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);
759 if (cfs_rq->avg.util_avg != 0) {
760 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
761 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
763 if (sa->util_avg > cap)
769 * If we wish to restore tuning via setting initial util,
770 * this is where we should do it.
772 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
775 if (entity_is_task(se)) {
776 struct task_struct *p = task_of(se);
777 if (p->sched_class != &fair_sched_class) {
779 * For !fair tasks do:
781 update_cfs_rq_load_avg(now, cfs_rq, false);
782 attach_entity_load_avg(cfs_rq, se);
783 switched_from_fair(rq, p);
785 * such that the next switched_to_fair() has the
788 se->avg.last_update_time = now;
793 update_cfs_rq_load_avg(now, cfs_rq, false);
794 attach_entity_load_avg(cfs_rq, se);
795 update_tg_load_avg(cfs_rq, false);
798 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq);
799 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq);
801 void init_entity_runnable_average(struct sched_entity *se)
804 void post_init_entity_util_avg(struct sched_entity *se)
807 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
810 #endif /* CONFIG_SMP */
813 * Update the current task's runtime statistics.
815 static void update_curr(struct cfs_rq *cfs_rq)
817 struct sched_entity *curr = cfs_rq->curr;
818 u64 now = rq_clock_task(rq_of(cfs_rq));
824 delta_exec = now - curr->exec_start;
825 if (unlikely((s64)delta_exec <= 0))
828 curr->exec_start = now;
830 schedstat_set(curr->statistics.exec_max,
831 max(delta_exec, curr->statistics.exec_max));
833 curr->sum_exec_runtime += delta_exec;
834 schedstat_add(cfs_rq, exec_clock, delta_exec);
836 curr->vruntime += calc_delta_fair(delta_exec, curr);
837 update_min_vruntime(cfs_rq);
839 if (entity_is_task(curr)) {
840 struct task_struct *curtask = task_of(curr);
842 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
843 cpuacct_charge(curtask, delta_exec);
844 account_group_exec_runtime(curtask, delta_exec);
847 account_cfs_rq_runtime(cfs_rq, delta_exec);
850 static void update_curr_fair(struct rq *rq)
852 update_curr(cfs_rq_of(&rq->curr->se));
856 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
858 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
862 * Task is being enqueued - update stats:
864 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
867 * Are we enqueueing a waiting task? (for current tasks
868 * a dequeue/enqueue event is a NOP)
870 if (se != cfs_rq->curr)
871 update_stats_wait_start(cfs_rq, se);
875 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
877 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
878 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
879 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
880 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
881 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
882 #ifdef CONFIG_SCHEDSTATS
883 if (entity_is_task(se)) {
884 trace_sched_stat_wait(task_of(se),
885 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
888 schedstat_set(se->statistics.wait_start, 0);
892 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
895 * Mark the end of the wait period if dequeueing a
898 if (se != cfs_rq->curr)
899 update_stats_wait_end(cfs_rq, se);
903 * We are picking a new current task - update its stats:
906 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
909 * We are starting a new run period:
911 se->exec_start = rq_clock_task(rq_of(cfs_rq));
914 /**************************************************
915 * Scheduling class queueing methods:
918 #ifdef CONFIG_NUMA_BALANCING
920 * Approximate time to scan a full NUMA task in ms. The task scan period is
921 * calculated based on the tasks virtual memory size and
922 * numa_balancing_scan_size.
924 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
925 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
927 /* Portion of address space to scan in MB */
928 unsigned int sysctl_numa_balancing_scan_size = 256;
930 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
931 unsigned int sysctl_numa_balancing_scan_delay = 1000;
933 static unsigned int task_nr_scan_windows(struct task_struct *p)
935 unsigned long rss = 0;
936 unsigned long nr_scan_pages;
939 * Calculations based on RSS as non-present and empty pages are skipped
940 * by the PTE scanner and NUMA hinting faults should be trapped based
943 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
944 rss = get_mm_rss(p->mm);
948 rss = round_up(rss, nr_scan_pages);
949 return rss / nr_scan_pages;
952 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
953 #define MAX_SCAN_WINDOW 2560
955 static unsigned int task_scan_min(struct task_struct *p)
957 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
958 unsigned int scan, floor;
959 unsigned int windows = 1;
961 if (scan_size < MAX_SCAN_WINDOW)
962 windows = MAX_SCAN_WINDOW / scan_size;
963 floor = 1000 / windows;
965 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
966 return max_t(unsigned int, floor, scan);
969 static unsigned int task_scan_max(struct task_struct *p)
971 unsigned int smin = task_scan_min(p);
974 /* Watch for min being lower than max due to floor calculations */
975 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
976 return max(smin, smax);
979 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
981 rq->nr_numa_running += (p->numa_preferred_nid != -1);
982 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
985 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
987 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
988 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
994 spinlock_t lock; /* nr_tasks, tasks */
999 nodemask_t active_nodes;
1000 unsigned long total_faults;
1002 * Faults_cpu is used to decide whether memory should move
1003 * towards the CPU. As a consequence, these stats are weighted
1004 * more by CPU use than by memory faults.
1006 unsigned long *faults_cpu;
1007 unsigned long faults[0];
1010 /* Shared or private faults. */
1011 #define NR_NUMA_HINT_FAULT_TYPES 2
1013 /* Memory and CPU locality */
1014 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1016 /* Averaged statistics, and temporary buffers. */
1017 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1019 pid_t task_numa_group_id(struct task_struct *p)
1021 return p->numa_group ? p->numa_group->gid : 0;
1025 * The averaged statistics, shared & private, memory & cpu,
1026 * occupy the first half of the array. The second half of the
1027 * array is for current counters, which are averaged into the
1028 * first set by task_numa_placement.
1030 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1032 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1035 static inline unsigned long task_faults(struct task_struct *p, int nid)
1037 if (!p->numa_faults)
1040 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1041 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1044 static inline unsigned long group_faults(struct task_struct *p, int nid)
1049 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1050 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1053 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1055 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1056 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1059 /* Handle placement on systems where not all nodes are directly connected. */
1060 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1061 int maxdist, bool task)
1063 unsigned long score = 0;
1067 * All nodes are directly connected, and the same distance
1068 * from each other. No need for fancy placement algorithms.
1070 if (sched_numa_topology_type == NUMA_DIRECT)
1074 * This code is called for each node, introducing N^2 complexity,
1075 * which should be ok given the number of nodes rarely exceeds 8.
1077 for_each_online_node(node) {
1078 unsigned long faults;
1079 int dist = node_distance(nid, node);
1082 * The furthest away nodes in the system are not interesting
1083 * for placement; nid was already counted.
1085 if (dist == sched_max_numa_distance || node == nid)
1089 * On systems with a backplane NUMA topology, compare groups
1090 * of nodes, and move tasks towards the group with the most
1091 * memory accesses. When comparing two nodes at distance
1092 * "hoplimit", only nodes closer by than "hoplimit" are part
1093 * of each group. Skip other nodes.
1095 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1099 /* Add up the faults from nearby nodes. */
1101 faults = task_faults(p, node);
1103 faults = group_faults(p, node);
1106 * On systems with a glueless mesh NUMA topology, there are
1107 * no fixed "groups of nodes". Instead, nodes that are not
1108 * directly connected bounce traffic through intermediate
1109 * nodes; a numa_group can occupy any set of nodes.
1110 * The further away a node is, the less the faults count.
1111 * This seems to result in good task placement.
1113 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1114 faults *= (sched_max_numa_distance - dist);
1115 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1125 * These return the fraction of accesses done by a particular task, or
1126 * task group, on a particular numa node. The group weight is given a
1127 * larger multiplier, in order to group tasks together that are almost
1128 * evenly spread out between numa nodes.
1130 static inline unsigned long task_weight(struct task_struct *p, int nid,
1133 unsigned long faults, total_faults;
1135 if (!p->numa_faults)
1138 total_faults = p->total_numa_faults;
1143 faults = task_faults(p, nid);
1144 faults += score_nearby_nodes(p, nid, dist, true);
1146 return 1000 * faults / total_faults;
1149 static inline unsigned long group_weight(struct task_struct *p, int nid,
1152 unsigned long faults, total_faults;
1157 total_faults = p->numa_group->total_faults;
1162 faults = group_faults(p, nid);
1163 faults += score_nearby_nodes(p, nid, dist, false);
1165 return 1000 * faults / total_faults;
1168 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1169 int src_nid, int dst_cpu)
1171 struct numa_group *ng = p->numa_group;
1172 int dst_nid = cpu_to_node(dst_cpu);
1173 int last_cpupid, this_cpupid;
1175 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1178 * Multi-stage node selection is used in conjunction with a periodic
1179 * migration fault to build a temporal task<->page relation. By using
1180 * a two-stage filter we remove short/unlikely relations.
1182 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1183 * a task's usage of a particular page (n_p) per total usage of this
1184 * page (n_t) (in a given time-span) to a probability.
1186 * Our periodic faults will sample this probability and getting the
1187 * same result twice in a row, given these samples are fully
1188 * independent, is then given by P(n)^2, provided our sample period
1189 * is sufficiently short compared to the usage pattern.
1191 * This quadric squishes small probabilities, making it less likely we
1192 * act on an unlikely task<->page relation.
1194 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1195 if (!cpupid_pid_unset(last_cpupid) &&
1196 cpupid_to_nid(last_cpupid) != dst_nid)
1199 /* Always allow migrate on private faults */
1200 if (cpupid_match_pid(p, last_cpupid))
1203 /* A shared fault, but p->numa_group has not been set up yet. */
1208 * Do not migrate if the destination is not a node that
1209 * is actively used by this numa group.
1211 if (!node_isset(dst_nid, ng->active_nodes))
1215 * Source is a node that is not actively used by this
1216 * numa group, while the destination is. Migrate.
1218 if (!node_isset(src_nid, ng->active_nodes))
1222 * Both source and destination are nodes in active
1223 * use by this numa group. Maximize memory bandwidth
1224 * by migrating from more heavily used groups, to less
1225 * heavily used ones, spreading the load around.
1226 * Use a 1/4 hysteresis to avoid spurious page movement.
1228 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1231 static unsigned long weighted_cpuload(const int cpu);
1232 static unsigned long source_load(int cpu, int type);
1233 static unsigned long target_load(int cpu, int type);
1234 static unsigned long capacity_of(int cpu);
1235 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1237 /* Cached statistics for all CPUs within a node */
1239 unsigned long nr_running;
1242 /* Total compute capacity of CPUs on a node */
1243 unsigned long compute_capacity;
1245 /* Approximate capacity in terms of runnable tasks on a node */
1246 unsigned long task_capacity;
1247 int has_free_capacity;
1251 * XXX borrowed from update_sg_lb_stats
1253 static void update_numa_stats(struct numa_stats *ns, int nid)
1255 int smt, cpu, cpus = 0;
1256 unsigned long capacity;
1258 memset(ns, 0, sizeof(*ns));
1259 for_each_cpu(cpu, cpumask_of_node(nid)) {
1260 struct rq *rq = cpu_rq(cpu);
1262 ns->nr_running += rq->nr_running;
1263 ns->load += weighted_cpuload(cpu);
1264 ns->compute_capacity += capacity_of(cpu);
1270 * If we raced with hotplug and there are no CPUs left in our mask
1271 * the @ns structure is NULL'ed and task_numa_compare() will
1272 * not find this node attractive.
1274 * We'll either bail at !has_free_capacity, or we'll detect a huge
1275 * imbalance and bail there.
1280 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1281 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1282 capacity = cpus / smt; /* cores */
1284 ns->task_capacity = min_t(unsigned, capacity,
1285 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1286 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1289 struct task_numa_env {
1290 struct task_struct *p;
1292 int src_cpu, src_nid;
1293 int dst_cpu, dst_nid;
1295 struct numa_stats src_stats, dst_stats;
1300 struct task_struct *best_task;
1305 static void task_numa_assign(struct task_numa_env *env,
1306 struct task_struct *p, long imp)
1309 put_task_struct(env->best_task);
1312 env->best_imp = imp;
1313 env->best_cpu = env->dst_cpu;
1316 static bool load_too_imbalanced(long src_load, long dst_load,
1317 struct task_numa_env *env)
1320 long orig_src_load, orig_dst_load;
1321 long src_capacity, dst_capacity;
1324 * The load is corrected for the CPU capacity available on each node.
1327 * ------------ vs ---------
1328 * src_capacity dst_capacity
1330 src_capacity = env->src_stats.compute_capacity;
1331 dst_capacity = env->dst_stats.compute_capacity;
1333 /* We care about the slope of the imbalance, not the direction. */
1334 if (dst_load < src_load)
1335 swap(dst_load, src_load);
1337 /* Is the difference below the threshold? */
1338 imb = dst_load * src_capacity * 100 -
1339 src_load * dst_capacity * env->imbalance_pct;
1344 * The imbalance is above the allowed threshold.
1345 * Compare it with the old imbalance.
1347 orig_src_load = env->src_stats.load;
1348 orig_dst_load = env->dst_stats.load;
1350 if (orig_dst_load < orig_src_load)
1351 swap(orig_dst_load, orig_src_load);
1353 old_imb = orig_dst_load * src_capacity * 100 -
1354 orig_src_load * dst_capacity * env->imbalance_pct;
1356 /* Would this change make things worse? */
1357 return (imb > old_imb);
1361 * This checks if the overall compute and NUMA accesses of the system would
1362 * be improved if the source tasks was migrated to the target dst_cpu taking
1363 * into account that it might be best if task running on the dst_cpu should
1364 * be exchanged with the source task
1366 static void task_numa_compare(struct task_numa_env *env,
1367 long taskimp, long groupimp)
1369 struct rq *src_rq = cpu_rq(env->src_cpu);
1370 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1371 struct task_struct *cur;
1372 long src_load, dst_load;
1374 long imp = env->p->numa_group ? groupimp : taskimp;
1376 int dist = env->dist;
1377 bool assigned = false;
1381 raw_spin_lock_irq(&dst_rq->lock);
1384 * No need to move the exiting task or idle task.
1386 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1390 * The task_struct must be protected here to protect the
1391 * p->numa_faults access in the task_weight since the
1392 * numa_faults could already be freed in the following path:
1393 * finish_task_switch()
1394 * --> put_task_struct()
1395 * --> __put_task_struct()
1396 * --> task_numa_free()
1398 get_task_struct(cur);
1401 raw_spin_unlock_irq(&dst_rq->lock);
1404 * Because we have preemption enabled we can get migrated around and
1405 * end try selecting ourselves (current == env->p) as a swap candidate.
1411 * "imp" is the fault differential for the source task between the
1412 * source and destination node. Calculate the total differential for
1413 * the source task and potential destination task. The more negative
1414 * the value is, the more rmeote accesses that would be expected to
1415 * be incurred if the tasks were swapped.
1418 /* Skip this swap candidate if cannot move to the source cpu */
1419 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1423 * If dst and source tasks are in the same NUMA group, or not
1424 * in any group then look only at task weights.
1426 if (cur->numa_group == env->p->numa_group) {
1427 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1428 task_weight(cur, env->dst_nid, dist);
1430 * Add some hysteresis to prevent swapping the
1431 * tasks within a group over tiny differences.
1433 if (cur->numa_group)
1437 * Compare the group weights. If a task is all by
1438 * itself (not part of a group), use the task weight
1441 if (cur->numa_group)
1442 imp += group_weight(cur, env->src_nid, dist) -
1443 group_weight(cur, env->dst_nid, dist);
1445 imp += task_weight(cur, env->src_nid, dist) -
1446 task_weight(cur, env->dst_nid, dist);
1450 if (imp <= env->best_imp && moveimp <= env->best_imp)
1454 /* Is there capacity at our destination? */
1455 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1456 !env->dst_stats.has_free_capacity)
1462 /* Balance doesn't matter much if we're running a task per cpu */
1463 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1464 dst_rq->nr_running == 1)
1468 * In the overloaded case, try and keep the load balanced.
1471 load = task_h_load(env->p);
1472 dst_load = env->dst_stats.load + load;
1473 src_load = env->src_stats.load - load;
1475 if (moveimp > imp && moveimp > env->best_imp) {
1477 * If the improvement from just moving env->p direction is
1478 * better than swapping tasks around, check if a move is
1479 * possible. Store a slightly smaller score than moveimp,
1480 * so an actually idle CPU will win.
1482 if (!load_too_imbalanced(src_load, dst_load, env)) {
1484 put_task_struct(cur);
1490 if (imp <= env->best_imp)
1494 load = task_h_load(cur);
1499 if (load_too_imbalanced(src_load, dst_load, env))
1503 * One idle CPU per node is evaluated for a task numa move.
1504 * Call select_idle_sibling to maybe find a better one.
1507 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1512 task_numa_assign(env, cur, imp);
1516 * The dst_rq->curr isn't assigned. The protection for task_struct is
1519 if (cur && !assigned)
1520 put_task_struct(cur);
1523 static void task_numa_find_cpu(struct task_numa_env *env,
1524 long taskimp, long groupimp)
1528 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1529 /* Skip this CPU if the source task cannot migrate */
1530 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1534 task_numa_compare(env, taskimp, groupimp);
1538 /* Only move tasks to a NUMA node less busy than the current node. */
1539 static bool numa_has_capacity(struct task_numa_env *env)
1541 struct numa_stats *src = &env->src_stats;
1542 struct numa_stats *dst = &env->dst_stats;
1544 if (src->has_free_capacity && !dst->has_free_capacity)
1548 * Only consider a task move if the source has a higher load
1549 * than the destination, corrected for CPU capacity on each node.
1551 * src->load dst->load
1552 * --------------------- vs ---------------------
1553 * src->compute_capacity dst->compute_capacity
1555 if (src->load * dst->compute_capacity * env->imbalance_pct >
1557 dst->load * src->compute_capacity * 100)
1563 static int task_numa_migrate(struct task_struct *p)
1565 struct task_numa_env env = {
1568 .src_cpu = task_cpu(p),
1569 .src_nid = task_node(p),
1571 .imbalance_pct = 112,
1577 struct sched_domain *sd;
1578 unsigned long taskweight, groupweight;
1580 long taskimp, groupimp;
1583 * Pick the lowest SD_NUMA domain, as that would have the smallest
1584 * imbalance and would be the first to start moving tasks about.
1586 * And we want to avoid any moving of tasks about, as that would create
1587 * random movement of tasks -- counter the numa conditions we're trying
1591 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1593 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1597 * Cpusets can break the scheduler domain tree into smaller
1598 * balance domains, some of which do not cross NUMA boundaries.
1599 * Tasks that are "trapped" in such domains cannot be migrated
1600 * elsewhere, so there is no point in (re)trying.
1602 if (unlikely(!sd)) {
1603 p->numa_preferred_nid = task_node(p);
1607 env.dst_nid = p->numa_preferred_nid;
1608 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1609 taskweight = task_weight(p, env.src_nid, dist);
1610 groupweight = group_weight(p, env.src_nid, dist);
1611 update_numa_stats(&env.src_stats, env.src_nid);
1612 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1613 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1614 update_numa_stats(&env.dst_stats, env.dst_nid);
1616 /* Try to find a spot on the preferred nid. */
1617 if (numa_has_capacity(&env))
1618 task_numa_find_cpu(&env, taskimp, groupimp);
1621 * Look at other nodes in these cases:
1622 * - there is no space available on the preferred_nid
1623 * - the task is part of a numa_group that is interleaved across
1624 * multiple NUMA nodes; in order to better consolidate the group,
1625 * we need to check other locations.
1627 if (env.best_cpu == -1 || (p->numa_group &&
1628 nodes_weight(p->numa_group->active_nodes) > 1)) {
1629 for_each_online_node(nid) {
1630 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1633 dist = node_distance(env.src_nid, env.dst_nid);
1634 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1636 taskweight = task_weight(p, env.src_nid, dist);
1637 groupweight = group_weight(p, env.src_nid, dist);
1640 /* Only consider nodes where both task and groups benefit */
1641 taskimp = task_weight(p, nid, dist) - taskweight;
1642 groupimp = group_weight(p, nid, dist) - groupweight;
1643 if (taskimp < 0 && groupimp < 0)
1648 update_numa_stats(&env.dst_stats, env.dst_nid);
1649 if (numa_has_capacity(&env))
1650 task_numa_find_cpu(&env, taskimp, groupimp);
1655 * If the task is part of a workload that spans multiple NUMA nodes,
1656 * and is migrating into one of the workload's active nodes, remember
1657 * this node as the task's preferred numa node, so the workload can
1659 * A task that migrated to a second choice node will be better off
1660 * trying for a better one later. Do not set the preferred node here.
1662 if (p->numa_group) {
1663 if (env.best_cpu == -1)
1668 if (node_isset(nid, p->numa_group->active_nodes))
1669 sched_setnuma(p, env.dst_nid);
1672 /* No better CPU than the current one was found. */
1673 if (env.best_cpu == -1)
1677 * Reset the scan period if the task is being rescheduled on an
1678 * alternative node to recheck if the tasks is now properly placed.
1680 p->numa_scan_period = task_scan_min(p);
1682 if (env.best_task == NULL) {
1683 ret = migrate_task_to(p, env.best_cpu);
1685 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1689 ret = migrate_swap(p, env.best_task);
1691 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1692 put_task_struct(env.best_task);
1696 /* Attempt to migrate a task to a CPU on the preferred node. */
1697 static void numa_migrate_preferred(struct task_struct *p)
1699 unsigned long interval = HZ;
1701 /* This task has no NUMA fault statistics yet */
1702 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1705 /* Periodically retry migrating the task to the preferred node */
1706 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1707 p->numa_migrate_retry = jiffies + interval;
1709 /* Success if task is already running on preferred CPU */
1710 if (task_node(p) == p->numa_preferred_nid)
1713 /* Otherwise, try migrate to a CPU on the preferred node */
1714 task_numa_migrate(p);
1718 * Find the nodes on which the workload is actively running. We do this by
1719 * tracking the nodes from which NUMA hinting faults are triggered. This can
1720 * be different from the set of nodes where the workload's memory is currently
1723 * The bitmask is used to make smarter decisions on when to do NUMA page
1724 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1725 * are added when they cause over 6/16 of the maximum number of faults, but
1726 * only removed when they drop below 3/16.
1728 static void update_numa_active_node_mask(struct numa_group *numa_group)
1730 unsigned long faults, max_faults = 0;
1733 for_each_online_node(nid) {
1734 faults = group_faults_cpu(numa_group, nid);
1735 if (faults > max_faults)
1736 max_faults = faults;
1739 for_each_online_node(nid) {
1740 faults = group_faults_cpu(numa_group, nid);
1741 if (!node_isset(nid, numa_group->active_nodes)) {
1742 if (faults > max_faults * 6 / 16)
1743 node_set(nid, numa_group->active_nodes);
1744 } else if (faults < max_faults * 3 / 16)
1745 node_clear(nid, numa_group->active_nodes);
1750 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1751 * increments. The more local the fault statistics are, the higher the scan
1752 * period will be for the next scan window. If local/(local+remote) ratio is
1753 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1754 * the scan period will decrease. Aim for 70% local accesses.
1756 #define NUMA_PERIOD_SLOTS 10
1757 #define NUMA_PERIOD_THRESHOLD 7
1760 * Increase the scan period (slow down scanning) if the majority of
1761 * our memory is already on our local node, or if the majority of
1762 * the page accesses are shared with other processes.
1763 * Otherwise, decrease the scan period.
1765 static void update_task_scan_period(struct task_struct *p,
1766 unsigned long shared, unsigned long private)
1768 unsigned int period_slot;
1772 unsigned long remote = p->numa_faults_locality[0];
1773 unsigned long local = p->numa_faults_locality[1];
1776 * If there were no record hinting faults then either the task is
1777 * completely idle or all activity is areas that are not of interest
1778 * to automatic numa balancing. Related to that, if there were failed
1779 * migration then it implies we are migrating too quickly or the local
1780 * node is overloaded. In either case, scan slower
1782 if (local + shared == 0 || p->numa_faults_locality[2]) {
1783 p->numa_scan_period = min(p->numa_scan_period_max,
1784 p->numa_scan_period << 1);
1786 p->mm->numa_next_scan = jiffies +
1787 msecs_to_jiffies(p->numa_scan_period);
1793 * Prepare to scale scan period relative to the current period.
1794 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1795 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1796 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1798 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1799 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1800 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1801 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1804 diff = slot * period_slot;
1806 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1809 * Scale scan rate increases based on sharing. There is an
1810 * inverse relationship between the degree of sharing and
1811 * the adjustment made to the scanning period. Broadly
1812 * speaking the intent is that there is little point
1813 * scanning faster if shared accesses dominate as it may
1814 * simply bounce migrations uselessly
1816 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1817 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1820 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1821 task_scan_min(p), task_scan_max(p));
1822 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1826 * Get the fraction of time the task has been running since the last
1827 * NUMA placement cycle. The scheduler keeps similar statistics, but
1828 * decays those on a 32ms period, which is orders of magnitude off
1829 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1830 * stats only if the task is so new there are no NUMA statistics yet.
1832 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1834 u64 runtime, delta, now;
1835 /* Use the start of this time slice to avoid calculations. */
1836 now = p->se.exec_start;
1837 runtime = p->se.sum_exec_runtime;
1839 if (p->last_task_numa_placement) {
1840 delta = runtime - p->last_sum_exec_runtime;
1841 *period = now - p->last_task_numa_placement;
1843 delta = p->se.avg.load_sum / p->se.load.weight;
1844 *period = LOAD_AVG_MAX;
1847 p->last_sum_exec_runtime = runtime;
1848 p->last_task_numa_placement = now;
1854 * Determine the preferred nid for a task in a numa_group. This needs to
1855 * be done in a way that produces consistent results with group_weight,
1856 * otherwise workloads might not converge.
1858 static int preferred_group_nid(struct task_struct *p, int nid)
1863 /* Direct connections between all NUMA nodes. */
1864 if (sched_numa_topology_type == NUMA_DIRECT)
1868 * On a system with glueless mesh NUMA topology, group_weight
1869 * scores nodes according to the number of NUMA hinting faults on
1870 * both the node itself, and on nearby nodes.
1872 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1873 unsigned long score, max_score = 0;
1874 int node, max_node = nid;
1876 dist = sched_max_numa_distance;
1878 for_each_online_node(node) {
1879 score = group_weight(p, node, dist);
1880 if (score > max_score) {
1889 * Finding the preferred nid in a system with NUMA backplane
1890 * interconnect topology is more involved. The goal is to locate
1891 * tasks from numa_groups near each other in the system, and
1892 * untangle workloads from different sides of the system. This requires
1893 * searching down the hierarchy of node groups, recursively searching
1894 * inside the highest scoring group of nodes. The nodemask tricks
1895 * keep the complexity of the search down.
