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 void attach_entity_cfs_rq(struct sched_entity *se);
725 * With new tasks being created, their initial util_avgs are extrapolated
726 * based on the cfs_rq's current util_avg:
728 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
730 * However, in many cases, the above util_avg does not give a desired
731 * value. Moreover, the sum of the util_avgs may be divergent, such
732 * as when the series is a harmonic series.
734 * To solve this problem, we also cap the util_avg of successive tasks to
735 * only 1/2 of the left utilization budget:
737 * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
739 * where n denotes the nth task.
741 * For example, a simplest series from the beginning would be like:
743 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
744 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
746 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
747 * if util_avg > util_avg_cap.
749 void post_init_entity_util_avg(struct sched_entity *se)
751 struct cfs_rq *cfs_rq = cfs_rq_of(se);
752 struct sched_avg *sa = &se->avg;
753 long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
756 if (cfs_rq->avg.util_avg != 0) {
757 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
758 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
760 if (sa->util_avg > cap)
766 * If we wish to restore tuning via setting initial util,
767 * this is where we should do it.
769 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
772 if (entity_is_task(se)) {
773 struct task_struct *p = task_of(se);
774 if (p->sched_class != &fair_sched_class) {
776 * For !fair tasks do:
778 update_cfs_rq_load_avg(now, cfs_rq, false);
779 attach_entity_load_avg(cfs_rq, se);
780 switched_from_fair(rq, p);
782 * such that the next switched_to_fair() has the
785 se->avg.last_update_time = cfs_rq_clock_task(cfs_rq);
790 attach_entity_cfs_rq(se);
793 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq);
794 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq);
796 void init_entity_runnable_average(struct sched_entity *se)
799 void post_init_entity_util_avg(struct sched_entity *se)
802 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
805 #endif /* CONFIG_SMP */
808 * Update the current task's runtime statistics.
810 static void update_curr(struct cfs_rq *cfs_rq)
812 struct sched_entity *curr = cfs_rq->curr;
813 u64 now = rq_clock_task(rq_of(cfs_rq));
819 delta_exec = now - curr->exec_start;
820 if (unlikely((s64)delta_exec <= 0))
823 curr->exec_start = now;
825 schedstat_set(curr->statistics.exec_max,
826 max(delta_exec, curr->statistics.exec_max));
828 curr->sum_exec_runtime += delta_exec;
829 schedstat_add(cfs_rq, exec_clock, delta_exec);
831 curr->vruntime += calc_delta_fair(delta_exec, curr);
832 update_min_vruntime(cfs_rq);
834 if (entity_is_task(curr)) {
835 struct task_struct *curtask = task_of(curr);
837 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
838 cpuacct_charge(curtask, delta_exec);
839 account_group_exec_runtime(curtask, delta_exec);
842 account_cfs_rq_runtime(cfs_rq, delta_exec);
845 static void update_curr_fair(struct rq *rq)
847 update_curr(cfs_rq_of(&rq->curr->se));
851 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
853 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
857 * Task is being enqueued - update stats:
859 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
862 * Are we enqueueing a waiting task? (for current tasks
863 * a dequeue/enqueue event is a NOP)
865 if (se != cfs_rq->curr)
866 update_stats_wait_start(cfs_rq, se);
870 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
872 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
873 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
874 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
875 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
876 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
877 #ifdef CONFIG_SCHEDSTATS
878 if (entity_is_task(se)) {
879 trace_sched_stat_wait(task_of(se),
880 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
883 schedstat_set(se->statistics.wait_start, 0);
887 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
890 * Mark the end of the wait period if dequeueing a
893 if (se != cfs_rq->curr)
894 update_stats_wait_end(cfs_rq, se);
898 * We are picking a new current task - update its stats:
901 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
904 * We are starting a new run period:
906 se->exec_start = rq_clock_task(rq_of(cfs_rq));
909 /**************************************************
910 * Scheduling class queueing methods:
913 #ifdef CONFIG_NUMA_BALANCING
915 * Approximate time to scan a full NUMA task in ms. The task scan period is
916 * calculated based on the tasks virtual memory size and
917 * numa_balancing_scan_size.
919 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
920 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
922 /* Portion of address space to scan in MB */
923 unsigned int sysctl_numa_balancing_scan_size = 256;
925 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
926 unsigned int sysctl_numa_balancing_scan_delay = 1000;
928 static unsigned int task_nr_scan_windows(struct task_struct *p)
930 unsigned long rss = 0;
931 unsigned long nr_scan_pages;
934 * Calculations based on RSS as non-present and empty pages are skipped
935 * by the PTE scanner and NUMA hinting faults should be trapped based
938 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
939 rss = get_mm_rss(p->mm);
943 rss = round_up(rss, nr_scan_pages);
944 return rss / nr_scan_pages;
947 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
948 #define MAX_SCAN_WINDOW 2560
950 static unsigned int task_scan_min(struct task_struct *p)
952 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
953 unsigned int scan, floor;
954 unsigned int windows = 1;
956 if (scan_size < MAX_SCAN_WINDOW)
957 windows = MAX_SCAN_WINDOW / scan_size;
958 floor = 1000 / windows;
960 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
961 return max_t(unsigned int, floor, scan);
964 static unsigned int task_scan_max(struct task_struct *p)
966 unsigned int smin = task_scan_min(p);
969 /* Watch for min being lower than max due to floor calculations */
970 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
971 return max(smin, smax);
974 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
976 rq->nr_numa_running += (p->numa_preferred_nid != -1);
977 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
980 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
982 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
983 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
989 spinlock_t lock; /* nr_tasks, tasks */
994 nodemask_t active_nodes;
995 unsigned long total_faults;
997 * Faults_cpu is used to decide whether memory should move
998 * towards the CPU. As a consequence, these stats are weighted
999 * more by CPU use than by memory faults.
1001 unsigned long *faults_cpu;
1002 unsigned long faults[0];
1005 /* Shared or private faults. */
1006 #define NR_NUMA_HINT_FAULT_TYPES 2
1008 /* Memory and CPU locality */
1009 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1011 /* Averaged statistics, and temporary buffers. */
1012 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1014 pid_t task_numa_group_id(struct task_struct *p)
1016 return p->numa_group ? p->numa_group->gid : 0;
1020 * The averaged statistics, shared & private, memory & cpu,
1021 * occupy the first half of the array. The second half of the
1022 * array is for current counters, which are averaged into the
1023 * first set by task_numa_placement.
1025 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1027 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1030 static inline unsigned long task_faults(struct task_struct *p, int nid)
1032 if (!p->numa_faults)
1035 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1036 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1039 static inline unsigned long group_faults(struct task_struct *p, int nid)
1044 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1045 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1048 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1050 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1051 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1054 /* Handle placement on systems where not all nodes are directly connected. */
1055 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1056 int maxdist, bool task)
1058 unsigned long score = 0;
1062 * All nodes are directly connected, and the same distance
1063 * from each other. No need for fancy placement algorithms.
1065 if (sched_numa_topology_type == NUMA_DIRECT)
1069 * This code is called for each node, introducing N^2 complexity,
1070 * which should be ok given the number of nodes rarely exceeds 8.
1072 for_each_online_node(node) {
1073 unsigned long faults;
1074 int dist = node_distance(nid, node);
1077 * The furthest away nodes in the system are not interesting
1078 * for placement; nid was already counted.
1080 if (dist == sched_max_numa_distance || node == nid)
1084 * On systems with a backplane NUMA topology, compare groups
1085 * of nodes, and move tasks towards the group with the most
1086 * memory accesses. When comparing two nodes at distance
1087 * "hoplimit", only nodes closer by than "hoplimit" are part
1088 * of each group. Skip other nodes.
1090 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1094 /* Add up the faults from nearby nodes. */
1096 faults = task_faults(p, node);
1098 faults = group_faults(p, node);
1101 * On systems with a glueless mesh NUMA topology, there are
1102 * no fixed "groups of nodes". Instead, nodes that are not
1103 * directly connected bounce traffic through intermediate
1104 * nodes; a numa_group can occupy any set of nodes.
1105 * The further away a node is, the less the faults count.
1106 * This seems to result in good task placement.
1108 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1109 faults *= (sched_max_numa_distance - dist);
1110 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1120 * These return the fraction of accesses done by a particular task, or
1121 * task group, on a particular numa node. The group weight is given a
1122 * larger multiplier, in order to group tasks together that are almost
1123 * evenly spread out between numa nodes.
1125 static inline unsigned long task_weight(struct task_struct *p, int nid,
1128 unsigned long faults, total_faults;
1130 if (!p->numa_faults)
1133 total_faults = p->total_numa_faults;
1138 faults = task_faults(p, nid);
1139 faults += score_nearby_nodes(p, nid, dist, true);
1141 return 1000 * faults / total_faults;
1144 static inline unsigned long group_weight(struct task_struct *p, int nid,
1147 unsigned long faults, total_faults;
1152 total_faults = p->numa_group->total_faults;
1157 faults = group_faults(p, nid);
1158 faults += score_nearby_nodes(p, nid, dist, false);
1160 return 1000 * faults / total_faults;
1163 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1164 int src_nid, int dst_cpu)
1166 struct numa_group *ng = p->numa_group;
1167 int dst_nid = cpu_to_node(dst_cpu);
1168 int last_cpupid, this_cpupid;
1170 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1173 * Multi-stage node selection is used in conjunction with a periodic
1174 * migration fault to build a temporal task<->page relation. By using
1175 * a two-stage filter we remove short/unlikely relations.
1177 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1178 * a task's usage of a particular page (n_p) per total usage of this
1179 * page (n_t) (in a given time-span) to a probability.
1181 * Our periodic faults will sample this probability and getting the
1182 * same result twice in a row, given these samples are fully
1183 * independent, is then given by P(n)^2, provided our sample period
1184 * is sufficiently short compared to the usage pattern.
1186 * This quadric squishes small probabilities, making it less likely we
1187 * act on an unlikely task<->page relation.
1189 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1190 if (!cpupid_pid_unset(last_cpupid) &&
1191 cpupid_to_nid(last_cpupid) != dst_nid)
1194 /* Always allow migrate on private faults */
1195 if (cpupid_match_pid(p, last_cpupid))
1198 /* A shared fault, but p->numa_group has not been set up yet. */
1203 * Do not migrate if the destination is not a node that
1204 * is actively used by this numa group.
1206 if (!node_isset(dst_nid, ng->active_nodes))
1210 * Source is a node that is not actively used by this
1211 * numa group, while the destination is. Migrate.
1213 if (!node_isset(src_nid, ng->active_nodes))
1217 * Both source and destination are nodes in active
1218 * use by this numa group. Maximize memory bandwidth
1219 * by migrating from more heavily used groups, to less
1220 * heavily used ones, spreading the load around.
1221 * Use a 1/4 hysteresis to avoid spurious page movement.
1223 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1226 static unsigned long weighted_cpuload(const int cpu);
1227 static unsigned long source_load(int cpu, int type);
1228 static unsigned long target_load(int cpu, int type);
1229 static unsigned long capacity_of(int cpu);
1230 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1232 /* Cached statistics for all CPUs within a node */
1234 unsigned long nr_running;
1237 /* Total compute capacity of CPUs on a node */
1238 unsigned long compute_capacity;
1240 /* Approximate capacity in terms of runnable tasks on a node */
1241 unsigned long task_capacity;
1242 int has_free_capacity;
1246 * XXX borrowed from update_sg_lb_stats
1248 static void update_numa_stats(struct numa_stats *ns, int nid)
1250 int smt, cpu, cpus = 0;
1251 unsigned long capacity;
1253 memset(ns, 0, sizeof(*ns));
1254 for_each_cpu(cpu, cpumask_of_node(nid)) {
1255 struct rq *rq = cpu_rq(cpu);
1257 ns->nr_running += rq->nr_running;
1258 ns->load += weighted_cpuload(cpu);
1259 ns->compute_capacity += capacity_of(cpu);
1265 * If we raced with hotplug and there are no CPUs left in our mask
1266 * the @ns structure is NULL'ed and task_numa_compare() will
1267 * not find this node attractive.
1269 * We'll either bail at !has_free_capacity, or we'll detect a huge
1270 * imbalance and bail there.
1275 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1276 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1277 capacity = cpus / smt; /* cores */
1279 ns->task_capacity = min_t(unsigned, capacity,
1280 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1281 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1284 struct task_numa_env {
1285 struct task_struct *p;
1287 int src_cpu, src_nid;
1288 int dst_cpu, dst_nid;
1290 struct numa_stats src_stats, dst_stats;
1295 struct task_struct *best_task;
1300 static void task_numa_assign(struct task_numa_env *env,
1301 struct task_struct *p, long imp)
1304 put_task_struct(env->best_task);
1307 env->best_imp = imp;
1308 env->best_cpu = env->dst_cpu;
1311 static bool load_too_imbalanced(long src_load, long dst_load,
1312 struct task_numa_env *env)
1315 long orig_src_load, orig_dst_load;
1316 long src_capacity, dst_capacity;
1319 * The load is corrected for the CPU capacity available on each node.
1322 * ------------ vs ---------
1323 * src_capacity dst_capacity
1325 src_capacity = env->src_stats.compute_capacity;
1326 dst_capacity = env->dst_stats.compute_capacity;
1328 /* We care about the slope of the imbalance, not the direction. */
1329 if (dst_load < src_load)
1330 swap(dst_load, src_load);
1332 /* Is the difference below the threshold? */
1333 imb = dst_load * src_capacity * 100 -
1334 src_load * dst_capacity * env->imbalance_pct;
1339 * The imbalance is above the allowed threshold.
1340 * Compare it with the old imbalance.
1342 orig_src_load = env->src_stats.load;
1343 orig_dst_load = env->dst_stats.load;
1345 if (orig_dst_load < orig_src_load)
1346 swap(orig_dst_load, orig_src_load);
1348 old_imb = orig_dst_load * src_capacity * 100 -
1349 orig_src_load * dst_capacity * env->imbalance_pct;
1351 /* Would this change make things worse? */
1352 return (imb > old_imb);
1356 * This checks if the overall compute and NUMA accesses of the system would
1357 * be improved if the source tasks was migrated to the target dst_cpu taking
1358 * into account that it might be best if task running on the dst_cpu should
1359 * be exchanged with the source task
1361 static void task_numa_compare(struct task_numa_env *env,
1362 long taskimp, long groupimp)
1364 struct rq *src_rq = cpu_rq(env->src_cpu);
1365 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1366 struct task_struct *cur;
1367 long src_load, dst_load;
1369 long imp = env->p->numa_group ? groupimp : taskimp;
1371 int dist = env->dist;
1372 bool assigned = false;
1376 raw_spin_lock_irq(&dst_rq->lock);
1379 * No need to move the exiting task or idle task.
1381 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1385 * The task_struct must be protected here to protect the
1386 * p->numa_faults access in the task_weight since the
1387 * numa_faults could already be freed in the following path:
1388 * finish_task_switch()
1389 * --> put_task_struct()
1390 * --> __put_task_struct()
1391 * --> task_numa_free()
1393 get_task_struct(cur);
1396 raw_spin_unlock_irq(&dst_rq->lock);
1399 * Because we have preemption enabled we can get migrated around and
1400 * end try selecting ourselves (current == env->p) as a swap candidate.
1406 * "imp" is the fault differential for the source task between the
1407 * source and destination node. Calculate the total differential for
1408 * the source task and potential destination task. The more negative
1409 * the value is, the more rmeote accesses that would be expected to
1410 * be incurred if the tasks were swapped.
1413 /* Skip this swap candidate if cannot move to the source cpu */
1414 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1418 * If dst and source tasks are in the same NUMA group, or not
1419 * in any group then look only at task weights.
1421 if (cur->numa_group == env->p->numa_group) {
1422 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1423 task_weight(cur, env->dst_nid, dist);
1425 * Add some hysteresis to prevent swapping the
1426 * tasks within a group over tiny differences.
1428 if (cur->numa_group)
1432 * Compare the group weights. If a task is all by
1433 * itself (not part of a group), use the task weight
1436 if (cur->numa_group)
1437 imp += group_weight(cur, env->src_nid, dist) -
1438 group_weight(cur, env->dst_nid, dist);
1440 imp += task_weight(cur, env->src_nid, dist) -
1441 task_weight(cur, env->dst_nid, dist);
1445 if (imp <= env->best_imp && moveimp <= env->best_imp)
1449 /* Is there capacity at our destination? */
1450 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1451 !env->dst_stats.has_free_capacity)
1457 /* Balance doesn't matter much if we're running a task per cpu */
1458 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1459 dst_rq->nr_running == 1)
1463 * In the overloaded case, try and keep the load balanced.
1466 load = task_h_load(env->p);
1467 dst_load = env->dst_stats.load + load;
1468 src_load = env->src_stats.load - load;
1470 if (moveimp > imp && moveimp > env->best_imp) {
1472 * If the improvement from just moving env->p direction is
1473 * better than swapping tasks around, check if a move is
1474 * possible. Store a slightly smaller score than moveimp,
1475 * so an actually idle CPU will win.
1477 if (!load_too_imbalanced(src_load, dst_load, env)) {
1479 put_task_struct(cur);
1485 if (imp <= env->best_imp)
1489 load = task_h_load(cur);
1494 if (load_too_imbalanced(src_load, dst_load, env))
1498 * One idle CPU per node is evaluated for a task numa move.
1499 * Call select_idle_sibling to maybe find a better one.
1502 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1507 task_numa_assign(env, cur, imp);
1511 * The dst_rq->curr isn't assigned. The protection for task_struct is
1514 if (cur && !assigned)
1515 put_task_struct(cur);
1518 static void task_numa_find_cpu(struct task_numa_env *env,
1519 long taskimp, long groupimp)
1523 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1524 /* Skip this CPU if the source task cannot migrate */
1525 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1529 task_numa_compare(env, taskimp, groupimp);
1533 /* Only move tasks to a NUMA node less busy than the current node. */
1534 static bool numa_has_capacity(struct task_numa_env *env)
1536 struct numa_stats *src = &env->src_stats;
1537 struct numa_stats *dst = &env->dst_stats;
1539 if (src->has_free_capacity && !dst->has_free_capacity)
1543 * Only consider a task move if the source has a higher load
1544 * than the destination, corrected for CPU capacity on each node.
1546 * src->load dst->load
1547 * --------------------- vs ---------------------
1548 * src->compute_capacity dst->compute_capacity
1550 if (src->load * dst->compute_capacity * env->imbalance_pct >
1552 dst->load * src->compute_capacity * 100)
1558 static int task_numa_migrate(struct task_struct *p)
1560 struct task_numa_env env = {
1563 .src_cpu = task_cpu(p),
1564 .src_nid = task_node(p),
1566 .imbalance_pct = 112,
1572 struct sched_domain *sd;
1573 unsigned long taskweight, groupweight;
1575 long taskimp, groupimp;
1578 * Pick the lowest SD_NUMA domain, as that would have the smallest
1579 * imbalance and would be the first to start moving tasks about.
1581 * And we want to avoid any moving of tasks about, as that would create
1582 * random movement of tasks -- counter the numa conditions we're trying
1586 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1588 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1592 * Cpusets can break the scheduler domain tree into smaller
1593 * balance domains, some of which do not cross NUMA boundaries.
1594 * Tasks that are "trapped" in such domains cannot be migrated
1595 * elsewhere, so there is no point in (re)trying.
1597 if (unlikely(!sd)) {
1598 p->numa_preferred_nid = task_node(p);
1602 env.dst_nid = p->numa_preferred_nid;
1603 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1604 taskweight = task_weight(p, env.src_nid, dist);
1605 groupweight = group_weight(p, env.src_nid, dist);
1606 update_numa_stats(&env.src_stats, env.src_nid);
1607 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1608 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1609 update_numa_stats(&env.dst_stats, env.dst_nid);
1611 /* Try to find a spot on the preferred nid. */
1612 if (numa_has_capacity(&env))
1613 task_numa_find_cpu(&env, taskimp, groupimp);
1616 * Look at other nodes in these cases:
1617 * - there is no space available on the preferred_nid
1618 * - the task is part of a numa_group that is interleaved across
1619 * multiple NUMA nodes; in order to better consolidate the group,
1620 * we need to check other locations.
1622 if (env.best_cpu == -1 || (p->numa_group &&
1623 nodes_weight(p->numa_group->active_nodes) > 1)) {
1624 for_each_online_node(nid) {
1625 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1628 dist = node_distance(env.src_nid, env.dst_nid);
1629 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1631 taskweight = task_weight(p, env.src_nid, dist);
1632 groupweight = group_weight(p, env.src_nid, dist);
1635 /* Only consider nodes where both task and groups benefit */
1636 taskimp = task_weight(p, nid, dist) - taskweight;
1637 groupimp = group_weight(p, nid, dist) - groupweight;
1638 if (taskimp < 0 && groupimp < 0)
1643 update_numa_stats(&env.dst_stats, env.dst_nid);
1644 if (numa_has_capacity(&env))
1645 task_numa_find_cpu(&env, taskimp, groupimp);
1650 * If the task is part of a workload that spans multiple NUMA nodes,
1651 * and is migrating into one of the workload's active nodes, remember
1652 * this node as the task's preferred numa node, so the workload can
1654 * A task that migrated to a second choice node will be better off
1655 * trying for a better one later. Do not set the preferred node here.
