))
8916 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8920 * If we're a completely isolated CPU, we don't play.
8922 if (on_null_domain(cpu_rq(cpu)))
8925 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8926 atomic_inc(&nohz.nr_cpus);
8927 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8930 static int sched_ilb_notifier(struct notifier_block *nfb,
8931 unsigned long action, void *hcpu)
8933 switch (action & ~CPU_TASKS_FROZEN) {
8935 nohz_balance_exit_idle(smp_processor_id());
8943 static DEFINE_SPINLOCK(balancing);
8946 * Scale the max load_balance interval with the number of CPUs in the system.
8947 * This trades load-balance latency on larger machines for less cross talk.
8949 void update_max_interval(void)
8951 max_load_balance_interval = HZ*num_online_cpus()/10;
8955 * It checks each scheduling domain to see if it is due to be balanced,
8956 * and initiates a balancing operation if so.
8958 * Balancing parameters are set up in init_sched_domains.
8960 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8962 int continue_balancing = 1;
8964 unsigned long interval;
8965 struct sched_domain *sd;
8966 /* Earliest time when we have to do rebalance again */
8967 unsigned long next_balance = jiffies + 60*HZ;
8968 int update_next_balance = 0;
8969 int need_serialize, need_decay = 0;
8972 update_blocked_averages(cpu);
8975 for_each_domain(cpu, sd) {
8977 * Decay the newidle max times here because this is a regular
8978 * visit to all the domains. Decay ~1% per second.
8980 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8981 sd->max_newidle_lb_cost =
8982 (sd->max_newidle_lb_cost * 253) / 256;
8983 sd->next_decay_max_lb_cost = jiffies + HZ;
8986 max_cost += sd->max_newidle_lb_cost;
8988 if (!(sd->flags & SD_LOAD_BALANCE))
8992 * Stop the load balance at this level. There is another
8993 * CPU in our sched group which is doing load balancing more
8996 if (!continue_balancing) {
9002 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9004 need_serialize = sd->flags & SD_SERIALIZE;
9005 if (need_serialize) {
9006 if (!spin_trylock(&balancing))
9010 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9011 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9013 * The LBF_DST_PINNED logic could have changed
9014 * env->dst_cpu, so we can't know our idle
9015 * state even if we migrated tasks. Update it.
9017 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9019 sd->last_balance = jiffies;
9020 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9023 spin_unlock(&balancing);
9025 if (time_after(next_balance, sd->last_balance + interval)) {
9026 next_balance = sd->last_balance + interval;
9027 update_next_balance = 1;
9032 * Ensure the rq-wide value also decays but keep it at a
9033 * reasonable floor to avoid funnies with rq->avg_idle.
9035 rq->max_idle_balance_cost =
9036 max((u64)sysctl_sched_migration_cost, max_cost);
9041 * next_balance will be updated only when there is a need.
9042 * When the cpu is attached to null domain for ex, it will not be
9045 if (likely(update_next_balance)) {
9046 rq->next_balance = next_balance;
9048 #ifdef CONFIG_NO_HZ_COMMON
9050 * If this CPU has been elected to perform the nohz idle
9051 * balance. Other idle CPUs have already rebalanced with
9052 * nohz_idle_balance() and nohz.next_balance has been
9053 * updated accordingly. This CPU is now running the idle load
9054 * balance for itself and we need to update the
9055 * nohz.next_balance accordingly.
9057 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9058 nohz.next_balance = rq->next_balance;
9063 #ifdef CONFIG_NO_HZ_COMMON
9065 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9066 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9068 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9070 int this_cpu = this_rq->cpu;
9073 /* Earliest time when we have to do rebalance again */
9074 unsigned long next_balance = jiffies + 60*HZ;
9075 int update_next_balance = 0;
9077 if (idle != CPU_IDLE ||
9078 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
9081 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9082 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9086 * If this cpu gets work to do, stop the load balancing
9087 * work being done for other cpus. Next load
9088 * balancing owner will pick it up.