1897 nodes = node_online_map;
1898 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1899 unsigned long max_faults = 0;
1900 nodemask_t max_group = NODE_MASK_NONE;
1903 /* Are there nodes at this distance from each other? */
1904 if (!find_numa_distance(dist))
1907 for_each_node_mask(a, nodes) {
1908 unsigned long faults = 0;
1909 nodemask_t this_group;
1910 nodes_clear(this_group);
1912 /* Sum group's NUMA faults; includes a==b case. */
1913 for_each_node_mask(b, nodes) {
1914 if (node_distance(a, b) < dist) {
1915 faults += group_faults(p, b);
1916 node_set(b, this_group);
1917 node_clear(b, nodes);
1921 /* Remember the top group. */
1922 if (faults > max_faults) {
1923 max_faults = faults;
1924 max_group = this_group;
1926 * subtle: at the smallest distance there is
1927 * just one node left in each "group", the
1928 * winner is the preferred nid.
1933 /* Next round, evaluate the nodes within max_group. */
1941 static void task_numa_placement(struct task_struct *p)
1943 int seq, nid, max_nid = -1, max_group_nid = -1;
1944 unsigned long max_faults = 0, max_group_faults = 0;
1945 unsigned long fault_types[2] = { 0, 0 };
1946 unsigned long total_faults;
1947 u64 runtime, period;
1948 spinlock_t *group_lock = NULL;
1951 * The p->mm->numa_scan_seq field gets updated without
1952 * exclusive access. Use READ_ONCE() here to ensure
1953 * that the field is read in a single access:
1955 seq = READ_ONCE(p->mm->numa_scan_seq);
1956 if (p->numa_scan_seq == seq)
1958 p->numa_scan_seq = seq;
1959 p->numa_scan_period_max = task_scan_max(p);
1961 total_faults = p->numa_faults_locality[0] +
1962 p->numa_faults_locality[1];
1963 runtime = numa_get_avg_runtime(p, &period);
1965 /* If the task is part of a group prevent parallel updates to group stats */
1966 if (p->numa_group) {
1967 group_lock = &p->numa_group->lock;
1968 spin_lock_irq(group_lock);
1971 /* Find the node with the highest number of faults */
1972 for_each_online_node(nid) {
1973 /* Keep track of the offsets in numa_faults array */
1974 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1975 unsigned long faults = 0, group_faults = 0;
1978 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1979 long diff, f_diff, f_weight;
1981 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1982 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1983 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1984 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1986 /* Decay existing window, copy faults since last scan */
1987 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1988 fault_types[priv] += p->numa_faults[membuf_idx];
1989 p->numa_faults[membuf_idx] = 0;
1992 * Normalize the faults_from, so all tasks in a group
1993 * count according to CPU use, instead of by the raw
1994 * number of faults. Tasks with little runtime have
1995 * little over-all impact on throughput, and thus their
1996 * faults are less important.
1998 f_weight = div64_u64(runtime << 16, period + 1);
1999 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2001 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2002 p->numa_faults[cpubuf_idx] = 0;
2004 p->numa_faults[mem_idx] += diff;
2005 p->numa_faults[cpu_idx] += f_diff;
2006 faults += p->numa_faults[mem_idx];
2007 p->total_numa_faults += diff;
2008 if (p->numa_group) {
2010 * safe because we can only change our own group
2012 * mem_idx represents the offset for a given
2013 * nid and priv in a specific region because it
2014 * is at the beginning of the numa_faults array.
2016 p->numa_group->faults[mem_idx] += diff;
2017 p->numa_group->faults_cpu[mem_idx] += f_diff;
2018 p->numa_group->total_faults += diff;
2019 group_faults += p->numa_group->faults[mem_idx];
2023 if (faults > max_faults) {
2024 max_faults = faults;
2028 if (group_faults > max_group_faults) {
2029 max_group_faults = group_faults;
2030 max_group_nid = nid;
2034 update_task_scan_period(p, fault_types[0], fault_types[1]);
2036 if (p->numa_group) {
2037 update_numa_active_node_mask(p->numa_group);
2038 spin_unlock_irq(group_lock);
2039 max_nid = preferred_group_nid(p, max_group_nid);
2043 /* Set the new preferred node */
2044 if (max_nid != p->numa_preferred_nid)
2045 sched_setnuma(p, max_nid);
2047 if (task_node(p) != p->numa_preferred_nid)
2048 numa_migrate_preferred(p);
2052 static inline int get_numa_group(struct numa_group *grp)
2054 return atomic_inc_not_zero(&grp->refcount);
2057 static inline void put_numa_group(struct numa_group *grp)
2059 if (atomic_dec_and_test(&grp->refcount))
2060 kfree_rcu(grp, rcu);
2063 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2066 struct numa_group *grp, *my_grp;
2067 struct task_struct *tsk;
2069 int cpu = cpupid_to_cpu(cpupid);
2072 if (unlikely(!p->numa_group)) {
2073 unsigned int size = sizeof(struct numa_group) +
2074 4*nr_node_ids*sizeof(unsigned long);
2076 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2080 atomic_set(&grp->refcount, 1);
2081 spin_lock_init(&grp->lock);
2083 /* Second half of the array tracks nids where faults happen */
2084 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2087 node_set(task_node(current), grp->active_nodes);
2089 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2090 grp->faults[i] = p->numa_faults[i];
2092 grp->total_faults = p->total_numa_faults;
2095 rcu_assign_pointer(p->numa_group, grp);
2099 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2101 if (!cpupid_match_pid(tsk, cpupid))
2104 grp = rcu_dereference(tsk->numa_group);
2108 my_grp = p->numa_group;
2113 * Only join the other group if its bigger; if we're the bigger group,
2114 * the other task will join us.
2116 if (my_grp->nr_tasks > grp->nr_tasks)
2120 * Tie-break on the grp address.
2122 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2125 /* Always join threads in the same process. */
2126 if (tsk->mm == current->mm)
2129 /* Simple filter to avoid false positives due to PID collisions */
2130 if (flags & TNF_SHARED)
2133 /* Update priv based on whether false sharing was detected */
2136 if (join && !get_numa_group(grp))
2144 BUG_ON(irqs_disabled());
2145 double_lock_irq(&my_grp->lock, &grp->lock);
2147 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2148 my_grp->faults[i] -= p->numa_faults[i];
2149 grp->faults[i] += p->numa_faults[i];
2151 my_grp->total_faults -= p->total_numa_faults;
2152 grp->total_faults += p->total_numa_faults;
2157 spin_unlock(&my_grp->lock);
2158 spin_unlock_irq(&grp->lock);
2160 rcu_assign_pointer(p->numa_group, grp);
2162 put_numa_group(my_grp);
2170 void task_numa_free(struct task_struct *p)
2172 struct numa_group *grp = p->numa_group;
2173 void *numa_faults = p->numa_faults;
2174 unsigned long flags;
2178 spin_lock_irqsave(&grp->lock, flags);
2179 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2180 grp->faults[i] -= p->numa_faults[i];
2181 grp->total_faults -= p->total_numa_faults;
2184 spin_unlock_irqrestore(&grp->lock, flags);
2185 RCU_INIT_POINTER(p->numa_group, NULL);
2186 put_numa_group(grp);
2189 p->numa_faults = NULL;
2194 * Got a PROT_NONE fault for a page on @node.
2196 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2198 struct task_struct *p = current;
2199 bool migrated = flags & TNF_MIGRATED;
2200 int cpu_node = task_node(current);
2201 int local = !!(flags & TNF_FAULT_LOCAL);
2204 if (!static_branch_likely(&sched_numa_balancing))
2207 /* for example, ksmd faulting in a user's mm */
2211 /* Allocate buffer to track faults on a per-node basis */
2212 if (unlikely(!p->numa_faults)) {
2213 int size = sizeof(*p->numa_faults) *
2214 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2216 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2217 if (!p->numa_faults)
2220 p->total_numa_faults = 0;
2221 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2225 * First accesses are treated as private, otherwise consider accesses
2226 * to be private if the accessing pid has not changed
2228 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2231 priv = cpupid_match_pid(p, last_cpupid);
2232 if (!priv && !(flags & TNF_NO_GROUP))
2233 task_numa_group(p, last_cpupid, flags, &priv);
2237 * If a workload spans multiple NUMA nodes, a shared fault that
2238 * occurs wholly within the set of nodes that the workload is
2239 * actively using should be counted as local. This allows the
2240 * scan rate to slow down when a workload has settled down.
2242 if (!priv && !local && p->numa_group &&
2243 node_isset(cpu_node, p->numa_group->active_nodes) &&
2244 node_isset(mem_node, p->numa_group->active_nodes))
2247 task_numa_placement(p);
2250 * Retry task to preferred node migration periodically, in case it
2251 * case it previously failed, or the scheduler moved us.
2253 if (time_after(jiffies, p->numa_migrate_retry))
2254 numa_migrate_preferred(p);
2257 p->numa_pages_migrated += pages;
2258 if (flags & TNF_MIGRATE_FAIL)
2259 p->numa_faults_locality[2] += pages;
2261 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2262 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2263 p->numa_faults_locality[local] += pages;
2266 static void reset_ptenuma_scan(struct task_struct *p)
2269 * We only did a read acquisition of the mmap sem, so
2270 * p->mm->numa_scan_seq is written to without exclusive access
2271 * and the update is not guaranteed to be atomic. That's not
2272 * much of an issue though, since this is just used for
2273 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2274 * expensive, to avoid any form of compiler optimizations:
2276 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2277 p->mm->numa_scan_offset = 0;
2281 * The expensive part of numa migration is done from task_work context.
2282 * Triggered from task_tick_numa().
2284 void task_numa_work(struct callback_head *work)
2286 unsigned long migrate, next_scan, now = jiffies;
2287 struct task_struct *p = current;
2288 struct mm_struct *mm = p->mm;
2289 struct vm_area_struct *vma;
2290 unsigned long start, end;
2291 unsigned long nr_pte_updates = 0;
2292 long pages, virtpages;
2294 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2296 work->next = work; /* protect against double add */
2298 * Who cares about NUMA placement when they're dying.
2300 * NOTE: make sure not to dereference p->mm before this check,
2301 * exit_task_work() happens _after_ exit_mm() so we could be called
2302 * without p->mm even though we still had it when we enqueued this
2305 if (p->flags & PF_EXITING)
2308 if (!mm->numa_next_scan) {
2309 mm->numa_next_scan = now +
2310 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2314 * Enforce maximal scan/migration frequency..
2316 migrate = mm->numa_next_scan;
2317 if (time_before(now, migrate))
2320 if (p->numa_scan_period == 0) {
2321 p->numa_scan_period_max = task_scan_max(p);
2322 p->numa_scan_period = task_scan_min(p);
2325 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2326 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2330 * Delay this task enough that another task of this mm will likely win
2331 * the next time around.
2333 p->node_stamp += 2 * TICK_NSEC;
2335 start = mm->numa_scan_offset;
2336 pages = sysctl_numa_balancing_scan_size;
2337 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2338 virtpages = pages * 8; /* Scan up to this much virtual space */
2343 down_read(&mm->mmap_sem);
2344 vma = find_vma(mm, start);
2346 reset_ptenuma_scan(p);
2350 for (; vma; vma = vma->vm_next) {
2351 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2352 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2357 * Shared library pages mapped by multiple processes are not
2358 * migrated as it is expected they are cache replicated. Avoid
2359 * hinting faults in read-only file-backed mappings or the vdso
2360 * as migrating the pages will be of marginal benefit.
2363 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2367 * Skip inaccessible VMAs to avoid any confusion between
2368 * PROT_NONE and NUMA hinting ptes
2370 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2374 start = max(start, vma->vm_start);
2375 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2376 end = min(end, vma->vm_end);
2377 nr_pte_updates = change_prot_numa(vma, start, end);
2380 * Try to scan sysctl_numa_balancing_size worth of
2381 * hpages that have at least one present PTE that
2382 * is not already pte-numa. If the VMA contains
2383 * areas that are unused or already full of prot_numa
2384 * PTEs, scan up to virtpages, to skip through those
2388 pages -= (end - start) >> PAGE_SHIFT;
2389 virtpages -= (end - start) >> PAGE_SHIFT;
2392 if (pages <= 0 || virtpages <= 0)
2396 } while (end != vma->vm_end);
2401 * It is possible to reach the end of the VMA list but the last few
2402 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2403 * would find the !migratable VMA on the next scan but not reset the
2404 * scanner to the start so check it now.
2407 mm->numa_scan_offset = start;
2409 reset_ptenuma_scan(p);
2410 up_read(&mm->mmap_sem);
2414 * Drive the periodic memory faults..
2416 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2418 struct callback_head *work = &curr->numa_work;
2422 * We don't care about NUMA placement if we don't have memory.
2424 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2428 * Using runtime rather than walltime has the dual advantage that
2429 * we (mostly) drive the selection from busy threads and that the
2430 * task needs to have done some actual work before we bother with
2433 now = curr->se.sum_exec_runtime;
2434 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2436 if (now > curr->node_stamp + period) {
2437 if (!curr->node_stamp)
2438 curr->numa_scan_period = task_scan_min(curr);
2439 curr->node_stamp += period;
2441 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2442 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2443 task_work_add(curr, work, true);
2448 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2452 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2456 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2459 #endif /* CONFIG_NUMA_BALANCING */
2462 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2464 update_load_add(&cfs_rq->load, se->load.weight);
2465 if (!parent_entity(se))
2466 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2468 if (entity_is_task(se)) {
2469 struct rq *rq = rq_of(cfs_rq);
2471 account_numa_enqueue(rq, task_of(se));
2472 list_add(&se->group_node, &rq->cfs_tasks);
2475 cfs_rq->nr_running++;
2479 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2481 update_load_sub(&cfs_rq->load, se->load.weight);
2482 if (!parent_entity(se))
2483 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2484 if (entity_is_task(se)) {
2485 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2486 list_del_init(&se->group_node);
2488 cfs_rq->nr_running--;
2491 #ifdef CONFIG_FAIR_GROUP_SCHED
2493 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2498 * Use this CPU's real-time load instead of the last load contribution
2499 * as the updating of the contribution is delayed, and we will use the
2500 * the real-time load to calc the share. See update_tg_load_avg().
2502 tg_weight = atomic_long_read(&tg->load_avg);
2503 tg_weight -= cfs_rq->tg_load_avg_contrib;
2504 tg_weight += cfs_rq->load.weight;
2509 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2511 long tg_weight, load, shares;
2513 tg_weight = calc_tg_weight(tg, cfs_rq);
2514 load = cfs_rq->load.weight;
2516 shares = (tg->shares * load);
2518 shares /= tg_weight;
2520 if (shares < MIN_SHARES)
2521 shares = MIN_SHARES;
2522 if (shares > tg->shares)
2523 shares = tg->shares;
2527 # else /* CONFIG_SMP */
2528 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2532 # endif /* CONFIG_SMP */
2533 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2534 unsigned long weight)
2537 /* commit outstanding execution time */
2538 if (cfs_rq->curr == se)
2539 update_curr(cfs_rq);
2540 account_entity_dequeue(cfs_rq, se);
2543 update_load_set(&se->load, weight);
2546 account_entity_enqueue(cfs_rq, se);
2549 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2551 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2553 struct task_group *tg;
2554 struct sched_entity *se;
2558 se = tg->se[cpu_of(rq_of(cfs_rq))];
2559 if (!se || throttled_hierarchy(cfs_rq))
2562 if (likely(se->load.weight == tg->shares))
2565 shares = calc_cfs_shares(cfs_rq, tg);
2567 reweight_entity(cfs_rq_of(se), se, shares);
2569 #else /* CONFIG_FAIR_GROUP_SCHED */
2570 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2573 #endif /* CONFIG_FAIR_GROUP_SCHED */
2576 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2577 static const u32 runnable_avg_yN_inv[] = {
2578 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2579 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2580 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2581 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2582 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2583 0x85aac367, 0x82cd8698,
2587 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2588 * over-estimates when re-combining.
2590 static const u32 runnable_avg_yN_sum[] = {
2591 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2592 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2593 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2598 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2600 static __always_inline u64 decay_load(u64 val, u64 n)
2602 unsigned int local_n;
2606 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2609 /* after bounds checking we can collapse to 32-bit */
2613 * As y^PERIOD = 1/2, we can combine
2614 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2615 * With a look-up table which covers y^n (n<PERIOD)
2617 * To achieve constant time decay_load.
2619 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2620 val >>= local_n / LOAD_AVG_PERIOD;
2621 local_n %= LOAD_AVG_PERIOD;
2624 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2629 * For updates fully spanning n periods, the contribution to runnable
2630 * average will be: \Sum 1024*y^n
2632 * We can compute this reasonably efficiently by combining:
2633 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2635 static u32 __compute_runnable_contrib(u64 n)
2639 if (likely(n <= LOAD_AVG_PERIOD))
2640 return runnable_avg_yN_sum[n];
2641 else if (unlikely(n >= LOAD_AVG_MAX_N))
2642 return LOAD_AVG_MAX;
2644 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2646 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2647 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2649 n -= LOAD_AVG_PERIOD;
2650 } while (n > LOAD_AVG_PERIOD);
2652 contrib = decay_load(contrib, n);
2653 return contrib + runnable_avg_yN_sum[n];
2656 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2657 #error "load tracking assumes 2^10 as unit"
2660 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2663 * We can represent the historical contribution to runnable average as the
2664 * coefficients of a geometric series. To do this we sub-divide our runnable
2665 * history into segments of approximately 1ms (1024us); label the segment that
2666 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2668 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2670 * (now) (~1ms ago) (~2ms ago)
2672 * Let u_i denote the fraction of p_i that the entity was runnable.
2674 * We then designate the fractions u_i as our co-efficients, yielding the
2675 * following representation of historical load:
2676 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2678 * We choose y based on the with of a reasonably scheduling period, fixing:
2681 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2682 * approximately half as much as the contribution to load within the last ms
2685 * When a period "rolls over" and we have new u_0`, multiplying the previous
2686 * sum again by y is sufficient to update:
2687 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2688 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2690 static __always_inline int
2691 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2692 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2694 u64 delta, scaled_delta, periods;
2696 unsigned int delta_w, scaled_delta_w, decayed = 0;
2697 unsigned long scale_freq, scale_cpu;
2699 delta = now - sa->last_update_time;
2701 * This should only happen when time goes backwards, which it
2702 * unfortunately does during sched clock init when we swap over to TSC.
2704 if ((s64)delta < 0) {
2705 sa->last_update_time = now;
2710 * Use 1024ns as the unit of measurement since it's a reasonable
2711 * approximation of 1us and fast to compute.
2716 sa->last_update_time = now;
2718 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2719 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2720 trace_sched_contrib_scale_f(cpu, scale_freq, scale_cpu);
2722 /* delta_w is the amount already accumulated against our next period */
2723 delta_w = sa->period_contrib;
2724 if (delta + delta_w >= 1024) {
2727 /* how much left for next period will start over, we don't know yet */
2728 sa->period_contrib = 0;
2731 * Now that we know we're crossing a period boundary, figure
2732 * out how much from delta we need to complete the current
2733 * period and accrue it.
2735 delta_w = 1024 - delta_w;
2736 scaled_delta_w = cap_scale(delta_w, scale_freq);
2738 sa->load_sum += weight * scaled_delta_w;
2740 cfs_rq->runnable_load_sum +=
2741 weight * scaled_delta_w;
2745 sa->util_sum += scaled_delta_w * scale_cpu;
2749 /* Figure out how many additional periods this update spans */
2750 periods = delta / 1024;
2753 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2755 cfs_rq->runnable_load_sum =
2756 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2758 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2760 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2761 contrib = __compute_runnable_contrib(periods);
2762 contrib = cap_scale(contrib, scale_freq);
2764 sa->load_sum += weight * contrib;
2766 cfs_rq->runnable_load_sum += weight * contrib;
2769 sa->util_sum += contrib * scale_cpu;
2772 /* Remainder of delta accrued against u_0` */
2773 scaled_delta = cap_scale(delta, scale_freq);
2775 sa->load_sum += weight * scaled_delta;
2777 cfs_rq->runnable_load_sum += weight * scaled_delta;
2780 sa->util_sum += scaled_delta * scale_cpu;
2782 sa->period_contrib += delta;
2785 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2787 cfs_rq->runnable_load_avg =
2788 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2790 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2796 #ifdef CONFIG_FAIR_GROUP_SCHED
2798 * update_tg_load_avg - update the tg's load avg
2799 * @cfs_rq: the cfs_rq whose avg changed
2800 * @force: update regardless of how small the difference
2802 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
2803 * However, because tg->load_avg is a global value there are performance
2806 * In order to avoid having to look at the other cfs_rq's, we use a
2807 * differential update where we store the last value we propagated. This in
2808 * turn allows skipping updates if the differential is 'small'.
2810 * Updating tg's load_avg is necessary before update_cfs_share() (which is
2811 * done) and effective_load() (which is not done because it is too costly).
2813 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2815 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2817 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2818 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2819 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2823 #else /* CONFIG_FAIR_GROUP_SCHED */
2824 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2825 #endif /* CONFIG_FAIR_GROUP_SCHED */
2827 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
2829 if (&this_rq()->cfs == cfs_rq) {
2831 * There are a few boundary cases this might miss but it should
2832 * get called often enough that that should (hopefully) not be
2833 * a real problem -- added to that it only calls on the local
2834 * CPU, so if we enqueue remotely we'll miss an update, but
2835 * the next tick/schedule should update.
2837 * It will not get called when we go idle, because the idle
2838 * thread is a different class (!fair), nor will the utilization
2839 * number include things like RT tasks.
2841 * As is, the util number is not freq-invariant (we'd have to
2842 * implement arch_scale_freq_capacity() for that).
2846 cpufreq_update_util(rq_of(cfs_rq), 0);
2850 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2853 * Unsigned subtract and clamp on underflow.
2855 * Explicitly do a load-store to ensure the intermediate value never hits
2856 * memory. This allows lockless observations without ever seeing the negative
2859 #define sub_positive(_ptr, _val) do { \
2860 typeof(_ptr) ptr = (_ptr); \
2861 typeof(*ptr) val = (_val); \
2862 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2866 WRITE_ONCE(*ptr, res); \
2870 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
2871 * @now: current time, as per cfs_rq_clock_task()
2872 * @cfs_rq: cfs_rq to update
2873 * @update_freq: should we call cfs_rq_util_change() or will the call do so
2875 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
2876 * avg. The immediate corollary is that all (fair) tasks must be attached, see
2877 * post_init_entity_util_avg().
2879 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
2881 * Returns true if the load decayed or we removed load.
2883 * Since both these conditions indicate a changed cfs_rq->avg.load we should
2884 * call update_tg_load_avg() when this function returns true.
2887 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
2889 struct sched_avg *sa = &cfs_rq->avg;
2890 int decayed, removed = 0, removed_util = 0;
2892 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2893 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2894 sub_positive(&sa->load_avg, r);
2895 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
2899 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2900 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2901 sub_positive(&sa->util_avg, r);
2902 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
2906 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2907 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2909 #ifndef CONFIG_64BIT
2911 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2914 /* Trace CPU load, unless cfs_rq belongs to a non-root task_group */
2915 if (cfs_rq == &rq_of(cfs_rq)->cfs)
2916 trace_sched_load_avg_cpu(cpu_of(rq_of(cfs_rq)), cfs_rq);
2918 if (update_freq && (decayed || removed_util))
2919 cfs_rq_util_change(cfs_rq);
2921 return decayed || removed;
2924 /* Update task and its cfs_rq load average */
2925 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2927 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2928 u64 now = cfs_rq_clock_task(cfs_rq);
2929 int cpu = cpu_of(rq_of(cfs_rq));
2932 * Track task load average for carrying it to new CPU after migrated, and
2933 * track group sched_entity load average for task_h_load calc in migration
2935 __update_load_avg(now, cpu, &se->avg,
2936 se->on_rq * scale_load_down(se->load.weight),
2937 cfs_rq->curr == se, NULL);
2939 if (update_cfs_rq_load_avg(now, cfs_rq, true) && update_tg)
2940 update_tg_load_avg(cfs_rq, 0);
2942 if (entity_is_task(se))
2943 trace_sched_load_avg_task(task_of(se), &se->avg);
2947 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
2948 * @cfs_rq: cfs_rq to attach to
2949 * @se: sched_entity to attach
2951 * Must call update_cfs_rq_load_avg() before this, since we rely on
2952 * cfs_rq->avg.last_update_time being current.
2954 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2956 if (!sched_feat(ATTACH_AGE_LOAD))
2960 * If we got migrated (either between CPUs or between cgroups) we'll
2961 * have aged the average right before clearing @last_update_time.
2963 if (se->avg.last_update_time) {
2964 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2965 &se->avg, 0, 0, NULL);
2968 * XXX: we could have just aged the entire load away if we've been
2969 * absent from the fair class for too long.
2974 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2975 cfs_rq->avg.load_avg += se->avg.load_avg;
2976 cfs_rq->avg.load_sum += se->avg.load_sum;
2977 cfs_rq->avg.util_avg += se->avg.util_avg;
2978 cfs_rq->avg.util_sum += se->avg.util_sum;
2980 cfs_rq_util_change(cfs_rq);
2984 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
2985 * @cfs_rq: cfs_rq to detach from
2986 * @se: sched_entity to detach
2988 * Must call update_cfs_rq_load_avg() before this, since we rely on
2989 * cfs_rq->avg.last_update_time being current.