1657 if (p->numa_group) {
1658 if (env.best_cpu == -1)
1663 if (node_isset(nid, p->numa_group->active_nodes))
1664 sched_setnuma(p, env.dst_nid);
1667 /* No better CPU than the current one was found. */
1668 if (env.best_cpu == -1)
1672 * Reset the scan period if the task is being rescheduled on an
1673 * alternative node to recheck if the tasks is now properly placed.
1675 p->numa_scan_period = task_scan_min(p);
1677 if (env.best_task == NULL) {
1678 ret = migrate_task_to(p, env.best_cpu);
1680 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1684 ret = migrate_swap(p, env.best_task);
1686 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1687 put_task_struct(env.best_task);
1691 /* Attempt to migrate a task to a CPU on the preferred node. */
1692 static void numa_migrate_preferred(struct task_struct *p)
1694 unsigned long interval = HZ;
1696 /* This task has no NUMA fault statistics yet */
1697 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1700 /* Periodically retry migrating the task to the preferred node */
1701 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1702 p->numa_migrate_retry = jiffies + interval;
1704 /* Success if task is already running on preferred CPU */
1705 if (task_node(p) == p->numa_preferred_nid)
1708 /* Otherwise, try migrate to a CPU on the preferred node */
1709 task_numa_migrate(p);
1713 * Find the nodes on which the workload is actively running. We do this by
1714 * tracking the nodes from which NUMA hinting faults are triggered. This can
1715 * be different from the set of nodes where the workload's memory is currently
1718 * The bitmask is used to make smarter decisions on when to do NUMA page
1719 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1720 * are added when they cause over 6/16 of the maximum number of faults, but
1721 * only removed when they drop below 3/16.
1723 static void update_numa_active_node_mask(struct numa_group *numa_group)
1725 unsigned long faults, max_faults = 0;
1728 for_each_online_node(nid) {
1729 faults = group_faults_cpu(numa_group, nid);
1730 if (faults > max_faults)
1731 max_faults = faults;
1734 for_each_online_node(nid) {
1735 faults = group_faults_cpu(numa_group, nid);
1736 if (!node_isset(nid, numa_group->active_nodes)) {
1737 if (faults > max_faults * 6 / 16)
1738 node_set(nid, numa_group->active_nodes);
1739 } else if (faults < max_faults * 3 / 16)
1740 node_clear(nid, numa_group->active_nodes);
1745 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1746 * increments. The more local the fault statistics are, the higher the scan
1747 * period will be for the next scan window. If local/(local+remote) ratio is
1748 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1749 * the scan period will decrease. Aim for 70% local accesses.
1751 #define NUMA_PERIOD_SLOTS 10
1752 #define NUMA_PERIOD_THRESHOLD 7
1755 * Increase the scan period (slow down scanning) if the majority of
1756 * our memory is already on our local node, or if the majority of
1757 * the page accesses are shared with other processes.
1758 * Otherwise, decrease the scan period.
1760 static void update_task_scan_period(struct task_struct *p,
1761 unsigned long shared, unsigned long private)
1763 unsigned int period_slot;
1767 unsigned long remote = p->numa_faults_locality[0];
1768 unsigned long local = p->numa_faults_locality[1];
1771 * If there were no record hinting faults then either the task is
1772 * completely idle or all activity is areas that are not of interest
1773 * to automatic numa balancing. Related to that, if there were failed
1774 * migration then it implies we are migrating too quickly or the local
1775 * node is overloaded. In either case, scan slower
1777 if (local + shared == 0 || p->numa_faults_locality[2]) {
1778 p->numa_scan_period = min(p->numa_scan_period_max,
1779 p->numa_scan_period << 1);
1781 p->mm->numa_next_scan = jiffies +
1782 msecs_to_jiffies(p->numa_scan_period);
1788 * Prepare to scale scan period relative to the current period.
1789 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1790 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1791 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1793 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1794 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1795 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1796 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1799 diff = slot * period_slot;
1801 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1804 * Scale scan rate increases based on sharing. There is an
1805 * inverse relationship between the degree of sharing and
1806 * the adjustment made to the scanning period. Broadly
1807 * speaking the intent is that there is little point
1808 * scanning faster if shared accesses dominate as it may
1809 * simply bounce migrations uselessly
1811 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1812 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1815 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1816 task_scan_min(p), task_scan_max(p));
1817 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1821 * Get the fraction of time the task has been running since the last
1822 * NUMA placement cycle. The scheduler keeps similar statistics, but
1823 * decays those on a 32ms period, which is orders of magnitude off
1824 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1825 * stats only if the task is so new there are no NUMA statistics yet.
1827 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1829 u64 runtime, delta, now;
1830 /* Use the start of this time slice to avoid calculations. */
1831 now = p->se.exec_start;
1832 runtime = p->se.sum_exec_runtime;
1834 if (p->last_task_numa_placement) {
1835 delta = runtime - p->last_sum_exec_runtime;
1836 *period = now - p->last_task_numa_placement;
1838 delta = p->se.avg.load_sum / p->se.load.weight;
1839 *period = LOAD_AVG_MAX;
1842 p->last_sum_exec_runtime = runtime;
1843 p->last_task_numa_placement = now;
1849 * Determine the preferred nid for a task in a numa_group. This needs to
1850 * be done in a way that produces consistent results with group_weight,
1851 * otherwise workloads might not converge.
1853 static int preferred_group_nid(struct task_struct *p, int nid)
1858 /* Direct connections between all NUMA nodes. */
1859 if (sched_numa_topology_type == NUMA_DIRECT)
1863 * On a system with glueless mesh NUMA topology, group_weight
1864 * scores nodes according to the number of NUMA hinting faults on
1865 * both the node itself, and on nearby nodes.
1867 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1868 unsigned long score, max_score = 0;
1869 int node, max_node = nid;
1871 dist = sched_max_numa_distance;
1873 for_each_online_node(node) {
1874 score = group_weight(p, node, dist);
1875 if (score > max_score) {
1884 * Finding the preferred nid in a system with NUMA backplane
1885 * interconnect topology is more involved. The goal is to locate
1886 * tasks from numa_groups near each other in the system, and
1887 * untangle workloads from different sides of the system. This requires
1888 * searching down the hierarchy of node groups, recursively searching
1889 * inside the highest scoring group of nodes. The nodemask tricks
1890 * keep the complexity of the search down.
1892 nodes = node_online_map;
1893 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1894 unsigned long max_faults = 0;
1895 nodemask_t max_group = NODE_MASK_NONE;
1898 /* Are there nodes at this distance from each other? */
1899 if (!find_numa_distance(dist))
1902 for_each_node_mask(a, nodes) {
1903 unsigned long faults = 0;
1904 nodemask_t this_group;
1905 nodes_clear(this_group);
1907 /* Sum group's NUMA faults; includes a==b case. */
1908 for_each_node_mask(b, nodes) {
1909 if (node_distance(a, b) < dist) {
1910 faults += group_faults(p, b);
1911 node_set(b, this_group);
1912 node_clear(b, nodes);
1916 /* Remember the top group. */
1917 if (faults > max_faults) {
1918 max_faults = faults;
1919 max_group = this_group;
1921 * subtle: at the smallest distance there is
1922 * just one node left in each "group", the
1923 * winner is the preferred nid.
1928 /* Next round, evaluate the nodes within max_group. */
1936 static void task_numa_placement(struct task_struct *p)
1938 int seq, nid, max_nid = -1, max_group_nid = -1;
1939 unsigned long max_faults = 0, max_group_faults = 0;
1940 unsigned long fault_types[2] = { 0, 0 };
1941 unsigned long total_faults;
1942 u64 runtime, period;
1943 spinlock_t *group_lock = NULL;
1946 * The p->mm->numa_scan_seq field gets updated without
1947 * exclusive access. Use READ_ONCE() here to ensure
1948 * that the field is read in a single access:
1950 seq = READ_ONCE(p->mm->numa_scan_seq);
1951 if (p->numa_scan_seq == seq)
1953 p->numa_scan_seq = seq;
1954 p->numa_scan_period_max = task_scan_max(p);
1956 total_faults = p->numa_faults_locality[0] +
1957 p->numa_faults_locality[1];
1958 runtime = numa_get_avg_runtime(p, &period);
1960 /* If the task is part of a group prevent parallel updates to group stats */
1961 if (p->numa_group) {
1962 group_lock = &p->numa_group->lock;
1963 spin_lock_irq(group_lock);
1966 /* Find the node with the highest number of faults */
1967 for_each_online_node(nid) {
1968 /* Keep track of the offsets in numa_faults array */
1969 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1970 unsigned long faults = 0, group_faults = 0;
1973 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1974 long diff, f_diff, f_weight;
1976 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1977 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1978 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1979 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1981 /* Decay existing window, copy faults since last scan */
1982 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1983 fault_types[priv] += p->numa_faults[membuf_idx];
1984 p->numa_faults[membuf_idx] = 0;
1987 * Normalize the faults_from, so all tasks in a group
1988 * count according to CPU use, instead of by the raw
1989 * number of faults. Tasks with little runtime have
1990 * little over-all impact on throughput, and thus their
1991 * faults are less important.
1993 f_weight = div64_u64(runtime << 16, period + 1);
1994 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1996 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1997 p->numa_faults[cpubuf_idx] = 0;
1999 p->numa_faults[mem_idx] += diff;
2000 p->numa_faults[cpu_idx] += f_diff;
2001 faults += p->numa_faults[mem_idx];
2002 p->total_numa_faults += diff;
2003 if (p->numa_group) {
2005 * safe because we can only change our own group
2007 * mem_idx represents the offset for a given
2008 * nid and priv in a specific region because it
2009 * is at the beginning of the numa_faults array.
2011 p->numa_group->faults[mem_idx] += diff;
2012 p->numa_group->faults_cpu[mem_idx] += f_diff;
2013 p->numa_group->total_faults += diff;
2014 group_faults += p->numa_group->faults[mem_idx];
2018 if (faults > max_faults) {
2019 max_faults = faults;
2023 if (group_faults > max_group_faults) {
2024 max_group_faults = group_faults;
2025 max_group_nid = nid;
2029 update_task_scan_period(p, fault_types[0], fault_types[1]);
2031 if (p->numa_group) {
2032 update_numa_active_node_mask(p->numa_group);
2033 spin_unlock_irq(group_lock);
2034 max_nid = preferred_group_nid(p, max_group_nid);
2038 /* Set the new preferred node */
2039 if (max_nid != p->numa_preferred_nid)
2040 sched_setnuma(p, max_nid);
2042 if (task_node(p) != p->numa_preferred_nid)
2043 numa_migrate_preferred(p);
2047 static inline int get_numa_group(struct numa_group *grp)
2049 return atomic_inc_not_zero(&grp->refcount);
2052 static inline void put_numa_group(struct numa_group *grp)
2054 if (atomic_dec_and_test(&grp->refcount))
2055 kfree_rcu(grp, rcu);
2058 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2061 struct numa_group *grp, *my_grp;
2062 struct task_struct *tsk;
2064 int cpu = cpupid_to_cpu(cpupid);
2067 if (unlikely(!p->numa_group)) {
2068 unsigned int size = sizeof(struct numa_group) +
2069 4*nr_node_ids*sizeof(unsigned long);
2071 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2075 atomic_set(&grp->refcount, 1);
2076 spin_lock_init(&grp->lock);
2078 /* Second half of the array tracks nids where faults happen */
2079 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2082 node_set(task_node(current), grp->active_nodes);
2084 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2085 grp->faults[i] = p->numa_faults[i];
2087 grp->total_faults = p->total_numa_faults;
2090 rcu_assign_pointer(p->numa_group, grp);
2094 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2096 if (!cpupid_match_pid(tsk, cpupid))
2099 grp = rcu_dereference(tsk->numa_group);
2103 my_grp = p->numa_group;
2108 * Only join the other group if its bigger; if we're the bigger group,
2109 * the other task will join us.
2111 if (my_grp->nr_tasks > grp->nr_tasks)
2115 * Tie-break on the grp address.
2117 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2120 /* Always join threads in the same process. */
2121 if (tsk->mm == current->mm)
2124 /* Simple filter to avoid false positives due to PID collisions */
2125 if (flags & TNF_SHARED)
2128 /* Update priv based on whether false sharing was detected */
2131 if (join && !get_numa_group(grp))
2139 BUG_ON(irqs_disabled());
2140 double_lock_irq(&my_grp->lock, &grp->lock);
2142 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2143 my_grp->faults[i] -= p->numa_faults[i];
2144 grp->faults[i] += p->numa_faults[i];
2146 my_grp->total_faults -= p->total_numa_faults;
2147 grp->total_faults += p->total_numa_faults;
2152 spin_unlock(&my_grp->lock);
2153 spin_unlock_irq(&grp->lock);
2155 rcu_assign_pointer(p->numa_group, grp);
2157 put_numa_group(my_grp);
2165 void task_numa_free(struct task_struct *p)
2167 struct numa_group *grp = p->numa_group;
2168 void *numa_faults = p->numa_faults;
2169 unsigned long flags;
2173 spin_lock_irqsave(&grp->lock, flags);
2174 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2175 grp->faults[i] -= p->numa_faults[i];
2176 grp->total_faults -= p->total_numa_faults;
2179 spin_unlock_irqrestore(&grp->lock, flags);
2180 RCU_INIT_POINTER(p->numa_group, NULL);
2181 put_numa_group(grp);
2184 p->numa_faults = NULL;
2189 * Got a PROT_NONE fault for a page on @node.
2191 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2193 struct task_struct *p = current;
2194 bool migrated = flags & TNF_MIGRATED;
2195 int cpu_node = task_node(current);
2196 int local = !!(flags & TNF_FAULT_LOCAL);
2199 if (!static_branch_likely(&sched_numa_balancing))
2202 /* for example, ksmd faulting in a user's mm */
2206 /* Allocate buffer to track faults on a per-node basis */
2207 if (unlikely(!p->numa_faults)) {
2208 int size = sizeof(*p->numa_faults) *
2209 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2211 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2212 if (!p->numa_faults)
2215 p->total_numa_faults = 0;
2216 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2220 * First accesses are treated as private, otherwise consider accesses
2221 * to be private if the accessing pid has not changed
2223 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2226 priv = cpupid_match_pid(p, last_cpupid);
2227 if (!priv && !(flags & TNF_NO_GROUP))
2228 task_numa_group(p, last_cpupid, flags, &priv);
2232 * If a workload spans multiple NUMA nodes, a shared fault that
2233 * occurs wholly within the set of nodes that the workload is
2234 * actively using should be counted as local. This allows the
2235 * scan rate to slow down when a workload has settled down.
2237 if (!priv && !local && p->numa_group &&
2238 node_isset(cpu_node, p->numa_group->active_nodes) &&
2239 node_isset(mem_node, p->numa_group->active_nodes))
2242 task_numa_placement(p);
2245 * Retry task to preferred node migration periodically, in case it
2246 * case it previously failed, or the scheduler moved us.
2248 if (time_after(jiffies, p->numa_migrate_retry))
2249 numa_migrate_preferred(p);
2252 p->numa_pages_migrated += pages;
2253 if (flags & TNF_MIGRATE_FAIL)
2254 p->numa_faults_locality[2] += pages;
2256 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2257 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2258 p->numa_faults_locality[local] += pages;
2261 static void reset_ptenuma_scan(struct task_struct *p)
2264 * We only did a read acquisition of the mmap sem, so
2265 * p->mm->numa_scan_seq is written to without exclusive access
2266 * and the update is not guaranteed to be atomic. That's not
2267 * much of an issue though, since this is just used for
2268 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2269 * expensive, to avoid any form of compiler optimizations:
2271 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2272 p->mm->numa_scan_offset = 0;
2276 * The expensive part of numa migration is done from task_work context.
2277 * Triggered from task_tick_numa().
2279 void task_numa_work(struct callback_head *work)
2281 unsigned long migrate, next_scan, now = jiffies;
2282 struct task_struct *p = current;
2283 struct mm_struct *mm = p->mm;
2284 struct vm_area_struct *vma;
2285 unsigned long start, end;
2286 unsigned long nr_pte_updates = 0;
2287 long pages, virtpages;
2289 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2291 work->next = work; /* protect against double add */
2293 * Who cares about NUMA placement when they're dying.
2295 * NOTE: make sure not to dereference p->mm before this check,
2296 * exit_task_work() happens _after_ exit_mm() so we could be called
2297 * without p->mm even though we still had it when we enqueued this
2300 if (p->flags & PF_EXITING)
2303 if (!mm->numa_next_scan) {
2304 mm->numa_next_scan = now +
2305 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2309 * Enforce maximal scan/migration frequency..
2311 migrate = mm->numa_next_scan;
2312 if (time_before(now, migrate))
2315 if (p->numa_scan_period == 0) {
2316 p->numa_scan_period_max = task_scan_max(p);
2317 p->numa_scan_period = task_scan_min(p);
2320 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2321 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2325 * Delay this task enough that another task of this mm will likely win
2326 * the next time around.
2328 p->node_stamp += 2 * TICK_NSEC;
2330 start = mm->numa_scan_offset;
2331 pages = sysctl_numa_balancing_scan_size;
2332 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2333 virtpages = pages * 8; /* Scan up to this much virtual space */
2338 down_read(&mm->mmap_sem);
2339 vma = find_vma(mm, start);
2341 reset_ptenuma_scan(p);
2345 for (; vma; vma = vma->vm_next) {
2346 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2347 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2352 * Shared library pages mapped by multiple processes are not
2353 * migrated as it is expected they are cache replicated. Avoid
2354 * hinting faults in read-only file-backed mappings or the vdso
2355 * as migrating the pages will be of marginal benefit.
2358 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2362 * Skip inaccessible VMAs to avoid any confusion between
2363 * PROT_NONE and NUMA hinting ptes
2365 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2369 start = max(start, vma->vm_start);
2370 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2371 end = min(end, vma->vm_end);
2372 nr_pte_updates = change_prot_numa(vma, start, end);
2375 * Try to scan sysctl_numa_balancing_size worth of
2376 * hpages that have at least one present PTE that
2377 * is not already pte-numa. If the VMA contains
2378 * areas that are unused or already full of prot_numa
2379 * PTEs, scan up to virtpages, to skip through those
2383 pages -= (end - start) >> PAGE_SHIFT;
2384 virtpages -= (end - start) >> PAGE_SHIFT;
2387 if (pages <= 0 || virtpages <= 0)
2391 } while (end != vma->vm_end);
2396 * It is possible to reach the end of the VMA list but the last few
2397 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2398 * would find the !migratable VMA on the next scan but not reset the
2399 * scanner to the start so check it now.
2402 mm->numa_scan_offset = start;
2404 reset_ptenuma_scan(p);
2405 up_read(&mm->mmap_sem);
2409 * Drive the periodic memory faults..
2411 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2413 struct callback_head *work = &curr->numa_work;
2417 * We don't care about NUMA placement if we don't have memory.
2419 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2423 * Using runtime rather than walltime has the dual advantage that
2424 * we (mostly) drive the selection from busy threads and that the
2425 * task needs to have done some actual work before we bother with
2428 now = curr->se.sum_exec_runtime;
2429 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2431 if (now > curr->node_stamp + period) {
2432 if (!curr->node_stamp)
2433 curr->numa_scan_period = task_scan_min(curr);
2434 curr->node_stamp += period;
2436 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2437 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2438 task_work_add(curr, work, true);
2443 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2447 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2451 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2454 #endif /* CONFIG_NUMA_BALANCING */
2457 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2459 update_load_add(&cfs_rq->load, se->load.weight);
2460 if (!parent_entity(se))
2461 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2463 if (entity_is_task(se)) {
2464 struct rq *rq = rq_of(cfs_rq);
2466 account_numa_enqueue(rq, task_of(se));
2467 list_add(&se->group_node, &rq->cfs_tasks);
2470 cfs_rq->nr_running++;
2474 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2476 update_load_sub(&cfs_rq->load, se->load.weight);
2477 if (!parent_entity(se))
2478 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2479 if (entity_is_task(se)) {
2480 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2481 list_del_init(&se->group_node);
2483 cfs_rq->nr_running--;
2486 #ifdef CONFIG_FAIR_GROUP_SCHED
2488 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2493 * Use this CPU's real-time load instead of the last load contribution
2494 * as the updating of the contribution is delayed, and we will use the
2495 * the real-time load to calc the share. See update_tg_load_avg().