9093 rq = cpu_rq(balance_cpu);
9096 * If time for next balance is due,
9099 if (time_after_eq(jiffies, rq->next_balance)) {
9100 raw_spin_lock_irq(&rq->lock);
9101 update_rq_clock(rq);
9102 update_idle_cpu_load(rq);
9103 raw_spin_unlock_irq(&rq->lock);
9104 rebalance_domains(rq, CPU_IDLE);
9107 if (time_after(next_balance, rq->next_balance)) {
9108 next_balance = rq->next_balance;
9109 update_next_balance = 1;
9114 * next_balance will be updated only when there is a need.
9115 * When the CPU is attached to null domain for ex, it will not be
9118 if (likely(update_next_balance))
9119 nohz.next_balance = next_balance;
9121 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
9125 * Current heuristic for kicking the idle load balancer in the presence
9126 * of an idle cpu in the system.
9127 * - This rq has more than one task.
9128 * - This rq has at least one CFS task and the capacity of the CPU is
9129 * significantly reduced because of RT tasks or IRQs.
9130 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
9131 * multiple busy cpu.
9132 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
9133 * domain span are idle.
9135 static inline bool nohz_kick_needed(struct rq *rq)
9137 unsigned long now = jiffies;
9138 struct sched_domain *sd;
9139 struct sched_group_capacity *sgc;
9140 int nr_busy, cpu = rq->cpu;
9143 if (unlikely(rq->idle_balance))
9147 * We may be recently in ticked or tickless idle mode. At the first
9148 * busy tick after returning from idle, we will update the busy stats.
9150 set_cpu_sd_state_busy();
9151 nohz_balance_exit_idle(cpu);
9154 * None are in tickless mode and hence no need for NOHZ idle load
9157 if (likely(!atomic_read(&nohz.nr_cpus)))
9160 if (time_before(now, nohz.next_balance))
9163 if (rq->nr_running >= 2 &&
9164 (!energy_aware() || cpu_overutilized(cpu)))
9168 sd = rcu_dereference(per_cpu(sd_busy, cpu));
9169 if (sd && !energy_aware()) {
9170 sgc = sd->groups->sgc;
9171 nr_busy = atomic_read(&sgc->nr_busy_cpus);
9180 sd = rcu_dereference(rq->sd);
9182 if ((rq->cfs.h_nr_running >= 1) &&
9183 check_cpu_capacity(rq, sd)) {
9189 sd = rcu_dereference(per_cpu(sd_asym, cpu));
9190 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
9191 sched_domain_span(sd)) < cpu)) {
9201 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
9205 * run_rebalance_domains is triggered when needed from the scheduler tick.
9206 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9208 static void run_rebalance_domains(struct softirq_action *h)
9210 struct rq *this_rq = this_rq();
9211 enum cpu_idle_type idle = this_rq->idle_balance ?
9212 CPU_IDLE : CPU_NOT_IDLE;
9215 * If this cpu has a pending nohz_balance_kick, then do the
9216 * balancing on behalf of the other idle cpus whose ticks are
9217 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9218 * give the idle cpus a chance to load balance. Else we may
9219 * load balance only within the local sched_domain hierarchy
9220 * and abort nohz_idle_balance altogether if we pull some load.