2991 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2993 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2994 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2995 cfs_rq->curr == se, NULL);
2997 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2998 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
2999 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3000 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3002 cfs_rq_util_change(cfs_rq);
3005 /* Add the load generated by se into cfs_rq's load average */
3007 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3009 struct sched_avg *sa = &se->avg;
3010 u64 now = cfs_rq_clock_task(cfs_rq);
3011 int migrated, decayed;
3013 migrated = !sa->last_update_time;
3015 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
3016 se->on_rq * scale_load_down(se->load.weight),
3017 cfs_rq->curr == se, NULL);
3020 decayed = update_cfs_rq_load_avg(now, cfs_rq, !migrated);
3022 cfs_rq->runnable_load_avg += sa->load_avg;
3023 cfs_rq->runnable_load_sum += sa->load_sum;
3026 attach_entity_load_avg(cfs_rq, se);
3028 if (decayed || migrated)
3029 update_tg_load_avg(cfs_rq, 0);
3032 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3034 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3036 update_load_avg(se, 1);
3038 cfs_rq->runnable_load_avg =
3039 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
3040 cfs_rq->runnable_load_sum =
3041 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
3044 #ifndef CONFIG_64BIT
3045 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3047 u64 last_update_time_copy;
3048 u64 last_update_time;
3051 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3053 last_update_time = cfs_rq->avg.last_update_time;
3054 } while (last_update_time != last_update_time_copy);
3056 return last_update_time;
3059 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3061 return cfs_rq->avg.last_update_time;
3066 * Synchronize entity load avg of dequeued entity without locking
3069 void sync_entity_load_avg(struct sched_entity *se)
3071 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3072 u64 last_update_time;
3074 last_update_time = cfs_rq_last_update_time(cfs_rq);
3075 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
3079 * Task first catches up with cfs_rq, and then subtract
3080 * itself from the cfs_rq (task must be off the queue now).
3082 void remove_entity_load_avg(struct sched_entity *se)
3084 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3087 * Newly created task or never used group entity should not be removed
3088 * from its (source) cfs_rq
3090 if (se->avg.last_update_time == 0)
3093 sync_entity_load_avg(se);
3094 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3095 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3099 * Update the rq's load with the elapsed running time before entering
3100 * idle. if the last scheduled task is not a CFS task, idle_enter will
3101 * be the only way to update the runnable statistic.
3103 void idle_enter_fair(struct rq *this_rq)
3108 * Update the rq's load with the elapsed idle time before a task is
3109 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
3110 * be the only way to update the runnable statistic.
3112 void idle_exit_fair(struct rq *this_rq)
3116 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3118 return cfs_rq->runnable_load_avg;
3121 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3123 return cfs_rq->avg.load_avg;
3126 static int idle_balance(struct rq *this_rq);
3128 #else /* CONFIG_SMP */
3130 static inline void update_load_avg(struct sched_entity *se, int update_tg)
3132 cpufreq_update_util(rq_of(cfs_rq_of(se)), 0);
3136 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3138 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3139 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3142 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3144 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3146 static inline int idle_balance(struct rq *rq)
3151 #endif /* CONFIG_SMP */
3153 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
3155 #ifdef CONFIG_SCHEDSTATS
3156 struct task_struct *tsk = NULL;
3158 if (entity_is_task(se))
3161 if (se->statistics.sleep_start) {
3162 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
3167 if (unlikely(delta > se->statistics.sleep_max))
3168 se->statistics.sleep_max = delta;
3170 se->statistics.sleep_start = 0;
3171 se->statistics.sum_sleep_runtime += delta;
3174 account_scheduler_latency(tsk, delta >> 10, 1);
3175 trace_sched_stat_sleep(tsk, delta);
3178 if (se->statistics.block_start) {
3179 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
3184 if (unlikely(delta > se->statistics.block_max))
3185 se->statistics.block_max = delta;
3187 se->statistics.block_start = 0;
3188 se->statistics.sum_sleep_runtime += delta;
3191 if (tsk->in_iowait) {
3192 se->statistics.iowait_sum += delta;
3193 se->statistics.iowait_count++;
3194 trace_sched_stat_iowait(tsk, delta);
3197 trace_sched_stat_blocked(tsk, delta);
3198 trace_sched_blocked_reason(tsk);
3201 * Blocking time is in units of nanosecs, so shift by
3202 * 20 to get a milliseconds-range estimation of the
3203 * amount of time that the task spent sleeping:
3205 if (unlikely(prof_on == SLEEP_PROFILING)) {
3206 profile_hits(SLEEP_PROFILING,
3207 (void *)get_wchan(tsk),
3210 account_scheduler_latency(tsk, delta >> 10, 0);
3216 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3218 #ifdef CONFIG_SCHED_DEBUG
3219 s64 d = se->vruntime - cfs_rq->min_vruntime;
3224 if (d > 3*sysctl_sched_latency)
3225 schedstat_inc(cfs_rq, nr_spread_over);
3230 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3232 u64 vruntime = cfs_rq->min_vruntime;
3235 * The 'current' period is already promised to the current tasks,
3236 * however the extra weight of the new task will slow them down a
3237 * little, place the new task so that it fits in the slot that
3238 * stays open at the end.
3240 if (initial && sched_feat(START_DEBIT))
3241 vruntime += sched_vslice(cfs_rq, se);
3243 /* sleeps up to a single latency don't count. */
3245 unsigned long thresh = sysctl_sched_latency;
3248 * Halve their sleep time's effect, to allow
3249 * for a gentler effect of sleepers:
3251 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3257 /* ensure we never gain time by being placed backwards. */
3258 se->vruntime = max_vruntime(se->vruntime, vruntime);
3261 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3264 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3267 * Update the normalized vruntime before updating min_vruntime
3268 * through calling update_curr().
3270 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3271 se->vruntime += cfs_rq->min_vruntime;
3274 * Update run-time statistics of the 'current'.
3276 update_curr(cfs_rq);
3277 enqueue_entity_load_avg(cfs_rq, se);
3278 account_entity_enqueue(cfs_rq, se);
3279 update_cfs_shares(cfs_rq);
3281 if (flags & ENQUEUE_WAKEUP) {
3282 place_entity(cfs_rq, se, 0);
3283 enqueue_sleeper(cfs_rq, se);
3286 update_stats_enqueue(cfs_rq, se);
3287 check_spread(cfs_rq, se);
3288 if (se != cfs_rq->curr)
3289 __enqueue_entity(cfs_rq, se);
3292 if (cfs_rq->nr_running == 1) {
3293 list_add_leaf_cfs_rq(cfs_rq);
3294 check_enqueue_throttle(cfs_rq);
3298 static void __clear_buddies_last(struct sched_entity *se)
3300 for_each_sched_entity(se) {
3301 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3302 if (cfs_rq->last != se)
3305 cfs_rq->last = NULL;
3309 static void __clear_buddies_next(struct sched_entity *se)
3311 for_each_sched_entity(se) {
3312 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3313 if (cfs_rq->next != se)
3316 cfs_rq->next = NULL;
3320 static void __clear_buddies_skip(struct sched_entity *se)
3322 for_each_sched_entity(se) {
3323 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3324 if (cfs_rq->skip != se)
3327 cfs_rq->skip = NULL;
3331 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3333 if (cfs_rq->last == se)
3334 __clear_buddies_last(se);
3336 if (cfs_rq->next == se)
3337 __clear_buddies_next(se);
3339 if (cfs_rq->skip == se)
3340 __clear_buddies_skip(se);
3343 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3346 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3349 * Update run-time statistics of the 'current'.
3351 update_curr(cfs_rq);
3352 dequeue_entity_load_avg(cfs_rq, se);
3354 update_stats_dequeue(cfs_rq, se);
3355 if (flags & DEQUEUE_SLEEP) {
3356 #ifdef CONFIG_SCHEDSTATS
3357 if (entity_is_task(se)) {
3358 struct task_struct *tsk = task_of(se);
3360 if (tsk->state & TASK_INTERRUPTIBLE)
3361 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3362 if (tsk->state & TASK_UNINTERRUPTIBLE)
3363 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3368 clear_buddies(cfs_rq, se);
3370 if (se != cfs_rq->curr)
3371 __dequeue_entity(cfs_rq, se);
3373 account_entity_dequeue(cfs_rq, se);
3376 * Normalize the entity after updating the min_vruntime because the
3377 * update can refer to the ->curr item and we need to reflect this
3378 * movement in our normalized position.
3380 if (!(flags & DEQUEUE_SLEEP))
3381 se->vruntime -= cfs_rq->min_vruntime;
3383 /* return excess runtime on last dequeue */
3384 return_cfs_rq_runtime(cfs_rq);
3386 update_min_vruntime(cfs_rq);
3387 update_cfs_shares(cfs_rq);
3391 * Preempt the current task with a newly woken task if needed:
3394 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3396 unsigned long ideal_runtime, delta_exec;
3397 struct sched_entity *se;
3400 ideal_runtime = sched_slice(cfs_rq, curr);
3401 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3402 if (delta_exec > ideal_runtime) {
3403 resched_curr(rq_of(cfs_rq));
3405 * The current task ran long enough, ensure it doesn't get
3406 * re-elected due to buddy favours.
3408 clear_buddies(cfs_rq, curr);
3413 * Ensure that a task that missed wakeup preemption by a
3414 * narrow margin doesn't have to wait for a full slice.
3415 * This also mitigates buddy induced latencies under load.
3417 if (delta_exec < sysctl_sched_min_granularity)
3420 se = __pick_first_entity(cfs_rq);
3421 delta = curr->vruntime - se->vruntime;
3426 if (delta > ideal_runtime)
3427 resched_curr(rq_of(cfs_rq));
3431 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3433 /* 'current' is not kept within the tree. */
3436 * Any task has to be enqueued before it get to execute on
3437 * a CPU. So account for the time it spent waiting on the
3440 update_stats_wait_end(cfs_rq, se);
3441 __dequeue_entity(cfs_rq, se);
3442 update_load_avg(se, 1);
3445 update_stats_curr_start(cfs_rq, se);
3447 #ifdef CONFIG_SCHEDSTATS
3449 * Track our maximum slice length, if the CPU's load is at
3450 * least twice that of our own weight (i.e. dont track it
3451 * when there are only lesser-weight tasks around):
3453 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3454 se->statistics.slice_max = max(se->statistics.slice_max,
3455 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3458 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3462 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3465 * Pick the next process, keeping these things in mind, in this order:
3466 * 1) keep things fair between processes/task groups
3467 * 2) pick the "next" process, since someone really wants that to run
3468 * 3) pick the "last" process, for cache locality
3469 * 4) do not run the "skip" process, if something else is available
3471 static struct sched_entity *
3472 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3474 struct sched_entity *left = __pick_first_entity(cfs_rq);
3475 struct sched_entity *se;
3478 * If curr is set we have to see if its left of the leftmost entity
3479 * still in the tree, provided there was anything in the tree at all.
3481 if (!left || (curr && entity_before(curr, left)))
3484 se = left; /* ideally we run the leftmost entity */
3487 * Avoid running the skip buddy, if running something else can
3488 * be done without getting too unfair.
3490 if (cfs_rq->skip == se) {
3491 struct sched_entity *second;
3494 second = __pick_first_entity(cfs_rq);
3496 second = __pick_next_entity(se);
3497 if (!second || (curr && entity_before(curr, second)))
3501 if (second && wakeup_preempt_entity(second, left) < 1)
3506 * Prefer last buddy, try to return the CPU to a preempted task.
3508 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3512 * Someone really wants this to run. If it's not unfair, run it.
3514 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3517 clear_buddies(cfs_rq, se);
3522 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3524 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3527 * If still on the runqueue then deactivate_task()
3528 * was not called and update_curr() has to be done:
3531 update_curr(cfs_rq);
3533 /* throttle cfs_rqs exceeding runtime */
3534 check_cfs_rq_runtime(cfs_rq);
3536 check_spread(cfs_rq, prev);
3538 update_stats_wait_start(cfs_rq, prev);
3539 /* Put 'current' back into the tree. */
3540 __enqueue_entity(cfs_rq, prev);
3541 /* in !on_rq case, update occurred at dequeue */
3542 update_load_avg(prev, 0);
3544 cfs_rq->curr = NULL;
3548 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3551 * Update run-time statistics of the 'current'.
3553 update_curr(cfs_rq);
3556 * Ensure that runnable average is periodically updated.
3558 update_load_avg(curr, 1);
3559 update_cfs_shares(cfs_rq);
3561 #ifdef CONFIG_SCHED_HRTICK
3563 * queued ticks are scheduled to match the slice, so don't bother
3564 * validating it and just reschedule.
3567 resched_curr(rq_of(cfs_rq));
3571 * don't let the period tick interfere with the hrtick preemption
3573 if (!sched_feat(DOUBLE_TICK) &&
3574 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3578 if (cfs_rq->nr_running > 1)
3579 check_preempt_tick(cfs_rq, curr);
3583 /**************************************************
3584 * CFS bandwidth control machinery
3587 #ifdef CONFIG_CFS_BANDWIDTH
3589 #ifdef HAVE_JUMP_LABEL
3590 static struct static_key __cfs_bandwidth_used;
3592 static inline bool cfs_bandwidth_used(void)
3594 return static_key_false(&__cfs_bandwidth_used);
3597 void cfs_bandwidth_usage_inc(void)
3599 static_key_slow_inc(&__cfs_bandwidth_used);
3602 void cfs_bandwidth_usage_dec(void)
3604 static_key_slow_dec(&__cfs_bandwidth_used);
3606 #else /* HAVE_JUMP_LABEL */
3607 static bool cfs_bandwidth_used(void)
3612 void cfs_bandwidth_usage_inc(void) {}
3613 void cfs_bandwidth_usage_dec(void) {}
3614 #endif /* HAVE_JUMP_LABEL */
3617 * default period for cfs group bandwidth.
3618 * default: 0.1s, units: nanoseconds
3620 static inline u64 default_cfs_period(void)
3622 return 100000000ULL;
3625 static inline u64 sched_cfs_bandwidth_slice(void)
3627 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3631 * Replenish runtime according to assigned quota and update expiration time.
3632 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3633 * additional synchronization around rq->lock.
3635 * requires cfs_b->lock
3637 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3641 if (cfs_b->quota == RUNTIME_INF)
3644 now = sched_clock_cpu(smp_processor_id());
3645 cfs_b->runtime = cfs_b->quota;
3646 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3649 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3651 return &tg->cfs_bandwidth;
3654 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3655 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3657 if (unlikely(cfs_rq->throttle_count))
3658 return cfs_rq->throttled_clock_task;
3660 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3663 /* returns 0 on failure to allocate runtime */
3664 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3666 struct task_group *tg = cfs_rq->tg;
3667 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3668 u64 amount = 0, min_amount, expires;
3670 /* note: this is a positive sum as runtime_remaining <= 0 */
3671 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3673 raw_spin_lock(&cfs_b->lock);
3674 if (cfs_b->quota == RUNTIME_INF)
3675 amount = min_amount;
3677 start_cfs_bandwidth(cfs_b);
3679 if (cfs_b->runtime > 0) {
3680 amount = min(cfs_b->runtime, min_amount);
3681 cfs_b->runtime -= amount;
3685 expires = cfs_b->runtime_expires;
3686 raw_spin_unlock(&cfs_b->lock);
3688 cfs_rq->runtime_remaining += amount;
3690 * we may have advanced our local expiration to account for allowed
3691 * spread between our sched_clock and the one on which runtime was
3694 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3695 cfs_rq->runtime_expires = expires;
3697 return cfs_rq->runtime_remaining > 0;
3701 * Note: This depends on the synchronization provided by sched_clock and the
3702 * fact that rq->clock snapshots this value.
3704 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3706 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3708 /* if the deadline is ahead of our clock, nothing to do */
3709 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3712 if (cfs_rq->runtime_remaining < 0)
3716 * If the local deadline has passed we have to consider the
3717 * possibility that our sched_clock is 'fast' and the global deadline
3718 * has not truly expired.
3720 * Fortunately we can check determine whether this the case by checking
3721 * whether the global deadline has advanced. It is valid to compare
3722 * cfs_b->runtime_expires without any locks since we only care about
3723 * exact equality, so a partial write will still work.
3726 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3727 /* extend local deadline, drift is bounded above by 2 ticks */
3728 cfs_rq->runtime_expires += TICK_NSEC;
3730 /* global deadline is ahead, expiration has passed */
3731 cfs_rq->runtime_remaining = 0;
3735 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3737 /* dock delta_exec before expiring quota (as it could span periods) */
3738 cfs_rq->runtime_remaining -= delta_exec;
3739 expire_cfs_rq_runtime(cfs_rq);
3741 if (likely(cfs_rq->runtime_remaining > 0))
3745 * if we're unable to extend our runtime we resched so that the active
3746 * hierarchy can be throttled
3748 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3749 resched_curr(rq_of(cfs_rq));
3752 static __always_inline
3753 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3755 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3758 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3761 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3763 return cfs_bandwidth_used() && cfs_rq->throttled;
3766 /* check whether cfs_rq, or any parent, is throttled */
3767 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3769 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3773 * Ensure that neither of the group entities corresponding to src_cpu or
3774 * dest_cpu are members of a throttled hierarchy when performing group
3775 * load-balance operations.
3777 static inline int throttled_lb_pair(struct task_group *tg,
3778 int src_cpu, int dest_cpu)
3780 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3782 src_cfs_rq = tg->cfs_rq[src_cpu];
3783 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3785 return throttled_hierarchy(src_cfs_rq) ||
3786 throttled_hierarchy(dest_cfs_rq);
3789 /* updated child weight may affect parent so we have to do this bottom up */
3790 static int tg_unthrottle_up(struct task_group *tg, void *data)
3792 struct rq *rq = data;
3793 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3795 cfs_rq->throttle_count--;
3797 if (!cfs_rq->throttle_count) {
3798 /* adjust cfs_rq_clock_task() */
3799 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3800 cfs_rq->throttled_clock_task;
3807 static int tg_throttle_down(struct task_group *tg, void *data)
3809 struct rq *rq = data;
3810 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3812 /* group is entering throttled state, stop time */
3813 if (!cfs_rq->throttle_count)
3814 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3815 cfs_rq->throttle_count++;
3820 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3822 struct rq *rq = rq_of(cfs_rq);
3823 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3824 struct sched_entity *se;
3825 long task_delta, dequeue = 1;
3828 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3830 /* freeze hierarchy runnable averages while throttled */
3832 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3835 task_delta = cfs_rq->h_nr_running;
3836 for_each_sched_entity(se) {
3837 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3838 /* throttled entity or throttle-on-deactivate */
3843 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3844 qcfs_rq->h_nr_running -= task_delta;
3846 if (qcfs_rq->load.weight)
3851 sub_nr_running(rq, task_delta);
3853 cfs_rq->throttled = 1;
3854 cfs_rq->throttled_clock = rq_clock(rq);
3855 raw_spin_lock(&cfs_b->lock);
3856 empty = list_empty(&cfs_b->throttled_cfs_rq);
3859 * Add to the _head_ of the list, so that an already-started
3860 * distribute_cfs_runtime will not see us
3862 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3865 * If we're the first throttled task, make sure the bandwidth
3869 start_cfs_bandwidth(cfs_b);
3871 raw_spin_unlock(&cfs_b->lock);
3874 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3876 struct rq *rq = rq_of(cfs_rq);
3877 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3878 struct sched_entity *se;
3882 se = cfs_rq->tg->se[cpu_of(rq)];
3884 cfs_rq->throttled = 0;
3886 update_rq_clock(rq);
3888 raw_spin_lock(&cfs_b->lock);
3889 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3890 list_del_rcu(&cfs_rq->throttled_list);
3891 raw_spin_unlock(&cfs_b->lock);
3893 /* update hierarchical throttle state */
3894 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3896 if (!cfs_rq->load.weight)
3899 task_delta = cfs_rq->h_nr_running;
3900 for_each_sched_entity(se) {
3904 cfs_rq = cfs_rq_of(se);
3906 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3907 cfs_rq->h_nr_running += task_delta;
3909 if (cfs_rq_throttled(cfs_rq))
3914 add_nr_running(rq, task_delta);
3916 /* determine whether we need to wake up potentially idle cpu */
3917 if (rq->curr == rq->idle && rq->cfs.nr_running)
3921 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3922 u64 remaining, u64 expires)
3924 struct cfs_rq *cfs_rq;
3926 u64 starting_runtime = remaining;
3929 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3931 struct rq *rq = rq_of(cfs_rq);
3933 raw_spin_lock(&rq->lock);
3934 if (!cfs_rq_throttled(cfs_rq))
3937 runtime = -cfs_rq->runtime_remaining + 1;
3938 if (runtime > remaining)
3939 runtime = remaining;
3940 remaining -= runtime;
3942 cfs_rq->runtime_remaining += runtime;
3943 cfs_rq->runtime_expires = expires;
3945 /* we check whether we're throttled above */
3946 if (cfs_rq->runtime_remaining > 0)
3947 unthrottle_cfs_rq(cfs_rq);
3950 raw_spin_unlock(&rq->lock);
3957 return starting_runtime - remaining;
3961 * Responsible for refilling a task_group's bandwidth and unthrottling its
3962 * cfs_rqs as appropriate. If there has been no activity within the last
3963 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3964 * used to track this state.
3966 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3968 u64 runtime, runtime_expires;
3971 /* no need to continue the timer with no bandwidth constraint */
3972 if (cfs_b->quota == RUNTIME_INF)
3973 goto out_deactivate;
3975 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3976 cfs_b->nr_periods += overrun;
3979 * idle depends on !throttled (for the case of a large deficit), and if
3980 * we're going inactive then everything else can be deferred
3982 if (cfs_b->idle && !throttled)
3983 goto out_deactivate;
3985 __refill_cfs_bandwidth_runtime(cfs_b);
3988 /* mark as potentially idle for the upcoming period */
3993 /* account preceding periods in which throttling occurred */
3994 cfs_b->nr_throttled += overrun;
3996 runtime_expires = cfs_b->runtime_expires;
3999 * This check is repeated as we are holding onto the new bandwidth while
4000 * we unthrottle. This can potentially race with an unthrottled group
4001 * trying to acquire new bandwidth from the global pool. This can result
4002 * in us over-using our runtime if it is all used during this loop, but
4003 * only by limited amounts in that extreme case.
4005 while (throttled && cfs_b->runtime > 0) {
4006 runtime = cfs_b->runtime;
4007 raw_spin_unlock(&cfs_b->lock);
4008 /* we can't nest cfs_b->lock while distributing bandwidth */
4009 runtime = distribute_cfs_runtime(cfs_b, runtime,
4011 raw_spin_lock(&cfs_b->lock);
4013 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4015 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4019 * While we are ensured activity in the period following an
4020 * unthrottle, this also covers the case in which the new bandwidth is
4021 * insufficient to cover the existing bandwidth deficit. (Forcing the
4022 * timer to remain active while there are any throttled entities.)
4032 /* a cfs_rq won't donate quota below this amount */
4033 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4034 /* minimum remaining period time to redistribute slack quota */
4035 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4036 /* how long we wait to gather additional slack before distributing */
4037 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4040 * Are we near the end of the current quota period?
4042 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4043 * hrtimer base being cleared by hrtimer_start. In the case of
4044 * migrate_hrtimers, base is never cleared, so we are fine.
4046 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4048 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4051 /* if the call-back is running a quota refresh is already occurring */
4052 if (hrtimer_callback_running(refresh_timer))
4055 /* is a quota refresh about to occur? */
4056 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4057 if (remaining < min_expire)
4063 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4065 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4067 /* if there's a quota refresh soon don't bother with slack */
4068 if (runtime_refresh_within(cfs_b, min_left))
4071 hrtimer_start(&cfs_b->slack_timer,
4072 ns_to_ktime(cfs_bandwidth_slack_period),
4076 /* we know any runtime found here is valid as update_curr() precedes return */
4077 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4079 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4080 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4082 if (slack_runtime <= 0)
4085 raw_spin_lock(&cfs_b->lock);
4086 if (cfs_b->quota != RUNTIME_INF &&
4087 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4088 cfs_b->runtime += slack_runtime;
4090 /* we are under rq->lock, defer unthrottling using a timer */
4091 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4092 !list_empty(&cfs_b->throttled_cfs_rq))
4093 start_cfs_slack_bandwidth(cfs_b);
4095 raw_spin_unlock(&cfs_b->lock);
4097 /* even if it's not valid for return we don't want to try again */
4098 cfs_rq->runtime_remaining -= slack_runtime;
4101 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4103 if (!cfs_bandwidth_used())
4106 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4109 __return_cfs_rq_runtime(cfs_rq);
4113 * This is done with a timer (instead of inline with bandwidth return) since
4114 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4116 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4118 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4121 /* confirm we're still not at a refresh boundary */
4122 raw_spin_lock(&cfs_b->lock);
4123 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4124 raw_spin_unlock(&cfs_b->lock);
4128 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4129 runtime = cfs_b->runtime;
4131 expires = cfs_b->runtime_expires;
4132 raw_spin_unlock(&cfs_b->lock);
4137 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4139 raw_spin_lock(&cfs_b->lock);
4140 if (expires == cfs_b->runtime_expires)
4141 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4142 raw_spin_unlock(&cfs_b->lock);
4146 * When a group wakes up we want to make sure that its quota is not already
4147 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4148 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4150 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4152 if (!cfs_bandwidth_used())
4155 /* Synchronize hierarchical throttle counter: */
4156 if (unlikely(!cfs_rq->throttle_uptodate)) {
4157 struct rq *rq = rq_of(cfs_rq);
4158 struct cfs_rq *pcfs_rq;
4159 struct task_group *tg;
4161 cfs_rq->throttle_uptodate = 1;
4163 /* Get closest up-to-date node, because leaves go first: */
4164 for (tg = cfs_rq->tg->parent; tg; tg = tg->parent) {
4165 pcfs_rq = tg->cfs_rq[cpu_of(rq)];
4166 if (pcfs_rq->throttle_uptodate)
4170 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4171 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4175 /* an active group must be handled by the update_curr()->put() path */
4176 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4179 /* ensure the group is not already throttled */
4180 if (cfs_rq_throttled(cfs_rq))
4183 /* update runtime allocation */
4184 account_cfs_rq_runtime(cfs_rq, 0);
4185 if (cfs_rq->runtime_remaining <= 0)
4186 throttle_cfs_rq(cfs_rq);
4189 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4190 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4192 if (!cfs_bandwidth_used())
4195 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4199 * it's possible for a throttled entity to be forced into a running
4200 * state (e.g. set_curr_task), in this case we're finished.