2497 tg_weight = atomic_long_read(&tg->load_avg);
2498 tg_weight -= cfs_rq->tg_load_avg_contrib;
2499 tg_weight += cfs_rq->load.weight;
2504 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2506 long tg_weight, load, shares;
2508 tg_weight = calc_tg_weight(tg, cfs_rq);
2509 load = cfs_rq->load.weight;
2511 shares = (tg->shares * load);
2513 shares /= tg_weight;
2515 if (shares < MIN_SHARES)
2516 shares = MIN_SHARES;
2517 if (shares > tg->shares)
2518 shares = tg->shares;
2522 # else /* CONFIG_SMP */
2523 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2527 # endif /* CONFIG_SMP */
2528 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2529 unsigned long weight)
2532 /* commit outstanding execution time */
2533 if (cfs_rq->curr == se)
2534 update_curr(cfs_rq);
2535 account_entity_dequeue(cfs_rq, se);
2538 update_load_set(&se->load, weight);
2541 account_entity_enqueue(cfs_rq, se);
2544 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2546 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2548 struct task_group *tg;
2549 struct sched_entity *se;
2553 se = tg->se[cpu_of(rq_of(cfs_rq))];
2554 if (!se || throttled_hierarchy(cfs_rq))
2557 if (likely(se->load.weight == tg->shares))
2560 shares = calc_cfs_shares(cfs_rq, tg);
2562 reweight_entity(cfs_rq_of(se), se, shares);
2564 #else /* CONFIG_FAIR_GROUP_SCHED */
2565 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2568 #endif /* CONFIG_FAIR_GROUP_SCHED */
2571 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2572 static const u32 runnable_avg_yN_inv[] = {
2573 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2574 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2575 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2576 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2577 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2578 0x85aac367, 0x82cd8698,
2582 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2583 * over-estimates when re-combining.
2585 static const u32 runnable_avg_yN_sum[] = {
2586 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2587 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2588 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2593 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2595 static __always_inline u64 decay_load(u64 val, u64 n)
2597 unsigned int local_n;
2601 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2604 /* after bounds checking we can collapse to 32-bit */
2608 * As y^PERIOD = 1/2, we can combine
2609 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2610 * With a look-up table which covers y^n (n<PERIOD)
2612 * To achieve constant time decay_load.
2614 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2615 val >>= local_n / LOAD_AVG_PERIOD;
2616 local_n %= LOAD_AVG_PERIOD;
2619 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2624 * For updates fully spanning n periods, the contribution to runnable
2625 * average will be: \Sum 1024*y^n
2627 * We can compute this reasonably efficiently by combining:
2628 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2630 static u32 __compute_runnable_contrib(u64 n)
2634 if (likely(n <= LOAD_AVG_PERIOD))
2635 return runnable_avg_yN_sum[n];
2636 else if (unlikely(n >= LOAD_AVG_MAX_N))
2637 return LOAD_AVG_MAX;
2639 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2641 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2642 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2644 n -= LOAD_AVG_PERIOD;
2645 } while (n > LOAD_AVG_PERIOD);
2647 contrib = decay_load(contrib, n);
2648 return contrib + runnable_avg_yN_sum[n];
2651 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2652 #error "load tracking assumes 2^10 as unit"
2655 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2658 * We can represent the historical contribution to runnable average as the
2659 * coefficients of a geometric series. To do this we sub-divide our runnable
2660 * history into segments of approximately 1ms (1024us); label the segment that
2661 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2663 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2665 * (now) (~1ms ago) (~2ms ago)
2667 * Let u_i denote the fraction of p_i that the entity was runnable.
2669 * We then designate the fractions u_i as our co-efficients, yielding the
2670 * following representation of historical load:
2671 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2673 * We choose y based on the with of a reasonably scheduling period, fixing:
2676 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2677 * approximately half as much as the contribution to load within the last ms
2680 * When a period "rolls over" and we have new u_0`, multiplying the previous
2681 * sum again by y is sufficient to update:
2682 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2683 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2685 static __always_inline int
2686 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2687 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2689 u64 delta, scaled_delta, periods;
2691 unsigned int delta_w, scaled_delta_w, decayed = 0;
2692 unsigned long scale_freq, scale_cpu;
2694 delta = now - sa->last_update_time;
2696 * This should only happen when time goes backwards, which it
2697 * unfortunately does during sched clock init when we swap over to TSC.
2699 if ((s64)delta < 0) {
2700 sa->last_update_time = now;
2705 * Use 1024ns as the unit of measurement since it's a reasonable
2706 * approximation of 1us and fast to compute.
2711 sa->last_update_time = now;
2713 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2714 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2715 trace_sched_contrib_scale_f(cpu, scale_freq, scale_cpu);
2717 /* delta_w is the amount already accumulated against our next period */
2718 delta_w = sa->period_contrib;
2719 if (delta + delta_w >= 1024) {
2722 /* how much left for next period will start over, we don't know yet */
2723 sa->period_contrib = 0;
2726 * Now that we know we're crossing a period boundary, figure
2727 * out how much from delta we need to complete the current
2728 * period and accrue it.
2730 delta_w = 1024 - delta_w;
2731 scaled_delta_w = cap_scale(delta_w, scale_freq);
2733 sa->load_sum += weight * scaled_delta_w;
2735 cfs_rq->runnable_load_sum +=
2736 weight * scaled_delta_w;
2740 sa->util_sum += scaled_delta_w * scale_cpu;
2744 /* Figure out how many additional periods this update spans */
2745 periods = delta / 1024;
2748 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2750 cfs_rq->runnable_load_sum =
2751 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2753 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2755 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2756 contrib = __compute_runnable_contrib(periods);
2757 contrib = cap_scale(contrib, scale_freq);
2759 sa->load_sum += weight * contrib;
2761 cfs_rq->runnable_load_sum += weight * contrib;
2764 sa->util_sum += contrib * scale_cpu;
2767 /* Remainder of delta accrued against u_0` */
2768 scaled_delta = cap_scale(delta, scale_freq);
2770 sa->load_sum += weight * scaled_delta;
2772 cfs_rq->runnable_load_sum += weight * scaled_delta;
2775 sa->util_sum += scaled_delta * scale_cpu;
2777 sa->period_contrib += delta;
2780 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2782 cfs_rq->runnable_load_avg =
2783 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2785 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2791 #ifdef CONFIG_FAIR_GROUP_SCHED
2793 * update_tg_load_avg - update the tg's load avg
2794 * @cfs_rq: the cfs_rq whose avg changed
2795 * @force: update regardless of how small the difference
2797 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
2798 * However, because tg->load_avg is a global value there are performance
2801 * In order to avoid having to look at the other cfs_rq's, we use a
2802 * differential update where we store the last value we propagated. This in
2803 * turn allows skipping updates if the differential is 'small'.
2805 * Updating tg's load_avg is necessary before update_cfs_share() (which is
2806 * done) and effective_load() (which is not done because it is too costly).
2808 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2810 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2812 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2813 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2814 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2818 #else /* CONFIG_FAIR_GROUP_SCHED */
2819 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2820 #endif /* CONFIG_FAIR_GROUP_SCHED */
2822 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
2824 if (&this_rq()->cfs == cfs_rq) {
2826 * There are a few boundary cases this might miss but it should
2827 * get called often enough that that should (hopefully) not be
2828 * a real problem -- added to that it only calls on the local
2829 * CPU, so if we enqueue remotely we'll miss an update, but
2830 * the next tick/schedule should update.
2832 * It will not get called when we go idle, because the idle
2833 * thread is a different class (!fair), nor will the utilization
2834 * number include things like RT tasks.
2836 * As is, the util number is not freq-invariant (we'd have to
2837 * implement arch_scale_freq_capacity() for that).
2841 cpufreq_update_util(rq_of(cfs_rq), 0);
2845 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2848 * Unsigned subtract and clamp on underflow.
2850 * Explicitly do a load-store to ensure the intermediate value never hits
2851 * memory. This allows lockless observations without ever seeing the negative
2854 #define sub_positive(_ptr, _val) do { \
2855 typeof(_ptr) ptr = (_ptr); \
2856 typeof(*ptr) val = (_val); \
2857 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2861 WRITE_ONCE(*ptr, res); \
2865 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
2866 * @now: current time, as per cfs_rq_clock_task()
2867 * @cfs_rq: cfs_rq to update
2868 * @update_freq: should we call cfs_rq_util_change() or will the call do so
2870 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
2871 * avg. The immediate corollary is that all (fair) tasks must be attached, see
2872 * post_init_entity_util_avg().
2874 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
2876 * Returns true if the load decayed or we removed load.
2878 * Since both these conditions indicate a changed cfs_rq->avg.load we should
2879 * call update_tg_load_avg() when this function returns true.
2882 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
2884 struct sched_avg *sa = &cfs_rq->avg;
2885 int decayed, removed = 0, removed_util = 0;
2887 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2888 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2889 sub_positive(&sa->load_avg, r);
2890 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
2894 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2895 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2896 sub_positive(&sa->util_avg, r);
2897 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
2901 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2902 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2904 #ifndef CONFIG_64BIT
2906 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2909 /* Trace CPU load, unless cfs_rq belongs to a non-root task_group */
2910 if (cfs_rq == &rq_of(cfs_rq)->cfs)
2911 trace_sched_load_avg_cpu(cpu_of(rq_of(cfs_rq)), cfs_rq);
2913 if (update_freq && (decayed || removed_util))
2914 cfs_rq_util_change(cfs_rq);
2916 return decayed || removed;
2920 * Optional action to be done while updating the load average
2922 #define UPDATE_TG 0x1
2923 #define SKIP_AGE_LOAD 0x2
2925 /* Update task and its cfs_rq load average */
2926 static inline void update_load_avg(struct sched_entity *se, int flags)
2928 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2929 u64 now = cfs_rq_clock_task(cfs_rq);
2930 int cpu = cpu_of(rq_of(cfs_rq));
2933 * Track task load average for carrying it to new CPU after migrated, and
2934 * track group sched_entity load average for task_h_load calc in migration
2936 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD)) {
2937 __update_load_avg(now, cpu, &se->avg,
2938 se->on_rq * scale_load_down(se->load.weight),
2939 cfs_rq->curr == se, NULL);
2942 if (update_cfs_rq_load_avg(now, cfs_rq, true) && (flags & UPDATE_TG))
2943 update_tg_load_avg(cfs_rq, 0);
2945 if (entity_is_task(se))
2946 trace_sched_load_avg_task(task_of(se), &se->avg);
2950 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
2951 * @cfs_rq: cfs_rq to attach to
2952 * @se: sched_entity to attach
2954 * Must call update_cfs_rq_load_avg() before this, since we rely on
2955 * cfs_rq->avg.last_update_time being current.
2957 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2959 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2960 cfs_rq->avg.load_avg += se->avg.load_avg;
2961 cfs_rq->avg.load_sum += se->avg.load_sum;
2962 cfs_rq->avg.util_avg += se->avg.util_avg;
2963 cfs_rq->avg.util_sum += se->avg.util_sum;
2965 cfs_rq_util_change(cfs_rq);
2969 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
2970 * @cfs_rq: cfs_rq to detach from
2971 * @se: sched_entity to detach
2973 * Must call update_cfs_rq_load_avg() before this, since we rely on
2974 * cfs_rq->avg.last_update_time being current.
2976 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2979 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2980 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
2981 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
2982 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
2984 cfs_rq_util_change(cfs_rq);
2987 /* Add the load generated by se into cfs_rq's load average */
2989 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2991 struct sched_avg *sa = &se->avg;
2993 cfs_rq->runnable_load_avg += sa->load_avg;
2994 cfs_rq->runnable_load_sum += sa->load_sum;
2996 if (!sa->last_update_time) {
2997 attach_entity_load_avg(cfs_rq, se);
2998 update_tg_load_avg(cfs_rq, 0);
3002 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3004 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3006 cfs_rq->runnable_load_avg =
3007 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
3008 cfs_rq->runnable_load_sum =
3009 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
3012 #ifndef CONFIG_64BIT
3013 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3015 u64 last_update_time_copy;
3016 u64 last_update_time;
3019 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3021 last_update_time = cfs_rq->avg.last_update_time;
3022 } while (last_update_time != last_update_time_copy);
3024 return last_update_time;
3027 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3029 return cfs_rq->avg.last_update_time;
3034 * Synchronize entity load avg of dequeued entity without locking
3037 void sync_entity_load_avg(struct sched_entity *se)
3039 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3040 u64 last_update_time;
3042 last_update_time = cfs_rq_last_update_time(cfs_rq);
3043 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
3047 * Task first catches up with cfs_rq, and then subtract
3048 * itself from the cfs_rq (task must be off the queue now).
3050 void remove_entity_load_avg(struct sched_entity *se)
3052 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3055 * Newly created task or never used group entity should not be removed
3056 * from its (source) cfs_rq
3058 if (se->avg.last_update_time == 0)
3061 sync_entity_load_avg(se);
3062 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3063 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3067 * Update the rq's load with the elapsed running time before entering
3068 * idle. if the last scheduled task is not a CFS task, idle_enter will
3069 * be the only way to update the runnable statistic.
3071 void idle_enter_fair(struct rq *this_rq)
3076 * Update the rq's load with the elapsed idle time before a task is
3077 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
3078 * be the only way to update the runnable statistic.
3080 void idle_exit_fair(struct rq *this_rq)
3084 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3086 return cfs_rq->runnable_load_avg;
3089 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3091 return cfs_rq->avg.load_avg;
3094 static int idle_balance(struct rq *this_rq);
3096 #else /* CONFIG_SMP */
3099 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
3104 #define UPDATE_TG 0x0
3105 #define SKIP_AGE_LOAD 0x0
3107 static inline void update_load_avg(struct sched_entity *se, int not_used1){}
3109 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3111 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3112 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3115 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3117 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3119 static inline int idle_balance(struct rq *rq)
3124 #endif /* CONFIG_SMP */
3126 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
3128 #ifdef CONFIG_SCHEDSTATS
3129 struct task_struct *tsk = NULL;
3131 if (entity_is_task(se))
3134 if (se->statistics.sleep_start) {
3135 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
3140 if (unlikely(delta > se->statistics.sleep_max))
3141 se->statistics.sleep_max = delta;
3143 se->statistics.sleep_start = 0;
3144 se->statistics.sum_sleep_runtime += delta;
3147 account_scheduler_latency(tsk, delta >> 10, 1);
3148 trace_sched_stat_sleep(tsk, delta);
3151 if (se->statistics.block_start) {
3152 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
3157 if (unlikely(delta > se->statistics.block_max))
3158 se->statistics.block_max = delta;
3160 se->statistics.block_start = 0;
3161 se->statistics.sum_sleep_runtime += delta;
3164 if (tsk->in_iowait) {
3165 se->statistics.iowait_sum += delta;
3166 se->statistics.iowait_count++;
3167 trace_sched_stat_iowait(tsk, delta);
3170 trace_sched_stat_blocked(tsk, delta);
3171 trace_sched_blocked_reason(tsk);
3174 * Blocking time is in units of nanosecs, so shift by
3175 * 20 to get a milliseconds-range estimation of the
3176 * amount of time that the task spent sleeping:
3178 if (unlikely(prof_on == SLEEP_PROFILING)) {
3179 profile_hits(SLEEP_PROFILING,
3180 (void *)get_wchan(tsk),
3183 account_scheduler_latency(tsk, delta >> 10, 0);
3189 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3191 #ifdef CONFIG_SCHED_DEBUG
3192 s64 d = se->vruntime - cfs_rq->min_vruntime;
3197 if (d > 3*sysctl_sched_latency)
3198 schedstat_inc(cfs_rq, nr_spread_over);
3203 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3205 u64 vruntime = cfs_rq->min_vruntime;
3208 * The 'current' period is already promised to the current tasks,
3209 * however the extra weight of the new task will slow them down a
3210 * little, place the new task so that it fits in the slot that
3211 * stays open at the end.
3213 if (initial && sched_feat(START_DEBIT))
3214 vruntime += sched_vslice(cfs_rq, se);
3216 /* sleeps up to a single latency don't count. */
3218 unsigned long thresh = sysctl_sched_latency;
3221 * Halve their sleep time's effect, to allow
3222 * for a gentler effect of sleepers:
3224 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3230 /* ensure we never gain time by being placed backwards. */
3231 se->vruntime = max_vruntime(se->vruntime, vruntime);
3234 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3237 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3240 * Update the normalized vruntime before updating min_vruntime
3241 * through calling update_curr().
3243 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3244 se->vruntime += cfs_rq->min_vruntime;
3247 * Update run-time statistics of the 'current'.
3249 update_curr(cfs_rq);
3250 update_load_avg(se, UPDATE_TG);
3251 enqueue_entity_load_avg(cfs_rq, se);
3252 account_entity_enqueue(cfs_rq, se);
3253 update_cfs_shares(cfs_rq);
3255 if (flags & ENQUEUE_WAKEUP) {
3256 place_entity(cfs_rq, se, 0);
3257 enqueue_sleeper(cfs_rq, se);
3260 update_stats_enqueue(cfs_rq, se);
3261 check_spread(cfs_rq, se);
3262 if (se != cfs_rq->curr)
3263 __enqueue_entity(cfs_rq, se);
3266 if (cfs_rq->nr_running == 1) {
3267 list_add_leaf_cfs_rq(cfs_rq);
3268 check_enqueue_throttle(cfs_rq);
3272 static void __clear_buddies_last(struct sched_entity *se)
3274 for_each_sched_entity(se) {
3275 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3276 if (cfs_rq->last != se)
3279 cfs_rq->last = NULL;
3283 static void __clear_buddies_next(struct sched_entity *se)
3285 for_each_sched_entity(se) {
3286 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3287 if (cfs_rq->next != se)
3290 cfs_rq->next = NULL;
3294 static void __clear_buddies_skip(struct sched_entity *se)
3296 for_each_sched_entity(se) {
3297 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3298 if (cfs_rq->skip != se)
3301 cfs_rq->skip = NULL;
3305 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3307 if (cfs_rq->last == se)
3308 __clear_buddies_last(se);
3310 if (cfs_rq->next == se)
3311 __clear_buddies_next(se);
3313 if (cfs_rq->skip == se)
3314 __clear_buddies_skip(se);
3317 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3320 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3323 * Update run-time statistics of the 'current'.
3325 update_curr(cfs_rq);
3326 update_load_avg(se, UPDATE_TG);
3327 dequeue_entity_load_avg(cfs_rq, se);
3329 update_stats_dequeue(cfs_rq, se);
3330 if (flags & DEQUEUE_SLEEP) {
3331 #ifdef CONFIG_SCHEDSTATS
3332 if (entity_is_task(se)) {
3333 struct task_struct *tsk = task_of(se);
3335 if (tsk->state & TASK_INTERRUPTIBLE)
3336 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3337 if (tsk->state & TASK_UNINTERRUPTIBLE)
3338 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3343 clear_buddies(cfs_rq, se);
3345 if (se != cfs_rq->curr)
3346 __dequeue_entity(cfs_rq, se);
3348 account_entity_dequeue(cfs_rq, se);
3351 * Normalize the entity after updating the min_vruntime because the
3352 * update can refer to the ->curr item and we need to reflect this
3353 * movement in our normalized position.
3355 if (!(flags & DEQUEUE_SLEEP))
3356 se->vruntime -= cfs_rq->min_vruntime;
3358 /* return excess runtime on last dequeue */
3359 return_cfs_rq_runtime(cfs_rq);
3361 update_min_vruntime(cfs_rq);
3362 update_cfs_shares(cfs_rq);
3366 * Preempt the current task with a newly woken task if needed:
3369 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3371 unsigned long ideal_runtime, delta_exec;
3372 struct sched_entity *se;
3375 ideal_runtime = sched_slice(cfs_rq, curr);
3376 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3377 if (delta_exec > ideal_runtime) {
3378 resched_curr(rq_of(cfs_rq));
3380 * The current task ran long enough, ensure it doesn't get
3381 * re-elected due to buddy favours.
3383 clear_buddies(cfs_rq, curr);
3388 * Ensure that a task that missed wakeup preemption by a
3389 * narrow margin doesn't have to wait for a full slice.
3390 * This also mitigates buddy induced latencies under load.
3392 if (delta_exec < sysctl_sched_min_granularity)
3395 se = __pick_first_entity(cfs_rq);
3396 delta = curr->vruntime - se->vruntime;
3401 if (delta > ideal_runtime)
3402 resched_curr(rq_of(cfs_rq));
3406 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3408 /* 'current' is not kept within the tree. */
3411 * Any task has to be enqueued before it get to execute on
3412 * a CPU. So account for the time it spent waiting on the
3415 update_stats_wait_end(cfs_rq, se);
3416 __dequeue_entity(cfs_rq, se);
3417 update_load_avg(se, UPDATE_TG);
3420 update_stats_curr_start(cfs_rq, se);
3422 #ifdef CONFIG_SCHEDSTATS
3424 * Track our maximum slice length, if the CPU's load is at
3425 * least twice that of our own weight (i.e. dont track it
3426 * when there are only lesser-weight tasks around):
3428 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3429 se->statistics.slice_max = max(se->statistics.slice_max,
3430 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3433 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3437 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3440 * Pick the next process, keeping these things in mind, in this order:
3441 * 1) keep things fair between processes/task groups
3442 * 2) pick the "next" process, since someone really wants that to run
3443 * 3) pick the "last" process, for cache locality
3444 * 4) do not run the "skip" process, if something else is available
3446 static struct sched_entity *
3447 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3449 struct sched_entity *left = __pick_first_entity(cfs_rq);
3450 struct sched_entity *se;
3453 * If curr is set we have to see if its left of the leftmost entity
3454 * still in the tree, provided there was anything in the tree at all.
3456 if (!left || (curr && entity_before(curr, left)))
3459 se = left; /* ideally we run the leftmost entity */
3462 * Avoid running the skip buddy, if running something else can
3463 * be done without getting too unfair.