9222 nohz_idle_balance(this_rq, idle);
9223 rebalance_domains(this_rq, idle);
9227 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9229 void trigger_load_balance(struct rq *rq)
9231 /* Don't need to rebalance while attached to NULL domain */
9232 if (unlikely(on_null_domain(rq)))
9235 if (time_after_eq(jiffies, rq->next_balance))
9236 raise_softirq(SCHED_SOFTIRQ);
9237 #ifdef CONFIG_NO_HZ_COMMON
9238 if (nohz_kick_needed(rq))
9239 nohz_balancer_kick();
9243 static void rq_online_fair(struct rq *rq)
9247 update_runtime_enabled(rq);
9250 static void rq_offline_fair(struct rq *rq)
9254 /* Ensure any throttled groups are reachable by pick_next_task */
9255 unthrottle_offline_cfs_rqs(rq);
9258 #endif /* CONFIG_SMP */
9261 * scheduler tick hitting a task of our scheduling class:
9263 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9265 struct cfs_rq *cfs_rq;
9266 struct sched_entity *se = &curr->se;
9268 for_each_sched_entity(se) {
9269 cfs_rq = cfs_rq_of(se);
9270 entity_tick(cfs_rq, se, queued);
9273 if (static_branch_unlikely(&sched_numa_balancing))
9274 task_tick_numa(rq, curr);
9277 if (!rq->rd->overutilized && cpu_overutilized(task_cpu(curr))) {
9278 rq->rd->overutilized = true;
9279 trace_sched_overutilized(true);
9282 rq->misfit_task = !task_fits_max(curr, rq->cpu);
9288 * called on fork with the child task as argument from the parent's context
9289 * - child not yet on the tasklist
9290 * - preemption disabled
9292 static void task_fork_fair(struct task_struct *p)
9294 struct cfs_rq *cfs_rq;
9295 struct sched_entity *se = &p->se, *curr;
9296 int this_cpu = smp_processor_id();
9297 struct rq *rq = this_rq();
9298 unsigned long flags;
9300 raw_spin_lock_irqsave(&rq->lock, flags);
9302 update_rq_clock(rq);
9304 cfs_rq = task_cfs_rq(current);
9305 curr = cfs_rq->curr;
9308 * Not only the cpu but also the task_group of the parent might have
9309 * been changed after parent->se.parent,cfs_rq were copied to
9310 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
9311 * of child point to valid ones.
9314 __set_task_cpu(p, this_cpu);
9317 update_curr(cfs_rq);
9320 se->vruntime = curr->vruntime;
9321 place_entity(cfs_rq, se, 1);
9323 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9325 * Upon rescheduling, sched_class::put_prev_task() will place
9326 * 'current' within the tree based on its new key value.
9328 swap(curr->vruntime, se->vruntime);
9332 se->vruntime -= cfs_rq->min_vruntime;
9334 raw_spin_unlock_irqrestore(&rq->lock, flags);
9338 * Priority of the task has changed. Check to see if we preempt
9342 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9344 if (!task_on_rq_queued(p))
9348 * Reschedule if we are currently running on this runqueue and
9349 * our priority decreased, or if we are not currently running on
9350 * this runqueue and our priority is higher than the current's
9352 if (rq->curr == p) {
9353 if (p->prio > oldprio)
9356 check_preempt_curr(rq, p, 0);
9359 static inline bool vruntime_normalized(struct task_struct *p)
9361 struct sched_entity *se = &p->se;
9364 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9365 * the dequeue_entity(.flags=0) will already have normalized the
9372 * When !on_rq, vruntime of the task has usually NOT been normalized.
9373 * But there are some cases where it has already been normalized:
9375 * - A forked child which is waiting for being woken up by
9376 * wake_up_new_task().
9377 * - A task which has been woken up by try_to_wake_up() and
9378 * waiting for actually being woken up by sched_ttwu_pending().
9380 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
9386 static void detach_task_cfs_rq(struct task_struct *p)
9388 struct sched_entity *se = &p->se;
9389 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9390 u64 now = cfs_rq_clock_task(cfs_rq);
9393 if (!vruntime_normalized(p)) {
9395 * Fix up our vruntime so that the current sleep doesn't
9396 * cause 'unlimited' sleep bonus.
9398 place_entity(cfs_rq, se, 0);
9399 se->vruntime -= cfs_rq->min_vruntime;
9402 /* Catch up with the cfs_rq and remove our load when we leave */
9403 tg_update = update_cfs_rq_load_avg(now, cfs_rq, false);
9404 detach_entity_load_avg(cfs_rq, se);
9406 update_tg_load_avg(cfs_rq, false);
9409 static void attach_task_cfs_rq(struct task_struct *p)
9411 struct sched_entity *se = &p->se;
9412 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9413 u64 now = cfs_rq_clock_task(cfs_rq);
9416 #ifdef CONFIG_FAIR_GROUP_SCHED
9418 * Since the real-depth could have been changed (only FAIR
9419 * class maintain depth value), reset depth properly.