4202 if (cfs_rq_throttled(cfs_rq))
4205 throttle_cfs_rq(cfs_rq);
4209 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4211 struct cfs_bandwidth *cfs_b =
4212 container_of(timer, struct cfs_bandwidth, slack_timer);
4214 do_sched_cfs_slack_timer(cfs_b);
4216 return HRTIMER_NORESTART;
4219 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4221 struct cfs_bandwidth *cfs_b =
4222 container_of(timer, struct cfs_bandwidth, period_timer);
4226 raw_spin_lock(&cfs_b->lock);
4228 overrun = hrtimer_forward_now(timer, cfs_b->period);
4232 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4235 cfs_b->period_active = 0;
4236 raw_spin_unlock(&cfs_b->lock);
4238 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4241 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4243 raw_spin_lock_init(&cfs_b->lock);
4245 cfs_b->quota = RUNTIME_INF;
4246 cfs_b->period = ns_to_ktime(default_cfs_period());
4248 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4249 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4250 cfs_b->period_timer.function = sched_cfs_period_timer;
4251 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4252 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4255 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4257 cfs_rq->runtime_enabled = 0;
4258 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4261 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4263 lockdep_assert_held(&cfs_b->lock);
4265 if (!cfs_b->period_active) {
4266 cfs_b->period_active = 1;
4267 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4268 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4272 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4274 /* init_cfs_bandwidth() was not called */
4275 if (!cfs_b->throttled_cfs_rq.next)
4278 hrtimer_cancel(&cfs_b->period_timer);
4279 hrtimer_cancel(&cfs_b->slack_timer);
4282 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4284 struct cfs_rq *cfs_rq;
4286 for_each_leaf_cfs_rq(rq, cfs_rq) {
4287 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4289 raw_spin_lock(&cfs_b->lock);
4290 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4291 raw_spin_unlock(&cfs_b->lock);
4295 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4297 struct cfs_rq *cfs_rq;
4299 for_each_leaf_cfs_rq(rq, cfs_rq) {
4300 if (!cfs_rq->runtime_enabled)
4304 * clock_task is not advancing so we just need to make sure
4305 * there's some valid quota amount
4307 cfs_rq->runtime_remaining = 1;
4309 * Offline rq is schedulable till cpu is completely disabled
4310 * in take_cpu_down(), so we prevent new cfs throttling here.
4312 cfs_rq->runtime_enabled = 0;
4314 if (cfs_rq_throttled(cfs_rq))
4315 unthrottle_cfs_rq(cfs_rq);
4319 #else /* CONFIG_CFS_BANDWIDTH */
4320 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4322 return rq_clock_task(rq_of(cfs_rq));
4325 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4326 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4327 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4328 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4330 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4335 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4340 static inline int throttled_lb_pair(struct task_group *tg,
4341 int src_cpu, int dest_cpu)
4346 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4348 #ifdef CONFIG_FAIR_GROUP_SCHED
4349 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4352 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4356 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4357 static inline void update_runtime_enabled(struct rq *rq) {}
4358 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4360 #endif /* CONFIG_CFS_BANDWIDTH */
4362 /**************************************************
4363 * CFS operations on tasks:
4366 #ifdef CONFIG_SCHED_HRTICK
4367 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4369 struct sched_entity *se = &p->se;
4370 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4372 WARN_ON(task_rq(p) != rq);
4374 if (cfs_rq->nr_running > 1) {
4375 u64 slice = sched_slice(cfs_rq, se);
4376 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4377 s64 delta = slice - ran;
4384 hrtick_start(rq, delta);
4389 * called from enqueue/dequeue and updates the hrtick when the
4390 * current task is from our class and nr_running is low enough
4393 static void hrtick_update(struct rq *rq)
4395 struct task_struct *curr = rq->curr;
4397 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4400 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4401 hrtick_start_fair(rq, curr);
4403 #else /* !CONFIG_SCHED_HRTICK */
4405 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4409 static inline void hrtick_update(struct rq *rq)
4415 static bool cpu_overutilized(int cpu);
4416 unsigned long boosted_cpu_util(int cpu);
4418 #define boosted_cpu_util(cpu) cpu_util(cpu)
4422 static void update_capacity_of(int cpu)
4424 unsigned long req_cap;
4429 /* Convert scale-invariant capacity to cpu. */
4430 req_cap = boosted_cpu_util(cpu);
4431 req_cap = req_cap * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
4432 set_cfs_cpu_capacity(cpu, true, req_cap);
4437 * The enqueue_task method is called before nr_running is
4438 * increased. Here we update the fair scheduling stats and
4439 * then put the task into the rbtree:
4442 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4444 struct cfs_rq *cfs_rq;
4445 struct sched_entity *se = &p->se;
4447 int task_new = flags & ENQUEUE_WAKEUP_NEW;
4448 int task_wakeup = flags & ENQUEUE_WAKEUP;
4452 * If in_iowait is set, the code below may not trigger any cpufreq
4453 * utilization updates, so do it here explicitly with the IOWAIT flag
4457 cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_IOWAIT);
4459 for_each_sched_entity(se) {
4462 cfs_rq = cfs_rq_of(se);
4463 enqueue_entity(cfs_rq, se, flags);
4466 * end evaluation on encountering a throttled cfs_rq
4468 * note: in the case of encountering a throttled cfs_rq we will
4469 * post the final h_nr_running increment below.
4471 if (cfs_rq_throttled(cfs_rq))
4473 cfs_rq->h_nr_running++;
4474 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4476 flags = ENQUEUE_WAKEUP;
4479 for_each_sched_entity(se) {
4480 cfs_rq = cfs_rq_of(se);
4481 cfs_rq->h_nr_running++;
4482 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4484 if (cfs_rq_throttled(cfs_rq))
4487 update_load_avg(se, 1);
4488 update_cfs_shares(cfs_rq);
4492 add_nr_running(rq, 1);
4497 * Update SchedTune accounting.
4499 * We do it before updating the CPU capacity to ensure the
4500 * boost value of the current task is accounted for in the
4501 * selection of the OPP.
4503 * We do it also in the case where we enqueue a throttled task;
4504 * we could argue that a throttled task should not boost a CPU,
4506 * a) properly implementing CPU boosting considering throttled
4507 * tasks will increase a lot the complexity of the solution
4508 * b) it's not easy to quantify the benefits introduced by
4509 * such a more complex solution.
4510 * Thus, for the time being we go for the simple solution and boost
4511 * also for throttled RQs.
4513 schedtune_enqueue_task(p, cpu_of(rq));
4516 walt_inc_cumulative_runnable_avg(rq, p);
4517 if (!task_new && !rq->rd->overutilized &&
4518 cpu_overutilized(rq->cpu)) {
4519 rq->rd->overutilized = true;
4520 trace_sched_overutilized(true);
4524 * We want to potentially trigger a freq switch
4525 * request only for tasks that are waking up; this is
4526 * because we get here also during load balancing, but
4527 * in these cases it seems wise to trigger as single
4528 * request after load balancing is done.
4530 if (task_new || task_wakeup)
4531 update_capacity_of(cpu_of(rq));
4534 #endif /* CONFIG_SMP */
4538 static void set_next_buddy(struct sched_entity *se);
4541 * The dequeue_task method is called before nr_running is
4542 * decreased. We remove the task from the rbtree and
4543 * update the fair scheduling stats:
4545 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4547 struct cfs_rq *cfs_rq;
4548 struct sched_entity *se = &p->se;
4549 int task_sleep = flags & DEQUEUE_SLEEP;
4551 for_each_sched_entity(se) {
4552 cfs_rq = cfs_rq_of(se);
4553 dequeue_entity(cfs_rq, se, flags);
4556 * end evaluation on encountering a throttled cfs_rq
4558 * note: in the case of encountering a throttled cfs_rq we will
4559 * post the final h_nr_running decrement below.
4561 if (cfs_rq_throttled(cfs_rq))
4563 cfs_rq->h_nr_running--;
4564 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4566 /* Don't dequeue parent if it has other entities besides us */
4567 if (cfs_rq->load.weight) {
4568 /* Avoid re-evaluating load for this entity: */
4569 se = parent_entity(se);
4571 * Bias pick_next to pick a task from this cfs_rq, as
4572 * p is sleeping when it is within its sched_slice.
4574 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4578 flags |= DEQUEUE_SLEEP;
4581 for_each_sched_entity(se) {
4582 cfs_rq = cfs_rq_of(se);
4583 cfs_rq->h_nr_running--;
4584 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4586 if (cfs_rq_throttled(cfs_rq))
4589 update_load_avg(se, 1);
4590 update_cfs_shares(cfs_rq);
4594 sub_nr_running(rq, 1);
4599 * Update SchedTune accounting
4601 * We do it before updating the CPU capacity to ensure the
4602 * boost value of the current task is accounted for in the
4603 * selection of the OPP.
4605 schedtune_dequeue_task(p, cpu_of(rq));
4608 walt_dec_cumulative_runnable_avg(rq, p);
4611 * We want to potentially trigger a freq switch
4612 * request only for tasks that are going to sleep;
4613 * this is because we get here also during load
4614 * balancing, but in these cases it seems wise to
4615 * trigger as single request after load balancing is
4619 if (rq->cfs.nr_running)
4620 update_capacity_of(cpu_of(rq));
4621 else if (sched_freq())
4622 set_cfs_cpu_capacity(cpu_of(rq), false, 0);
4626 #endif /* CONFIG_SMP */
4634 * per rq 'load' arrray crap; XXX kill this.
4638 * The exact cpuload at various idx values, calculated at every tick would be
4639 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4641 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4642 * on nth tick when cpu may be busy, then we have:
4643 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4644 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4646 * decay_load_missed() below does efficient calculation of
4647 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4648 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4650 * The calculation is approximated on a 128 point scale.
4651 * degrade_zero_ticks is the number of ticks after which load at any
4652 * particular idx is approximated to be zero.
4653 * degrade_factor is a precomputed table, a row for each load idx.
4654 * Each column corresponds to degradation factor for a power of two ticks,
4655 * based on 128 point scale.
4657 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4658 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4660 * With this power of 2 load factors, we can degrade the load n times
4661 * by looking at 1 bits in n and doing as many mult/shift instead of
4662 * n mult/shifts needed by the exact degradation.
4664 #define DEGRADE_SHIFT 7
4665 static const unsigned char
4666 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4667 static const unsigned char
4668 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4669 {0, 0, 0, 0, 0, 0, 0, 0},
4670 {64, 32, 8, 0, 0, 0, 0, 0},
4671 {96, 72, 40, 12, 1, 0, 0},
4672 {112, 98, 75, 43, 15, 1, 0},
4673 {120, 112, 98, 76, 45, 16, 2} };
4676 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4677 * would be when CPU is idle and so we just decay the old load without
4678 * adding any new load.
4680 static unsigned long
4681 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4685 if (!missed_updates)
4688 if (missed_updates >= degrade_zero_ticks[idx])
4692 return load >> missed_updates;
4694 while (missed_updates) {
4695 if (missed_updates % 2)
4696 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4698 missed_updates >>= 1;
4705 * Update rq->cpu_load[] statistics. This function is usually called every
4706 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4707 * every tick. We fix it up based on jiffies.
4709 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4710 unsigned long pending_updates)
4714 this_rq->nr_load_updates++;
4716 /* Update our load: */
4717 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4718 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4719 unsigned long old_load, new_load;
4721 /* scale is effectively 1 << i now, and >> i divides by scale */
4723 old_load = this_rq->cpu_load[i];
4724 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4725 new_load = this_load;
4727 * Round up the averaging division if load is increasing. This
4728 * prevents us from getting stuck on 9 if the load is 10, for
4731 if (new_load > old_load)
4732 new_load += scale - 1;
4734 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4737 sched_avg_update(this_rq);
4740 /* Used instead of source_load when we know the type == 0 */
4741 static unsigned long weighted_cpuload(const int cpu)
4743 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4746 #ifdef CONFIG_NO_HZ_COMMON
4748 * There is no sane way to deal with nohz on smp when using jiffies because the
4749 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4750 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4752 * Therefore we cannot use the delta approach from the regular tick since that
4753 * would seriously skew the load calculation. However we'll make do for those
4754 * updates happening while idle (nohz_idle_balance) or coming out of idle
4755 * (tick_nohz_idle_exit).
4757 * This means we might still be one tick off for nohz periods.
4761 * Called from nohz_idle_balance() to update the load ratings before doing the
4764 static void update_idle_cpu_load(struct rq *this_rq)
4766 unsigned long curr_jiffies = READ_ONCE(jiffies);
4767 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4768 unsigned long pending_updates;
4771 * bail if there's load or we're actually up-to-date.
4773 if (load || curr_jiffies == this_rq->last_load_update_tick)
4776 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4777 this_rq->last_load_update_tick = curr_jiffies;
4779 __update_cpu_load(this_rq, load, pending_updates);
4783 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4785 void update_cpu_load_nohz(void)
4787 struct rq *this_rq = this_rq();
4788 unsigned long curr_jiffies = READ_ONCE(jiffies);
4789 unsigned long pending_updates;
4791 if (curr_jiffies == this_rq->last_load_update_tick)
4794 raw_spin_lock(&this_rq->lock);
4795 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4796 if (pending_updates) {
4797 this_rq->last_load_update_tick = curr_jiffies;
4799 * We were idle, this means load 0, the current load might be
4800 * !0 due to remote wakeups and the sort.
4802 __update_cpu_load(this_rq, 0, pending_updates);
4804 raw_spin_unlock(&this_rq->lock);
4806 #endif /* CONFIG_NO_HZ */
4809 * Called from scheduler_tick()
4811 void update_cpu_load_active(struct rq *this_rq)
4813 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4815 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4817 this_rq->last_load_update_tick = jiffies;
4818 __update_cpu_load(this_rq, load, 1);
4822 * Return a low guess at the load of a migration-source cpu weighted
4823 * according to the scheduling class and "nice" value.
4825 * We want to under-estimate the load of migration sources, to
4826 * balance conservatively.
4828 static unsigned long source_load(int cpu, int type)
4830 struct rq *rq = cpu_rq(cpu);
4831 unsigned long total = weighted_cpuload(cpu);
4833 if (type == 0 || !sched_feat(LB_BIAS))
4836 return min(rq->cpu_load[type-1], total);
4840 * Return a high guess at the load of a migration-target cpu weighted
4841 * according to the scheduling class and "nice" value.
4843 static unsigned long target_load(int cpu, int type)
4845 struct rq *rq = cpu_rq(cpu);
4846 unsigned long total = weighted_cpuload(cpu);
4848 if (type == 0 || !sched_feat(LB_BIAS))
4851 return max(rq->cpu_load[type-1], total);
4855 static unsigned long cpu_avg_load_per_task(int cpu)
4857 struct rq *rq = cpu_rq(cpu);
4858 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4859 unsigned long load_avg = weighted_cpuload(cpu);
4862 return load_avg / nr_running;
4867 static void record_wakee(struct task_struct *p)
4870 * Rough decay (wiping) for cost saving, don't worry
4871 * about the boundary, really active task won't care
4874 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4875 current->wakee_flips >>= 1;
4876 current->wakee_flip_decay_ts = jiffies;
4879 if (current->last_wakee != p) {
4880 current->last_wakee = p;
4881 current->wakee_flips++;
4885 static void task_waking_fair(struct task_struct *p)
4887 struct sched_entity *se = &p->se;
4888 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4891 #ifndef CONFIG_64BIT
4892 u64 min_vruntime_copy;
4895 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4897 min_vruntime = cfs_rq->min_vruntime;
4898 } while (min_vruntime != min_vruntime_copy);
4900 min_vruntime = cfs_rq->min_vruntime;
4903 se->vruntime -= min_vruntime;
4907 #ifdef CONFIG_FAIR_GROUP_SCHED
4909 * effective_load() calculates the load change as seen from the root_task_group
4911 * Adding load to a group doesn't make a group heavier, but can cause movement
4912 * of group shares between cpus. Assuming the shares were perfectly aligned one
4913 * can calculate the shift in shares.
4915 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4916 * on this @cpu and results in a total addition (subtraction) of @wg to the
4917 * total group weight.
4919 * Given a runqueue weight distribution (rw_i) we can compute a shares
4920 * distribution (s_i) using:
4922 * s_i = rw_i / \Sum rw_j (1)
4924 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4925 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4926 * shares distribution (s_i):
4928 * rw_i = { 2, 4, 1, 0 }
4929 * s_i = { 2/7, 4/7, 1/7, 0 }
4931 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4932 * task used to run on and the CPU the waker is running on), we need to
4933 * compute the effect of waking a task on either CPU and, in case of a sync
4934 * wakeup, compute the effect of the current task going to sleep.
4936 * So for a change of @wl to the local @cpu with an overall group weight change
4937 * of @wl we can compute the new shares distribution (s'_i) using:
4939 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4941 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4942 * differences in waking a task to CPU 0. The additional task changes the
4943 * weight and shares distributions like:
4945 * rw'_i = { 3, 4, 1, 0 }
4946 * s'_i = { 3/8, 4/8, 1/8, 0 }
4948 * We can then compute the difference in effective weight by using:
4950 * dw_i = S * (s'_i - s_i) (3)
4952 * Where 'S' is the group weight as seen by its parent.
4954 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4955 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4956 * 4/7) times the weight of the group.
4958 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4960 struct sched_entity *se = tg->se[cpu];
4962 if (!tg->parent) /* the trivial, non-cgroup case */
4965 for_each_sched_entity(se) {
4966 struct cfs_rq *cfs_rq = se->my_q;
4967 long W, w = cfs_rq_load_avg(cfs_rq);
4972 * W = @wg + \Sum rw_j
4974 W = wg + atomic_long_read(&tg->load_avg);
4976 /* Ensure \Sum rw_j >= rw_i */
4977 W -= cfs_rq->tg_load_avg_contrib;
4986 * wl = S * s'_i; see (2)
4989 wl = (w * (long)tg->shares) / W;
4994 * Per the above, wl is the new se->load.weight value; since
4995 * those are clipped to [MIN_SHARES, ...) do so now. See
4996 * calc_cfs_shares().
4998 if (wl < MIN_SHARES)
5002 * wl = dw_i = S * (s'_i - s_i); see (3)
5004 wl -= se->avg.load_avg;
5007 * Recursively apply this logic to all parent groups to compute
5008 * the final effective load change on the root group. Since
5009 * only the @tg group gets extra weight, all parent groups can
5010 * only redistribute existing shares. @wl is the shift in shares
5011 * resulting from this level per the above.
5020 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
5028 * Returns the current capacity of cpu after applying both
5029 * cpu and freq scaling.
5031 unsigned long capacity_curr_of(int cpu)
5033 return cpu_rq(cpu)->cpu_capacity_orig *
5034 arch_scale_freq_capacity(NULL, cpu)
5035 >> SCHED_CAPACITY_SHIFT;
5038 static inline bool energy_aware(void)
5040 return sched_feat(ENERGY_AWARE);
5044 struct sched_group *sg_top;
5045 struct sched_group *sg_cap;
5052 struct task_struct *task;
5067 * __cpu_norm_util() returns the cpu util relative to a specific capacity,
5068 * i.e. it's busy ratio, in the range [0..SCHED_LOAD_SCALE] which is useful for
5069 * energy calculations. Using the scale-invariant util returned by
5070 * cpu_util() and approximating scale-invariant util by:
5072 * util ~ (curr_freq/max_freq)*1024 * capacity_orig/1024 * running_time/time
5074 * the normalized util can be found using the specific capacity.
5076 * capacity = capacity_orig * curr_freq/max_freq
5078 * norm_util = running_time/time ~ util/capacity
5080 static unsigned long __cpu_norm_util(int cpu, unsigned long capacity, int delta)
5082 int util = __cpu_util(cpu, delta);
5084 if (util >= capacity)
5085 return SCHED_CAPACITY_SCALE;
5087 return (util << SCHED_CAPACITY_SHIFT)/capacity;
5090 static int calc_util_delta(struct energy_env *eenv, int cpu)
5092 if (cpu == eenv->src_cpu)
5093 return -eenv->util_delta;
5094 if (cpu == eenv->dst_cpu)
5095 return eenv->util_delta;
5100 unsigned long group_max_util(struct energy_env *eenv)
5103 unsigned long max_util = 0;
5105 for_each_cpu(i, sched_group_cpus(eenv->sg_cap)) {
5106 delta = calc_util_delta(eenv, i);
5107 max_util = max(max_util, __cpu_util(i, delta));
5114 * group_norm_util() returns the approximated group util relative to it's
5115 * current capacity (busy ratio) in the range [0..SCHED_LOAD_SCALE] for use in
5116 * energy calculations. Since task executions may or may not overlap in time in
5117 * the group the true normalized util is between max(cpu_norm_util(i)) and
5118 * sum(cpu_norm_util(i)) when iterating over all cpus in the group, i. The
5119 * latter is used as the estimate as it leads to a more pessimistic energy
5120 * estimate (more busy).
5123 long group_norm_util(struct energy_env *eenv, struct sched_group *sg)
5126 unsigned long util_sum = 0;
5127 unsigned long capacity = sg->sge->cap_states[eenv->cap_idx].cap;
5129 for_each_cpu(i, sched_group_cpus(sg)) {
5130 delta = calc_util_delta(eenv, i);
5131 util_sum += __cpu_norm_util(i, capacity, delta);
5134 if (util_sum > SCHED_CAPACITY_SCALE)
5135 return SCHED_CAPACITY_SCALE;
5139 static int find_new_capacity(struct energy_env *eenv,
5140 const struct sched_group_energy * const sge)
5143 unsigned long util = group_max_util(eenv);
5145 for (idx = 0; idx < sge->nr_cap_states; idx++) {
5146 if (sge->cap_states[idx].cap >= util)
5150 eenv->cap_idx = idx;
5155 static int group_idle_state(struct sched_group *sg)
5157 int i, state = INT_MAX;
5159 /* Find the shallowest idle state in the sched group. */
5160 for_each_cpu(i, sched_group_cpus(sg))
5161 state = min(state, idle_get_state_idx(cpu_rq(i)));
5163 /* Take non-cpuidle idling into account (active idle/arch_cpu_idle()) */
5170 * sched_group_energy(): Computes the absolute energy consumption of cpus
5171 * belonging to the sched_group including shared resources shared only by
5172 * members of the group. Iterates over all cpus in the hierarchy below the
5173 * sched_group starting from the bottom working it's way up before going to
5174 * the next cpu until all cpus are covered at all levels. The current
5175 * implementation is likely to gather the same util statistics multiple times.
5176 * This can probably be done in a faster but more complex way.
5177 * Note: sched_group_energy() may fail when racing with sched_domain updates.
5179 static int sched_group_energy(struct energy_env *eenv)
5181 struct sched_domain *sd;
5182 int cpu, total_energy = 0;
5183 struct cpumask visit_cpus;
5184 struct sched_group *sg;
5186 WARN_ON(!eenv->sg_top->sge);
5188 cpumask_copy(&visit_cpus, sched_group_cpus(eenv->sg_top));
5190 while (!cpumask_empty(&visit_cpus)) {
5191 struct sched_group *sg_shared_cap = NULL;
5193 cpu = cpumask_first(&visit_cpus);
5196 * Is the group utilization affected by cpus outside this
5199 sd = rcu_dereference(per_cpu(sd_scs, cpu));
5203 * We most probably raced with hotplug; returning a
5204 * wrong energy estimation is better than entering an
5210 sg_shared_cap = sd->parent->groups;
5212 for_each_domain(cpu, sd) {
5215 /* Has this sched_domain already been visited? */
5216 if (sd->child && group_first_cpu(sg) != cpu)
5220 unsigned long group_util;
5221 int sg_busy_energy, sg_idle_energy;
5222 int cap_idx, idle_idx;
5224 if (sg_shared_cap && sg_shared_cap->group_weight >= sg->group_weight)
5225 eenv->sg_cap = sg_shared_cap;
5229 cap_idx = find_new_capacity(eenv, sg->sge);
5231 if (sg->group_weight == 1) {
5232 /* Remove capacity of src CPU (before task move) */
5233 if (eenv->util_delta == 0 &&
5234 cpumask_test_cpu(eenv->src_cpu, sched_group_cpus(sg))) {
5235 eenv->cap.before = sg->sge->cap_states[cap_idx].cap;
5236 eenv->cap.delta -= eenv->cap.before;
5238 /* Add capacity of dst CPU (after task move) */
5239 if (eenv->util_delta != 0 &&
5240 cpumask_test_cpu(eenv->dst_cpu, sched_group_cpus(sg))) {
5241 eenv->cap.after = sg->sge->cap_states[cap_idx].cap;
5242 eenv->cap.delta += eenv->cap.after;
5246 idle_idx = group_idle_state(sg);
5247 group_util = group_norm_util(eenv, sg);
5248 sg_busy_energy = (group_util * sg->sge->cap_states[cap_idx].power)
5249 >> SCHED_CAPACITY_SHIFT;
5250 sg_idle_energy = ((SCHED_LOAD_SCALE-group_util)
5251 * sg->sge->idle_states[idle_idx].power)
5252 >> SCHED_CAPACITY_SHIFT;
5254 total_energy += sg_busy_energy + sg_idle_energy;
5257 cpumask_xor(&visit_cpus, &visit_cpus, sched_group_cpus(sg));
5259 if (cpumask_equal(sched_group_cpus(sg), sched_group_cpus(eenv->sg_top)))
5262 } while (sg = sg->next, sg != sd->groups);
5265 cpumask_clear_cpu(cpu, &visit_cpus);
5269 eenv->energy = total_energy;
5273 static inline bool cpu_in_sg(struct sched_group *sg, int cpu)
5275 return cpu != -1 && cpumask_test_cpu(cpu, sched_group_cpus(sg));
5279 * energy_diff(): Estimate the energy impact of changing the utilization
5280 * distribution. eenv specifies the change: utilisation amount, source, and
5281 * destination cpu. Source or destination cpu may be -1 in which case the
5282 * utilization is removed from or added to the system (e.g. task wake-up). If
5283 * both are specified, the utilization is migrated.
5285 static inline int __energy_diff(struct energy_env *eenv)
5287 struct sched_domain *sd;
5288 struct sched_group *sg;
5289 int sd_cpu = -1, energy_before = 0, energy_after = 0;
5292 struct energy_env eenv_before = {
5294 .src_cpu = eenv->src_cpu,
5295 .dst_cpu = eenv->dst_cpu,
5296 .nrg = { 0, 0, 0, 0},
5300 if (eenv->src_cpu == eenv->dst_cpu)
5303 sd_cpu = (eenv->src_cpu != -1) ? eenv->src_cpu : eenv->dst_cpu;
5304 sd = rcu_dereference(per_cpu(sd_ea, sd_cpu));
5307 return 0; /* Error */
5312 if (cpu_in_sg(sg, eenv->src_cpu) || cpu_in_sg(sg, eenv->dst_cpu)) {
5313 eenv_before.sg_top = eenv->sg_top = sg;
5315 if (sched_group_energy(&eenv_before))
5316 return 0; /* Invalid result abort */
5317 energy_before += eenv_before.energy;
5319 /* Keep track of SRC cpu (before) capacity */
5320 eenv->cap.before = eenv_before.cap.before;
5321 eenv->cap.delta = eenv_before.cap.delta;
5323 if (sched_group_energy(eenv))
5324 return 0; /* Invalid result abort */
5325 energy_after += eenv->energy;
5327 } while (sg = sg->next, sg != sd->groups);
5329 eenv->nrg.before = energy_before;
5330 eenv->nrg.after = energy_after;
5331 eenv->nrg.diff = eenv->nrg.after - eenv->nrg.before;
5334 trace_sched_energy_diff(eenv->task,
5335 eenv->src_cpu, eenv->dst_cpu, eenv->util_delta,
5336 eenv->nrg.before, eenv->nrg.after, eenv->nrg.diff,
5337 eenv->cap.before, eenv->cap.after, eenv->cap.delta,
5338 eenv->nrg.delta, eenv->payoff);
5341 * Dead-zone margin preventing too many migrations.