3465 if (cfs_rq->skip == se) {
3466 struct sched_entity *second;
3469 second = __pick_first_entity(cfs_rq);
3471 second = __pick_next_entity(se);
3472 if (!second || (curr && entity_before(curr, second)))
3476 if (second && wakeup_preempt_entity(second, left) < 1)
3481 * Prefer last buddy, try to return the CPU to a preempted task.
3483 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3487 * Someone really wants this to run. If it's not unfair, run it.
3489 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3492 clear_buddies(cfs_rq, se);
3497 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3499 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3502 * If still on the runqueue then deactivate_task()
3503 * was not called and update_curr() has to be done:
3506 update_curr(cfs_rq);
3508 /* throttle cfs_rqs exceeding runtime */
3509 check_cfs_rq_runtime(cfs_rq);
3511 check_spread(cfs_rq, prev);
3513 update_stats_wait_start(cfs_rq, prev);
3514 /* Put 'current' back into the tree. */
3515 __enqueue_entity(cfs_rq, prev);
3516 /* in !on_rq case, update occurred at dequeue */
3517 update_load_avg(prev, 0);
3519 cfs_rq->curr = NULL;
3523 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3526 * Update run-time statistics of the 'current'.
3528 update_curr(cfs_rq);
3531 * Ensure that runnable average is periodically updated.
3533 update_load_avg(curr, UPDATE_TG);
3534 update_cfs_shares(cfs_rq);
3536 #ifdef CONFIG_SCHED_HRTICK
3538 * queued ticks are scheduled to match the slice, so don't bother
3539 * validating it and just reschedule.
3542 resched_curr(rq_of(cfs_rq));
3546 * don't let the period tick interfere with the hrtick preemption
3548 if (!sched_feat(DOUBLE_TICK) &&
3549 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3553 if (cfs_rq->nr_running > 1)
3554 check_preempt_tick(cfs_rq, curr);
3558 /**************************************************
3559 * CFS bandwidth control machinery
3562 #ifdef CONFIG_CFS_BANDWIDTH
3564 #ifdef HAVE_JUMP_LABEL
3565 static struct static_key __cfs_bandwidth_used;
3567 static inline bool cfs_bandwidth_used(void)
3569 return static_key_false(&__cfs_bandwidth_used);
3572 void cfs_bandwidth_usage_inc(void)
3574 static_key_slow_inc(&__cfs_bandwidth_used);
3577 void cfs_bandwidth_usage_dec(void)
3579 static_key_slow_dec(&__cfs_bandwidth_used);
3581 #else /* HAVE_JUMP_LABEL */
3582 static bool cfs_bandwidth_used(void)
3587 void cfs_bandwidth_usage_inc(void) {}
3588 void cfs_bandwidth_usage_dec(void) {}
3589 #endif /* HAVE_JUMP_LABEL */
3592 * default period for cfs group bandwidth.
3593 * default: 0.1s, units: nanoseconds
3595 static inline u64 default_cfs_period(void)
3597 return 100000000ULL;
3600 static inline u64 sched_cfs_bandwidth_slice(void)
3602 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3606 * Replenish runtime according to assigned quota and update expiration time.
3607 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3608 * additional synchronization around rq->lock.
3610 * requires cfs_b->lock
3612 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3616 if (cfs_b->quota == RUNTIME_INF)
3619 now = sched_clock_cpu(smp_processor_id());
3620 cfs_b->runtime = cfs_b->quota;
3621 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3624 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3626 return &tg->cfs_bandwidth;
3629 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3630 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3632 if (unlikely(cfs_rq->throttle_count))
3633 return cfs_rq->throttled_clock_task;
3635 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3638 /* returns 0 on failure to allocate runtime */
3639 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3641 struct task_group *tg = cfs_rq->tg;
3642 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3643 u64 amount = 0, min_amount, expires;
3645 /* note: this is a positive sum as runtime_remaining <= 0 */
3646 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3648 raw_spin_lock(&cfs_b->lock);
3649 if (cfs_b->quota == RUNTIME_INF)
3650 amount = min_amount;
3652 start_cfs_bandwidth(cfs_b);
3654 if (cfs_b->runtime > 0) {
3655 amount = min(cfs_b->runtime, min_amount);
3656 cfs_b->runtime -= amount;
3660 expires = cfs_b->runtime_expires;
3661 raw_spin_unlock(&cfs_b->lock);
3663 cfs_rq->runtime_remaining += amount;
3665 * we may have advanced our local expiration to account for allowed
3666 * spread between our sched_clock and the one on which runtime was
3669 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3670 cfs_rq->runtime_expires = expires;
3672 return cfs_rq->runtime_remaining > 0;
3676 * Note: This depends on the synchronization provided by sched_clock and the
3677 * fact that rq->clock snapshots this value.
3679 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3681 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3683 /* if the deadline is ahead of our clock, nothing to do */
3684 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3687 if (cfs_rq->runtime_remaining < 0)
3691 * If the local deadline has passed we have to consider the
3692 * possibility that our sched_clock is 'fast' and the global deadline
3693 * has not truly expired.
3695 * Fortunately we can check determine whether this the case by checking
3696 * whether the global deadline has advanced. It is valid to compare
3697 * cfs_b->runtime_expires without any locks since we only care about
3698 * exact equality, so a partial write will still work.
3701 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3702 /* extend local deadline, drift is bounded above by 2 ticks */
3703 cfs_rq->runtime_expires += TICK_NSEC;
3705 /* global deadline is ahead, expiration has passed */
3706 cfs_rq->runtime_remaining = 0;
3710 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3712 /* dock delta_exec before expiring quota (as it could span periods) */
3713 cfs_rq->runtime_remaining -= delta_exec;
3714 expire_cfs_rq_runtime(cfs_rq);
3716 if (likely(cfs_rq->runtime_remaining > 0))
3720 * if we're unable to extend our runtime we resched so that the active
3721 * hierarchy can be throttled
3723 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3724 resched_curr(rq_of(cfs_rq));
3727 static __always_inline
3728 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3730 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3733 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3736 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3738 return cfs_bandwidth_used() && cfs_rq->throttled;
3741 /* check whether cfs_rq, or any parent, is throttled */
3742 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3744 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3748 * Ensure that neither of the group entities corresponding to src_cpu or
3749 * dest_cpu are members of a throttled hierarchy when performing group
3750 * load-balance operations.
3752 static inline int throttled_lb_pair(struct task_group *tg,
3753 int src_cpu, int dest_cpu)
3755 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3757 src_cfs_rq = tg->cfs_rq[src_cpu];
3758 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3760 return throttled_hierarchy(src_cfs_rq) ||
3761 throttled_hierarchy(dest_cfs_rq);
3764 /* updated child weight may affect parent so we have to do this bottom up */
3765 static int tg_unthrottle_up(struct task_group *tg, void *data)
3767 struct rq *rq = data;
3768 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3770 cfs_rq->throttle_count--;
3772 if (!cfs_rq->throttle_count) {
3773 /* adjust cfs_rq_clock_task() */
3774 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3775 cfs_rq->throttled_clock_task;
3782 static int tg_throttle_down(struct task_group *tg, void *data)
3784 struct rq *rq = data;
3785 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3787 /* group is entering throttled state, stop time */
3788 if (!cfs_rq->throttle_count)
3789 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3790 cfs_rq->throttle_count++;
3795 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3797 struct rq *rq = rq_of(cfs_rq);
3798 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3799 struct sched_entity *se;
3800 long task_delta, dequeue = 1;
3803 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3805 /* freeze hierarchy runnable averages while throttled */
3807 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3810 task_delta = cfs_rq->h_nr_running;
3811 for_each_sched_entity(se) {
3812 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3813 /* throttled entity or throttle-on-deactivate */
3818 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3819 qcfs_rq->h_nr_running -= task_delta;
3821 if (qcfs_rq->load.weight)
3826 sub_nr_running(rq, task_delta);
3828 cfs_rq->throttled = 1;
3829 cfs_rq->throttled_clock = rq_clock(rq);
3830 raw_spin_lock(&cfs_b->lock);
3831 empty = list_empty(&cfs_b->throttled_cfs_rq);
3834 * Add to the _head_ of the list, so that an already-started
3835 * distribute_cfs_runtime will not see us
3837 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3840 * If we're the first throttled task, make sure the bandwidth
3844 start_cfs_bandwidth(cfs_b);
3846 raw_spin_unlock(&cfs_b->lock);
3849 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3851 struct rq *rq = rq_of(cfs_rq);
3852 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3853 struct sched_entity *se;
3857 se = cfs_rq->tg->se[cpu_of(rq)];
3859 cfs_rq->throttled = 0;
3861 update_rq_clock(rq);
3863 raw_spin_lock(&cfs_b->lock);
3864 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3865 list_del_rcu(&cfs_rq->throttled_list);
3866 raw_spin_unlock(&cfs_b->lock);
3868 /* update hierarchical throttle state */
3869 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3871 if (!cfs_rq->load.weight)
3874 task_delta = cfs_rq->h_nr_running;
3875 for_each_sched_entity(se) {
3879 cfs_rq = cfs_rq_of(se);
3881 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3882 cfs_rq->h_nr_running += task_delta;
3884 if (cfs_rq_throttled(cfs_rq))
3889 add_nr_running(rq, task_delta);
3891 /* determine whether we need to wake up potentially idle cpu */
3892 if (rq->curr == rq->idle && rq->cfs.nr_running)
3896 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3897 u64 remaining, u64 expires)
3899 struct cfs_rq *cfs_rq;
3901 u64 starting_runtime = remaining;
3904 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3906 struct rq *rq = rq_of(cfs_rq);
3908 raw_spin_lock(&rq->lock);
3909 if (!cfs_rq_throttled(cfs_rq))
3912 runtime = -cfs_rq->runtime_remaining + 1;
3913 if (runtime > remaining)
3914 runtime = remaining;
3915 remaining -= runtime;
3917 cfs_rq->runtime_remaining += runtime;
3918 cfs_rq->runtime_expires = expires;
3920 /* we check whether we're throttled above */
3921 if (cfs_rq->runtime_remaining > 0)
3922 unthrottle_cfs_rq(cfs_rq);
3925 raw_spin_unlock(&rq->lock);
3932 return starting_runtime - remaining;
3936 * Responsible for refilling a task_group's bandwidth and unthrottling its
3937 * cfs_rqs as appropriate. If there has been no activity within the last
3938 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3939 * used to track this state.
3941 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3943 u64 runtime, runtime_expires;
3946 /* no need to continue the timer with no bandwidth constraint */
3947 if (cfs_b->quota == RUNTIME_INF)
3948 goto out_deactivate;
3950 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3951 cfs_b->nr_periods += overrun;
3954 * idle depends on !throttled (for the case of a large deficit), and if
3955 * we're going inactive then everything else can be deferred
3957 if (cfs_b->idle && !throttled)
3958 goto out_deactivate;
3960 __refill_cfs_bandwidth_runtime(cfs_b);
3963 /* mark as potentially idle for the upcoming period */
3968 /* account preceding periods in which throttling occurred */
3969 cfs_b->nr_throttled += overrun;
3971 runtime_expires = cfs_b->runtime_expires;
3974 * This check is repeated as we are holding onto the new bandwidth while
3975 * we unthrottle. This can potentially race with an unthrottled group
3976 * trying to acquire new bandwidth from the global pool. This can result
3977 * in us over-using our runtime if it is all used during this loop, but
3978 * only by limited amounts in that extreme case.
3980 while (throttled && cfs_b->runtime > 0) {
3981 runtime = cfs_b->runtime;
3982 raw_spin_unlock(&cfs_b->lock);
3983 /* we can't nest cfs_b->lock while distributing bandwidth */
3984 runtime = distribute_cfs_runtime(cfs_b, runtime,
3986 raw_spin_lock(&cfs_b->lock);
3988 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3990 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3994 * While we are ensured activity in the period following an
3995 * unthrottle, this also covers the case in which the new bandwidth is
3996 * insufficient to cover the existing bandwidth deficit. (Forcing the
3997 * timer to remain active while there are any throttled entities.)
4007 /* a cfs_rq won't donate quota below this amount */
4008 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4009 /* minimum remaining period time to redistribute slack quota */
4010 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4011 /* how long we wait to gather additional slack before distributing */
4012 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4015 * Are we near the end of the current quota period?
4017 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4018 * hrtimer base being cleared by hrtimer_start. In the case of
4019 * migrate_hrtimers, base is never cleared, so we are fine.
4021 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4023 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4026 /* if the call-back is running a quota refresh is already occurring */
4027 if (hrtimer_callback_running(refresh_timer))
4030 /* is a quota refresh about to occur? */
4031 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4032 if (remaining < min_expire)
4038 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4040 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4042 /* if there's a quota refresh soon don't bother with slack */
4043 if (runtime_refresh_within(cfs_b, min_left))
4046 hrtimer_start(&cfs_b->slack_timer,
4047 ns_to_ktime(cfs_bandwidth_slack_period),
4051 /* we know any runtime found here is valid as update_curr() precedes return */
4052 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4054 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4055 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4057 if (slack_runtime <= 0)
4060 raw_spin_lock(&cfs_b->lock);
4061 if (cfs_b->quota != RUNTIME_INF &&
4062 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4063 cfs_b->runtime += slack_runtime;
4065 /* we are under rq->lock, defer unthrottling using a timer */
4066 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4067 !list_empty(&cfs_b->throttled_cfs_rq))
4068 start_cfs_slack_bandwidth(cfs_b);
4070 raw_spin_unlock(&cfs_b->lock);
4072 /* even if it's not valid for return we don't want to try again */
4073 cfs_rq->runtime_remaining -= slack_runtime;
4076 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4078 if (!cfs_bandwidth_used())
4081 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4084 __return_cfs_rq_runtime(cfs_rq);
4088 * This is done with a timer (instead of inline with bandwidth return) since
4089 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4091 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4093 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4096 /* confirm we're still not at a refresh boundary */
4097 raw_spin_lock(&cfs_b->lock);
4098 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4099 raw_spin_unlock(&cfs_b->lock);
4103 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4104 runtime = cfs_b->runtime;
4106 expires = cfs_b->runtime_expires;
4107 raw_spin_unlock(&cfs_b->lock);
4112 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4114 raw_spin_lock(&cfs_b->lock);
4115 if (expires == cfs_b->runtime_expires)
4116 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4117 raw_spin_unlock(&cfs_b->lock);
4121 * When a group wakes up we want to make sure that its quota is not already
4122 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4123 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4125 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4127 if (!cfs_bandwidth_used())
4130 /* Synchronize hierarchical throttle counter: */
4131 if (unlikely(!cfs_rq->throttle_uptodate)) {
4132 struct rq *rq = rq_of(cfs_rq);
4133 struct cfs_rq *pcfs_rq;
4134 struct task_group *tg;
4136 cfs_rq->throttle_uptodate = 1;
4138 /* Get closest up-to-date node, because leaves go first: */
4139 for (tg = cfs_rq->tg->parent; tg; tg = tg->parent) {
4140 pcfs_rq = tg->cfs_rq[cpu_of(rq)];
4141 if (pcfs_rq->throttle_uptodate)
4145 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4146 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4150 /* an active group must be handled by the update_curr()->put() path */
4151 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4154 /* ensure the group is not already throttled */
4155 if (cfs_rq_throttled(cfs_rq))
4158 /* update runtime allocation */
4159 account_cfs_rq_runtime(cfs_rq, 0);
4160 if (cfs_rq->runtime_remaining <= 0)
4161 throttle_cfs_rq(cfs_rq);
4164 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4165 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4167 if (!cfs_bandwidth_used())
4170 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4174 * it's possible for a throttled entity to be forced into a running
4175 * state (e.g. set_curr_task), in this case we're finished.
4177 if (cfs_rq_throttled(cfs_rq))
4180 throttle_cfs_rq(cfs_rq);
4184 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4186 struct cfs_bandwidth *cfs_b =
4187 container_of(timer, struct cfs_bandwidth, slack_timer);
4189 do_sched_cfs_slack_timer(cfs_b);
4191 return HRTIMER_NORESTART;
4194 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4196 struct cfs_bandwidth *cfs_b =
4197 container_of(timer, struct cfs_bandwidth, period_timer);
4201 raw_spin_lock(&cfs_b->lock);
4203 overrun = hrtimer_forward_now(timer, cfs_b->period);
4207 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4210 cfs_b->period_active = 0;
4211 raw_spin_unlock(&cfs_b->lock);
4213 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4216 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4218 raw_spin_lock_init(&cfs_b->lock);
4220 cfs_b->quota = RUNTIME_INF;
4221 cfs_b->period = ns_to_ktime(default_cfs_period());
4223 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4224 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4225 cfs_b->period_timer.function = sched_cfs_period_timer;
4226 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4227 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4230 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4232 cfs_rq->runtime_enabled = 0;
4233 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4236 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4238 lockdep_assert_held(&cfs_b->lock);
4240 if (!cfs_b->period_active) {
4241 cfs_b->period_active = 1;
4242 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4243 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4247 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4249 /* init_cfs_bandwidth() was not called */
4250 if (!cfs_b->throttled_cfs_rq.next)
4253 hrtimer_cancel(&cfs_b->period_timer);
4254 hrtimer_cancel(&cfs_b->slack_timer);
4257 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4259 struct cfs_rq *cfs_rq;
4261 for_each_leaf_cfs_rq(rq, cfs_rq) {
4262 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4264 raw_spin_lock(&cfs_b->lock);
4265 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4266 raw_spin_unlock(&cfs_b->lock);
4270 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4272 struct cfs_rq *cfs_rq;
4274 for_each_leaf_cfs_rq(rq, cfs_rq) {
4275 if (!cfs_rq->runtime_enabled)
4279 * clock_task is not advancing so we just need to make sure
4280 * there's some valid quota amount
4282 cfs_rq->runtime_remaining = 1;
4284 * Offline rq is schedulable till cpu is completely disabled
4285 * in take_cpu_down(), so we prevent new cfs throttling here.
4287 cfs_rq->runtime_enabled = 0;
4289 if (cfs_rq_throttled(cfs_rq))
4290 unthrottle_cfs_rq(cfs_rq);
4294 #else /* CONFIG_CFS_BANDWIDTH */
4295 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4297 return rq_clock_task(rq_of(cfs_rq));
4300 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4301 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4302 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4303 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4305 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4310 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4315 static inline int throttled_lb_pair(struct task_group *tg,
4316 int src_cpu, int dest_cpu)
4321 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4323 #ifdef CONFIG_FAIR_GROUP_SCHED
4324 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4327 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4331 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4332 static inline void update_runtime_enabled(struct rq *rq) {}
4333 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4335 #endif /* CONFIG_CFS_BANDWIDTH */
4337 /**************************************************
4338 * CFS operations on tasks:
4341 #ifdef CONFIG_SCHED_HRTICK
4342 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4344 struct sched_entity *se = &p->se;
4345 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4347 WARN_ON(task_rq(p) != rq);
4349 if (cfs_rq->nr_running > 1) {
4350 u64 slice = sched_slice(cfs_rq, se);
4351 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4352 s64 delta = slice - ran;
4359 hrtick_start(rq, delta);
4364 * called from enqueue/dequeue and updates the hrtick when the
4365 * current task is from our class and nr_running is low enough
4368 static void hrtick_update(struct rq *rq)
4370 struct task_struct *curr = rq->curr;
4372 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4375 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4376 hrtick_start_fair(rq, curr);
4378 #else /* !CONFIG_SCHED_HRTICK */
4380 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4384 static inline void hrtick_update(struct rq *rq)
4390 static bool cpu_overutilized(int cpu);
4391 unsigned long boosted_cpu_util(int cpu);
4393 #define boosted_cpu_util(cpu) cpu_util(cpu)
4397 static void update_capacity_of(int cpu)
4399 unsigned long req_cap;
4404 /* Convert scale-invariant capacity to cpu. */
4405 req_cap = boosted_cpu_util(cpu);
4406 req_cap = req_cap * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
4407 set_cfs_cpu_capacity(cpu, true, req_cap);
4412 * The enqueue_task method is called before nr_running is
4413 * increased. Here we update the fair scheduling stats and
4414 * then put the task into the rbtree:
4417 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4419 struct cfs_rq *cfs_rq;
4420 struct sched_entity *se = &p->se;
4422 int task_new = flags & ENQUEUE_WAKEUP_NEW;
4423 int task_wakeup = flags & ENQUEUE_WAKEUP;
4427 * If in_iowait is set, the code below may not trigger any cpufreq
4428 * utilization updates, so do it here explicitly with the IOWAIT flag
4432 cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_IOWAIT);
4434 for_each_sched_entity(se) {
4437 cfs_rq = cfs_rq_of(se);
4438 enqueue_entity(cfs_rq, se, flags);
4441 * end evaluation on encountering a throttled cfs_rq
4443 * note: in the case of encountering a throttled cfs_rq we will
4444 * post the final h_nr_running increment below.
4446 if (cfs_rq_throttled(cfs_rq))
4448 cfs_rq->h_nr_running++;
4449 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4451 flags = ENQUEUE_WAKEUP;
4454 for_each_sched_entity(se) {
4455 cfs_rq = cfs_rq_of(se);
4456 cfs_rq->h_nr_running++;
4457 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4459 if (cfs_rq_throttled(cfs_rq))
4462 update_load_avg(se, UPDATE_TG);
4463 update_cfs_shares(cfs_rq);
4467 add_nr_running(rq, 1);
4472 * Update SchedTune accounting.