9421 se->depth = se->parent ? se->parent->depth + 1 : 0;
9424 /* Synchronize task with its cfs_rq */
9425 tg_update = update_cfs_rq_load_avg(now, cfs_rq, false);
9426 attach_entity_load_avg(cfs_rq, se);
9428 update_tg_load_avg(cfs_rq, false);
9430 if (!vruntime_normalized(p))
9431 se->vruntime += cfs_rq->min_vruntime;
9434 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9436 detach_task_cfs_rq(p);
9439 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9441 attach_task_cfs_rq(p);
9443 if (task_on_rq_queued(p)) {
9445 * We were most likely switched from sched_rt, so
9446 * kick off the schedule if running, otherwise just see
9447 * if we can still preempt the current task.
9452 check_preempt_curr(rq, p, 0);
9456 /* Account for a task changing its policy or group.
9458 * This routine is mostly called to set cfs_rq->curr field when a task
9459 * migrates between groups/classes.
9461 static void set_curr_task_fair(struct rq *rq)
9463 struct sched_entity *se = &rq->curr->se;
9465 for_each_sched_entity(se) {
9466 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9468 set_next_entity(cfs_rq, se);
9469 /* ensure bandwidth has been allocated on our new cfs_rq */
9470 account_cfs_rq_runtime(cfs_rq, 0);
9474 void init_cfs_rq(struct cfs_rq *cfs_rq)
9476 cfs_rq->tasks_timeline = RB_ROOT;
9477 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9478 #ifndef CONFIG_64BIT
9479 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9482 atomic_long_set(&cfs_rq->removed_load_avg, 0);
9483 atomic_long_set(&cfs_rq->removed_util_avg, 0);
9487 #ifdef CONFIG_FAIR_GROUP_SCHED
9488 static void task_move_group_fair(struct task_struct *p)
9490 detach_task_cfs_rq(p);
9491 set_task_rq(p, task_cpu(p));
9494 /* Tell se's cfs_rq has been changed -- migrated */
9495 p->se.avg.last_update_time = 0;
9497 attach_task_cfs_rq(p);
9500 void free_fair_sched_group(struct task_group *tg)
9504 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9506 for_each_possible_cpu(i) {
9508 kfree(tg->cfs_rq[i]);
9511 remove_entity_load_avg(tg->se[i]);
9520 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9522 struct sched_entity *se;
9523 struct cfs_rq *cfs_rq;
9527 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9530 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9534 tg->shares = NICE_0_LOAD;
9536 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9538 for_each_possible_cpu(i) {
9541 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9542 GFP_KERNEL, cpu_to_node(i));
9546 se = kzalloc_node(sizeof(struct sched_entity),
9547 GFP_KERNEL, cpu_to_node(i));
9551 init_cfs_rq(cfs_rq);
9552 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9553 init_entity_runnable_average(se);
9555 raw_spin_lock_irq(&rq->lock);
9556 post_init_entity_util_avg(se);
9557 raw_spin_unlock_irq(&rq->lock);
9568 void unregister_fair_sched_group(struct task_group *tg, int cpu)
9570 struct rq *rq = cpu_rq(cpu);
9571 unsigned long flags;
9574 * Only empty task groups can be destroyed; so we can speculatively
9575 * check on_list without danger of it being re-added.