5344 margin = eenv->nrg.before >> 6; /* ~1.56% */
5346 diff = eenv->nrg.after - eenv->nrg.before;
5348 eenv->nrg.diff = (abs(diff) < margin) ? 0 : eenv->nrg.diff;
5350 return eenv->nrg.diff;
5353 #ifdef CONFIG_SCHED_TUNE
5355 struct target_nrg schedtune_target_nrg;
5358 * System energy normalization
5359 * Returns the normalized value, in the range [0..SCHED_LOAD_SCALE],
5360 * corresponding to the specified energy variation.
5363 normalize_energy(int energy_diff)
5366 #ifdef CONFIG_SCHED_DEBUG
5369 /* Check for boundaries */
5370 max_delta = schedtune_target_nrg.max_power;
5371 max_delta -= schedtune_target_nrg.min_power;
5372 WARN_ON(abs(energy_diff) >= max_delta);
5375 /* Do scaling using positive numbers to increase the range */
5376 normalized_nrg = (energy_diff < 0) ? -energy_diff : energy_diff;
5378 /* Scale by energy magnitude */
5379 normalized_nrg <<= SCHED_LOAD_SHIFT;
5381 /* Normalize on max energy for target platform */
5382 normalized_nrg = reciprocal_divide(
5383 normalized_nrg, schedtune_target_nrg.rdiv);
5385 return (energy_diff < 0) ? -normalized_nrg : normalized_nrg;
5389 energy_diff(struct energy_env *eenv)
5391 int boost = schedtune_task_boost(eenv->task);
5394 /* Conpute "absolute" energy diff */
5395 __energy_diff(eenv);
5397 /* Return energy diff when boost margin is 0 */
5399 return eenv->nrg.diff;
5401 /* Compute normalized energy diff */
5402 nrg_delta = normalize_energy(eenv->nrg.diff);
5403 eenv->nrg.delta = nrg_delta;
5405 eenv->payoff = schedtune_accept_deltas(
5411 * When SchedTune is enabled, the energy_diff() function will return
5412 * the computed energy payoff value. Since the energy_diff() return
5413 * value is expected to be negative by its callers, this evaluation
5414 * function return a negative value each time the evaluation return a
5415 * positive payoff, which is the condition for the acceptance of
5416 * a scheduling decision
5418 return -eenv->payoff;
5420 #else /* CONFIG_SCHED_TUNE */
5421 #define energy_diff(eenv) __energy_diff(eenv)
5425 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5426 * A waker of many should wake a different task than the one last awakened
5427 * at a frequency roughly N times higher than one of its wakees. In order
5428 * to determine whether we should let the load spread vs consolodating to
5429 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
5430 * partner, and a factor of lls_size higher frequency in the other. With
5431 * both conditions met, we can be relatively sure that the relationship is
5432 * non-monogamous, with partner count exceeding socket size. Waker/wakee
5433 * being client/server, worker/dispatcher, interrupt source or whatever is
5434 * irrelevant, spread criteria is apparent partner count exceeds socket size.
5436 static int wake_wide(struct task_struct *p)
5438 unsigned int master = current->wakee_flips;
5439 unsigned int slave = p->wakee_flips;
5440 int factor = this_cpu_read(sd_llc_size);
5443 swap(master, slave);
5444 if (slave < factor || master < slave * factor)
5449 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5450 int prev_cpu, int sync)
5452 s64 this_load, load;
5453 s64 this_eff_load, prev_eff_load;
5455 struct task_group *tg;
5456 unsigned long weight;
5460 this_cpu = smp_processor_id();
5461 load = source_load(prev_cpu, idx);
5462 this_load = target_load(this_cpu, idx);
5465 * If sync wakeup then subtract the (maximum possible)
5466 * effect of the currently running task from the load
5467 * of the current CPU:
5470 tg = task_group(current);
5471 weight = current->se.avg.load_avg;
5473 this_load += effective_load(tg, this_cpu, -weight, -weight);
5474 load += effective_load(tg, prev_cpu, 0, -weight);
5478 weight = p->se.avg.load_avg;
5481 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5482 * due to the sync cause above having dropped this_load to 0, we'll
5483 * always have an imbalance, but there's really nothing you can do
5484 * about that, so that's good too.
5486 * Otherwise check if either cpus are near enough in load to allow this
5487 * task to be woken on this_cpu.
5489 this_eff_load = 100;
5490 this_eff_load *= capacity_of(prev_cpu);
5492 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5493 prev_eff_load *= capacity_of(this_cpu);
5495 if (this_load > 0) {
5496 this_eff_load *= this_load +
5497 effective_load(tg, this_cpu, weight, weight);
5499 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5502 balanced = this_eff_load <= prev_eff_load;
5504 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5509 schedstat_inc(sd, ttwu_move_affine);
5510 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5515 static inline unsigned long task_util(struct task_struct *p)
5517 #ifdef CONFIG_SCHED_WALT
5518 if (!walt_disabled && sysctl_sched_use_walt_task_util) {
5519 unsigned long demand = p->ravg.demand;
5520 return (demand << 10) / walt_ravg_window;
5523 return p->se.avg.util_avg;
5526 static inline unsigned long boosted_task_util(struct task_struct *task);
5528 static inline bool __task_fits(struct task_struct *p, int cpu, int util)
5530 unsigned long capacity = capacity_of(cpu);
5532 util += boosted_task_util(p);
5534 return (capacity * 1024) > (util * capacity_margin);
5537 static inline bool task_fits_max(struct task_struct *p, int cpu)
5539 unsigned long capacity = capacity_of(cpu);
5540 unsigned long max_capacity = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5542 if (capacity == max_capacity)
5545 if (capacity * capacity_margin > max_capacity * 1024)
5548 return __task_fits(p, cpu, 0);
5551 static bool cpu_overutilized(int cpu)
5553 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5556 #ifdef CONFIG_SCHED_TUNE
5559 schedtune_margin(unsigned long signal, long boost)
5561 long long margin = 0;
5564 * Signal proportional compensation (SPC)
5566 * The Boost (B) value is used to compute a Margin (M) which is
5567 * proportional to the complement of the original Signal (S):
5568 * M = B * (SCHED_LOAD_SCALE - S), if B is positive
5569 * M = B * S, if B is negative
5570 * The obtained M could be used by the caller to "boost" S.
5573 margin = SCHED_LOAD_SCALE - signal;
5576 margin = -signal * boost;
5578 * Fast integer division by constant:
5579 * Constant : (C) = 100
5580 * Precision : 0.1% (P) = 0.1
5581 * Reference : C * 100 / P (R) = 100000
5584 * Shift bits : ceil(log(R,2)) (S) = 17
5585 * Mult const : round(2^S/C) (M) = 1311
5598 schedtune_cpu_margin(unsigned long util, int cpu)
5600 int boost = schedtune_cpu_boost(cpu);
5605 return schedtune_margin(util, boost);
5609 schedtune_task_margin(struct task_struct *task)
5611 int boost = schedtune_task_boost(task);
5618 util = task_util(task);
5619 margin = schedtune_margin(util, boost);
5624 #else /* CONFIG_SCHED_TUNE */
5627 schedtune_cpu_margin(unsigned long util, int cpu)
5633 schedtune_task_margin(struct task_struct *task)
5638 #endif /* CONFIG_SCHED_TUNE */
5641 boosted_cpu_util(int cpu)
5643 unsigned long util = cpu_util(cpu);
5644 long margin = schedtune_cpu_margin(util, cpu);
5646 trace_sched_boost_cpu(cpu, util, margin);
5648 return util + margin;
5651 static inline unsigned long
5652 boosted_task_util(struct task_struct *task)
5654 unsigned long util = task_util(task);
5655 long margin = schedtune_task_margin(task);
5657 trace_sched_boost_task(task, util, margin);
5659 return util + margin;
5662 static int cpu_util_wake(int cpu, struct task_struct *p);
5664 static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
5666 return capacity_orig_of(cpu) - cpu_util_wake(cpu, p);
5670 * find_idlest_group finds and returns the least busy CPU group within the
5673 static struct sched_group *
5674 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5675 int this_cpu, int sd_flag)
5677 struct sched_group *idlest = NULL, *group = sd->groups;
5678 struct sched_group *most_spare_sg = NULL;
5679 unsigned long min_load = ULONG_MAX, this_load = 0;
5680 unsigned long most_spare = 0, this_spare = 0;
5681 int load_idx = sd->forkexec_idx;
5682 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5684 if (sd_flag & SD_BALANCE_WAKE)
5685 load_idx = sd->wake_idx;
5688 unsigned long load, avg_load, spare_cap, max_spare_cap;
5692 /* Skip over this group if it has no CPUs allowed */
5693 if (!cpumask_intersects(sched_group_cpus(group),
5694 tsk_cpus_allowed(p)))
5697 local_group = cpumask_test_cpu(this_cpu,
5698 sched_group_cpus(group));
5701 * Tally up the load of all CPUs in the group and find
5702 * the group containing the CPU with most spare capacity.
5707 for_each_cpu(i, sched_group_cpus(group)) {
5708 /* Bias balancing toward cpus of our domain */
5710 load = source_load(i, load_idx);
5712 load = target_load(i, load_idx);
5716 spare_cap = capacity_spare_wake(i, p);
5718 if (spare_cap > max_spare_cap)
5719 max_spare_cap = spare_cap;
5722 /* Adjust by relative CPU capacity of the group */
5723 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5726 this_load = avg_load;
5727 this_spare = max_spare_cap;
5729 if (avg_load < min_load) {
5730 min_load = avg_load;
5734 if (most_spare < max_spare_cap) {
5735 most_spare = max_spare_cap;
5736 most_spare_sg = group;
5739 } while (group = group->next, group != sd->groups);
5742 * The cross-over point between using spare capacity or least load
5743 * is too conservative for high utilization tasks on partially
5744 * utilized systems if we require spare_capacity > task_util(p),
5745 * so we allow for some task stuffing by using
5746 * spare_capacity > task_util(p)/2.
5748 if (this_spare > task_util(p) / 2 &&
5749 imbalance*this_spare > 100*most_spare)
5751 else if (most_spare > task_util(p) / 2)
5752 return most_spare_sg;
5754 if (!idlest || 100*this_load < imbalance*min_load)
5760 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5763 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5765 unsigned long load, min_load = ULONG_MAX;
5766 unsigned int min_exit_latency = UINT_MAX;
5767 u64 latest_idle_timestamp = 0;
5768 int least_loaded_cpu = this_cpu;
5769 int shallowest_idle_cpu = -1;
5772 /* Check if we have any choice: */
5773 if (group->group_weight == 1)
5774 return cpumask_first(sched_group_cpus(group));
5776 /* Traverse only the allowed CPUs */
5777 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5779 struct rq *rq = cpu_rq(i);
5780 struct cpuidle_state *idle = idle_get_state(rq);
5781 if (idle && idle->exit_latency < min_exit_latency) {
5783 * We give priority to a CPU whose idle state
5784 * has the smallest exit latency irrespective
5785 * of any idle timestamp.
5787 min_exit_latency = idle->exit_latency;
5788 latest_idle_timestamp = rq->idle_stamp;
5789 shallowest_idle_cpu = i;
5790 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5791 rq->idle_stamp > latest_idle_timestamp) {
5793 * If equal or no active idle state, then
5794 * the most recently idled CPU might have
5797 latest_idle_timestamp = rq->idle_stamp;
5798 shallowest_idle_cpu = i;
5800 } else if (shallowest_idle_cpu == -1) {
5801 load = weighted_cpuload(i);
5802 if (load < min_load || (load == min_load && i == this_cpu)) {
5804 least_loaded_cpu = i;
5809 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5813 * Try and locate an idle CPU in the sched_domain.
5815 static int select_idle_sibling(struct task_struct *p, int prev, int target)
5817 struct sched_domain *sd;
5818 struct sched_group *sg;
5819 int best_idle_cpu = -1;
5820 int best_idle_cstate = INT_MAX;
5821 unsigned long best_idle_capacity = ULONG_MAX;
5823 if (!sysctl_sched_cstate_aware) {
5824 if (idle_cpu(target))
5828 * If the prevous cpu is cache affine and idle, don't be stupid.
5830 if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev))
5835 * Otherwise, iterate the domains and find an elegible idle cpu.
5837 sd = rcu_dereference(per_cpu(sd_llc, target));
5838 for_each_lower_domain(sd) {
5842 if (!cpumask_intersects(sched_group_cpus(sg),
5843 tsk_cpus_allowed(p)))
5846 if (sysctl_sched_cstate_aware) {
5847 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg)) {
5848 int idle_idx = idle_get_state_idx(cpu_rq(i));
5849 unsigned long new_usage = boosted_task_util(p);
5850 unsigned long capacity_orig = capacity_orig_of(i);
5852 if (new_usage > capacity_orig || !idle_cpu(i))
5855 if (i == target && new_usage <= capacity_curr_of(target))
5858 if (idle_idx < best_idle_cstate &&
5859 capacity_orig <= best_idle_capacity) {
5861 best_idle_cstate = idle_idx;
5862 best_idle_capacity = capacity_orig;
5866 for_each_cpu(i, sched_group_cpus(sg)) {
5867 if (i == target || !idle_cpu(i))
5871 target = cpumask_first_and(sched_group_cpus(sg),
5872 tsk_cpus_allowed(p));
5877 } while (sg != sd->groups);
5880 if (best_idle_cpu >= 0)
5881 target = best_idle_cpu;
5887 static int start_cpu(bool boosted)
5889 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
5891 RCU_LOCKDEP_WARN(rcu_read_lock_sched_held(),
5892 "sched RCU must be held");
5894 return boosted ? rd->max_cap_orig_cpu : rd->min_cap_orig_cpu;
5897 static inline int find_best_target(struct task_struct *p, bool boosted, bool prefer_idle)
5899 int target_cpu = -1;
5900 unsigned long target_util = prefer_idle ? ULONG_MAX : 0;
5901 unsigned long backup_capacity = ULONG_MAX;
5902 int best_idle_cpu = -1;
5903 int best_idle_cstate = INT_MAX;
5904 int backup_cpu = -1;
5905 unsigned long min_util = boosted_task_util(p);
5906 struct sched_domain *sd;
5907 struct sched_group *sg;
5908 int cpu = start_cpu(boosted);
5913 sd = rcu_dereference(per_cpu(sd_ea, cpu));
5923 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg)) {
5924 unsigned long cur_capacity, new_util;
5930 * p's blocked utilization is still accounted for on prev_cpu
5931 * so prev_cpu will receive a negative bias due to the double
5932 * accounting. However, the blocked utilization may be zero.
5934 new_util = cpu_util(i) + task_util(p);
5937 * Ensure minimum capacity to grant the required boost.
5938 * The target CPU can be already at a capacity level higher
5939 * than the one required to boost the task.
5941 new_util = max(min_util, new_util);
5943 if (new_util > capacity_orig_of(i))
5946 #ifdef CONFIG_SCHED_WALT
5947 if (walt_cpu_high_irqload(i))
5952 * Unconditionally favoring tasks that prefer idle cpus to
5955 if (idle_cpu(i) && prefer_idle)
5958 cur_capacity = capacity_curr_of(i);
5960 if (new_util < cur_capacity) {
5961 if (cpu_rq(i)->nr_running) {
5963 * Find a target cpu with the lowest/highest
5964 * utilization if prefer_idle/!prefer_idle.
5966 if ((prefer_idle && target_util > new_util) ||
5967 (!prefer_idle && target_util < new_util)) {
5968 target_util = new_util;
5971 } else if (!prefer_idle) {
5972 int idle_idx = idle_get_state_idx(cpu_rq(i));
5974 if (best_idle_cpu < 0 ||
5975 (sysctl_sched_cstate_aware &&
5976 best_idle_cstate > idle_idx)) {
5977 best_idle_cstate = idle_idx;
5981 } else if (backup_capacity > cur_capacity) {
5982 /* Find a backup cpu with least capacity. */
5983 backup_capacity = cur_capacity;
5987 } while (sg = sg->next, sg != sd->groups);
5990 target_cpu = best_idle_cpu >= 0 ? best_idle_cpu : backup_cpu;
5996 * cpu_util_wake: Compute cpu utilization with any contributions from
5997 * the waking task p removed.
5999 static int cpu_util_wake(int cpu, struct task_struct *p)
6001 unsigned long util, capacity;
6003 /* Task has no contribution or is new */
6004 if (cpu != task_cpu(p) || !p->se.avg.last_update_time)
6005 return cpu_util(cpu);
6007 capacity = capacity_orig_of(cpu);
6008 util = max_t(long, cpu_rq(cpu)->cfs.avg.util_avg - task_util(p), 0);
6010 return (util >= capacity) ? capacity : util;
6014 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6015 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6017 * In that case WAKE_AFFINE doesn't make sense and we'll let
6018 * BALANCE_WAKE sort things out.
6020 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6022 long min_cap, max_cap;
6024 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6025 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity.val;
6027 /* Minimum capacity is close to max, no need to abort wake_affine */
6028 if (max_cap - min_cap < max_cap >> 3)
6031 /* Bring task utilization in sync with prev_cpu */
6032 sync_entity_load_avg(&p->se);
6034 return min_cap * 1024 < task_util(p) * capacity_margin;
6037 static int select_energy_cpu_brute(struct task_struct *p, int prev_cpu, int sync)
6039 struct sched_domain *sd;
6040 int target_cpu = prev_cpu, tmp_target;
6041 bool boosted, prefer_idle;
6043 if (sysctl_sched_sync_hint_enable && sync) {
6044 int cpu = smp_processor_id();
6046 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
6051 #ifdef CONFIG_CGROUP_SCHEDTUNE
6052 boosted = schedtune_task_boost(p) > 0;
6053 prefer_idle = schedtune_prefer_idle(p) > 0;
6055 boosted = get_sysctl_sched_cfs_boost() > 0;
6059 sd = rcu_dereference(per_cpu(sd_ea, prev_cpu));
6060 /* Find a cpu with sufficient capacity */
6061 tmp_target = find_best_target(p, boosted, prefer_idle);
6065 if (tmp_target >= 0) {
6066 target_cpu = tmp_target;
6067 if ((boosted || prefer_idle) && idle_cpu(target_cpu))
6071 if (target_cpu != prev_cpu) {
6072 struct energy_env eenv = {
6073 .util_delta = task_util(p),
6074 .src_cpu = prev_cpu,
6075 .dst_cpu = target_cpu,
6079 /* Not enough spare capacity on previous cpu */
6080 if (cpu_overutilized(prev_cpu))
6083 if (energy_diff(&eenv) >= 0)
6084 target_cpu = prev_cpu;
6093 * select_task_rq_fair: Select target runqueue for the waking task in domains
6094 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6095 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6097 * Balances load by selecting the idlest cpu in the idlest group, or under
6098 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
6100 * Returns the target cpu number.
6102 * preempt must be disabled.
6105 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6107 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
6108 int cpu = smp_processor_id();
6109 int new_cpu = prev_cpu;
6110 int want_affine = 0;
6111 int sync = wake_flags & WF_SYNC;
6113 if (sd_flag & SD_BALANCE_WAKE)
6114 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
6115 && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
6117 if (energy_aware() && !(cpu_rq(prev_cpu)->rd->overutilized))
6118 return select_energy_cpu_brute(p, prev_cpu, sync);
6121 for_each_domain(cpu, tmp) {
6122 if (!(tmp->flags & SD_LOAD_BALANCE))
6126 * If both cpu and prev_cpu are part of this domain,
6127 * cpu is a valid SD_WAKE_AFFINE target.
6129 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6130 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6135 if (tmp->flags & sd_flag)
6137 else if (!want_affine)
6142 sd = NULL; /* Prefer wake_affine over balance flags */
6143 if (cpu != prev_cpu && wake_affine(affine_sd, p, prev_cpu, sync))
6148 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
6149 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6152 struct sched_group *group;
6155 if (!(sd->flags & sd_flag)) {
6160 group = find_idlest_group(sd, p, cpu, sd_flag);
6166 new_cpu = find_idlest_cpu(group, p, cpu);
6167 if (new_cpu == -1 || new_cpu == cpu) {
6168 /* Now try balancing at a lower domain level of cpu */
6173 /* Now try balancing at a lower domain level of new_cpu */
6175 weight = sd->span_weight;
6177 for_each_domain(cpu, tmp) {
6178 if (weight <= tmp->span_weight)
6180 if (tmp->flags & sd_flag)
6183 /* while loop will break here if sd == NULL */
6191 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
6192 * cfs_rq_of(p) references at time of call are still valid and identify the
6193 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
6194 * other assumptions, including the state of rq->lock, should be made.
6196 static void migrate_task_rq_fair(struct task_struct *p)
6199 * We are supposed to update the task to "current" time, then its up to date
6200 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
6201 * what current time is, so simply throw away the out-of-date time. This
6202 * will result in the wakee task is less decayed, but giving the wakee more
6203 * load sounds not bad.
6205 remove_entity_load_avg(&p->se);
6207 /* Tell new CPU we are migrated */
6208 p->se.avg.last_update_time = 0;
6210 /* We have migrated, no longer consider this task hot */
6211 p->se.exec_start = 0;
6214 static void task_dead_fair(struct task_struct *p)
6216 remove_entity_load_avg(&p->se);
6219 #define task_fits_max(p, cpu) true
6220 #endif /* CONFIG_SMP */
6222 static unsigned long
6223 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
6225 unsigned long gran = sysctl_sched_wakeup_granularity;
6228 * Since its curr running now, convert the gran from real-time
6229 * to virtual-time in his units.
6231 * By using 'se' instead of 'curr' we penalize light tasks, so
6232 * they get preempted easier. That is, if 'se' < 'curr' then
6233 * the resulting gran will be larger, therefore penalizing the
6234 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6235 * be smaller, again penalizing the lighter task.
6237 * This is especially important for buddies when the leftmost
6238 * task is higher priority than the buddy.
6240 return calc_delta_fair(gran, se);
6244 * Should 'se' preempt 'curr'.
6258 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6260 s64 gran, vdiff = curr->vruntime - se->vruntime;
6265 gran = wakeup_gran(curr, se);
6272 static void set_last_buddy(struct sched_entity *se)
6274 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6277 for_each_sched_entity(se)
6278 cfs_rq_of(se)->last = se;
6281 static void set_next_buddy(struct sched_entity *se)
6283 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6286 for_each_sched_entity(se)
6287 cfs_rq_of(se)->next = se;
6290 static void set_skip_buddy(struct sched_entity *se)
6292 for_each_sched_entity(se)
6293 cfs_rq_of(se)->skip = se;
6297 * Preempt the current task with a newly woken task if needed:
6299 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6301 struct task_struct *curr = rq->curr;
6302 struct sched_entity *se = &curr->se, *pse = &p->se;
6303 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6304 int scale = cfs_rq->nr_running >= sched_nr_latency;
6305 int next_buddy_marked = 0;
6307 if (unlikely(se == pse))
6311 * This is possible from callers such as attach_tasks(), in which we
6312 * unconditionally check_prempt_curr() after an enqueue (which may have
6313 * lead to a throttle). This both saves work and prevents false
6314 * next-buddy nomination below.
6316 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6319 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6320 set_next_buddy(pse);
6321 next_buddy_marked = 1;
6325 * We can come here with TIF_NEED_RESCHED already set from new task
6328 * Note: this also catches the edge-case of curr being in a throttled
6329 * group (e.g. via set_curr_task), since update_curr() (in the
6330 * enqueue of curr) will have resulted in resched being set. This
6331 * prevents us from potentially nominating it as a false LAST_BUDDY
6334 if (test_tsk_need_resched(curr))
6337 /* Idle tasks are by definition preempted by non-idle tasks. */
6338 if (unlikely(curr->policy == SCHED_IDLE) &&
6339 likely(p->policy != SCHED_IDLE))
6343 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6344 * is driven by the tick):
6346 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6349 find_matching_se(&se, &pse);
6350 update_curr(cfs_rq_of(se));
6352 if (wakeup_preempt_entity(se, pse) == 1) {
6354 * Bias pick_next to pick the sched entity that is
6355 * triggering this preemption.
6357 if (!next_buddy_marked)
6358 set_next_buddy(pse);
6367 * Only set the backward buddy when the current task is still
6368 * on the rq. This can happen when a wakeup gets interleaved
6369 * with schedule on the ->pre_schedule() or idle_balance()
6370 * point, either of which can * drop the rq lock.
6372 * Also, during early boot the idle thread is in the fair class,
6373 * for obvious reasons its a bad idea to schedule back to it.
6375 if (unlikely(!se->on_rq || curr == rq->idle))
6378 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6382 static struct task_struct *
6383 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
6385 struct cfs_rq *cfs_rq = &rq->cfs;
6386 struct sched_entity *se;
6387 struct task_struct *p;
6391 #ifdef CONFIG_FAIR_GROUP_SCHED
6392 if (!cfs_rq->nr_running)
6395 if (prev->sched_class != &fair_sched_class)
6399 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6400 * likely that a next task is from the same cgroup as the current.
6402 * Therefore attempt to avoid putting and setting the entire cgroup
6403 * hierarchy, only change the part that actually changes.
6407 struct sched_entity *curr = cfs_rq->curr;
6410 * Since we got here without doing put_prev_entity() we also
6411 * have to consider cfs_rq->curr. If it is still a runnable
6412 * entity, update_curr() will update its vruntime, otherwise
6413 * forget we've ever seen it.
6417 update_curr(cfs_rq);
6422 * This call to check_cfs_rq_runtime() will do the
6423 * throttle and dequeue its entity in the parent(s).
6424 * Therefore the 'simple' nr_running test will indeed
6427 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
6431 se = pick_next_entity(cfs_rq, curr);
6432 cfs_rq = group_cfs_rq(se);
6438 * Since we haven't yet done put_prev_entity and if the selected task
6439 * is a different task than we started out with, try and touch the
6440 * least amount of cfs_rqs.