4474 * We do it before updating the CPU capacity to ensure the
4475 * boost value of the current task is accounted for in the
4476 * selection of the OPP.
4478 * We do it also in the case where we enqueue a throttled task;
4479 * we could argue that a throttled task should not boost a CPU,
4481 * a) properly implementing CPU boosting considering throttled
4482 * tasks will increase a lot the complexity of the solution
4483 * b) it's not easy to quantify the benefits introduced by
4484 * such a more complex solution.
4485 * Thus, for the time being we go for the simple solution and boost
4486 * also for throttled RQs.
4488 schedtune_enqueue_task(p, cpu_of(rq));
4491 walt_inc_cumulative_runnable_avg(rq, p);
4492 if (!task_new && !rq->rd->overutilized &&
4493 cpu_overutilized(rq->cpu)) {
4494 rq->rd->overutilized = true;
4495 trace_sched_overutilized(true);
4499 * We want to potentially trigger a freq switch
4500 * request only for tasks that are waking up; this is
4501 * because we get here also during load balancing, but
4502 * in these cases it seems wise to trigger as single
4503 * request after load balancing is done.
4505 if (task_new || task_wakeup)
4506 update_capacity_of(cpu_of(rq));
4509 #endif /* CONFIG_SMP */
4513 static void set_next_buddy(struct sched_entity *se);
4516 * The dequeue_task method is called before nr_running is
4517 * decreased. We remove the task from the rbtree and
4518 * update the fair scheduling stats:
4520 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4522 struct cfs_rq *cfs_rq;
4523 struct sched_entity *se = &p->se;
4524 int task_sleep = flags & DEQUEUE_SLEEP;
4526 for_each_sched_entity(se) {
4527 cfs_rq = cfs_rq_of(se);
4528 dequeue_entity(cfs_rq, se, flags);
4531 * end evaluation on encountering a throttled cfs_rq
4533 * note: in the case of encountering a throttled cfs_rq we will
4534 * post the final h_nr_running decrement below.
4536 if (cfs_rq_throttled(cfs_rq))
4538 cfs_rq->h_nr_running--;
4539 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4541 /* Don't dequeue parent if it has other entities besides us */
4542 if (cfs_rq->load.weight) {
4543 /* Avoid re-evaluating load for this entity: */
4544 se = parent_entity(se);
4546 * Bias pick_next to pick a task from this cfs_rq, as
4547 * p is sleeping when it is within its sched_slice.
4549 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4553 flags |= DEQUEUE_SLEEP;
4556 for_each_sched_entity(se) {
4557 cfs_rq = cfs_rq_of(se);
4558 cfs_rq->h_nr_running--;
4559 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4561 if (cfs_rq_throttled(cfs_rq))
4564 update_load_avg(se, UPDATE_TG);
4565 update_cfs_shares(cfs_rq);
4569 sub_nr_running(rq, 1);
4574 * Update SchedTune accounting
4576 * We do it before updating the CPU capacity to ensure the
4577 * boost value of the current task is accounted for in the
4578 * selection of the OPP.
4580 schedtune_dequeue_task(p, cpu_of(rq));
4583 walt_dec_cumulative_runnable_avg(rq, p);
4586 * We want to potentially trigger a freq switch
4587 * request only for tasks that are going to sleep;
4588 * this is because we get here also during load
4589 * balancing, but in these cases it seems wise to
4590 * trigger as single request after load balancing is
4594 if (rq->cfs.nr_running)
4595 update_capacity_of(cpu_of(rq));
4596 else if (sched_freq())
4597 set_cfs_cpu_capacity(cpu_of(rq), false, 0);
4601 #endif /* CONFIG_SMP */
4609 * per rq 'load' arrray crap; XXX kill this.
4613 * The exact cpuload at various idx values, calculated at every tick would be
4614 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4616 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4617 * on nth tick when cpu may be busy, then we have:
4618 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4619 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4621 * decay_load_missed() below does efficient calculation of
4622 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4623 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4625 * The calculation is approximated on a 128 point scale.
4626 * degrade_zero_ticks is the number of ticks after which load at any
4627 * particular idx is approximated to be zero.
4628 * degrade_factor is a precomputed table, a row for each load idx.
4629 * Each column corresponds to degradation factor for a power of two ticks,
4630 * based on 128 point scale.
4632 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4633 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4635 * With this power of 2 load factors, we can degrade the load n times
4636 * by looking at 1 bits in n and doing as many mult/shift instead of
4637 * n mult/shifts needed by the exact degradation.
4639 #define DEGRADE_SHIFT 7
4640 static const unsigned char
4641 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4642 static const unsigned char
4643 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4644 {0, 0, 0, 0, 0, 0, 0, 0},
4645 {64, 32, 8, 0, 0, 0, 0, 0},
4646 {96, 72, 40, 12, 1, 0, 0},
4647 {112, 98, 75, 43, 15, 1, 0},
4648 {120, 112, 98, 76, 45, 16, 2} };
4651 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4652 * would be when CPU is idle and so we just decay the old load without
4653 * adding any new load.
4655 static unsigned long
4656 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4660 if (!missed_updates)
4663 if (missed_updates >= degrade_zero_ticks[idx])
4667 return load >> missed_updates;
4669 while (missed_updates) {
4670 if (missed_updates % 2)
4671 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4673 missed_updates >>= 1;
4680 * Update rq->cpu_load[] statistics. This function is usually called every
4681 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4682 * every tick. We fix it up based on jiffies.
4684 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4685 unsigned long pending_updates)
4689 this_rq->nr_load_updates++;
4691 /* Update our load: */
4692 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4693 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4694 unsigned long old_load, new_load;
4696 /* scale is effectively 1 << i now, and >> i divides by scale */
4698 old_load = this_rq->cpu_load[i];
4699 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4700 new_load = this_load;
4702 * Round up the averaging division if load is increasing. This
4703 * prevents us from getting stuck on 9 if the load is 10, for
4706 if (new_load > old_load)
4707 new_load += scale - 1;
4709 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4712 sched_avg_update(this_rq);
4715 /* Used instead of source_load when we know the type == 0 */
4716 static unsigned long weighted_cpuload(const int cpu)
4718 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4721 #ifdef CONFIG_NO_HZ_COMMON
4723 * There is no sane way to deal with nohz on smp when using jiffies because the
4724 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4725 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4727 * Therefore we cannot use the delta approach from the regular tick since that
4728 * would seriously skew the load calculation. However we'll make do for those
4729 * updates happening while idle (nohz_idle_balance) or coming out of idle
4730 * (tick_nohz_idle_exit).
4732 * This means we might still be one tick off for nohz periods.
4736 * Called from nohz_idle_balance() to update the load ratings before doing the
4739 static void update_idle_cpu_load(struct rq *this_rq)
4741 unsigned long curr_jiffies = READ_ONCE(jiffies);
4742 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4743 unsigned long pending_updates;
4746 * bail if there's load or we're actually up-to-date.
4748 if (load || curr_jiffies == this_rq->last_load_update_tick)
4751 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4752 this_rq->last_load_update_tick = curr_jiffies;
4754 __update_cpu_load(this_rq, load, pending_updates);
4758 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4760 void update_cpu_load_nohz(void)
4762 struct rq *this_rq = this_rq();
4763 unsigned long curr_jiffies = READ_ONCE(jiffies);
4764 unsigned long pending_updates;
4766 if (curr_jiffies == this_rq->last_load_update_tick)
4769 raw_spin_lock(&this_rq->lock);
4770 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4771 if (pending_updates) {
4772 this_rq->last_load_update_tick = curr_jiffies;
4774 * We were idle, this means load 0, the current load might be
4775 * !0 due to remote wakeups and the sort.
4777 __update_cpu_load(this_rq, 0, pending_updates);
4779 raw_spin_unlock(&this_rq->lock);
4781 #endif /* CONFIG_NO_HZ */
4784 * Called from scheduler_tick()
4786 void update_cpu_load_active(struct rq *this_rq)
4788 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4790 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4792 this_rq->last_load_update_tick = jiffies;
4793 __update_cpu_load(this_rq, load, 1);
4797 * Return a low guess at the load of a migration-source cpu weighted
4798 * according to the scheduling class and "nice" value.
4800 * We want to under-estimate the load of migration sources, to
4801 * balance conservatively.
4803 static unsigned long source_load(int cpu, int type)
4805 struct rq *rq = cpu_rq(cpu);
4806 unsigned long total = weighted_cpuload(cpu);
4808 if (type == 0 || !sched_feat(LB_BIAS))
4811 return min(rq->cpu_load[type-1], total);
4815 * Return a high guess at the load of a migration-target cpu weighted
4816 * according to the scheduling class and "nice" value.
4818 static unsigned long target_load(int cpu, int type)
4820 struct rq *rq = cpu_rq(cpu);
4821 unsigned long total = weighted_cpuload(cpu);
4823 if (type == 0 || !sched_feat(LB_BIAS))
4826 return max(rq->cpu_load[type-1], total);
4830 static unsigned long cpu_avg_load_per_task(int cpu)
4832 struct rq *rq = cpu_rq(cpu);
4833 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4834 unsigned long load_avg = weighted_cpuload(cpu);
4837 return load_avg / nr_running;
4842 static void record_wakee(struct task_struct *p)
4845 * Rough decay (wiping) for cost saving, don't worry
4846 * about the boundary, really active task won't care
4849 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4850 current->wakee_flips >>= 1;
4851 current->wakee_flip_decay_ts = jiffies;
4854 if (current->last_wakee != p) {
4855 current->last_wakee = p;
4856 current->wakee_flips++;
4860 static void task_waking_fair(struct task_struct *p)
4862 struct sched_entity *se = &p->se;
4863 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4866 #ifndef CONFIG_64BIT
4867 u64 min_vruntime_copy;
4870 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4872 min_vruntime = cfs_rq->min_vruntime;
4873 } while (min_vruntime != min_vruntime_copy);
4875 min_vruntime = cfs_rq->min_vruntime;
4878 se->vruntime -= min_vruntime;
4882 #ifdef CONFIG_FAIR_GROUP_SCHED
4884 * effective_load() calculates the load change as seen from the root_task_group
4886 * Adding load to a group doesn't make a group heavier, but can cause movement
4887 * of group shares between cpus. Assuming the shares were perfectly aligned one
4888 * can calculate the shift in shares.
4890 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4891 * on this @cpu and results in a total addition (subtraction) of @wg to the
4892 * total group weight.
4894 * Given a runqueue weight distribution (rw_i) we can compute a shares
4895 * distribution (s_i) using:
4897 * s_i = rw_i / \Sum rw_j (1)
4899 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4900 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4901 * shares distribution (s_i):
4903 * rw_i = { 2, 4, 1, 0 }
4904 * s_i = { 2/7, 4/7, 1/7, 0 }
4906 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4907 * task used to run on and the CPU the waker is running on), we need to
4908 * compute the effect of waking a task on either CPU and, in case of a sync
4909 * wakeup, compute the effect of the current task going to sleep.
4911 * So for a change of @wl to the local @cpu with an overall group weight change
4912 * of @wl we can compute the new shares distribution (s'_i) using:
4914 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4916 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4917 * differences in waking a task to CPU 0. The additional task changes the
4918 * weight and shares distributions like:
4920 * rw'_i = { 3, 4, 1, 0 }
4921 * s'_i = { 3/8, 4/8, 1/8, 0 }
4923 * We can then compute the difference in effective weight by using:
4925 * dw_i = S * (s'_i - s_i) (3)
4927 * Where 'S' is the group weight as seen by its parent.
4929 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4930 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4931 * 4/7) times the weight of the group.
4933 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4935 struct sched_entity *se = tg->se[cpu];
4937 if (!tg->parent) /* the trivial, non-cgroup case */
4940 for_each_sched_entity(se) {
4941 struct cfs_rq *cfs_rq = se->my_q;
4942 long W, w = cfs_rq_load_avg(cfs_rq);
4947 * W = @wg + \Sum rw_j
4949 W = wg + atomic_long_read(&tg->load_avg);
4951 /* Ensure \Sum rw_j >= rw_i */
4952 W -= cfs_rq->tg_load_avg_contrib;
4961 * wl = S * s'_i; see (2)
4964 wl = (w * (long)tg->shares) / W;
4969 * Per the above, wl is the new se->load.weight value; since
4970 * those are clipped to [MIN_SHARES, ...) do so now. See
4971 * calc_cfs_shares().
4973 if (wl < MIN_SHARES)
4977 * wl = dw_i = S * (s'_i - s_i); see (3)
4979 wl -= se->avg.load_avg;
4982 * Recursively apply this logic to all parent groups to compute
4983 * the final effective load change on the root group. Since
4984 * only the @tg group gets extra weight, all parent groups can
4985 * only redistribute existing shares. @wl is the shift in shares
4986 * resulting from this level per the above.
4995 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
5003 * Returns the current capacity of cpu after applying both
5004 * cpu and freq scaling.
5006 unsigned long capacity_curr_of(int cpu)
5008 return cpu_rq(cpu)->cpu_capacity_orig *
5009 arch_scale_freq_capacity(NULL, cpu)
5010 >> SCHED_CAPACITY_SHIFT;
5013 static inline bool energy_aware(void)
5015 return sched_feat(ENERGY_AWARE);
5019 struct sched_group *sg_top;
5020 struct sched_group *sg_cap;
5027 struct task_struct *task;
5042 * __cpu_norm_util() returns the cpu util relative to a specific capacity,
5043 * i.e. it's busy ratio, in the range [0..SCHED_LOAD_SCALE] which is useful for
5044 * energy calculations. Using the scale-invariant util returned by
5045 * cpu_util() and approximating scale-invariant util by:
5047 * util ~ (curr_freq/max_freq)*1024 * capacity_orig/1024 * running_time/time
5049 * the normalized util can be found using the specific capacity.
5051 * capacity = capacity_orig * curr_freq/max_freq
5053 * norm_util = running_time/time ~ util/capacity
5055 static unsigned long __cpu_norm_util(int cpu, unsigned long capacity, int delta)
5057 int util = __cpu_util(cpu, delta);
5059 if (util >= capacity)
5060 return SCHED_CAPACITY_SCALE;
5062 return (util << SCHED_CAPACITY_SHIFT)/capacity;
5065 static int calc_util_delta(struct energy_env *eenv, int cpu)
5067 if (cpu == eenv->src_cpu)
5068 return -eenv->util_delta;
5069 if (cpu == eenv->dst_cpu)
5070 return eenv->util_delta;
5075 unsigned long group_max_util(struct energy_env *eenv)
5078 unsigned long max_util = 0;
5080 for_each_cpu(i, sched_group_cpus(eenv->sg_cap)) {
5081 delta = calc_util_delta(eenv, i);
5082 max_util = max(max_util, __cpu_util(i, delta));
5089 * group_norm_util() returns the approximated group util relative to it's
5090 * current capacity (busy ratio) in the range [0..SCHED_LOAD_SCALE] for use in
5091 * energy calculations. Since task executions may or may not overlap in time in
5092 * the group the true normalized util is between max(cpu_norm_util(i)) and
5093 * sum(cpu_norm_util(i)) when iterating over all cpus in the group, i. The
5094 * latter is used as the estimate as it leads to a more pessimistic energy
5095 * estimate (more busy).
5098 long group_norm_util(struct energy_env *eenv, struct sched_group *sg)
5101 unsigned long util_sum = 0;
5102 unsigned long capacity = sg->sge->cap_states[eenv->cap_idx].cap;
5104 for_each_cpu(i, sched_group_cpus(sg)) {
5105 delta = calc_util_delta(eenv, i);
5106 util_sum += __cpu_norm_util(i, capacity, delta);
5109 if (util_sum > SCHED_CAPACITY_SCALE)
5110 return SCHED_CAPACITY_SCALE;
5114 static int find_new_capacity(struct energy_env *eenv,
5115 const struct sched_group_energy * const sge)
5118 unsigned long util = group_max_util(eenv);
5120 for (idx = 0; idx < sge->nr_cap_states; idx++) {
5121 if (sge->cap_states[idx].cap >= util)
5125 eenv->cap_idx = idx;
5130 static int group_idle_state(struct sched_group *sg)
5132 int i, state = INT_MAX;
5134 /* Find the shallowest idle state in the sched group. */
5135 for_each_cpu(i, sched_group_cpus(sg))
5136 state = min(state, idle_get_state_idx(cpu_rq(i)));
5138 /* Take non-cpuidle idling into account (active idle/arch_cpu_idle()) */
5145 * sched_group_energy(): Computes the absolute energy consumption of cpus
5146 * belonging to the sched_group including shared resources shared only by
5147 * members of the group. Iterates over all cpus in the hierarchy below the
5148 * sched_group starting from the bottom working it's way up before going to
5149 * the next cpu until all cpus are covered at all levels. The current
5150 * implementation is likely to gather the same util statistics multiple times.
5151 * This can probably be done in a faster but more complex way.
5152 * Note: sched_group_energy() may fail when racing with sched_domain updates.
5154 static int sched_group_energy(struct energy_env *eenv)
5156 struct sched_domain *sd;
5157 int cpu, total_energy = 0;
5158 struct cpumask visit_cpus;
5159 struct sched_group *sg;
5161 WARN_ON(!eenv->sg_top->sge);
5163 cpumask_copy(&visit_cpus, sched_group_cpus(eenv->sg_top));
5165 while (!cpumask_empty(&visit_cpus)) {
5166 struct sched_group *sg_shared_cap = NULL;
5168 cpu = cpumask_first(&visit_cpus);
5171 * Is the group utilization affected by cpus outside this
5174 sd = rcu_dereference(per_cpu(sd_scs, cpu));
5178 * We most probably raced with hotplug; returning a
5179 * wrong energy estimation is better than entering an
5185 sg_shared_cap = sd->parent->groups;
5187 for_each_domain(cpu, sd) {
5190 /* Has this sched_domain already been visited? */
5191 if (sd->child && group_first_cpu(sg) != cpu)
5195 unsigned long group_util;
5196 int sg_busy_energy, sg_idle_energy;
5197 int cap_idx, idle_idx;
5199 if (sg_shared_cap && sg_shared_cap->group_weight >= sg->group_weight)
5200 eenv->sg_cap = sg_shared_cap;
5204 cap_idx = find_new_capacity(eenv, sg->sge);
5206 if (sg->group_weight == 1) {
5207 /* Remove capacity of src CPU (before task move) */
5208 if (eenv->util_delta == 0 &&
5209 cpumask_test_cpu(eenv->src_cpu, sched_group_cpus(sg))) {
5210 eenv->cap.before = sg->sge->cap_states[cap_idx].cap;
5211 eenv->cap.delta -= eenv->cap.before;
5213 /* Add capacity of dst CPU (after task move) */
5214 if (eenv->util_delta != 0 &&
5215 cpumask_test_cpu(eenv->dst_cpu, sched_group_cpus(sg))) {
5216 eenv->cap.after = sg->sge->cap_states[cap_idx].cap;
5217 eenv->cap.delta += eenv->cap.after;
5221 idle_idx = group_idle_state(sg);
5222 group_util = group_norm_util(eenv, sg);
5223 sg_busy_energy = (group_util * sg->sge->cap_states[cap_idx].power)
5224 >> SCHED_CAPACITY_SHIFT;
5225 sg_idle_energy = ((SCHED_LOAD_SCALE-group_util)
5226 * sg->sge->idle_states[idle_idx].power)
5227 >> SCHED_CAPACITY_SHIFT;
5229 total_energy += sg_busy_energy + sg_idle_energy;
5232 cpumask_xor(&visit_cpus, &visit_cpus, sched_group_cpus(sg));
5234 if (cpumask_equal(sched_group_cpus(sg), sched_group_cpus(eenv->sg_top)))
5237 } while (sg = sg->next, sg != sd->groups);
5240 cpumask_clear_cpu(cpu, &visit_cpus);
5244 eenv->energy = total_energy;
5248 static inline bool cpu_in_sg(struct sched_group *sg, int cpu)
5250 return cpu != -1 && cpumask_test_cpu(cpu, sched_group_cpus(sg));
5254 * energy_diff(): Estimate the energy impact of changing the utilization
5255 * distribution. eenv specifies the change: utilisation amount, source, and
5256 * destination cpu. Source or destination cpu may be -1 in which case the
5257 * utilization is removed from or added to the system (e.g. task wake-up). If
5258 * both are specified, the utilization is migrated.