9577 if (!tg->cfs_rq[cpu]->on_list)
9580 raw_spin_lock_irqsave(&rq->lock, flags);
9581 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9582 raw_spin_unlock_irqrestore(&rq->lock, flags);
9585 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9586 struct sched_entity *se, int cpu,
9587 struct sched_entity *parent)
9589 struct rq *rq = cpu_rq(cpu);
9593 init_cfs_rq_runtime(cfs_rq);
9595 tg->cfs_rq[cpu] = cfs_rq;
9598 /* se could be NULL for root_task_group */
9603 se->cfs_rq = &rq->cfs;
9606 se->cfs_rq = parent->my_q;
9607 se->depth = parent->depth + 1;
9611 /* guarantee group entities always have weight */
9612 update_load_set(&se->load, NICE_0_LOAD);
9613 se->parent = parent;
9616 static DEFINE_MUTEX(shares_mutex);
9618 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9621 unsigned long flags;
9624 * We can't change the weight of the root cgroup.
9629 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9631 mutex_lock(&shares_mutex);
9632 if (tg->shares == shares)
9635 tg->shares = shares;
9636 for_each_possible_cpu(i) {
9637 struct rq *rq = cpu_rq(i);
9638 struct sched_entity *se;
9641 /* Propagate contribution to hierarchy */
9642 raw_spin_lock_irqsave(&rq->lock, flags);
9644 /* Possible calls to update_curr() need rq clock */
9645 update_rq_clock(rq);
9646 for_each_sched_entity(se)
9647 update_cfs_shares(group_cfs_rq(se));
9648 raw_spin_unlock_irqrestore(&rq->lock, flags);
9652 mutex_unlock(&shares_mutex);
9655 #else /* CONFIG_FAIR_GROUP_SCHED */
9657 void free_fair_sched_group(struct task_group *tg) { }
9659 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9664 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
9666 #endif /* CONFIG_FAIR_GROUP_SCHED */
9669 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9671 struct sched_entity *se = &task->se;
9672 unsigned int rr_interval = 0;
9675 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9678 if (rq->cfs.load.weight)
9679 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9685 * All the scheduling class methods:
9687 const struct sched_class fair_sched_class = {
9688 .next = &idle_sched_class,
9689 .enqueue_task = enqueue_task_fair,
9690 .dequeue_task = dequeue_task_fair,
9691 .yield_task = yield_task_fair,
9692 .yield_to_task = yield_to_task_fair,
9694 .check_preempt_curr = check_preempt_wakeup,
9696 .pick_next_task = pick_next_task_fair,
9697 .put_prev_task = put_prev_task_fair,
9700 .select_task_rq = select_task_rq_fair,
9701 .migrate_task_rq = migrate_task_rq_fair,
9703 .rq_online = rq_online_fair,
9704 .rq_offline = rq_offline_fair,
9706 .task_waking = task_waking_fair,
9707 .task_dead = task_dead_fair,
9708 .set_cpus_allowed = set_cpus_allowed_common,
9711 .set_curr_task = set_curr_task_fair,
9712 .task_tick = task_tick_fair,
9713 .task_fork = task_fork_fair,
9715 .prio_changed = prio_changed_fair,
9716 .switched_from = switched_from_fair,
9717 .switched_to = switched_to_fair,
9719 .get_rr_interval = get_rr_interval_fair,
9721 .update_curr = update_curr_fair,
9723 #ifdef CONFIG_FAIR_GROUP_SCHED
9724 .task_move_group = task_move_group_fair,
9728 #ifdef CONFIG_SCHED_DEBUG
9729 void print_cfs_stats(struct seq_file *m, int cpu)
9731 struct cfs_rq *cfs_rq;
9734 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
9735 print_cfs_rq(m, cpu, cfs_rq);
9739 #ifdef CONFIG_NUMA_BALANCING
9740 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9743 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9745 for_each_online_node(node) {
9746 if (p->numa_faults) {
9747 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9748 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9750 if (p->numa_group) {
9751 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9752 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9754 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9757 #endif /* CONFIG_NUMA_BALANCING */
9758 #endif /* CONFIG_SCHED_DEBUG */
9760 __init void init_sched_fair_class(void)
9763 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9765 #ifdef CONFIG_NO_HZ_COMMON
9766 nohz.next_balance = jiffies;
9767 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
9768 cpu_notifier(sched_ilb_notifier, 0);