6443 struct sched_entity *pse = &prev->se;
6445 while (!(cfs_rq = is_same_group(se, pse))) {
6446 int se_depth = se->depth;
6447 int pse_depth = pse->depth;
6449 if (se_depth <= pse_depth) {
6450 put_prev_entity(cfs_rq_of(pse), pse);
6451 pse = parent_entity(pse);
6453 if (se_depth >= pse_depth) {
6454 set_next_entity(cfs_rq_of(se), se);
6455 se = parent_entity(se);
6459 put_prev_entity(cfs_rq, pse);
6460 set_next_entity(cfs_rq, se);
6463 if (hrtick_enabled(rq))
6464 hrtick_start_fair(rq, p);
6466 rq->misfit_task = !task_fits_max(p, rq->cpu);
6473 if (!cfs_rq->nr_running)
6476 put_prev_task(rq, prev);
6479 se = pick_next_entity(cfs_rq, NULL);
6480 set_next_entity(cfs_rq, se);
6481 cfs_rq = group_cfs_rq(se);
6486 if (hrtick_enabled(rq))
6487 hrtick_start_fair(rq, p);
6489 rq->misfit_task = !task_fits_max(p, rq->cpu);
6494 rq->misfit_task = 0;
6496 * This is OK, because current is on_cpu, which avoids it being picked
6497 * for load-balance and preemption/IRQs are still disabled avoiding
6498 * further scheduler activity on it and we're being very careful to
6499 * re-start the picking loop.
6501 lockdep_unpin_lock(&rq->lock);
6502 new_tasks = idle_balance(rq);
6503 lockdep_pin_lock(&rq->lock);
6505 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6506 * possible for any higher priority task to appear. In that case we
6507 * must re-start the pick_next_entity() loop.
6519 * Account for a descheduled task:
6521 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6523 struct sched_entity *se = &prev->se;
6524 struct cfs_rq *cfs_rq;
6526 for_each_sched_entity(se) {
6527 cfs_rq = cfs_rq_of(se);
6528 put_prev_entity(cfs_rq, se);
6533 * sched_yield() is very simple
6535 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6537 static void yield_task_fair(struct rq *rq)
6539 struct task_struct *curr = rq->curr;
6540 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6541 struct sched_entity *se = &curr->se;
6544 * Are we the only task in the tree?
6546 if (unlikely(rq->nr_running == 1))
6549 clear_buddies(cfs_rq, se);
6551 if (curr->policy != SCHED_BATCH) {
6552 update_rq_clock(rq);
6554 * Update run-time statistics of the 'current'.
6556 update_curr(cfs_rq);
6558 * Tell update_rq_clock() that we've just updated,
6559 * so we don't do microscopic update in schedule()
6560 * and double the fastpath cost.
6562 rq_clock_skip_update(rq, true);
6568 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6570 struct sched_entity *se = &p->se;
6572 /* throttled hierarchies are not runnable */
6573 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6576 /* Tell the scheduler that we'd really like pse to run next. */
6579 yield_task_fair(rq);
6585 /**************************************************
6586 * Fair scheduling class load-balancing methods.
6590 * The purpose of load-balancing is to achieve the same basic fairness the
6591 * per-cpu scheduler provides, namely provide a proportional amount of compute
6592 * time to each task. This is expressed in the following equation:
6594 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6596 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6597 * W_i,0 is defined as:
6599 * W_i,0 = \Sum_j w_i,j (2)
6601 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6602 * is derived from the nice value as per prio_to_weight[].
6604 * The weight average is an exponential decay average of the instantaneous
6607 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6609 * C_i is the compute capacity of cpu i, typically it is the
6610 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6611 * can also include other factors [XXX].
6613 * To achieve this balance we define a measure of imbalance which follows
6614 * directly from (1):
6616 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6618 * We them move tasks around to minimize the imbalance. In the continuous
6619 * function space it is obvious this converges, in the discrete case we get
6620 * a few fun cases generally called infeasible weight scenarios.
6623 * - infeasible weights;
6624 * - local vs global optima in the discrete case. ]
6629 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6630 * for all i,j solution, we create a tree of cpus that follows the hardware
6631 * topology where each level pairs two lower groups (or better). This results
6632 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6633 * tree to only the first of the previous level and we decrease the frequency
6634 * of load-balance at each level inv. proportional to the number of cpus in
6640 * \Sum { --- * --- * 2^i } = O(n) (5)
6642 * `- size of each group
6643 * | | `- number of cpus doing load-balance
6645 * `- sum over all levels
6647 * Coupled with a limit on how many tasks we can migrate every balance pass,
6648 * this makes (5) the runtime complexity of the balancer.
6650 * An important property here is that each CPU is still (indirectly) connected
6651 * to every other cpu in at most O(log n) steps:
6653 * The adjacency matrix of the resulting graph is given by:
6656 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6659 * And you'll find that:
6661 * A^(log_2 n)_i,j != 0 for all i,j (7)
6663 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6664 * The task movement gives a factor of O(m), giving a convergence complexity
6667 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6672 * In order to avoid CPUs going idle while there's still work to do, new idle
6673 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6674 * tree itself instead of relying on other CPUs to bring it work.
6676 * This adds some complexity to both (5) and (8) but it reduces the total idle
6684 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6687 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6692 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6694 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6696 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6699 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6700 * rewrite all of this once again.]
6703 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6705 enum fbq_type { regular, remote, all };
6714 #define LBF_ALL_PINNED 0x01
6715 #define LBF_NEED_BREAK 0x02
6716 #define LBF_DST_PINNED 0x04
6717 #define LBF_SOME_PINNED 0x08
6720 struct sched_domain *sd;
6728 struct cpumask *dst_grpmask;
6730 enum cpu_idle_type idle;
6732 unsigned int src_grp_nr_running;
6733 /* The set of CPUs under consideration for load-balancing */
6734 struct cpumask *cpus;
6739 unsigned int loop_break;
6740 unsigned int loop_max;
6742 enum fbq_type fbq_type;
6743 enum group_type busiest_group_type;
6744 struct list_head tasks;
6748 * Is this task likely cache-hot:
6750 static int task_hot(struct task_struct *p, struct lb_env *env)
6754 lockdep_assert_held(&env->src_rq->lock);
6756 if (p->sched_class != &fair_sched_class)
6759 if (unlikely(p->policy == SCHED_IDLE))
6763 * Buddy candidates are cache hot:
6765 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6766 (&p->se == cfs_rq_of(&p->se)->next ||
6767 &p->se == cfs_rq_of(&p->se)->last))
6770 if (sysctl_sched_migration_cost == -1)
6772 if (sysctl_sched_migration_cost == 0)
6775 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6777 return delta < (s64)sysctl_sched_migration_cost;
6780 #ifdef CONFIG_NUMA_BALANCING
6782 * Returns 1, if task migration degrades locality
6783 * Returns 0, if task migration improves locality i.e migration preferred.
6784 * Returns -1, if task migration is not affected by locality.
6786 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6788 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6789 unsigned long src_faults, dst_faults;
6790 int src_nid, dst_nid;
6792 if (!static_branch_likely(&sched_numa_balancing))
6795 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6798 src_nid = cpu_to_node(env->src_cpu);
6799 dst_nid = cpu_to_node(env->dst_cpu);
6801 if (src_nid == dst_nid)
6804 /* Migrating away from the preferred node is always bad. */
6805 if (src_nid == p->numa_preferred_nid) {
6806 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6812 /* Encourage migration to the preferred node. */
6813 if (dst_nid == p->numa_preferred_nid)
6817 src_faults = group_faults(p, src_nid);
6818 dst_faults = group_faults(p, dst_nid);
6820 src_faults = task_faults(p, src_nid);
6821 dst_faults = task_faults(p, dst_nid);
6824 return dst_faults < src_faults;
6828 static inline int migrate_degrades_locality(struct task_struct *p,
6836 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6839 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6843 lockdep_assert_held(&env->src_rq->lock);
6846 * We do not migrate tasks that are:
6847 * 1) throttled_lb_pair, or
6848 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6849 * 3) running (obviously), or
6850 * 4) are cache-hot on their current CPU.
6852 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6855 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6858 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6860 env->flags |= LBF_SOME_PINNED;
6863 * Remember if this task can be migrated to any other cpu in
6864 * our sched_group. We may want to revisit it if we couldn't
6865 * meet load balance goals by pulling other tasks on src_cpu.
6867 * Also avoid computing new_dst_cpu if we have already computed
6868 * one in current iteration.
6870 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6873 /* Prevent to re-select dst_cpu via env's cpus */
6874 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6875 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6876 env->flags |= LBF_DST_PINNED;
6877 env->new_dst_cpu = cpu;
6885 /* Record that we found atleast one task that could run on dst_cpu */
6886 env->flags &= ~LBF_ALL_PINNED;
6888 if (task_running(env->src_rq, p)) {
6889 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6894 * Aggressive migration if:
6895 * 1) destination numa is preferred
6896 * 2) task is cache cold, or
6897 * 3) too many balance attempts have failed.
6899 tsk_cache_hot = migrate_degrades_locality(p, env);
6900 if (tsk_cache_hot == -1)
6901 tsk_cache_hot = task_hot(p, env);
6903 if (tsk_cache_hot <= 0 ||
6904 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6905 if (tsk_cache_hot == 1) {
6906 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6907 schedstat_inc(p, se.statistics.nr_forced_migrations);
6912 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6917 * detach_task() -- detach the task for the migration specified in env
6919 static void detach_task(struct task_struct *p, struct lb_env *env)
6921 lockdep_assert_held(&env->src_rq->lock);
6923 deactivate_task(env->src_rq, p, 0);
6924 p->on_rq = TASK_ON_RQ_MIGRATING;
6925 double_lock_balance(env->src_rq, env->dst_rq);
6926 set_task_cpu(p, env->dst_cpu);
6927 double_unlock_balance(env->src_rq, env->dst_rq);
6931 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6932 * part of active balancing operations within "domain".
6934 * Returns a task if successful and NULL otherwise.
6936 static struct task_struct *detach_one_task(struct lb_env *env)
6938 struct task_struct *p, *n;
6940 lockdep_assert_held(&env->src_rq->lock);
6942 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6943 if (!can_migrate_task(p, env))
6946 detach_task(p, env);
6949 * Right now, this is only the second place where
6950 * lb_gained[env->idle] is updated (other is detach_tasks)
6951 * so we can safely collect stats here rather than
6952 * inside detach_tasks().
6954 schedstat_inc(env->sd, lb_gained[env->idle]);
6960 static const unsigned int sched_nr_migrate_break = 32;
6963 * detach_tasks() -- tries to detach up to imbalance weighted load from
6964 * busiest_rq, as part of a balancing operation within domain "sd".
6966 * Returns number of detached tasks if successful and 0 otherwise.
6968 static int detach_tasks(struct lb_env *env)
6970 struct list_head *tasks = &env->src_rq->cfs_tasks;
6971 struct task_struct *p;
6975 lockdep_assert_held(&env->src_rq->lock);
6977 if (env->imbalance <= 0)
6980 while (!list_empty(tasks)) {
6982 * We don't want to steal all, otherwise we may be treated likewise,
6983 * which could at worst lead to a livelock crash.
6985 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6988 p = list_first_entry(tasks, struct task_struct, se.group_node);
6991 /* We've more or less seen every task there is, call it quits */
6992 if (env->loop > env->loop_max)
6995 /* take a breather every nr_migrate tasks */
6996 if (env->loop > env->loop_break) {
6997 env->loop_break += sched_nr_migrate_break;
6998 env->flags |= LBF_NEED_BREAK;
7002 if (!can_migrate_task(p, env))
7005 load = task_h_load(p);
7007 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7010 if ((load / 2) > env->imbalance)
7013 detach_task(p, env);
7014 list_add(&p->se.group_node, &env->tasks);
7017 env->imbalance -= load;
7019 #ifdef CONFIG_PREEMPT
7021 * NEWIDLE balancing is a source of latency, so preemptible
7022 * kernels will stop after the first task is detached to minimize
7023 * the critical section.
7025 if (env->idle == CPU_NEWLY_IDLE)
7030 * We only want to steal up to the prescribed amount of
7033 if (env->imbalance <= 0)
7038 list_move_tail(&p->se.group_node, tasks);
7042 * Right now, this is one of only two places we collect this stat
7043 * so we can safely collect detach_one_task() stats here rather
7044 * than inside detach_one_task().
7046 schedstat_add(env->sd, lb_gained[env->idle], detached);
7052 * attach_task() -- attach the task detached by detach_task() to its new rq.
7054 static void attach_task(struct rq *rq, struct task_struct *p)
7056 lockdep_assert_held(&rq->lock);
7058 BUG_ON(task_rq(p) != rq);
7059 p->on_rq = TASK_ON_RQ_QUEUED;
7060 activate_task(rq, p, 0);
7061 check_preempt_curr(rq, p, 0);
7065 * attach_one_task() -- attaches the task returned from detach_one_task() to
7068 static void attach_one_task(struct rq *rq, struct task_struct *p)
7070 raw_spin_lock(&rq->lock);
7073 * We want to potentially raise target_cpu's OPP.
7075 update_capacity_of(cpu_of(rq));
7076 raw_spin_unlock(&rq->lock);
7080 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7083 static void attach_tasks(struct lb_env *env)
7085 struct list_head *tasks = &env->tasks;
7086 struct task_struct *p;
7088 raw_spin_lock(&env->dst_rq->lock);
7090 while (!list_empty(tasks)) {
7091 p = list_first_entry(tasks, struct task_struct, se.group_node);
7092 list_del_init(&p->se.group_node);
7094 attach_task(env->dst_rq, p);
7098 * We want to potentially raise env.dst_cpu's OPP.
7100 update_capacity_of(env->dst_cpu);
7102 raw_spin_unlock(&env->dst_rq->lock);
7105 #ifdef CONFIG_FAIR_GROUP_SCHED
7106 static void update_blocked_averages(int cpu)
7108 struct rq *rq = cpu_rq(cpu);
7109 struct cfs_rq *cfs_rq;
7110 unsigned long flags;
7112 raw_spin_lock_irqsave(&rq->lock, flags);
7113 update_rq_clock(rq);
7116 * Iterates the task_group tree in a bottom up fashion, see
7117 * list_add_leaf_cfs_rq() for details.
7119 for_each_leaf_cfs_rq(rq, cfs_rq) {
7120 /* throttled entities do not contribute to load */
7121 if (throttled_hierarchy(cfs_rq))
7124 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq,
7126 update_tg_load_avg(cfs_rq, 0);
7128 raw_spin_unlock_irqrestore(&rq->lock, flags);
7132 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7133 * This needs to be done in a top-down fashion because the load of a child
7134 * group is a fraction of its parents load.
7136 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7138 struct rq *rq = rq_of(cfs_rq);
7139 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7140 unsigned long now = jiffies;
7143 if (cfs_rq->last_h_load_update == now)
7146 cfs_rq->h_load_next = NULL;
7147 for_each_sched_entity(se) {
7148 cfs_rq = cfs_rq_of(se);
7149 cfs_rq->h_load_next = se;
7150 if (cfs_rq->last_h_load_update == now)
7155 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7156 cfs_rq->last_h_load_update = now;
7159 while ((se = cfs_rq->h_load_next) != NULL) {
7160 load = cfs_rq->h_load;
7161 load = div64_ul(load * se->avg.load_avg,
7162 cfs_rq_load_avg(cfs_rq) + 1);
7163 cfs_rq = group_cfs_rq(se);
7164 cfs_rq->h_load = load;
7165 cfs_rq->last_h_load_update = now;
7169 static unsigned long task_h_load(struct task_struct *p)
7171 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7173 update_cfs_rq_h_load(cfs_rq);
7174 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7175 cfs_rq_load_avg(cfs_rq) + 1);
7178 static inline void update_blocked_averages(int cpu)
7180 struct rq *rq = cpu_rq(cpu);
7181 struct cfs_rq *cfs_rq = &rq->cfs;
7182 unsigned long flags;
7184 raw_spin_lock_irqsave(&rq->lock, flags);
7185 update_rq_clock(rq);
7186 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true);
7187 raw_spin_unlock_irqrestore(&rq->lock, flags);
7190 static unsigned long task_h_load(struct task_struct *p)
7192 return p->se.avg.load_avg;
7196 /********** Helpers for find_busiest_group ************************/
7199 * sg_lb_stats - stats of a sched_group required for load_balancing
7201 struct sg_lb_stats {
7202 unsigned long avg_load; /*Avg load across the CPUs of the group */
7203 unsigned long group_load; /* Total load over the CPUs of the group */
7204 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7205 unsigned long load_per_task;
7206 unsigned long group_capacity;
7207 unsigned long group_util; /* Total utilization of the group */
7208 unsigned int sum_nr_running; /* Nr tasks running in the group */
7209 unsigned int idle_cpus;
7210 unsigned int group_weight;
7211 enum group_type group_type;
7212 int group_no_capacity;
7213 int group_misfit_task; /* A cpu has a task too big for its capacity */
7214 #ifdef CONFIG_NUMA_BALANCING
7215 unsigned int nr_numa_running;
7216 unsigned int nr_preferred_running;
7221 * sd_lb_stats - Structure to store the statistics of a sched_domain
7222 * during load balancing.
7224 struct sd_lb_stats {
7225 struct sched_group *busiest; /* Busiest group in this sd */
7226 struct sched_group *local; /* Local group in this sd */
7227 unsigned long total_load; /* Total load of all groups in sd */
7228 unsigned long total_capacity; /* Total capacity of all groups in sd */
7229 unsigned long avg_load; /* Average load across all groups in sd */
7231 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7232 struct sg_lb_stats local_stat; /* Statistics of the local group */
7235 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7238 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7239 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7240 * We must however clear busiest_stat::avg_load because
7241 * update_sd_pick_busiest() reads this before assignment.
7243 *sds = (struct sd_lb_stats){
7247 .total_capacity = 0UL,
7250 .sum_nr_running = 0,
7251 .group_type = group_other,
7257 * get_sd_load_idx - Obtain the load index for a given sched domain.
7258 * @sd: The sched_domain whose load_idx is to be obtained.
7259 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7261 * Return: The load index.
7263 static inline int get_sd_load_idx(struct sched_domain *sd,
7264 enum cpu_idle_type idle)
7270 load_idx = sd->busy_idx;
7273 case CPU_NEWLY_IDLE:
7274 load_idx = sd->newidle_idx;
7277 load_idx = sd->idle_idx;
7284 static unsigned long scale_rt_capacity(int cpu)
7286 struct rq *rq = cpu_rq(cpu);
7287 u64 total, used, age_stamp, avg;
7291 * Since we're reading these variables without serialization make sure
7292 * we read them once before doing sanity checks on them.
7294 age_stamp = READ_ONCE(rq->age_stamp);
7295 avg = READ_ONCE(rq->rt_avg);
7296 delta = __rq_clock_broken(rq) - age_stamp;
7298 if (unlikely(delta < 0))
7301 total = sched_avg_period() + delta;
7303 used = div_u64(avg, total);
7306 * deadline bandwidth is defined at system level so we must
7307 * weight this bandwidth with the max capacity of the system.
7308 * As a reminder, avg_bw is 20bits width and
7309 * scale_cpu_capacity is 10 bits width
7311 used += div_u64(rq->dl.avg_bw, arch_scale_cpu_capacity(NULL, cpu));
7313 if (likely(used < SCHED_CAPACITY_SCALE))
7314 return SCHED_CAPACITY_SCALE - used;
7319 void init_max_cpu_capacity(struct max_cpu_capacity *mcc)
7321 raw_spin_lock_init(&mcc->lock);
7326 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7328 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7329 struct sched_group *sdg = sd->groups;
7330 struct max_cpu_capacity *mcc;
7331 unsigned long max_capacity;
7333 unsigned long flags;
7335 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7337 mcc = &cpu_rq(cpu)->rd->max_cpu_capacity;
7339 raw_spin_lock_irqsave(&mcc->lock, flags);
7340 max_capacity = mcc->val;
7341 max_cap_cpu = mcc->cpu;
7343 if ((max_capacity > capacity && max_cap_cpu == cpu) ||
7344 (max_capacity < capacity)) {
7345 mcc->val = capacity;
7347 #ifdef CONFIG_SCHED_DEBUG
7348 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7349 printk_deferred(KERN_INFO "CPU%d: update max cpu_capacity %lu\n",
7354 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7356 skip_unlock: __attribute__ ((unused));
7357 capacity *= scale_rt_capacity(cpu);
7358 capacity >>= SCHED_CAPACITY_SHIFT;
7363 cpu_rq(cpu)->cpu_capacity = capacity;
7364 sdg->sgc->capacity = capacity;
7365 sdg->sgc->max_capacity = capacity;
7366 sdg->sgc->min_capacity = capacity;
7369 void update_group_capacity(struct sched_domain *sd, int cpu)
7371 struct sched_domain *child = sd->child;
7372 struct sched_group *group, *sdg = sd->groups;
7373 unsigned long capacity, max_capacity, min_capacity;
7374 unsigned long interval;
7376 interval = msecs_to_jiffies(sd->balance_interval);
7377 interval = clamp(interval, 1UL, max_load_balance_interval);
7378 sdg->sgc->next_update = jiffies + interval;
7381 update_cpu_capacity(sd, cpu);
7387 min_capacity = ULONG_MAX;
7389 if (child->flags & SD_OVERLAP) {
7391 * SD_OVERLAP domains cannot assume that child groups
7392 * span the current group.
7395 for_each_cpu(cpu, sched_group_cpus(sdg)) {
7396 struct sched_group_capacity *sgc;
7397 struct rq *rq = cpu_rq(cpu);
7400 * build_sched_domains() -> init_sched_groups_capacity()
7401 * gets here before we've attached the domains to the
7404 * Use capacity_of(), which is set irrespective of domains
7405 * in update_cpu_capacity().
7407 * This avoids capacity from being 0 and
7408 * causing divide-by-zero issues on boot.
7410 if (unlikely(!rq->sd)) {
7411 capacity += capacity_of(cpu);
7413 sgc = rq->sd->groups->sgc;
7414 capacity += sgc->capacity;
7417 max_capacity = max(capacity, max_capacity);
7418 min_capacity = min(capacity, min_capacity);
7422 * !SD_OVERLAP domains can assume that child groups
7423 * span the current group.
7426 group = child->groups;
7428 struct sched_group_capacity *sgc = group->sgc;
7430 capacity += sgc->capacity;
7431 max_capacity = max(sgc->max_capacity, max_capacity);
7432 min_capacity = min(sgc->min_capacity, min_capacity);
7433 group = group->next;
7434 } while (group != child->groups);
7437 sdg->sgc->capacity = capacity;
7438 sdg->sgc->max_capacity = max_capacity;
7439 sdg->sgc->min_capacity = min_capacity;
7443 * Check whether the capacity of the rq has been noticeably reduced by side
7444 * activity. The imbalance_pct is used for the threshold.
7445 * Return true is the capacity is reduced
7448 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7450 return ((rq->cpu_capacity * sd->imbalance_pct) <
7451 (rq->cpu_capacity_orig * 100));
7455 * Group imbalance indicates (and tries to solve) the problem where balancing
7456 * groups is inadequate due to tsk_cpus_allowed() constraints.
7458 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7459 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7462 * { 0 1 2 3 } { 4 5 6 7 }
7465 * If we were to balance group-wise we'd place two tasks in the first group and
7466 * two tasks in the second group. Clearly this is undesired as it will overload
7467 * cpu 3 and leave one of the cpus in the second group unused.
7469 * The current solution to this issue is detecting the skew in the first group
7470 * by noticing the lower domain failed to reach balance and had difficulty
7471 * moving tasks due to affinity constraints.
7473 * When this is so detected; this group becomes a candidate for busiest; see
7474 * update_sd_pick_busiest(). And calculate_imbalance() and
7475 * find_busiest_group() avoid some of the usual balance conditions to allow it
7476 * to create an effective group imbalance.
7478 * This is a somewhat tricky proposition since the next run might not find the
7479 * group imbalance and decide the groups need to be balanced again. A most
7480 * subtle and fragile situation.
7483 static inline int sg_imbalanced(struct sched_group *group)
7485 return group->sgc->imbalance;
7489 * group_has_capacity returns true if the group has spare capacity that could
7490 * be used by some tasks.
7491 * We consider that a group has spare capacity if the * number of task is
7492 * smaller than the number of CPUs or if the utilization is lower than the
7493 * available capacity for CFS tasks.
7494 * For the latter, we use a threshold to stabilize the state, to take into
7495 * account the variance of the tasks' load and to return true if the available
7496 * capacity in meaningful for the load balancer.
7497 * As an example, an available capacity of 1% can appear but it doesn't make
7498 * any benefit for the load balance.
7501 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7503 if (sgs->sum_nr_running < sgs->group_weight)
7506 if ((sgs->group_capacity * 100) >
7507 (sgs->group_util * env->sd->imbalance_pct))
7514 * group_is_overloaded returns true if the group has more tasks than it can
7516 * group_is_overloaded is not equals to !group_has_capacity because a group
7517 * with the exact right number of tasks, has no more spare capacity but is not
7518 * overloaded so both group_has_capacity and group_is_overloaded return
7522 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7524 if (sgs->sum_nr_running <= sgs->group_weight)
7527 if ((sgs->group_capacity * 100) <
7528 (sgs->group_util * env->sd->imbalance_pct))
7536 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7537 * per-cpu capacity than sched_group ref.
7540 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7542 return sg->sgc->max_capacity + capacity_margin - SCHED_LOAD_SCALE <
7543 ref->sgc->max_capacity;
7547 group_type group_classify(struct sched_group *group,
7548 struct sg_lb_stats *sgs)
7550 if (sgs->group_no_capacity)
7551 return group_overloaded;
7553 if (sg_imbalanced(group))
7554 return group_imbalanced;
7556 if (sgs->group_misfit_task)
7557 return group_misfit_task;
7563 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7564 * @env: The load balancing environment.
7565 * @group: sched_group whose statistics are to be updated.
7566 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7567 * @local_group: Does group contain this_cpu.
7568 * @sgs: variable to hold the statistics for this group.
7569 * @overload: Indicate more than one runnable task for any CPU.
7570 * @overutilized: Indicate overutilization for any CPU.