5260 static inline int __energy_diff(struct energy_env *eenv)
5262 struct sched_domain *sd;
5263 struct sched_group *sg;
5264 int sd_cpu = -1, energy_before = 0, energy_after = 0;
5267 struct energy_env eenv_before = {
5269 .src_cpu = eenv->src_cpu,
5270 .dst_cpu = eenv->dst_cpu,
5271 .nrg = { 0, 0, 0, 0},
5275 if (eenv->src_cpu == eenv->dst_cpu)
5278 sd_cpu = (eenv->src_cpu != -1) ? eenv->src_cpu : eenv->dst_cpu;
5279 sd = rcu_dereference(per_cpu(sd_ea, sd_cpu));
5282 return 0; /* Error */
5287 if (cpu_in_sg(sg, eenv->src_cpu) || cpu_in_sg(sg, eenv->dst_cpu)) {
5288 eenv_before.sg_top = eenv->sg_top = sg;
5290 if (sched_group_energy(&eenv_before))
5291 return 0; /* Invalid result abort */
5292 energy_before += eenv_before.energy;
5294 /* Keep track of SRC cpu (before) capacity */
5295 eenv->cap.before = eenv_before.cap.before;
5296 eenv->cap.delta = eenv_before.cap.delta;
5298 if (sched_group_energy(eenv))
5299 return 0; /* Invalid result abort */
5300 energy_after += eenv->energy;
5302 } while (sg = sg->next, sg != sd->groups);
5304 eenv->nrg.before = energy_before;
5305 eenv->nrg.after = energy_after;
5306 eenv->nrg.diff = eenv->nrg.after - eenv->nrg.before;
5309 trace_sched_energy_diff(eenv->task,
5310 eenv->src_cpu, eenv->dst_cpu, eenv->util_delta,
5311 eenv->nrg.before, eenv->nrg.after, eenv->nrg.diff,
5312 eenv->cap.before, eenv->cap.after, eenv->cap.delta,
5313 eenv->nrg.delta, eenv->payoff);
5316 * Dead-zone margin preventing too many migrations.
5319 margin = eenv->nrg.before >> 6; /* ~1.56% */
5321 diff = eenv->nrg.after - eenv->nrg.before;
5323 eenv->nrg.diff = (abs(diff) < margin) ? 0 : eenv->nrg.diff;
5325 return eenv->nrg.diff;
5328 #ifdef CONFIG_SCHED_TUNE
5330 struct target_nrg schedtune_target_nrg;
5333 * System energy normalization
5334 * Returns the normalized value, in the range [0..SCHED_LOAD_SCALE],
5335 * corresponding to the specified energy variation.
5338 normalize_energy(int energy_diff)
5341 #ifdef CONFIG_SCHED_DEBUG
5344 /* Check for boundaries */
5345 max_delta = schedtune_target_nrg.max_power;
5346 max_delta -= schedtune_target_nrg.min_power;
5347 WARN_ON(abs(energy_diff) >= max_delta);
5350 /* Do scaling using positive numbers to increase the range */
5351 normalized_nrg = (energy_diff < 0) ? -energy_diff : energy_diff;
5353 /* Scale by energy magnitude */
5354 normalized_nrg <<= SCHED_LOAD_SHIFT;
5356 /* Normalize on max energy for target platform */
5357 normalized_nrg = reciprocal_divide(
5358 normalized_nrg, schedtune_target_nrg.rdiv);
5360 return (energy_diff < 0) ? -normalized_nrg : normalized_nrg;
5364 energy_diff(struct energy_env *eenv)
5366 int boost = schedtune_task_boost(eenv->task);
5369 /* Conpute "absolute" energy diff */
5370 __energy_diff(eenv);
5372 /* Return energy diff when boost margin is 0 */
5374 return eenv->nrg.diff;
5376 /* Compute normalized energy diff */
5377 nrg_delta = normalize_energy(eenv->nrg.diff);
5378 eenv->nrg.delta = nrg_delta;
5380 eenv->payoff = schedtune_accept_deltas(
5386 * When SchedTune is enabled, the energy_diff() function will return
5387 * the computed energy payoff value. Since the energy_diff() return
5388 * value is expected to be negative by its callers, this evaluation
5389 * function return a negative value each time the evaluation return a
5390 * positive payoff, which is the condition for the acceptance of
5391 * a scheduling decision
5393 return -eenv->payoff;
5395 #else /* CONFIG_SCHED_TUNE */
5396 #define energy_diff(eenv) __energy_diff(eenv)
5400 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5401 * A waker of many should wake a different task than the one last awakened
5402 * at a frequency roughly N times higher than one of its wakees. In order
5403 * to determine whether we should let the load spread vs consolodating to
5404 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
5405 * partner, and a factor of lls_size higher frequency in the other. With
5406 * both conditions met, we can be relatively sure that the relationship is
5407 * non-monogamous, with partner count exceeding socket size. Waker/wakee
5408 * being client/server, worker/dispatcher, interrupt source or whatever is
5409 * irrelevant, spread criteria is apparent partner count exceeds socket size.
5411 static int wake_wide(struct task_struct *p)
5413 unsigned int master = current->wakee_flips;
5414 unsigned int slave = p->wakee_flips;
5415 int factor = this_cpu_read(sd_llc_size);
5418 swap(master, slave);
5419 if (slave < factor || master < slave * factor)
5424 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5425 int prev_cpu, int sync)
5427 s64 this_load, load;
5428 s64 this_eff_load, prev_eff_load;
5430 struct task_group *tg;
5431 unsigned long weight;
5435 this_cpu = smp_processor_id();
5436 load = source_load(prev_cpu, idx);
5437 this_load = target_load(this_cpu, idx);
5440 * If sync wakeup then subtract the (maximum possible)
5441 * effect of the currently running task from the load
5442 * of the current CPU:
5445 tg = task_group(current);
5446 weight = current->se.avg.load_avg;
5448 this_load += effective_load(tg, this_cpu, -weight, -weight);
5449 load += effective_load(tg, prev_cpu, 0, -weight);
5453 weight = p->se.avg.load_avg;
5456 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5457 * due to the sync cause above having dropped this_load to 0, we'll
5458 * always have an imbalance, but there's really nothing you can do
5459 * about that, so that's good too.
5461 * Otherwise check if either cpus are near enough in load to allow this
5462 * task to be woken on this_cpu.
5464 this_eff_load = 100;
5465 this_eff_load *= capacity_of(prev_cpu);
5467 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5468 prev_eff_load *= capacity_of(this_cpu);
5470 if (this_load > 0) {
5471 this_eff_load *= this_load +
5472 effective_load(tg, this_cpu, weight, weight);
5474 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5477 balanced = this_eff_load <= prev_eff_load;
5479 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5484 schedstat_inc(sd, ttwu_move_affine);
5485 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5490 static inline unsigned long task_util(struct task_struct *p)
5492 #ifdef CONFIG_SCHED_WALT
5493 if (!walt_disabled && sysctl_sched_use_walt_task_util) {
5494 unsigned long demand = p->ravg.demand;
5495 return (demand << 10) / walt_ravg_window;
5498 return p->se.avg.util_avg;
5501 static inline unsigned long boosted_task_util(struct task_struct *task);
5503 static inline bool __task_fits(struct task_struct *p, int cpu, int util)
5505 unsigned long capacity = capacity_of(cpu);
5507 util += boosted_task_util(p);
5509 return (capacity * 1024) > (util * capacity_margin);
5512 static inline bool task_fits_max(struct task_struct *p, int cpu)
5514 unsigned long capacity = capacity_of(cpu);
5515 unsigned long max_capacity = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5517 if (capacity == max_capacity)
5520 if (capacity * capacity_margin > max_capacity * 1024)
5523 return __task_fits(p, cpu, 0);
5526 static bool cpu_overutilized(int cpu)
5528 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5531 #ifdef CONFIG_SCHED_TUNE
5534 schedtune_margin(unsigned long signal, long boost)
5536 long long margin = 0;
5539 * Signal proportional compensation (SPC)
5541 * The Boost (B) value is used to compute a Margin (M) which is
5542 * proportional to the complement of the original Signal (S):
5543 * M = B * (SCHED_LOAD_SCALE - S), if B is positive
5544 * M = B * S, if B is negative
5545 * The obtained M could be used by the caller to "boost" S.
5548 margin = SCHED_LOAD_SCALE - signal;
5551 margin = -signal * boost;
5553 * Fast integer division by constant:
5554 * Constant : (C) = 100
5555 * Precision : 0.1% (P) = 0.1
5556 * Reference : C * 100 / P (R) = 100000
5559 * Shift bits : ceil(log(R,2)) (S) = 17
5560 * Mult const : round(2^S/C) (M) = 1311
5573 schedtune_cpu_margin(unsigned long util, int cpu)
5575 int boost = schedtune_cpu_boost(cpu);
5580 return schedtune_margin(util, boost);
5584 schedtune_task_margin(struct task_struct *task)
5586 int boost = schedtune_task_boost(task);
5593 util = task_util(task);
5594 margin = schedtune_margin(util, boost);
5599 #else /* CONFIG_SCHED_TUNE */
5602 schedtune_cpu_margin(unsigned long util, int cpu)
5608 schedtune_task_margin(struct task_struct *task)
5613 #endif /* CONFIG_SCHED_TUNE */
5616 boosted_cpu_util(int cpu)
5618 unsigned long util = cpu_util(cpu);
5619 long margin = schedtune_cpu_margin(util, cpu);
5621 trace_sched_boost_cpu(cpu, util, margin);
5623 return util + margin;
5626 static inline unsigned long
5627 boosted_task_util(struct task_struct *task)
5629 unsigned long util = task_util(task);
5630 long margin = schedtune_task_margin(task);
5632 trace_sched_boost_task(task, util, margin);
5634 return util + margin;
5637 static int cpu_util_wake(int cpu, struct task_struct *p);
5639 static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
5641 return capacity_orig_of(cpu) - cpu_util_wake(cpu, p);
5645 * find_idlest_group finds and returns the least busy CPU group within the
5648 static struct sched_group *
5649 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5650 int this_cpu, int sd_flag)
5652 struct sched_group *idlest = NULL, *group = sd->groups;
5653 struct sched_group *most_spare_sg = NULL;
5654 unsigned long min_load = ULONG_MAX, this_load = 0;
5655 unsigned long most_spare = 0, this_spare = 0;
5656 int load_idx = sd->forkexec_idx;
5657 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5659 if (sd_flag & SD_BALANCE_WAKE)
5660 load_idx = sd->wake_idx;
5663 unsigned long load, avg_load, spare_cap, max_spare_cap;
5667 /* Skip over this group if it has no CPUs allowed */
5668 if (!cpumask_intersects(sched_group_cpus(group),
5669 tsk_cpus_allowed(p)))
5672 local_group = cpumask_test_cpu(this_cpu,
5673 sched_group_cpus(group));
5676 * Tally up the load of all CPUs in the group and find
5677 * the group containing the CPU with most spare capacity.
5682 for_each_cpu(i, sched_group_cpus(group)) {
5683 /* Bias balancing toward cpus of our domain */
5685 load = source_load(i, load_idx);
5687 load = target_load(i, load_idx);
5691 spare_cap = capacity_spare_wake(i, p);
5693 if (spare_cap > max_spare_cap)
5694 max_spare_cap = spare_cap;
5697 /* Adjust by relative CPU capacity of the group */
5698 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5701 this_load = avg_load;
5702 this_spare = max_spare_cap;
5704 if (avg_load < min_load) {
5705 min_load = avg_load;
5709 if (most_spare < max_spare_cap) {
5710 most_spare = max_spare_cap;
5711 most_spare_sg = group;
5714 } while (group = group->next, group != sd->groups);
5717 * The cross-over point between using spare capacity or least load
5718 * is too conservative for high utilization tasks on partially
5719 * utilized systems if we require spare_capacity > task_util(p),
5720 * so we allow for some task stuffing by using
5721 * spare_capacity > task_util(p)/2.
5723 if (this_spare > task_util(p) / 2 &&
5724 imbalance*this_spare > 100*most_spare)
5726 else if (most_spare > task_util(p) / 2)
5727 return most_spare_sg;
5729 if (!idlest || 100*this_load < imbalance*min_load)
5735 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5738 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5740 unsigned long load, min_load = ULONG_MAX;
5741 unsigned int min_exit_latency = UINT_MAX;
5742 u64 latest_idle_timestamp = 0;
5743 int least_loaded_cpu = this_cpu;
5744 int shallowest_idle_cpu = -1;
5747 /* Check if we have any choice: */
5748 if (group->group_weight == 1)
5749 return cpumask_first(sched_group_cpus(group));
5751 /* Traverse only the allowed CPUs */
5752 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5754 struct rq *rq = cpu_rq(i);
5755 struct cpuidle_state *idle = idle_get_state(rq);
5756 if (idle && idle->exit_latency < min_exit_latency) {
5758 * We give priority to a CPU whose idle state
5759 * has the smallest exit latency irrespective
5760 * of any idle timestamp.
5762 min_exit_latency = idle->exit_latency;
5763 latest_idle_timestamp = rq->idle_stamp;
5764 shallowest_idle_cpu = i;
5765 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5766 rq->idle_stamp > latest_idle_timestamp) {
5768 * If equal or no active idle state, then
5769 * the most recently idled CPU might have
5772 latest_idle_timestamp = rq->idle_stamp;
5773 shallowest_idle_cpu = i;
5775 } else if (shallowest_idle_cpu == -1) {
5776 load = weighted_cpuload(i);
5777 if (load < min_load || (load == min_load && i == this_cpu)) {
5779 least_loaded_cpu = i;
5784 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5788 * Try and locate an idle CPU in the sched_domain.
5790 static int select_idle_sibling(struct task_struct *p, int prev, int target)
5792 struct sched_domain *sd;
5793 struct sched_group *sg;
5794 int best_idle_cpu = -1;
5795 int best_idle_cstate = INT_MAX;
5796 unsigned long best_idle_capacity = ULONG_MAX;
5798 if (!sysctl_sched_cstate_aware) {
5799 if (idle_cpu(target))
5803 * If the prevous cpu is cache affine and idle, don't be stupid.
5805 if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev))
5810 * Otherwise, iterate the domains and find an elegible idle cpu.
5812 sd = rcu_dereference(per_cpu(sd_llc, target));
5813 for_each_lower_domain(sd) {
5817 if (!cpumask_intersects(sched_group_cpus(sg),
5818 tsk_cpus_allowed(p)))
5821 if (sysctl_sched_cstate_aware) {
5822 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg)) {
5823 int idle_idx = idle_get_state_idx(cpu_rq(i));
5824 unsigned long new_usage = boosted_task_util(p);
5825 unsigned long capacity_orig = capacity_orig_of(i);
5827 if (new_usage > capacity_orig || !idle_cpu(i))
5830 if (i == target && new_usage <= capacity_curr_of(target))
5833 if (idle_idx < best_idle_cstate &&
5834 capacity_orig <= best_idle_capacity) {
5836 best_idle_cstate = idle_idx;
5837 best_idle_capacity = capacity_orig;
5841 for_each_cpu(i, sched_group_cpus(sg)) {
5842 if (i == target || !idle_cpu(i))
5846 target = cpumask_first_and(sched_group_cpus(sg),
5847 tsk_cpus_allowed(p));
5852 } while (sg != sd->groups);
5855 if (best_idle_cpu >= 0)
5856 target = best_idle_cpu;
5862 static int start_cpu(bool boosted)
5864 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
5866 RCU_LOCKDEP_WARN(rcu_read_lock_sched_held(),
5867 "sched RCU must be held");
5869 return boosted ? rd->max_cap_orig_cpu : rd->min_cap_orig_cpu;
5872 static inline int find_best_target(struct task_struct *p, bool boosted, bool prefer_idle)
5874 int target_cpu = -1;
5875 unsigned long target_util = prefer_idle ? ULONG_MAX : 0;
5876 unsigned long backup_capacity = ULONG_MAX;
5877 int best_idle_cpu = -1;
5878 int best_idle_cstate = INT_MAX;
5879 int backup_cpu = -1;
5880 unsigned long min_util = boosted_task_util(p);
5881 struct sched_domain *sd;
5882 struct sched_group *sg;
5883 int cpu = start_cpu(boosted);
5888 sd = rcu_dereference(per_cpu(sd_ea, cpu));
5898 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg)) {
5899 unsigned long cur_capacity, new_util;
5905 * p's blocked utilization is still accounted for on prev_cpu
5906 * so prev_cpu will receive a negative bias due to the double
5907 * accounting. However, the blocked utilization may be zero.
5909 new_util = cpu_util(i) + task_util(p);
5912 * Ensure minimum capacity to grant the required boost.
5913 * The target CPU can be already at a capacity level higher
5914 * than the one required to boost the task.
5916 new_util = max(min_util, new_util);
5918 if (new_util > capacity_orig_of(i))
5921 #ifdef CONFIG_SCHED_WALT
5922 if (walt_cpu_high_irqload(i))
5927 * Unconditionally favoring tasks that prefer idle cpus to
5930 if (idle_cpu(i) && prefer_idle)
5933 cur_capacity = capacity_curr_of(i);
5935 if (new_util < cur_capacity) {
5936 if (cpu_rq(i)->nr_running) {
5938 * Find a target cpu with the lowest/highest
5939 * utilization if prefer_idle/!prefer_idle.
5941 if ((prefer_idle && target_util > new_util) ||
5942 (!prefer_idle && target_util < new_util)) {
5943 target_util = new_util;
5946 } else if (!prefer_idle) {
5947 int idle_idx = idle_get_state_idx(cpu_rq(i));
5949 if (best_idle_cpu < 0 ||
5950 (sysctl_sched_cstate_aware &&
5951 best_idle_cstate > idle_idx)) {
5952 best_idle_cstate = idle_idx;
5956 } else if (backup_capacity > cur_capacity) {
5957 /* Find a backup cpu with least capacity. */
5958 backup_capacity = cur_capacity;
5962 } while (sg = sg->next, sg != sd->groups);
5965 target_cpu = best_idle_cpu >= 0 ? best_idle_cpu : backup_cpu;
5971 * cpu_util_wake: Compute cpu utilization with any contributions from
5972 * the waking task p removed.
5974 static int cpu_util_wake(int cpu, struct task_struct *p)
5976 unsigned long util, capacity;
5978 /* Task has no contribution or is new */
5979 if (cpu != task_cpu(p) || !p->se.avg.last_update_time)
5980 return cpu_util(cpu);
5982 capacity = capacity_orig_of(cpu);
5983 util = max_t(long, cpu_rq(cpu)->cfs.avg.util_avg - task_util(p), 0);
5985 return (util >= capacity) ? capacity : util;
5989 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
5990 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
5992 * In that case WAKE_AFFINE doesn't make sense and we'll let
5993 * BALANCE_WAKE sort things out.
5995 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
5997 long min_cap, max_cap;
5999 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6000 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity.val;
6002 /* Minimum capacity is close to max, no need to abort wake_affine */
6003 if (max_cap - min_cap < max_cap >> 3)
6006 /* Bring task utilization in sync with prev_cpu */
6007 sync_entity_load_avg(&p->se);
6009 return min_cap * 1024 < task_util(p) * capacity_margin;
6012 static int select_energy_cpu_brute(struct task_struct *p, int prev_cpu, int sync)
6014 struct sched_domain *sd;
6015 int target_cpu = prev_cpu, tmp_target;
6016 bool boosted, prefer_idle;
6018 if (sysctl_sched_sync_hint_enable && sync) {
6019 int cpu = smp_processor_id();
6021 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
6026 #ifdef CONFIG_CGROUP_SCHEDTUNE
6027 boosted = schedtune_task_boost(p) > 0;
6028 prefer_idle = schedtune_prefer_idle(p) > 0;
6030 boosted = get_sysctl_sched_cfs_boost() > 0;
6034 sd = rcu_dereference(per_cpu(sd_ea, prev_cpu));
6035 /* Find a cpu with sufficient capacity */
6036 tmp_target = find_best_target(p, boosted, prefer_idle);
6040 if (tmp_target >= 0) {
6041 target_cpu = tmp_target;
6042 if ((boosted || prefer_idle) && idle_cpu(target_cpu))
6046 if (target_cpu != prev_cpu) {
6047 struct energy_env eenv = {
6048 .util_delta = task_util(p),
6049 .src_cpu = prev_cpu,
6050 .dst_cpu = target_cpu,
6054 /* Not enough spare capacity on previous cpu */
6055 if (cpu_overutilized(prev_cpu))
6058 if (energy_diff(&eenv) >= 0)
6059 target_cpu = prev_cpu;
6068 * select_task_rq_fair: Select target runqueue for the waking task in domains
6069 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6070 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6072 * Balances load by selecting the idlest cpu in the idlest group, or under
6073 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
6075 * Returns the target cpu number.
6077 * preempt must be disabled.
6080 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6082 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
6083 int cpu = smp_processor_id();
6084 int new_cpu = prev_cpu;
6085 int want_affine = 0;
6086 int sync = wake_flags & WF_SYNC;
6088 if (sd_flag & SD_BALANCE_WAKE)
6089 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
6090 && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
6092 if (energy_aware() && !(cpu_rq(prev_cpu)->rd->overutilized))
6093 return select_energy_cpu_brute(p, prev_cpu, sync);
6096 for_each_domain(cpu, tmp) {
6097 if (!(tmp->flags & SD_LOAD_BALANCE))
6101 * If both cpu and prev_cpu are part of this domain,
6102 * cpu is a valid SD_WAKE_AFFINE target.