7572 static inline void update_sg_lb_stats(struct lb_env *env,
7573 struct sched_group *group, int load_idx,
7574 int local_group, struct sg_lb_stats *sgs,
7575 bool *overload, bool *overutilized)
7580 memset(sgs, 0, sizeof(*sgs));
7582 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7583 struct rq *rq = cpu_rq(i);
7585 /* Bias balancing toward cpus of our domain */
7587 load = target_load(i, load_idx);
7589 load = source_load(i, load_idx);
7591 sgs->group_load += load;
7592 sgs->group_util += cpu_util(i);
7593 sgs->sum_nr_running += rq->cfs.h_nr_running;
7595 nr_running = rq->nr_running;
7599 #ifdef CONFIG_NUMA_BALANCING
7600 sgs->nr_numa_running += rq->nr_numa_running;
7601 sgs->nr_preferred_running += rq->nr_preferred_running;
7603 sgs->sum_weighted_load += weighted_cpuload(i);
7605 * No need to call idle_cpu() if nr_running is not 0
7607 if (!nr_running && idle_cpu(i))
7610 if (cpu_overutilized(i)) {
7611 *overutilized = true;
7612 if (!sgs->group_misfit_task && rq->misfit_task)
7613 sgs->group_misfit_task = capacity_of(i);
7617 /* Adjust by relative CPU capacity of the group */
7618 sgs->group_capacity = group->sgc->capacity;
7619 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7621 if (sgs->sum_nr_running)
7622 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7624 sgs->group_weight = group->group_weight;
7626 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7627 sgs->group_type = group_classify(group, sgs);
7631 * update_sd_pick_busiest - return 1 on busiest group
7632 * @env: The load balancing environment.
7633 * @sds: sched_domain statistics
7634 * @sg: sched_group candidate to be checked for being the busiest
7635 * @sgs: sched_group statistics
7637 * Determine if @sg is a busier group than the previously selected
7640 * Return: %true if @sg is a busier group than the previously selected
7641 * busiest group. %false otherwise.
7643 static bool update_sd_pick_busiest(struct lb_env *env,
7644 struct sd_lb_stats *sds,
7645 struct sched_group *sg,
7646 struct sg_lb_stats *sgs)
7648 struct sg_lb_stats *busiest = &sds->busiest_stat;
7650 if (sgs->group_type > busiest->group_type)
7653 if (sgs->group_type < busiest->group_type)
7657 * Candidate sg doesn't face any serious load-balance problems
7658 * so don't pick it if the local sg is already filled up.
7660 if (sgs->group_type == group_other &&
7661 !group_has_capacity(env, &sds->local_stat))
7664 if (sgs->avg_load <= busiest->avg_load)
7667 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
7671 * Candidate sg has no more than one task per CPU and
7672 * has higher per-CPU capacity. Migrating tasks to less
7673 * capable CPUs may harm throughput. Maximize throughput,
7674 * power/energy consequences are not considered.
7676 if (sgs->sum_nr_running <= sgs->group_weight &&
7677 group_smaller_cpu_capacity(sds->local, sg))
7681 /* This is the busiest node in its class. */
7682 if (!(env->sd->flags & SD_ASYM_PACKING))
7686 * ASYM_PACKING needs to move all the work to the lowest
7687 * numbered CPUs in the group, therefore mark all groups
7688 * higher than ourself as busy.
7690 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
7694 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
7701 #ifdef CONFIG_NUMA_BALANCING
7702 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7704 if (sgs->sum_nr_running > sgs->nr_numa_running)
7706 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7711 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7713 if (rq->nr_running > rq->nr_numa_running)
7715 if (rq->nr_running > rq->nr_preferred_running)
7720 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7725 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7729 #endif /* CONFIG_NUMA_BALANCING */
7732 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7733 * @env: The load balancing environment.
7734 * @sds: variable to hold the statistics for this sched_domain.
7736 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7738 struct sched_domain *child = env->sd->child;
7739 struct sched_group *sg = env->sd->groups;
7740 struct sg_lb_stats tmp_sgs;
7741 int load_idx, prefer_sibling = 0;
7742 bool overload = false, overutilized = false;
7744 if (child && child->flags & SD_PREFER_SIBLING)
7747 load_idx = get_sd_load_idx(env->sd, env->idle);
7750 struct sg_lb_stats *sgs = &tmp_sgs;
7753 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
7756 sgs = &sds->local_stat;
7758 if (env->idle != CPU_NEWLY_IDLE ||
7759 time_after_eq(jiffies, sg->sgc->next_update))
7760 update_group_capacity(env->sd, env->dst_cpu);
7763 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7764 &overload, &overutilized);
7770 * In case the child domain prefers tasks go to siblings
7771 * first, lower the sg capacity so that we'll try
7772 * and move all the excess tasks away. We lower the capacity
7773 * of a group only if the local group has the capacity to fit
7774 * these excess tasks. The extra check prevents the case where
7775 * you always pull from the heaviest group when it is already
7776 * under-utilized (possible with a large weight task outweighs
7777 * the tasks on the system).
7779 if (prefer_sibling && sds->local &&
7780 group_has_capacity(env, &sds->local_stat) &&
7781 (sgs->sum_nr_running > 1)) {
7782 sgs->group_no_capacity = 1;
7783 sgs->group_type = group_classify(sg, sgs);
7787 * Ignore task groups with misfit tasks if local group has no
7788 * capacity or if per-cpu capacity isn't higher.
7790 if (sgs->group_type == group_misfit_task &&
7791 (!group_has_capacity(env, &sds->local_stat) ||
7792 !group_smaller_cpu_capacity(sg, sds->local)))
7793 sgs->group_type = group_other;
7795 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7797 sds->busiest_stat = *sgs;
7801 /* Now, start updating sd_lb_stats */
7802 sds->total_load += sgs->group_load;
7803 sds->total_capacity += sgs->group_capacity;
7806 } while (sg != env->sd->groups);
7808 if (env->sd->flags & SD_NUMA)
7809 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7811 env->src_grp_nr_running = sds->busiest_stat.sum_nr_running;
7813 if (!env->sd->parent) {
7814 /* update overload indicator if we are at root domain */
7815 if (env->dst_rq->rd->overload != overload)
7816 env->dst_rq->rd->overload = overload;
7818 /* Update over-utilization (tipping point, U >= 0) indicator */
7819 if (env->dst_rq->rd->overutilized != overutilized) {
7820 env->dst_rq->rd->overutilized = overutilized;
7821 trace_sched_overutilized(overutilized);
7824 if (!env->dst_rq->rd->overutilized && overutilized) {
7825 env->dst_rq->rd->overutilized = true;
7826 trace_sched_overutilized(true);
7833 * check_asym_packing - Check to see if the group is packed into the
7836 * This is primarily intended to used at the sibling level. Some
7837 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7838 * case of POWER7, it can move to lower SMT modes only when higher
7839 * threads are idle. When in lower SMT modes, the threads will
7840 * perform better since they share less core resources. Hence when we
7841 * have idle threads, we want them to be the higher ones.
7843 * This packing function is run on idle threads. It checks to see if
7844 * the busiest CPU in this domain (core in the P7 case) has a higher
7845 * CPU number than the packing function is being run on. Here we are
7846 * assuming lower CPU number will be equivalent to lower a SMT thread
7849 * Return: 1 when packing is required and a task should be moved to
7850 * this CPU. The amount of the imbalance is returned in *imbalance.
7852 * @env: The load balancing environment.
7853 * @sds: Statistics of the sched_domain which is to be packed
7855 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7859 if (!(env->sd->flags & SD_ASYM_PACKING))
7865 busiest_cpu = group_first_cpu(sds->busiest);
7866 if (env->dst_cpu > busiest_cpu)
7869 env->imbalance = DIV_ROUND_CLOSEST(
7870 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7871 SCHED_CAPACITY_SCALE);
7877 * fix_small_imbalance - Calculate the minor imbalance that exists
7878 * amongst the groups of a sched_domain, during
7880 * @env: The load balancing environment.
7881 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7884 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7886 unsigned long tmp, capa_now = 0, capa_move = 0;
7887 unsigned int imbn = 2;
7888 unsigned long scaled_busy_load_per_task;
7889 struct sg_lb_stats *local, *busiest;
7891 local = &sds->local_stat;
7892 busiest = &sds->busiest_stat;
7894 if (!local->sum_nr_running)
7895 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7896 else if (busiest->load_per_task > local->load_per_task)
7899 scaled_busy_load_per_task =
7900 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7901 busiest->group_capacity;
7903 if (busiest->avg_load + scaled_busy_load_per_task >=
7904 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7905 env->imbalance = busiest->load_per_task;
7910 * OK, we don't have enough imbalance to justify moving tasks,
7911 * however we may be able to increase total CPU capacity used by
7915 capa_now += busiest->group_capacity *
7916 min(busiest->load_per_task, busiest->avg_load);
7917 capa_now += local->group_capacity *
7918 min(local->load_per_task, local->avg_load);
7919 capa_now /= SCHED_CAPACITY_SCALE;
7921 /* Amount of load we'd subtract */
7922 if (busiest->avg_load > scaled_busy_load_per_task) {
7923 capa_move += busiest->group_capacity *
7924 min(busiest->load_per_task,
7925 busiest->avg_load - scaled_busy_load_per_task);
7928 /* Amount of load we'd add */
7929 if (busiest->avg_load * busiest->group_capacity <
7930 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7931 tmp = (busiest->avg_load * busiest->group_capacity) /
7932 local->group_capacity;
7934 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7935 local->group_capacity;
7937 capa_move += local->group_capacity *
7938 min(local->load_per_task, local->avg_load + tmp);
7939 capa_move /= SCHED_CAPACITY_SCALE;
7941 /* Move if we gain throughput */
7942 if (capa_move > capa_now)
7943 env->imbalance = busiest->load_per_task;
7947 * calculate_imbalance - Calculate the amount of imbalance present within the
7948 * groups of a given sched_domain during load balance.
7949 * @env: load balance environment
7950 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7952 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7954 unsigned long max_pull, load_above_capacity = ~0UL;
7955 struct sg_lb_stats *local, *busiest;
7957 local = &sds->local_stat;
7958 busiest = &sds->busiest_stat;
7960 if (busiest->group_type == group_imbalanced) {
7962 * In the group_imb case we cannot rely on group-wide averages
7963 * to ensure cpu-load equilibrium, look at wider averages. XXX
7965 busiest->load_per_task =
7966 min(busiest->load_per_task, sds->avg_load);
7970 * In the presence of smp nice balancing, certain scenarios can have
7971 * max load less than avg load(as we skip the groups at or below
7972 * its cpu_capacity, while calculating max_load..)
7974 if (busiest->avg_load <= sds->avg_load ||
7975 local->avg_load >= sds->avg_load) {
7976 /* Misfitting tasks should be migrated in any case */
7977 if (busiest->group_type == group_misfit_task) {
7978 env->imbalance = busiest->group_misfit_task;
7983 * Busiest group is overloaded, local is not, use the spare
7984 * cycles to maximize throughput
7986 if (busiest->group_type == group_overloaded &&
7987 local->group_type <= group_misfit_task) {
7988 env->imbalance = busiest->load_per_task;
7993 return fix_small_imbalance(env, sds);
7997 * If there aren't any idle cpus, avoid creating some.
7999 if (busiest->group_type == group_overloaded &&
8000 local->group_type == group_overloaded) {
8001 load_above_capacity = busiest->sum_nr_running *
8003 if (load_above_capacity > busiest->group_capacity)
8004 load_above_capacity -= busiest->group_capacity;
8006 load_above_capacity = ~0UL;
8010 * We're trying to get all the cpus to the average_load, so we don't
8011 * want to push ourselves above the average load, nor do we wish to
8012 * reduce the max loaded cpu below the average load. At the same time,
8013 * we also don't want to reduce the group load below the group capacity
8014 * (so that we can implement power-savings policies etc). Thus we look
8015 * for the minimum possible imbalance.
8017 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8019 /* How much load to actually move to equalise the imbalance */
8020 env->imbalance = min(
8021 max_pull * busiest->group_capacity,
8022 (sds->avg_load - local->avg_load) * local->group_capacity
8023 ) / SCHED_CAPACITY_SCALE;
8025 /* Boost imbalance to allow misfit task to be balanced. */
8026 if (busiest->group_type == group_misfit_task)
8027 env->imbalance = max_t(long, env->imbalance,
8028 busiest->group_misfit_task);
8031 * if *imbalance is less than the average load per runnable task
8032 * there is no guarantee that any tasks will be moved so we'll have
8033 * a think about bumping its value to force at least one task to be
8036 if (env->imbalance < busiest->load_per_task)
8037 return fix_small_imbalance(env, sds);
8040 /******* find_busiest_group() helpers end here *********************/
8043 * find_busiest_group - Returns the busiest group within the sched_domain
8044 * if there is an imbalance. If there isn't an imbalance, and
8045 * the user has opted for power-savings, it returns a group whose
8046 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
8047 * such a group exists.
8049 * Also calculates the amount of weighted load which should be moved
8050 * to restore balance.
8052 * @env: The load balancing environment.
8054 * Return: - The busiest group if imbalance exists.
8055 * - If no imbalance and user has opted for power-savings balance,
8056 * return the least loaded group whose CPUs can be
8057 * put to idle by rebalancing its tasks onto our group.
8059 static struct sched_group *find_busiest_group(struct lb_env *env)
8061 struct sg_lb_stats *local, *busiest;
8062 struct sd_lb_stats sds;
8064 init_sd_lb_stats(&sds);
8067 * Compute the various statistics relavent for load balancing at
8070 update_sd_lb_stats(env, &sds);
8072 if (energy_aware() && !env->dst_rq->rd->overutilized)
8075 local = &sds.local_stat;
8076 busiest = &sds.busiest_stat;
8078 /* ASYM feature bypasses nice load balance check */
8079 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
8080 check_asym_packing(env, &sds))
8083 /* There is no busy sibling group to pull tasks from */
8084 if (!sds.busiest || busiest->sum_nr_running == 0)
8087 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8088 / sds.total_capacity;
8091 * If the busiest group is imbalanced the below checks don't
8092 * work because they assume all things are equal, which typically
8093 * isn't true due to cpus_allowed constraints and the like.
8095 if (busiest->group_type == group_imbalanced)
8098 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
8099 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
8100 busiest->group_no_capacity)
8103 /* Misfitting tasks should be dealt with regardless of the avg load */
8104 if (busiest->group_type == group_misfit_task) {
8109 * If the local group is busier than the selected busiest group
8110 * don't try and pull any tasks.
8112 if (local->avg_load >= busiest->avg_load)
8116 * Don't pull any tasks if this group is already above the domain
8119 if (local->avg_load >= sds.avg_load)
8122 if (env->idle == CPU_IDLE) {
8124 * This cpu is idle. If the busiest group is not overloaded
8125 * and there is no imbalance between this and busiest group
8126 * wrt idle cpus, it is balanced. The imbalance becomes
8127 * significant if the diff is greater than 1 otherwise we
8128 * might end up to just move the imbalance on another group
8130 if ((busiest->group_type != group_overloaded) &&
8131 (local->idle_cpus <= (busiest->idle_cpus + 1)) &&
8132 !group_smaller_cpu_capacity(sds.busiest, sds.local))
8136 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8137 * imbalance_pct to be conservative.
8139 if (100 * busiest->avg_load <=
8140 env->sd->imbalance_pct * local->avg_load)
8145 env->busiest_group_type = busiest->group_type;
8146 /* Looks like there is an imbalance. Compute it */
8147 calculate_imbalance(env, &sds);
8156 * find_busiest_queue - find the busiest runqueue among the cpus in group.
8158 static struct rq *find_busiest_queue(struct lb_env *env,
8159 struct sched_group *group)
8161 struct rq *busiest = NULL, *rq;
8162 unsigned long busiest_load = 0, busiest_capacity = 1;
8165 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
8166 unsigned long capacity, wl;
8170 rt = fbq_classify_rq(rq);
8173 * We classify groups/runqueues into three groups:
8174 * - regular: there are !numa tasks
8175 * - remote: there are numa tasks that run on the 'wrong' node
8176 * - all: there is no distinction
8178 * In order to avoid migrating ideally placed numa tasks,
8179 * ignore those when there's better options.
8181 * If we ignore the actual busiest queue to migrate another
8182 * task, the next balance pass can still reduce the busiest
8183 * queue by moving tasks around inside the node.
8185 * If we cannot move enough load due to this classification
8186 * the next pass will adjust the group classification and
8187 * allow migration of more tasks.
8189 * Both cases only affect the total convergence complexity.
8191 if (rt > env->fbq_type)
8194 capacity = capacity_of(i);
8196 wl = weighted_cpuload(i);
8199 * When comparing with imbalance, use weighted_cpuload()
8200 * which is not scaled with the cpu capacity.
8203 if (rq->nr_running == 1 && wl > env->imbalance &&
8204 !check_cpu_capacity(rq, env->sd) &&
8205 env->busiest_group_type != group_misfit_task)
8209 * For the load comparisons with the other cpu's, consider
8210 * the weighted_cpuload() scaled with the cpu capacity, so
8211 * that the load can be moved away from the cpu that is
8212 * potentially running at a lower capacity.
8214 * Thus we're looking for max(wl_i / capacity_i), crosswise
8215 * multiplication to rid ourselves of the division works out
8216 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8217 * our previous maximum.
8219 if (wl * busiest_capacity > busiest_load * capacity) {
8221 busiest_capacity = capacity;
8230 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8231 * so long as it is large enough.
8233 #define MAX_PINNED_INTERVAL 512
8235 /* Working cpumask for load_balance and load_balance_newidle. */
8236 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
8238 static int need_active_balance(struct lb_env *env)
8240 struct sched_domain *sd = env->sd;
8242 if (env->idle == CPU_NEWLY_IDLE) {
8245 * ASYM_PACKING needs to force migrate tasks from busy but
8246 * higher numbered CPUs in order to pack all tasks in the
8247 * lowest numbered CPUs.
8249 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
8254 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8255 * It's worth migrating the task if the src_cpu's capacity is reduced
8256 * because of other sched_class or IRQs if more capacity stays
8257 * available on dst_cpu.
8259 if ((env->idle != CPU_NOT_IDLE) &&
8260 (env->src_rq->cfs.h_nr_running == 1)) {
8261 if ((check_cpu_capacity(env->src_rq, sd)) &&
8262 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8266 if ((capacity_of(env->src_cpu) < capacity_of(env->dst_cpu)) &&
8267 env->src_rq->cfs.h_nr_running == 1 &&
8268 cpu_overutilized(env->src_cpu) &&
8269 !cpu_overutilized(env->dst_cpu)) {
8273 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8276 static int active_load_balance_cpu_stop(void *data);
8278 static int should_we_balance(struct lb_env *env)
8280 struct sched_group *sg = env->sd->groups;
8281 struct cpumask *sg_cpus, *sg_mask;
8282 int cpu, balance_cpu = -1;
8285 * In the newly idle case, we will allow all the cpu's
8286 * to do the newly idle load balance.
8288 if (env->idle == CPU_NEWLY_IDLE)
8291 sg_cpus = sched_group_cpus(sg);
8292 sg_mask = sched_group_mask(sg);
8293 /* Try to find first idle cpu */
8294 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
8295 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
8302 if (balance_cpu == -1)
8303 balance_cpu = group_balance_cpu(sg);
8306 * First idle cpu or the first cpu(busiest) in this sched group
8307 * is eligible for doing load balancing at this and above domains.
8309 return balance_cpu == env->dst_cpu;
8313 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8314 * tasks if there is an imbalance.
8316 static int load_balance(int this_cpu, struct rq *this_rq,
8317 struct sched_domain *sd, enum cpu_idle_type idle,
8318 int *continue_balancing)
8320 int ld_moved, cur_ld_moved, active_balance = 0;
8321 struct sched_domain *sd_parent = sd->parent;
8322 struct sched_group *group;
8324 unsigned long flags;
8325 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8327 struct lb_env env = {
8329 .dst_cpu = this_cpu,
8331 .dst_grpmask = sched_group_cpus(sd->groups),
8333 .loop_break = sched_nr_migrate_break,
8336 .tasks = LIST_HEAD_INIT(env.tasks),
8340 * For NEWLY_IDLE load_balancing, we don't need to consider
8341 * other cpus in our group
8343 if (idle == CPU_NEWLY_IDLE)
8344 env.dst_grpmask = NULL;
8346 cpumask_copy(cpus, cpu_active_mask);
8348 schedstat_inc(sd, lb_count[idle]);
8351 if (!should_we_balance(&env)) {
8352 *continue_balancing = 0;
8356 group = find_busiest_group(&env);
8358 schedstat_inc(sd, lb_nobusyg[idle]);
8362 busiest = find_busiest_queue(&env, group);
8364 schedstat_inc(sd, lb_nobusyq[idle]);
8368 BUG_ON(busiest == env.dst_rq);
8370 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
8372 env.src_cpu = busiest->cpu;
8373 env.src_rq = busiest;
8376 if (busiest->nr_running > 1) {
8378 * Attempt to move tasks. If find_busiest_group has found
8379 * an imbalance but busiest->nr_running <= 1, the group is
8380 * still unbalanced. ld_moved simply stays zero, so it is
8381 * correctly treated as an imbalance.
8383 env.flags |= LBF_ALL_PINNED;
8384 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8387 raw_spin_lock_irqsave(&busiest->lock, flags);
8390 * cur_ld_moved - load moved in current iteration
8391 * ld_moved - cumulative load moved across iterations
8393 cur_ld_moved = detach_tasks(&env);
8395 * We want to potentially lower env.src_cpu's OPP.
8398 update_capacity_of(env.src_cpu);
8401 * We've detached some tasks from busiest_rq. Every
8402 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8403 * unlock busiest->lock, and we are able to be sure
8404 * that nobody can manipulate the tasks in parallel.
8405 * See task_rq_lock() family for the details.
8408 raw_spin_unlock(&busiest->lock);
8412 ld_moved += cur_ld_moved;
8415 local_irq_restore(flags);
8417 if (env.flags & LBF_NEED_BREAK) {
8418 env.flags &= ~LBF_NEED_BREAK;
8423 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8424 * us and move them to an alternate dst_cpu in our sched_group
8425 * where they can run. The upper limit on how many times we
8426 * iterate on same src_cpu is dependent on number of cpus in our
8429 * This changes load balance semantics a bit on who can move
8430 * load to a given_cpu. In addition to the given_cpu itself
8431 * (or a ilb_cpu acting on its behalf where given_cpu is
8432 * nohz-idle), we now have balance_cpu in a position to move
8433 * load to given_cpu. In rare situations, this may cause
8434 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8435 * _independently_ and at _same_ time to move some load to
8436 * given_cpu) causing exceess load to be moved to given_cpu.
8437 * This however should not happen so much in practice and
8438 * moreover subsequent load balance cycles should correct the
8439 * excess load moved.
8441 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8443 /* Prevent to re-select dst_cpu via env's cpus */
8444 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8446 env.dst_rq = cpu_rq(env.new_dst_cpu);
8447 env.dst_cpu = env.new_dst_cpu;
8448 env.flags &= ~LBF_DST_PINNED;
8450 env.loop_break = sched_nr_migrate_break;
8453 * Go back to "more_balance" rather than "redo" since we
8454 * need to continue with same src_cpu.
8460 * We failed to reach balance because of affinity.
8463 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8465 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8466 *group_imbalance = 1;
8469 /* All tasks on this runqueue were pinned by CPU affinity */
8470 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8471 cpumask_clear_cpu(cpu_of(busiest), cpus);
8472 if (!cpumask_empty(cpus)) {
8474 env.loop_break = sched_nr_migrate_break;
8477 goto out_all_pinned;
8482 schedstat_inc(sd, lb_failed[idle]);
8484 * Increment the failure counter only on periodic balance.
8485 * We do not want newidle balance, which can be very
8486 * frequent, pollute the failure counter causing
8487 * excessive cache_hot migrations and active balances.
8489 if (idle != CPU_NEWLY_IDLE)
8490 if (env.src_grp_nr_running > 1)
8491 sd->nr_balance_failed++;
8493 if (need_active_balance(&env)) {
8494 raw_spin_lock_irqsave(&busiest->lock, flags);
8496 /* don't kick the active_load_balance_cpu_stop,
8497 * if the curr task on busiest cpu can't be
8500 if (!cpumask_test_cpu(this_cpu,
8501 tsk_cpus_allowed(busiest->curr))) {
8502 raw_spin_unlock_irqrestore(&busiest->lock,
8504 env.flags |= LBF_ALL_PINNED;
8505 goto out_one_pinned;
8509 * ->active_balance synchronizes accesses to
8510 * ->active_balance_work. Once set, it's cleared
8511 * only after active load balance is finished.
8513 if (!busiest->active_balance) {
8514 busiest->active_balance = 1;
8515 busiest->push_cpu = this_cpu;
8518 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8520 if (active_balance) {
8521 stop_one_cpu_nowait(cpu_of(busiest),
8522 active_load_balance_cpu_stop, busiest,
8523 &busiest->active_balance_work);
8527 * We've kicked active balancing, reset the failure
8530 sd->nr_balance_failed = sd->cache_nice_tries+1;
8533 sd->nr_balance_failed = 0;
8535 if (likely(!active_balance)) {
8536 /* We were unbalanced, so reset the balancing interval */
8537 sd->balance_interval = sd->min_interval;
8540 * If we've begun active balancing, start to back off. This
8541 * case may not be covered by the all_pinned logic if there
8542 * is only 1 task on the busy runqueue (because we don't call
8545 if (sd->balance_interval < sd->max_interval)
8546 sd->balance_interval *= 2;
8553 * We reach balance although we may have faced some affinity
8554 * constraints. Clear the imbalance flag if it was set.
8557 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8559 if (*group_imbalance)
8560 *group_imbalance = 0;
8565 * We reach balance because all tasks are pinned at this level so
8566 * we can't migrate them. Let the imbalance flag set so parent level
8567 * can try to migrate them.
8569 schedstat_inc(sd, lb_balanced[idle]);
8571 sd->nr_balance_failed = 0;
8574 /* tune up the balancing interval */
8575 if (((env.flags & LBF_ALL_PINNED) &&
8576 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8577 (sd->balance_interval < sd->max_interval))
8578 sd->balance_interval *= 2;
8585 static inline unsigned long
8586 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8588 unsigned long interval = sd->balance_interval;
8591 interval *= sd->busy_factor;
8593 /* scale ms to jiffies */
8594 interval = msecs_to_jiffies(interval);
8595 interval = clamp(interval, 1UL, max_load_balance_interval);
8601 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
8603 unsigned long interval, next;
8605 interval = get_sd_balance_interval(sd, cpu_busy);
8606 next = sd->last_balance + interval;
8608 if (time_after(*next_balance, next))
8609 *next_balance = next;
8613 * idle_balance is called by schedule() if this_cpu is about to become
8614 * idle. Attempts to pull tasks from other CPUs.
8616 static int idle_balance(struct rq *this_rq)
8618 unsigned long next_balance = jiffies + HZ;
8619 int this_cpu = this_rq->cpu;
8620 struct sched_domain *sd;
8621 int pulled_task = 0;
8623 long removed_util=0;
8625 idle_enter_fair(this_rq);
8628 * We must set idle_stamp _before_ calling idle_balance(), such that we
8629 * measure the duration of idle_balance() as idle time.
8631 this_rq->idle_stamp = rq_clock(this_rq);
8633 if (!energy_aware() &&
8634 (this_rq->avg_idle < sysctl_sched_migration_cost ||
8635 !this_rq->rd->overload)) {
8637 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8639 update_next_balance(sd, 0, &next_balance);
8645 raw_spin_unlock(&this_rq->lock);
8648 * If removed_util_avg is !0 we most probably migrated some task away
8649 * from this_cpu. In this case we might be willing to trigger an OPP
8650 * update, but we want to do so if we don't find anybody else to pull
8651 * here (we will trigger an OPP update with the pulled task's enqueue
8654 * Record removed_util before calling update_blocked_averages, and use
8655 * it below (before returning) to see if an OPP update is required.
8657 removed_util = atomic_long_read(&(this_rq->cfs).removed_util_avg);
8658 update_blocked_averages(this_cpu);
8660 for_each_domain(this_cpu, sd) {
8661 int continue_balancing = 1;
8662 u64 t0, domain_cost;
8664 if (!(sd->flags & SD_LOAD_BALANCE))
8667 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8668 update_next_balance(sd, 0, &next_balance);
8672 if (sd->flags & SD_BALANCE_NEWIDLE) {
8673 t0 = sched_clock_cpu(this_cpu);
8675 pulled_task = load_balance(this_cpu, this_rq,
8677 &continue_balancing);
8679 domain_cost = sched_clock_cpu(this_cpu) - t0;
8680 if (domain_cost > sd->max_newidle_lb_cost)
8681 sd->max_newidle_lb_cost = domain_cost;
8683 curr_cost += domain_cost;
8686 update_next_balance(sd, 0, &next_balance);
8689 * Stop searching for tasks to pull if there are
8690 * now runnable tasks on this rq.
8692 if (pulled_task || this_rq->nr_running > 0)
8697 raw_spin_lock(&this_rq->lock);
8699 if (curr_cost > this_rq->max_idle_balance_cost)
8700 this_rq->max_idle_balance_cost = curr_cost;
8703 * While browsing the domains, we released the rq lock, a task could
8704 * have been enqueued in the meantime. Since we're not going idle,
8705 * pretend we pulled a task.
8707 if (this_rq->cfs.h_nr_running && !pulled_task)
8711 /* Move the next balance forward */
8712 if (time_after(this_rq->next_balance, next_balance))
8713 this_rq->next_balance = next_balance;
8715 /* Is there a task of a high priority class? */
8716 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8720 idle_exit_fair(this_rq);
8721 this_rq->idle_stamp = 0;
8722 } else if (removed_util) {
8724 * No task pulled and someone has been migrated away.
8725 * Good case to trigger an OPP update.
8727 update_capacity_of(this_cpu);
8734 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8735 * running tasks off the busiest CPU onto idle CPUs. It requires at
8736 * least 1 task to be running on each physical CPU where possible, and
8737 * avoids physical / logical imbalances.
8739 static int active_load_balance_cpu_stop(void *data)
8741 struct rq *busiest_rq = data;
8742 int busiest_cpu = cpu_of(busiest_rq);
8743 int target_cpu = busiest_rq->push_cpu;
8744 struct rq *target_rq = cpu_rq(target_cpu);
8745 struct sched_domain *sd;
8746 struct task_struct *p = NULL;
8748 raw_spin_lock_irq(&busiest_rq->lock);
8750 /* make sure the requested cpu hasn't gone down in the meantime */
8751 if (unlikely(busiest_cpu != smp_processor_id() ||
8752 !busiest_rq->active_balance))
8755 /* Is there any task to move? */
8756 if (busiest_rq->nr_running <= 1)
8760 * This condition is "impossible", if it occurs
8761 * we need to fix it. Originally reported by
8762 * Bjorn Helgaas on a 128-cpu setup.
8764 BUG_ON(busiest_rq == target_rq);
8766 /* Search for an sd spanning us and the target CPU. */
8768 for_each_domain(target_cpu, sd) {
8769 if ((sd->flags & SD_LOAD_BALANCE) &&
8770 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8775 struct lb_env env = {
8777 .dst_cpu = target_cpu,
8778 .dst_rq = target_rq,
8779 .src_cpu = busiest_rq->cpu,
8780 .src_rq = busiest_rq,
8784 schedstat_inc(sd, alb_count);
8786 p = detach_one_task(&env);
8788 schedstat_inc(sd, alb_pushed);
8790 * We want to potentially lower env.src_cpu's OPP.
8792 update_capacity_of(env.src_cpu);
8795 schedstat_inc(sd, alb_failed);
8799 busiest_rq->active_balance = 0;
8800 raw_spin_unlock(&busiest_rq->lock);
8803 attach_one_task(target_rq, p);
8810 static inline int on_null_domain(struct rq *rq)
8812 return unlikely(!rcu_dereference_sched(rq->sd));
8815 #ifdef CONFIG_NO_HZ_COMMON
8817 * idle load balancing details
8818 * - When one of the busy CPUs notice that there may be an idle rebalancing
8819 * needed, they will kick the idle load balancer, which then does idle
8820 * load balancing for all the idle CPUs.
8823 cpumask_var_t idle_cpus_mask;
8825 unsigned long next_balance; /* in jiffy units */
8826 } nohz ____cacheline_aligned;
8828 static inline int find_new_ilb(void)
8830 int ilb = cpumask_first(nohz.idle_cpus_mask);
8832 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8839 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8840 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8841 * CPU (if there is one).
8843 static void nohz_balancer_kick(void)
8847 nohz.next_balance++;
8849 ilb_cpu = find_new_ilb();
8851 if (ilb_cpu >= nr_cpu_ids)
8854 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8857 * Use smp_send_reschedule() instead of resched_cpu().
8858 * This way we generate a sched IPI on the target cpu which
8859 * is idle. And the softirq performing nohz idle load balance
8860 * will be run before returning from the IPI.
8862 smp_send_reschedule(ilb_cpu);
8866 static inline void nohz_balance_exit_idle(int cpu)
8868 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8870 * Completely isolated CPUs don't ever set, so we must test.
8872 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8873 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8874 atomic_dec(&nohz.nr_cpus);
8876 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8880 static inline void set_cpu_sd_state_busy(void)
8882 struct sched_domain *sd;
8883 int cpu = smp_processor_id();
8886 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8888 if (!sd || !sd->nohz_idle)
8892 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
8897 void set_cpu_sd_state_idle(void)
8899 struct sched_domain *sd;
8900 int cpu = smp_processor_id();
8903 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8905 if (!sd || sd->nohz_idle)
8909 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
8915 * This routine will record that the cpu is going idle with tick stopped.
8916 * This info will be used in performing idle load balancing in the future.
8918 void nohz_balance_enter_idle(int cpu)
8921 * If this cpu is going down, then nothing needs to be done.
8923 if (!cpu_active(cpu))
8926 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8930 * If we're a completely isolated CPU, we don't play.
8932 if (on_null_domain(cpu_rq(cpu)))
8935 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8936 atomic_inc(&nohz.nr_cpus);
8937 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8940 static int sched_ilb_notifier(struct notifier_block *nfb,
8941 unsigned long action, void *hcpu)
8943 switch (action & ~CPU_TASKS_FROZEN) {
8945 nohz_balance_exit_idle(smp_processor_id());
8953 static DEFINE_SPINLOCK(balancing);
8956 * Scale the max load_balance interval with the number of CPUs in the system.
8957 * This trades load-balance latency on larger machines for less cross talk.
8959 void update_max_interval(void)
8961 max_load_balance_interval = HZ*num_online_cpus()/10;
8965 * It checks each scheduling domain to see if it is due to be balanced,
8966 * and initiates a balancing operation if so.
8968 * Balancing parameters are set up in init_sched_domains.
8970 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8972 int continue_balancing = 1;
8974 unsigned long interval;
8975 struct sched_domain *sd;
8976 /* Earliest time when we have to do rebalance again */
8977 unsigned long next_balance = jiffies + 60*HZ;
8978 int update_next_balance = 0;
8979 int need_serialize, need_decay = 0;
8982 update_blocked_averages(cpu);
8985 for_each_domain(cpu, sd) {
8987 * Decay the newidle max times here because this is a regular
8988 * visit to all the domains. Decay ~1% per second.
8990 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8991 sd->max_newidle_lb_cost =
8992 (sd->max_newidle_lb_cost * 253) / 256;
8993 sd->next_decay_max_lb_cost = jiffies + HZ;
8996 max_cost += sd->max_newidle_lb_cost;
8998 if (!(sd->flags & SD_LOAD_BALANCE))
9002 * Stop the load balance at this level. There is another
9003 * CPU in our sched group which is doing load balancing more
9006 if (!continue_balancing) {
9012 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9014 need_serialize = sd->flags & SD_SERIALIZE;
9015 if (need_serialize) {
9016 if (!spin_trylock(&balancing))
9020 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9021 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9023 * The LBF_DST_PINNED logic could have changed
9024 * env->dst_cpu, so we can't know our idle
9025 * state even if we migrated tasks. Update it.
9027 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9029 sd->last_balance = jiffies;
9030 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9033 spin_unlock(&balancing);
9035 if (time_after(next_balance, sd->last_balance + interval)) {
9036 next_balance = sd->last_balance + interval;
9037 update_next_balance = 1;
9042 * Ensure the rq-wide value also decays but keep it at a
9043 * reasonable floor to avoid funnies with rq->avg_idle.
9045 rq->max_idle_balance_cost =
9046 max((u64)sysctl_sched_migration_cost, max_cost);
9051 * next_balance will be updated only when there is a need.
9052 * When the cpu is attached to null domain for ex, it will not be
9055 if (likely(update_next_balance)) {
9056 rq->next_balance = next_balance;
9058 #ifdef CONFIG_NO_HZ_COMMON
9060 * If this CPU has been elected to perform the nohz idle
9061 * balance. Other idle CPUs have already rebalanced with
9062 * nohz_idle_balance() and nohz.next_balance has been
9063 * updated accordingly. This CPU is now running the idle load
9064 * balance for itself and we need to update the
9065 * nohz.next_balance accordingly.
9067 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9068 nohz.next_balance = rq->next_balance;
9073 #ifdef CONFIG_NO_HZ_COMMON
9075 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9076 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9078 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9080 int this_cpu = this_rq->cpu;
9083 /* Earliest time when we have to do rebalance again */
9084 unsigned long next_balance = jiffies + 60*HZ;
9085 int update_next_balance = 0;
9087 if (idle != CPU_IDLE ||
9088 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
9091 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9092 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9096 * If this cpu gets work to do, stop the load balancing
9097 * work being done for other cpus. Next load
9098 * balancing owner will pick it up.
9103 rq = cpu_rq(balance_cpu);
9106 * If time for next balance is due,
9109 if (time_after_eq(jiffies, rq->next_balance)) {
9110 raw_spin_lock_irq(&rq->lock);
9111 update_rq_clock(rq);
9112 update_idle_cpu_load(rq);
9113 raw_spin_unlock_irq(&rq->lock);
9114 rebalance_domains(rq, CPU_IDLE);
9117 if (time_after(next_balance, rq->next_balance)) {
9118 next_balance = rq->next_balance;
9119 update_next_balance = 1;
9124 * next_balance will be updated only when there is a need.
9125 * When the CPU is attached to null domain for ex, it will not be
9128 if (likely(update_next_balance))
9129 nohz.next_balance = next_balance;
9131 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
9135 * Current heuristic for kicking the idle load balancer in the presence
9136 * of an idle cpu in the system.
9137 * - This rq has more than one task.
9138 * - This rq has at least one CFS task and the capacity of the CPU is
9139 * significantly reduced because of RT tasks or IRQs.
9140 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
9141 * multiple busy cpu.
9142 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
9143 * domain span are idle.
9145 static inline bool nohz_kick_needed(struct rq *rq)
9147 unsigned long now = jiffies;
9148 struct sched_domain *sd;
9149 struct sched_group_capacity *sgc;
9150 int nr_busy, cpu = rq->cpu;
9153 if (unlikely(rq->idle_balance))
9157 * We may be recently in ticked or tickless idle mode. At the first
9158 * busy tick after returning from idle, we will update the busy stats.
9160 set_cpu_sd_state_busy();
9161 nohz_balance_exit_idle(cpu);
9164 * None are in tickless mode and hence no need for NOHZ idle load
9167 if (likely(!atomic_read(&nohz.nr_cpus)))
9170 if (time_before(now, nohz.next_balance))
9173 if (rq->nr_running >= 2 &&
9174 (!energy_aware() || cpu_overutilized(cpu)))
9178 sd = rcu_dereference(per_cpu(sd_busy, cpu));
9179 if (sd && !energy_aware()) {
9180 sgc = sd->groups->sgc;
9181 nr_busy = atomic_read(&sgc->nr_busy_cpus);
9190 sd = rcu_dereference(rq->sd);
9192 if ((rq->cfs.h_nr_running >= 1) &&
9193 check_cpu_capacity(rq, sd)) {
9199 sd = rcu_dereference(per_cpu(sd_asym, cpu));
9200 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
9201 sched_domain_span(sd)) < cpu)) {
9211 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
9215 * run_rebalance_domains is triggered when needed from the scheduler tick.
9216 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9218 static void run_rebalance_domains(struct softirq_action *h)
9220 struct rq *this_rq = this_rq();
9221 enum cpu_idle_type idle = this_rq->idle_balance ?
9222 CPU_IDLE : CPU_NOT_IDLE;
9225 * If this cpu has a pending nohz_balance_kick, then do the
9226 * balancing on behalf of the other idle cpus whose ticks are
9227 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9228 * give the idle cpus a chance to load balance. Else we may
9229 * load balance only within the local sched_domain hierarchy
9230 * and abort nohz_idle_balance altogether if we pull some load.
9232 nohz_idle_balance(this_rq, idle);
9233 rebalance_domains(this_rq, idle);
9237 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9239 void trigger_load_balance(struct rq *rq)
9241 /* Don't need to rebalance while attached to NULL domain */
9242 if (unlikely(on_null_domain(rq)))
9245 if (time_after_eq(jiffies, rq->next_balance))
9246 raise_softirq(SCHED_SOFTIRQ);
9247 #ifdef CONFIG_NO_HZ_COMMON
9248 if (nohz_kick_needed(rq))
9249 nohz_balancer_kick();
9253 static void rq_online_fair(struct rq *rq)
9257 update_runtime_enabled(rq);
9260 static void rq_offline_fair(struct rq *rq)
9264 /* Ensure any throttled groups are reachable by pick_next_task */
9265 unthrottle_offline_cfs_rqs(rq);
9268 #endif /* CONFIG_SMP */
9271 * scheduler tick hitting a task of our scheduling class:
9273 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9275 struct cfs_rq *cfs_rq;
9276 struct sched_entity *se = &curr->se;
9278 for_each_sched_entity(se) {
9279 cfs_rq = cfs_rq_of(se);
9280 entity_tick(cfs_rq, se, queued);
9283 if (static_branch_unlikely(&sched_numa_balancing))
9284 task_tick_numa(rq, curr);
9287 if (!rq->rd->overutilized && cpu_overutilized(task_cpu(curr))) {
9288 rq->rd->overutilized = true;
9289 trace_sched_overutilized(true);
9292 rq->misfit_task = !task_fits_max(curr, rq->cpu);
9298 * called on fork with the child task as argument from the parent's context
9299 * - child not yet on the tasklist
9300 * - preemption disabled
9302 static void task_fork_fair(struct task_struct *p)
9304 struct cfs_rq *cfs_rq;
9305 struct sched_entity *se = &p->se, *curr;
9306 int this_cpu = smp_processor_id();
9307 struct rq *rq = this_rq();
9308 unsigned long flags;
9310 raw_spin_lock_irqsave(&rq->lock, flags);
9312 update_rq_clock(rq);
9314 cfs_rq = task_cfs_rq(current);
9315 curr = cfs_rq->curr;
9318 * Not only the cpu but also the task_group of the parent might have
9319 * been changed after parent->se.parent,cfs_rq were copied to
9320 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
9321 * of child point to valid ones.
9324 __set_task_cpu(p, this_cpu);
9327 update_curr(cfs_rq);
9330 se->vruntime = curr->vruntime;
9331 place_entity(cfs_rq, se, 1);
9333 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9335 * Upon rescheduling, sched_class::put_prev_task() will place
9336 * 'current' within the tree based on its new key value.
9338 swap(curr->vruntime, se->vruntime);
9342 se->vruntime -= cfs_rq->min_vruntime;
9344 raw_spin_unlock_irqrestore(&rq->lock, flags);
9348 * Priority of the task has changed. Check to see if we preempt
9352 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9354 if (!task_on_rq_queued(p))
9358 * Reschedule if we are currently running on this runqueue and
9359 * our priority decreased, or if we are not currently running on
9360 * this runqueue and our priority is higher than the current's
9362 if (rq->curr == p) {
9363 if (p->prio > oldprio)
9366 check_preempt_curr(rq, p, 0);
9369 static inline bool vruntime_normalized(struct task_struct *p)
9371 struct sched_entity *se = &p->se;
9374 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9375 * the dequeue_entity(.flags=0) will already have normalized the
9382 * When !on_rq, vruntime of the task has usually NOT been normalized.
9383 * But there are some cases where it has already been normalized:
9385 * - A forked child which is waiting for being woken up by
9386 * wake_up_new_task().
9387 * - A task which has been woken up by try_to_wake_up() and
9388 * waiting for actually being woken up by sched_ttwu_pending().
9390 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
9396 static void detach_task_cfs_rq(struct task_struct *p)
9398 struct sched_entity *se = &p->se;
9399 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9400 u64 now = cfs_rq_clock_task(cfs_rq);
9402 if (!vruntime_normalized(p)) {
9404 * Fix up our vruntime so that the current sleep doesn't
9405 * cause 'unlimited' sleep bonus.
9407 place_entity(cfs_rq, se, 0);
9408 se->vruntime -= cfs_rq->min_vruntime;
9411 /* Catch up with the cfs_rq and remove our load when we leave */
9412 update_cfs_rq_load_avg(now, cfs_rq, false);
9413 detach_entity_load_avg(cfs_rq, se);
9414 update_tg_load_avg(cfs_rq, false);
9417 static void attach_task_cfs_rq(struct task_struct *p)
9419 struct sched_entity *se = &p->se;
9420 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9421 u64 now = cfs_rq_clock_task(cfs_rq);
9423 #ifdef CONFIG_FAIR_GROUP_SCHED
9425 * Since the real-depth could have been changed (only FAIR
9426 * class maintain depth value), reset depth properly.
9428 se->depth = se->parent ? se->parent->depth + 1 : 0;
9431 /* Synchronize task with its cfs_rq */
9432 update_cfs_rq_load_avg(now, cfs_rq, false);
9433 attach_entity_load_avg(cfs_rq, se);
9434 update_tg_load_avg(cfs_rq, false);
9436 if (!vruntime_normalized(p))
9437 se->vruntime += cfs_rq->min_vruntime;
9440 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9442 detach_task_cfs_rq(p);
9445 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9447 attach_task_cfs_rq(p);
9449 if (task_on_rq_queued(p)) {
9451 * We were most likely switched from sched_rt, so
9452 * kick off the schedule if running, otherwise just see
9453 * if we can still preempt the current task.
9458 check_preempt_curr(rq, p, 0);
9462 /* Account for a task changing its policy or group.
9464 * This routine is mostly called to set cfs_rq->curr field when a task
9465 * migrates between groups/classes.
9467 static void set_curr_task_fair(struct rq *rq)
9469 struct sched_entity *se = &rq->curr->se;
9471 for_each_sched_entity(se) {
9472 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9474 set_next_entity(cfs_rq, se);
9475 /* ensure bandwidth has been allocated on our new cfs_rq */
9476 account_cfs_rq_runtime(cfs_rq, 0);
9480 void init_cfs_rq(struct cfs_rq *cfs_rq)
9482 cfs_rq->tasks_timeline = RB_ROOT;
9483 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9484 #ifndef CONFIG_64BIT
9485 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9488 atomic_long_set(&cfs_rq->removed_load_avg, 0);
9489 atomic_long_set(&cfs_rq->removed_util_avg, 0);
9493 #ifdef CONFIG_FAIR_GROUP_SCHED
9494 static void task_move_group_fair(struct task_struct *p)
9496 detach_task_cfs_rq(p);
9497 set_task_rq(p, task_cpu(p));
9500 /* Tell se's cfs_rq has been changed -- migrated */
9501 p->se.avg.last_update_time = 0;
9503 attach_task_cfs_rq(p);
9506 void free_fair_sched_group(struct task_group *tg)
9510 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9512 for_each_possible_cpu(i) {
9514 kfree(tg->cfs_rq[i]);
9517 remove_entity_load_avg(tg->se[i]);
9526 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9528 struct sched_entity *se;
9529 struct cfs_rq *cfs_rq;
9533 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9536 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9540 tg->shares = NICE_0_LOAD;
9542 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9544 for_each_possible_cpu(i) {
9547 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9548 GFP_KERNEL, cpu_to_node(i));
9552 se = kzalloc_node(sizeof(struct sched_entity),
9553 GFP_KERNEL, cpu_to_node(i));
9557 init_cfs_rq(cfs_rq);
9558 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9559 init_entity_runnable_average(se);
9561 raw_spin_lock_irq(&rq->lock);
9562 post_init_entity_util_avg(se);
9563 raw_spin_unlock_irq(&rq->lock);
9574 void unregister_fair_sched_group(struct task_group *tg, int cpu)
9576 struct rq *rq = cpu_rq(cpu);
9577 unsigned long flags;
9580 * Only empty task groups can be destroyed; so we can speculatively
9581 * check on_list without danger of it being re-added.
9583 if (!tg->cfs_rq[cpu]->on_list)
9586 raw_spin_lock_irqsave(&rq->lock, flags);
9587 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9588 raw_spin_unlock_irqrestore(&rq->lock, flags);
9591 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9592 struct sched_entity *se, int cpu,
9593 struct sched_entity *parent)
9595 struct rq *rq = cpu_rq(cpu);
9599 init_cfs_rq_runtime(cfs_rq);
9601 tg->cfs_rq[cpu] = cfs_rq;
9604 /* se could be NULL for root_task_group */
9609 se->cfs_rq = &rq->cfs;
9612 se->cfs_rq = parent->my_q;
9613 se->depth = parent->depth + 1;
9617 /* guarantee group entities always have weight */
9618 update_load_set(&se->load, NICE_0_LOAD);
9619 se->parent = parent;
9622 static DEFINE_MUTEX(shares_mutex);
9624 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9627 unsigned long flags;
9630 * We can't change the weight of the root cgroup.
9635 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9637 mutex_lock(&shares_mutex);
9638 if (tg->shares == shares)
9641 tg->shares = shares;
9642 for_each_possible_cpu(i) {
9643 struct rq *rq = cpu_rq(i);
9644 struct sched_entity *se;
9647 /* Propagate contribution to hierarchy */
9648 raw_spin_lock_irqsave(&rq->lock, flags);
9650 /* Possible calls to update_curr() need rq clock */
9651 update_rq_clock(rq);
9652 for_each_sched_entity(se)
9653 update_cfs_shares(group_cfs_rq(se));
9654 raw_spin_unlock_irqrestore(&rq->lock, flags);
9658 mutex_unlock(&shares_mutex);
9661 #else /* CONFIG_FAIR_GROUP_SCHED */
9663 void free_fair_sched_group(struct task_group *tg) { }
9665 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9670 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
9672 #endif /* CONFIG_FAIR_GROUP_SCHED */
9675 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9677 struct sched_entity *se = &task->se;
9678 unsigned int rr_interval = 0;
9681 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9684 if (rq->cfs.load.weight)
9685 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9691 * All the scheduling class methods:
9693 const struct sched_class fair_sched_class = {
9694 .next = &idle_sched_class,
9695 .enqueue_task = enqueue_task_fair,
9696 .dequeue_task = dequeue_task_fair,
9697 .yield_task = yield_task_fair,
9698 .yield_to_task = yield_to_task_fair,
9700 .check_preempt_curr = check_preempt_wakeup,
9702 .pick_next_task = pick_next_task_fair,
9703 .put_prev_task = put_prev_task_fair,
9706 .select_task_rq = select_task_rq_fair,
9707 .migrate_task_rq = migrate_task_rq_fair,
9709 .rq_online = rq_online_fair,
9710 .rq_offline = rq_offline_fair,
9712 .task_waking = task_waking_fair,
9713 .task_dead = task_dead_fair,
9714 .set_cpus_allowed = set_cpus_allowed_common,
9717 .set_curr_task = set_curr_task_fair,
9718 .task_tick = task_tick_fair,
9719 .task_fork = task_fork_fair,
9721 .prio_changed = prio_changed_fair,
9722 .switched_from = switched_from_fair,
9723 .switched_to = switched_to_fair,
9725 .get_rr_interval = get_rr_interval_fair,
9727 .update_curr = update_curr_fair,
9729 #ifdef CONFIG_FAIR_GROUP_SCHED
9730 .task_move_group = task_move_group_fair,
9734 #ifdef CONFIG_SCHED_DEBUG
9735 void print_cfs_stats(struct seq_file *m, int cpu)
9737 struct cfs_rq *cfs_rq;
9740 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
9741 print_cfs_rq(m, cpu, cfs_rq);
9745 #ifdef CONFIG_NUMA_BALANCING
9746 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9749 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9751 for_each_online_node(node) {
9752 if (p->numa_faults) {
9753 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9754 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9756 if (p->numa_group) {
9757 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9758 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9760 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9763 #endif /* CONFIG_NUMA_BALANCING */
9764 #endif /* CONFIG_SCHED_DEBUG */
9766 __init void init_sched_fair_class(void)
9769 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9771 #ifdef CONFIG_NO_HZ_COMMON
9772 nohz.next_balance = jiffies;
9773 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
9774 cpu_notifier(sched_ilb_notifier, 0);