6104 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6105 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6110 if (tmp->flags & sd_flag)
6112 else if (!want_affine)
6117 sd = NULL; /* Prefer wake_affine over balance flags */
6118 if (cpu != prev_cpu && wake_affine(affine_sd, p, prev_cpu, sync))
6123 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
6124 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6127 struct sched_group *group;
6130 if (!(sd->flags & sd_flag)) {
6135 group = find_idlest_group(sd, p, cpu, sd_flag);
6141 new_cpu = find_idlest_cpu(group, p, cpu);
6142 if (new_cpu == -1 || new_cpu == cpu) {
6143 /* Now try balancing at a lower domain level of cpu */
6148 /* Now try balancing at a lower domain level of new_cpu */
6150 weight = sd->span_weight;
6152 for_each_domain(cpu, tmp) {
6153 if (weight <= tmp->span_weight)
6155 if (tmp->flags & sd_flag)
6158 /* while loop will break here if sd == NULL */
6166 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
6167 * cfs_rq_of(p) references at time of call are still valid and identify the
6168 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
6169 * other assumptions, including the state of rq->lock, should be made.
6171 static void migrate_task_rq_fair(struct task_struct *p)
6174 * We are supposed to update the task to "current" time, then its up to date
6175 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
6176 * what current time is, so simply throw away the out-of-date time. This
6177 * will result in the wakee task is less decayed, but giving the wakee more
6178 * load sounds not bad.
6180 remove_entity_load_avg(&p->se);
6182 /* Tell new CPU we are migrated */
6183 p->se.avg.last_update_time = 0;
6185 /* We have migrated, no longer consider this task hot */
6186 p->se.exec_start = 0;
6189 static void task_dead_fair(struct task_struct *p)
6191 remove_entity_load_avg(&p->se);
6194 #define task_fits_max(p, cpu) true
6195 #endif /* CONFIG_SMP */
6197 static unsigned long
6198 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
6200 unsigned long gran = sysctl_sched_wakeup_granularity;
6203 * Since its curr running now, convert the gran from real-time
6204 * to virtual-time in his units.
6206 * By using 'se' instead of 'curr' we penalize light tasks, so
6207 * they get preempted easier. That is, if 'se' < 'curr' then
6208 * the resulting gran will be larger, therefore penalizing the
6209 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6210 * be smaller, again penalizing the lighter task.
6212 * This is especially important for buddies when the leftmost
6213 * task is higher priority than the buddy.
6215 return calc_delta_fair(gran, se);
6219 * Should 'se' preempt 'curr'.
6233 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6235 s64 gran, vdiff = curr->vruntime - se->vruntime;
6240 gran = wakeup_gran(curr, se);
6247 static void set_last_buddy(struct sched_entity *se)
6249 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6252 for_each_sched_entity(se)
6253 cfs_rq_of(se)->last = se;
6256 static void set_next_buddy(struct sched_entity *se)
6258 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6261 for_each_sched_entity(se)
6262 cfs_rq_of(se)->next = se;
6265 static void set_skip_buddy(struct sched_entity *se)
6267 for_each_sched_entity(se)
6268 cfs_rq_of(se)->skip = se;
6272 * Preempt the current task with a newly woken task if needed:
6274 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6276 struct task_struct *curr = rq->curr;
6277 struct sched_entity *se = &curr->se, *pse = &p->se;
6278 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6279 int scale = cfs_rq->nr_running >= sched_nr_latency;
6280 int next_buddy_marked = 0;
6282 if (unlikely(se == pse))
6286 * This is possible from callers such as attach_tasks(), in which we
6287 * unconditionally check_prempt_curr() after an enqueue (which may have
6288 * lead to a throttle). This both saves work and prevents false
6289 * next-buddy nomination below.
6291 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6294 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6295 set_next_buddy(pse);
6296 next_buddy_marked = 1;
6300 * We can come here with TIF_NEED_RESCHED already set from new task
6303 * Note: this also catches the edge-case of curr being in a throttled
6304 * group (e.g. via set_curr_task), since update_curr() (in the
6305 * enqueue of curr) will have resulted in resched being set. This
6306 * prevents us from potentially nominating it as a false LAST_BUDDY
6309 if (test_tsk_need_resched(curr))
6312 /* Idle tasks are by definition preempted by non-idle tasks. */
6313 if (unlikely(curr->policy == SCHED_IDLE) &&
6314 likely(p->policy != SCHED_IDLE))
6318 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6319 * is driven by the tick):
6321 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6324 find_matching_se(&se, &pse);
6325 update_curr(cfs_rq_of(se));
6327 if (wakeup_preempt_entity(se, pse) == 1) {
6329 * Bias pick_next to pick the sched entity that is
6330 * triggering this preemption.
6332 if (!next_buddy_marked)
6333 set_next_buddy(pse);
6342 * Only set the backward buddy when the current task is still
6343 * on the rq. This can happen when a wakeup gets interleaved
6344 * with schedule on the ->pre_schedule() or idle_balance()
6345 * point, either of which can * drop the rq lock.
6347 * Also, during early boot the idle thread is in the fair class,
6348 * for obvious reasons its a bad idea to schedule back to it.
6350 if (unlikely(!se->on_rq || curr == rq->idle))
6353 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6357 static struct task_struct *
6358 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
6360 struct cfs_rq *cfs_rq = &rq->cfs;
6361 struct sched_entity *se;
6362 struct task_struct *p;
6366 #ifdef CONFIG_FAIR_GROUP_SCHED
6367 if (!cfs_rq->nr_running)
6370 if (prev->sched_class != &fair_sched_class)
6374 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6375 * likely that a next task is from the same cgroup as the current.
6377 * Therefore attempt to avoid putting and setting the entire cgroup
6378 * hierarchy, only change the part that actually changes.
6382 struct sched_entity *curr = cfs_rq->curr;
6385 * Since we got here without doing put_prev_entity() we also
6386 * have to consider cfs_rq->curr. If it is still a runnable
6387 * entity, update_curr() will update its vruntime, otherwise
6388 * forget we've ever seen it.
6392 update_curr(cfs_rq);
6397 * This call to check_cfs_rq_runtime() will do the
6398 * throttle and dequeue its entity in the parent(s).
6399 * Therefore the 'simple' nr_running test will indeed
6402 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
6406 se = pick_next_entity(cfs_rq, curr);
6407 cfs_rq = group_cfs_rq(se);
6413 * Since we haven't yet done put_prev_entity and if the selected task
6414 * is a different task than we started out with, try and touch the
6415 * least amount of cfs_rqs.
6418 struct sched_entity *pse = &prev->se;
6420 while (!(cfs_rq = is_same_group(se, pse))) {
6421 int se_depth = se->depth;
6422 int pse_depth = pse->depth;
6424 if (se_depth <= pse_depth) {
6425 put_prev_entity(cfs_rq_of(pse), pse);
6426 pse = parent_entity(pse);
6428 if (se_depth >= pse_depth) {
6429 set_next_entity(cfs_rq_of(se), se);
6430 se = parent_entity(se);
6434 put_prev_entity(cfs_rq, pse);
6435 set_next_entity(cfs_rq, se);
6438 if (hrtick_enabled(rq))
6439 hrtick_start_fair(rq, p);
6441 rq->misfit_task = !task_fits_max(p, rq->cpu);
6448 if (!cfs_rq->nr_running)
6451 put_prev_task(rq, prev);
6454 se = pick_next_entity(cfs_rq, NULL);
6455 set_next_entity(cfs_rq, se);
6456 cfs_rq = group_cfs_rq(se);
6461 if (hrtick_enabled(rq))
6462 hrtick_start_fair(rq, p);
6464 rq->misfit_task = !task_fits_max(p, rq->cpu);
6469 rq->misfit_task = 0;
6471 * This is OK, because current is on_cpu, which avoids it being picked
6472 * for load-balance and preemption/IRQs are still disabled avoiding
6473 * further scheduler activity on it and we're being very careful to
6474 * re-start the picking loop.
6476 lockdep_unpin_lock(&rq->lock);
6477 new_tasks = idle_balance(rq);
6478 lockdep_pin_lock(&rq->lock);
6480 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6481 * possible for any higher priority task to appear. In that case we
6482 * must re-start the pick_next_entity() loop.
6494 * Account for a descheduled task:
6496 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6498 struct sched_entity *se = &prev->se;
6499 struct cfs_rq *cfs_rq;
6501 for_each_sched_entity(se) {
6502 cfs_rq = cfs_rq_of(se);
6503 put_prev_entity(cfs_rq, se);
6508 * sched_yield() is very simple
6510 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6512 static void yield_task_fair(struct rq *rq)
6514 struct task_struct *curr = rq->curr;
6515 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6516 struct sched_entity *se = &curr->se;
6519 * Are we the only task in the tree?
6521 if (unlikely(rq->nr_running == 1))
6524 clear_buddies(cfs_rq, se);
6526 if (curr->policy != SCHED_BATCH) {
6527 update_rq_clock(rq);
6529 * Update run-time statistics of the 'current'.
6531 update_curr(cfs_rq);
6533 * Tell update_rq_clock() that we've just updated,
6534 * so we don't do microscopic update in schedule()
6535 * and double the fastpath cost.
6537 rq_clock_skip_update(rq, true);
6543 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6545 struct sched_entity *se = &p->se;
6547 /* throttled hierarchies are not runnable */
6548 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6551 /* Tell the scheduler that we'd really like pse to run next. */
6554 yield_task_fair(rq);
6560 /**************************************************
6561 * Fair scheduling class load-balancing methods.
6565 * The purpose of load-balancing is to achieve the same basic fairness the
6566 * per-cpu scheduler provides, namely provide a proportional amount of compute
6567 * time to each task. This is expressed in the following equation:
6569 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6571 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6572 * W_i,0 is defined as:
6574 * W_i,0 = \Sum_j w_i,j (2)
6576 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6577 * is derived from the nice value as per prio_to_weight[].
6579 * The weight average is an exponential decay average of the instantaneous
6582 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6584 * C_i is the compute capacity of cpu i, typically it is the
6585 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6586 * can also include other factors [XXX].
6588 * To achieve this balance we define a measure of imbalance which follows
6589 * directly from (1):
6591 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6593 * We them move tasks around to minimize the imbalance. In the continuous
6594 * function space it is obvious this converges, in the discrete case we get
6595 * a few fun cases generally called infeasible weight scenarios.
6598 * - infeasible weights;
6599 * - local vs global optima in the discrete case. ]
6604 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6605 * for all i,j solution, we create a tree of cpus that follows the hardware
6606 * topology where each level pairs two lower groups (or better). This results
6607 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6608 * tree to only the first of the previous level and we decrease the frequency
6609 * of load-balance at each level inv. proportional to the number of cpus in
6615 * \Sum { --- * --- * 2^i } = O(n) (5)
6617 * `- size of each group
6618 * | | `- number of cpus doing load-balance
6620 * `- sum over all levels
6622 * Coupled with a limit on how many tasks we can migrate every balance pass,
6623 * this makes (5) the runtime complexity of the balancer.
6625 * An important property here is that each CPU is still (indirectly) connected
6626 * to every other cpu in at most O(log n) steps:
6628 * The adjacency matrix of the resulting graph is given by:
6631 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6634 * And you'll find that:
6636 * A^(log_2 n)_i,j != 0 for all i,j (7)
6638 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6639 * The task movement gives a factor of O(m), giving a convergence complexity
6642 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6647 * In order to avoid CPUs going idle while there's still work to do, new idle
6648 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6649 * tree itself instead of relying on other CPUs to bring it work.
6651 * This adds some complexity to both (5) and (8) but it reduces the total idle
6659 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6662 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6667 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6669 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6671 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6674 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6675 * rewrite all of this once again.]
6678 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6680 enum fbq_type { regular, remote, all };
6689 #define LBF_ALL_PINNED 0x01
6690 #define LBF_NEED_BREAK 0x02
6691 #define LBF_DST_PINNED 0x04
6692 #define LBF_SOME_PINNED 0x08
6695 struct sched_domain *sd;
6703 struct cpumask *dst_grpmask;
6705 enum cpu_idle_type idle;
6707 unsigned int src_grp_nr_running;
6708 /* The set of CPUs under consideration for load-balancing */
6709 struct cpumask *cpus;
6714 unsigned int loop_break;
6715 unsigned int loop_max;
6717 enum fbq_type fbq_type;
6718 enum group_type busiest_group_type;
6719 struct list_head tasks;
6723 * Is this task likely cache-hot:
6725 static int task_hot(struct task_struct *p, struct lb_env *env)
6729 lockdep_assert_held(&env->src_rq->lock);
6731 if (p->sched_class != &fair_sched_class)
6734 if (unlikely(p->policy == SCHED_IDLE))
6738 * Buddy candidates are cache hot:
6740 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6741 (&p->se == cfs_rq_of(&p->se)->next ||
6742 &p->se == cfs_rq_of(&p->se)->last))
6745 if (sysctl_sched_migration_cost == -1)
6747 if (sysctl_sched_migration_cost == 0)
6750 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6752 return delta < (s64)sysctl_sched_migration_cost;
6755 #ifdef CONFIG_NUMA_BALANCING
6757 * Returns 1, if task migration degrades locality
6758 * Returns 0, if task migration improves locality i.e migration preferred.
6759 * Returns -1, if task migration is not affected by locality.
6761 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6763 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6764 unsigned long src_faults, dst_faults;
6765 int src_nid, dst_nid;
6767 if (!static_branch_likely(&sched_numa_balancing))
6770 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6773 src_nid = cpu_to_node(env->src_cpu);
6774 dst_nid = cpu_to_node(env->dst_cpu);
6776 if (src_nid == dst_nid)
6779 /* Migrating away from the preferred node is always bad. */
6780 if (src_nid == p->numa_preferred_nid) {
6781 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6787 /* Encourage migration to the preferred node. */
6788 if (dst_nid == p->numa_preferred_nid)
6792 src_faults = group_faults(p, src_nid);
6793 dst_faults = group_faults(p, dst_nid);
6795 src_faults = task_faults(p, src_nid);
6796 dst_faults = task_faults(p, dst_nid);
6799 return dst_faults < src_faults;
6803 static inline int migrate_degrades_locality(struct task_struct *p,
6811 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6814 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6818 lockdep_assert_held(&env->src_rq->lock);
6821 * We do not migrate tasks that are:
6822 * 1) throttled_lb_pair, or
6823 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6824 * 3) running (obviously), or
6825 * 4) are cache-hot on their current CPU.
6827 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6830 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6833 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6835 env->flags |= LBF_SOME_PINNED;
6838 * Remember if this task can be migrated to any other cpu in
6839 * our sched_group. We may want to revisit it if we couldn't
6840 * meet load balance goals by pulling other tasks on src_cpu.
6842 * Also avoid computing new_dst_cpu if we have already computed
6843 * one in current iteration.
6845 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6848 /* Prevent to re-select dst_cpu via env's cpus */
6849 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6850 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6851 env->flags |= LBF_DST_PINNED;
6852 env->new_dst_cpu = cpu;
6860 /* Record that we found atleast one task that could run on dst_cpu */
6861 env->flags &= ~LBF_ALL_PINNED;
6863 if (task_running(env->src_rq, p)) {
6864 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6869 * Aggressive migration if:
6870 * 1) destination numa is preferred
6871 * 2) task is cache cold, or
6872 * 3) too many balance attempts have failed.
6874 tsk_cache_hot = migrate_degrades_locality(p, env);
6875 if (tsk_cache_hot == -1)
6876 tsk_cache_hot = task_hot(p, env);
6878 if (tsk_cache_hot <= 0 ||
6879 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6880 if (tsk_cache_hot == 1) {
6881 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6882 schedstat_inc(p, se.statistics.nr_forced_migrations);
6887 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6892 * detach_task() -- detach the task for the migration specified in env
6894 static void detach_task(struct task_struct *p, struct lb_env *env)
6896 lockdep_assert_held(&env->src_rq->lock);
6898 deactivate_task(env->src_rq, p, 0);
6899 p->on_rq = TASK_ON_RQ_MIGRATING;
6900 double_lock_balance(env->src_rq, env->dst_rq);
6901 set_task_cpu(p, env->dst_cpu);
6902 double_unlock_balance(env->src_rq, env->dst_rq);
6906 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6907 * part of active balancing operations within "domain".
6909 * Returns a task if successful and NULL otherwise.
6911 static struct task_struct *detach_one_task(struct lb_env *env)
6913 struct task_struct *p, *n;
6915 lockdep_assert_held(&env->src_rq->lock);
6917 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6918 if (!can_migrate_task(p, env))
6921 detach_task(p, env);
6924 * Right now, this is only the second place where
6925 * lb_gained[env->idle] is updated (other is detach_tasks)
6926 * so we can safely collect stats here rather than
6927 * inside detach_tasks().
6929 schedstat_inc(env->sd, lb_gained[env->idle]);
6935 static const unsigned int sched_nr_migrate_break = 32;
6938 * detach_tasks() -- tries to detach up to imbalance weighted load from
6939 * busiest_rq, as part of a balancing operation within domain "sd".
6941 * Returns number of detached tasks if successful and 0 otherwise.
6943 static int detach_tasks(struct lb_env *env)
6945 struct list_head *tasks = &env->src_rq->cfs_tasks;
6946 struct task_struct *p;
6950 lockdep_assert_held(&env->src_rq->lock);
6952 if (env->imbalance <= 0)
6955 while (!list_empty(tasks)) {
6957 * We don't want to steal all, otherwise we may be treated likewise,
6958 * which could at worst lead to a livelock crash.
6960 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6963 p = list_first_entry(tasks, struct task_struct, se.group_node);
6966 /* We've more or less seen every task there is, call it quits */
6967 if (env->loop > env->loop_max)
6970 /* take a breather every nr_migrate tasks */
6971 if (env->loop > env->loop_break) {
6972 env->loop_break += sched_nr_migrate_break;
6973 env->flags |= LBF_NEED_BREAK;
6977 if (!can_migrate_task(p, env))
6980 load = task_h_load(p);
6982 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6985 if ((load / 2) > env->imbalance)
6988 detach_task(p, env);
6989 list_add(&p->se.group_node, &env->tasks);
6992 env->imbalance -= load;
6994 #ifdef CONFIG_PREEMPT
6996 * NEWIDLE balancing is a source of latency, so preemptible
6997 * kernels will stop after the first task is detached to minimize
6998 * the critical section.
7000 if (env->idle == CPU_NEWLY_IDLE)
7005 * We only want to steal up to the prescribed amount of
7008 if (env->imbalance <= 0)
7013 list_move_tail(&p->se.group_node, tasks);
7017 * Right now, this is one of only two places we collect this stat
7018 * so we can safely collect detach_one_task() stats here rather
7019 * than inside detach_one_task().
7021 schedstat_add(env->sd, lb_gained[env->idle], detached);
7027 * attach_task() -- attach the task detached by detach_task() to its new rq.
7029 static void attach_task(struct rq *rq, struct task_struct *p)
7031 lockdep_assert_held(&rq->lock);
7033 BUG_ON(task_rq(p) != rq);
7034 p->on_rq = TASK_ON_RQ_QUEUED;
7035 activate_task(rq, p, 0);
7036 check_preempt_curr(rq, p, 0);
7040 * attach_one_task() -- attaches the task returned from detach_one_task() to
7043 static void attach_one_task(struct rq *rq, struct task_struct *p)
7045 raw_spin_lock(&rq->lock);
7048 * We want to potentially raise target_cpu's OPP.
7050 update_capacity_of(cpu_of(rq));
7051 raw_spin_unlock(&rq->lock);
7055 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7058 static void attach_tasks(struct lb_env *env)
7060 struct list_head *tasks = &env->tasks;
7061 struct task_struct *p;
7063 raw_spin_lock(&env->dst_rq->lock);
7065 while (!list_empty(tasks)) {
7066 p = list_first_entry(tasks, struct task_struct, se.group_node);
7067 list_del_init(&p->se.group_node);
7069 attach_task(env->dst_rq, p);
7073 * We want to potentially raise env.dst_cpu's OPP.
7075 update_capacity_of(env->dst_cpu);
7077 raw_spin_unlock(&env->dst_rq->lock);
7080 #ifdef CONFIG_FAIR_GROUP_SCHED
7081 static void update_blocked_averages(int cpu)
7083 struct rq *rq = cpu_rq(cpu);
7084 struct cfs_rq *cfs_rq;
7085 unsigned long flags;
7087 raw_spin_lock_irqsave(&rq->lock, flags);
7088 update_rq_clock(rq);
7091 * Iterates the task_group tree in a bottom up fashion, see
7092 * list_add_leaf_cfs_rq() for details.
7094 for_each_leaf_cfs_rq(rq, cfs_rq) {
7095 /* throttled entities do not contribute to load */
7096 if (throttled_hierarchy(cfs_rq))
7099 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq,
7101 update_tg_load_avg(cfs_rq, 0);
7103 raw_spin_unlock_irqrestore(&rq->lock, flags);
7107 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7108 * This needs to be done in a top-down fashion because the load of a child
7109 * group is a fraction of its parents load.
7111 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7113 struct rq *rq = rq_of(cfs_rq);
7114 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7115 unsigned long now = jiffies;
7118 if (cfs_rq->last_h_load_update == now)
7121 cfs_rq->h_load_next = NULL;
7122 for_each_sched_entity(se) {
7123 cfs_rq = cfs_rq_of(se);
7124 cfs_rq->h_load_next = se;
7125 if (cfs_rq->last_h_load_update == now)
7130 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7131 cfs_rq->last_h_load_update = now;
7134 while ((se = cfs_rq->h_load_next) != NULL) {
7135 load = cfs_rq->h_load;
7136 load = div64_ul(load * se->avg.load_avg,
7137 cfs_rq_load_avg(cfs_rq) + 1);
7138 cfs_rq = group_cfs_rq(se);
7139 cfs_rq->h_load = load;
7140 cfs_rq->last_h_load_update = now;
7144 static unsigned long task_h_load(struct task_struct *p)
7146 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7148 update_cfs_rq_h_load(cfs_rq);
7149 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7150 cfs_rq_load_avg(cfs_rq) + 1);
7153 static inline void update_blocked_averages(int cpu)
7155 struct rq *rq = cpu_rq(cpu);
7156 struct cfs_rq *cfs_rq = &rq->cfs;
7157 unsigned long flags;
7159 raw_spin_lock_irqsave(&rq->lock, flags);
7160 update_rq_clock(rq);
7161 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true);
7162 raw_spin_unlock_irqrestore(&rq->lock, flags);
7165 static unsigned long task_h_load(struct task_struct *p)
7167 return p->se.avg.load_avg;
7171 /********** Helpers for find_busiest_group ************************/
7174 * sg_lb_stats - stats of a sched_group required for load_balancing
7176 struct sg_lb_stats {
7177 unsigned long avg_load; /*Avg load across the CPUs of the group */
7178 unsigned long group_load; /* Total load over the CPUs of the group */
7179 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7180 unsigned long load_per_task;
7181 unsigned long group_capacity;
7182 unsigned long group_util; /* Total utilization of the group */
7183 unsigned int sum_nr_running; /* Nr tasks running in the group */
7184 unsigned int idle_cpus;
7185 unsigned int group_weight;
7186 enum group_type group_type;
7187 int group_no_capacity;
7188 int group_misfit_task; /* A cpu has a task too big for its capacity */
7189 #ifdef CONFIG_NUMA_BALANCING
7190 unsigned int nr_numa_running;
7191 unsigned int nr_preferred_running;
7196 * sd_lb_stats - Structure to store the statistics of a sched_domain
7197 * during load balancing.
7199 struct sd_lb_stats {
7200 struct sched_group *busiest; /* Busiest group in this sd */
7201 struct sched_group *local; /* Local group in this sd */
7202 unsigned long total_load; /* Total load of all groups in sd */
7203 unsigned long total_capacity; /* Total capacity of all groups in sd */
7204 unsigned long avg_load; /* Average load across all groups in sd */
7206 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7207 struct sg_lb_stats local_stat; /* Statistics of the local group */
7210 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7213 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7214 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7215 * We must however clear busiest_stat::avg_load because
7216 * update_sd_pick_busiest() reads this before assignment.
7218 *sds = (struct sd_lb_stats){
7222 .total_capacity = 0UL,
7225 .sum_nr_running = 0,
7226 .group_type = group_other,
7232 * get_sd_load_idx - Obtain the load index for a given sched domain.
7233 * @sd: The sched_domain whose load_idx is to be obtained.
7234 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7236 * Return: The load index.
7238 static inline int get_sd_load_idx(struct sched_domain *sd,
7239 enum cpu_idle_type idle)
7245 load_idx = sd->busy_idx;
7248 case CPU_NEWLY_IDLE:
7249 load_idx = sd->newidle_idx;
7252 load_idx = sd->idle_idx;
7259 static unsigned long scale_rt_capacity(int cpu)
7261 struct rq *rq = cpu_rq(cpu);
7262 u64 total, used, age_stamp, avg;
7266 * Since we're reading these variables without serialization make sure
7267 * we read them once before doing sanity checks on them.
7269 age_stamp = READ_ONCE(rq->age_stamp);
7270 avg = READ_ONCE(rq->rt_avg);
7271 delta = __rq_clock_broken(rq) - age_stamp;
7273 if (unlikely(delta < 0))
7276 total = sched_avg_period() + delta;
7278 used = div_u64(avg, total);
7281 * deadline bandwidth is defined at system level so we must
7282 * weight this bandwidth with the max capacity of the system.
7283 * As a reminder, avg_bw is 20bits width and
7284 * scale_cpu_capacity is 10 bits width
7286 used += div_u64(rq->dl.avg_bw, arch_scale_cpu_capacity(NULL, cpu));
7288 if (likely(used < SCHED_CAPACITY_SCALE))
7289 return SCHED_CAPACITY_SCALE - used;
7294 void init_max_cpu_capacity(struct max_cpu_capacity *mcc)
7296 raw_spin_lock_init(&mcc->lock);
7301 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7303 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7304 struct sched_group *sdg = sd->groups;
7305 struct max_cpu_capacity *mcc;
7306 unsigned long max_capacity;
7308 unsigned long flags;
7310 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7312 mcc = &cpu_rq(cpu)->rd->max_cpu_capacity;
7314 raw_spin_lock_irqsave(&mcc->lock, flags);
7315 max_capacity = mcc->val;
7316 max_cap_cpu = mcc->cpu;
7318 if ((max_capacity > capacity && max_cap_cpu == cpu) ||
7319 (max_capacity < capacity)) {
7320 mcc->val = capacity;
7322 #ifdef CONFIG_SCHED_DEBUG
7323 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7324 printk_deferred(KERN_INFO "CPU%d: update max cpu_capacity %lu\n",
7329 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7331 skip_unlock: __attribute__ ((unused));
7332 capacity *= scale_rt_capacity(cpu);
7333 capacity >>= SCHED_CAPACITY_SHIFT;
7338 cpu_rq(cpu)->cpu_capacity = capacity;
7339 sdg->sgc->capacity = capacity;
7340 sdg->sgc->max_capacity = capacity;
7341 sdg->sgc->min_capacity = capacity;
7344 void update_group_capacity(struct sched_domain *sd, int cpu)
7346 struct sched_domain *child = sd->child;
7347 struct sched_group *group, *sdg = sd->groups;
7348 unsigned long capacity, max_capacity, min_capacity;
7349 unsigned long interval;
7351 interval = msecs_to_jiffies(sd->balance_interval);
7352 interval = clamp(interval, 1UL, max_load_balance_interval);
7353 sdg->sgc->next_update = jiffies + interval;
7356 update_cpu_capacity(sd, cpu);
7362 min_capacity = ULONG_MAX;
7364 if (child->flags & SD_OVERLAP) {
7366 * SD_OVERLAP domains cannot assume that child groups
7367 * span the current group.
7370 for_each_cpu(cpu, sched_group_cpus(sdg)) {
7371 struct sched_group_capacity *sgc;
7372 struct rq *rq = cpu_rq(cpu);
7375 * build_sched_domains() -> init_sched_groups_capacity()
7376 * gets here before we've attached the domains to the
7379 * Use capacity_of(), which is set irrespective of domains
7380 * in update_cpu_capacity().
7382 * This avoids capacity from being 0 and
7383 * causing divide-by-zero issues on boot.
7385 if (unlikely(!rq->sd)) {
7386 capacity += capacity_of(cpu);
7388 sgc = rq->sd->groups->sgc;
7389 capacity += sgc->capacity;
7392 max_capacity = max(capacity, max_capacity);
7393 min_capacity = min(capacity, min_capacity);
7397 * !SD_OVERLAP domains can assume that child groups
7398 * span the current group.
7401 group = child->groups;
7403 struct sched_group_capacity *sgc = group->sgc;
7405 capacity += sgc->capacity;
7406 max_capacity = max(sgc->max_capacity, max_capacity);
7407 min_capacity = min(sgc->min_capacity, min_capacity);
7408 group = group->next;
7409 } while (group != child->groups);
7412 sdg->sgc->capacity = capacity;
7413 sdg->sgc->max_capacity = max_capacity;
7414 sdg->sgc->min_capacity = min_capacity;
7418 * Check whether the capacity of the rq has been noticeably reduced by side
7419 * activity. The imbalance_pct is used for the threshold.
7420 * Return true is the capacity is reduced
7423 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7425 return ((rq->cpu_capacity * sd->imbalance_pct) <
7426 (rq->cpu_capacity_orig * 100));
7430 * Group imbalance indicates (and tries to solve) the problem where balancing
7431 * groups is inadequate due to tsk_cpus_allowed() constraints.
7433 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7434 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7437 * { 0 1 2 3 } { 4 5 6 7 }
7440 * If we were to balance group-wise we'd place two tasks in the first group and
7441 * two tasks in the second group. Clearly this is undesired as it will overload
7442 * cpu 3 and leave one of the cpus in the second group unused.
7444 * The current solution to this issue is detecting the skew in the first group
7445 * by noticing the lower domain failed to reach balance and had difficulty
7446 * moving tasks due to affinity constraints.
7448 * When this is so detected; this group becomes a candidate for busiest; see
7449 * update_sd_pick_busiest(). And calculate_imbalance() and
7450 * find_busiest_group() avoid some of the usual balance conditions to allow it
7451 * to create an effective group imbalance.
7453 * This is a somewhat tricky proposition since the next run might not find the
7454 * group imbalance and decide the groups need to be balanced again. A most
7455 * subtle and fragile situation.
7458 static inline int sg_imbalanced(struct sched_group *group)
7460 return group->sgc->imbalance;
7464 * group_has_capacity returns true if the group has spare capacity that could
7465 * be used by some tasks.
7466 * We consider that a group has spare capacity if the * number of task is
7467 * smaller than the number of CPUs or if the utilization is lower than the
7468 * available capacity for CFS tasks.
7469 * For the latter, we use a threshold to stabilize the state, to take into
7470 * account the variance of the tasks' load and to return true if the available
7471 * capacity in meaningful for the load balancer.
7472 * As an example, an available capacity of 1% can appear but it doesn't make
7473 * any benefit for the load balance.
7476 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7478 if (sgs->sum_nr_running < sgs->group_weight)
7481 if ((sgs->group_capacity * 100) >
7482 (sgs->group_util * env->sd->imbalance_pct))
7489 * group_is_overloaded returns true if the group has more tasks than it can
7491 * group_is_overloaded is not equals to !group_has_capacity because a group
7492 * with the exact right number of tasks, has no more spare capacity but is not
7493 * overloaded so both group_has_capacity and group_is_overloaded return
7497 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7499 if (sgs->sum_nr_running <= sgs->group_weight)
7502 if ((sgs->group_capacity * 100) <
7503 (sgs->group_util * env->sd->imbalance_pct))
7511 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7512 * per-cpu capacity than sched_group ref.
7515 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7517 return sg->sgc->max_capacity + capacity_margin - SCHED_LOAD_SCALE <
7518 ref->sgc->max_capacity;
7522 group_type group_classify(struct sched_group *group,
7523 struct sg_lb_stats *sgs)
7525 if (sgs->group_no_capacity)
7526 return group_overloaded;
7528 if (sg_imbalanced(group))
7529 return group_imbalanced;
7531 if (sgs->group_misfit_task)
7532 return group_misfit_task;
7538 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7539 * @env: The load balancing environment.
7540 * @group: sched_group whose statistics are to be updated.
7541 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7542 * @local_group: Does group contain this_cpu.
7543 * @sgs: variable to hold the statistics for this group.
7544 * @overload: Indicate more than one runnable task for any CPU.
7545 * @overutilized: Indicate overutilization for any CPU.
7547 static inline void update_sg_lb_stats(struct lb_env *env,
7548 struct sched_group *group, int load_idx,
7549 int local_group, struct sg_lb_stats *sgs,
7550 bool *overload, bool *overutilized)
7555 memset(sgs, 0, sizeof(*sgs));
7557 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7558 struct rq *rq = cpu_rq(i);
7560 /* Bias balancing toward cpus of our domain */
7562 load = target_load(i, load_idx);
7564 load = source_load(i, load_idx);
7566 sgs->group_load += load;
7567 sgs->group_util += cpu_util(i);
7568 sgs->sum_nr_running += rq->cfs.h_nr_running;
7570 nr_running = rq->nr_running;
7574 #ifdef CONFIG_NUMA_BALANCING
7575 sgs->nr_numa_running += rq->nr_numa_running;
7576 sgs->nr_preferred_running += rq->nr_preferred_running;
7578 sgs->sum_weighted_load += weighted_cpuload(i);
7580 * No need to call idle_cpu() if nr_running is not 0
7582 if (!nr_running && idle_cpu(i))
7585 if (cpu_overutilized(i)) {
7586 *overutilized = true;
7587 if (!sgs->group_misfit_task && rq->misfit_task)
7588 sgs->group_misfit_task = capacity_of(i);
7592 /* Adjust by relative CPU capacity of the group */
7593 sgs->group_capacity = group->sgc->capacity;
7594 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7596 if (sgs->sum_nr_running)
7597 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7599 sgs->group_weight = group->group_weight;
7601 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7602 sgs->group_type = group_classify(group, sgs);
7606 * update_sd_pick_busiest - return 1 on busiest group
7607 * @env: The load balancing environment.
7608 * @sds: sched_domain statistics
7609 * @sg: sched_group candidate to be checked for being the busiest
7610 * @sgs: sched_group statistics
7612 * Determine if @sg is a busier group than the previously selected
7615 * Return: %true if @sg is a busier group than the previously selected
7616 * busiest group. %false otherwise.
7618 static bool update_sd_pick_busiest(struct lb_env *env,
7619 struct sd_lb_stats *sds,
7620 struct sched_group *sg,
7621 struct sg_lb_stats *sgs)
7623 struct sg_lb_stats *busiest = &sds->busiest_stat;
7625 if (sgs->group_type > busiest->group_type)
7628 if (sgs->group_type < busiest->group_type)
7632 * Candidate sg doesn't face any serious load-balance problems
7633 * so don't pick it if the local sg is already filled up.
7635 if (sgs->group_type == group_other &&
7636 !group_has_capacity(env, &sds->local_stat))
7639 if (sgs->avg_load <= busiest->avg_load)
7642 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
7646 * Candidate sg has no more than one task per CPU and
7647 * has higher per-CPU capacity. Migrating tasks to less
7648 * capable CPUs may harm throughput. Maximize throughput,
7649 * power/energy consequences are not considered.
7651 if (sgs->sum_nr_running <= sgs->group_weight &&
7652 group_smaller_cpu_capacity(sds->local, sg))
7656 /* This is the busiest node in its class. */
7657 if (!(env->sd->flags & SD_ASYM_PACKING))
7661 * ASYM_PACKING needs to move all the work to the lowest
7662 * numbered CPUs in the group, therefore mark all groups
7663 * higher than ourself as busy.
7665 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
7669 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
7676 #ifdef CONFIG_NUMA_BALANCING
7677 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7679 if (sgs->sum_nr_running > sgs->nr_numa_running)
7681 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7686 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7688 if (rq->nr_running > rq->nr_numa_running)
7690 if (rq->nr_running > rq->nr_preferred_running)
7695 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7700 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7704 #endif /* CONFIG_NUMA_BALANCING */
7707 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7708 * @env: The load balancing environment.
7709 * @sds: variable to hold the statistics for this sched_domain.
7711 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7713 struct sched_domain *child = env->sd->child;
7714 struct sched_group *sg = env->sd->groups;
7715 struct sg_lb_stats tmp_sgs;
7716 int load_idx, prefer_sibling = 0;
7717 bool overload = false, overutilized = false;
7719 if (child && child->flags & SD_PREFER_SIBLING)
7722 load_idx = get_sd_load_idx(env->sd, env->idle);
7725 struct sg_lb_stats *sgs = &tmp_sgs;
7728 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
7731 sgs = &sds->local_stat;
7733 if (env->idle != CPU_NEWLY_IDLE ||
7734 time_after_eq(jiffies, sg->sgc->next_update))
7735 update_group_capacity(env->sd, env->dst_cpu);
7738 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7739 &overload, &overutilized);
7745 * In case the child domain prefers tasks go to siblings
7746 * first, lower the sg capacity so that we'll try
7747 * and move all the excess tasks away. We lower the capacity
7748 * of a group only if the local group has the capacity to fit
7749 * these excess tasks. The extra check prevents the case where
7750 * you always pull from the heaviest group when it is already
7751 * under-utilized (possible with a large weight task outweighs
7752 * the tasks on the system).
7754 if (prefer_sibling && sds->local &&
7755 group_has_capacity(env, &sds->local_stat) &&
7756 (sgs->sum_nr_running > 1)) {
7757 sgs->group_no_capacity = 1;
7758 sgs->group_type = group_classify(sg, sgs);
7762 * Ignore task groups with misfit tasks if local group has no
7763 * capacity or if per-cpu capacity isn't higher.
7765 if (sgs->group_type == group_misfit_task &&
7766 (!group_has_capacity(env, &sds->local_stat) ||
7767 !group_smaller_cpu_capacity(sg, sds->local)))
7768 sgs->group_type = group_other;
7770 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7772 sds->busiest_stat = *sgs;
7776 /* Now, start updating sd_lb_stats */
7777 sds->total_load += sgs->group_load;
7778 sds->total_capacity += sgs->group_capacity;
7781 } while (sg != env->sd->groups);
7783 if (env->sd->flags & SD_NUMA)
7784 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7786 env->src_grp_nr_running = sds->busiest_stat.sum_nr_running;
7788 if (!env->sd->parent) {
7789 /* update overload indicator if we are at root domain */
7790 if (env->dst_rq->rd->overload != overload)
7791 env->dst_rq->rd->overload = overload;
7793 /* Update over-utilization (tipping point, U >= 0) indicator */
7794 if (env->dst_rq->rd->overutilized != overutilized) {
7795 env->dst_rq->rd->overutilized = overutilized;
7796 trace_sched_overutilized(overutilized);
7799 if (!env->dst_rq->rd->overutilized && overutilized) {
7800 env->dst_rq->rd->overutilized = true;
7801 trace_sched_overutilized(true);
7808 * check_asym_packing - Check to see if the group is packed into the
7811 * This is primarily intended to used at the sibling level. Some
7812 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7813 * case of POWER7, it can move to lower SMT modes only when higher
7814 * threads are idle. When in lower SMT modes, the threads will
7815 * perform better since they share less core resources. Hence when we
7816 * have idle threads, we want them to be the higher ones.
7818 * This packing function is run on idle threads. It checks to see if
7819 * the busiest CPU in this domain (core in the P7 case) has a higher
7820 * CPU number than the packing function is being run on. Here we are
7821 * assuming lower CPU number will be equivalent to lower a SMT thread
7824 * Return: 1 when packing is required and a task should be moved to
7825 * this CPU. The amount of the imbalance is returned in *imbalance.
7827 * @env: The load balancing environment.
7828 * @sds: Statistics of the sched_domain which is to be packed
7830 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7834 if (!(env->sd->flags & SD_ASYM_PACKING))
7840 busiest_cpu = group_first_cpu(sds->busiest);
7841 if (env->dst_cpu > busiest_cpu)
7844 env->imbalance = DIV_ROUND_CLOSEST(
7845 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7846 SCHED_CAPACITY_SCALE);
7852 * fix_small_imbalance - Calculate the minor imbalance that exists
7853 * amongst the groups of a sched_domain, during
7855 * @env: The load balancing environment.
7856 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7859 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7861 unsigned long tmp, capa_now = 0, capa_move = 0;
7862 unsigned int imbn = 2;
7863 unsigned long scaled_busy_load_per_task;
7864 struct sg_lb_stats *local, *busiest;
7866 local = &sds->local_stat;
7867 busiest = &sds->busiest_stat;
7869 if (!local->sum_nr_running)
7870 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7871 else if (busiest->load_per_task > local->load_per_task)
7874 scaled_busy_load_per_task =
7875 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7876 busiest->group_capacity;
7878 if (busiest->avg_load + scaled_busy_load_per_task >=
7879 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7880 env->imbalance = busiest->load_per_task;
7885 * OK, we don't have enough imbalance to justify moving tasks,
7886 * however we may be able to increase total CPU capacity used by
7890 capa_now += busiest->group_capacity *
7891 min(busiest->load_per_task, busiest->avg_load);
7892 capa_now += local->group_capacity *
7893 min(local->load_per_task, local->avg_load);
7894 capa_now /= SCHED_CAPACITY_SCALE;
7896 /* Amount of load we'd subtract */
7897 if (busiest->avg_load > scaled_busy_load_per_task) {
7898 capa_move += busiest->group_capacity *
7899 min(busiest->load_per_task,
7900 busiest->avg_load - scaled_busy_load_per_task);
7903 /* Amount of load we'd add */
7904 if (busiest->avg_load * busiest->group_capacity <
7905 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7906 tmp = (busiest->avg_load * busiest->group_capacity) /
7907 local->group_capacity;
7909 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7910 local->group_capacity;
7912 capa_move += local->group_capacity *
7913 min(local->load_per_task, local->avg_load + tmp);
7914 capa_move /= SCHED_CAPACITY_SCALE;
7916 /* Move if we gain throughput */
7917 if (capa_move > capa_now)
7918 env->imbalance = busiest->load_per_task;
7922 * calculate_imbalance - Calculate the amount of imbalance present within the
7923 * groups of a given sched_domain during load balance.
7924 * @env: load balance environment
7925 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7927 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7929 unsigned long max_pull, load_above_capacity = ~0UL;
7930 struct sg_lb_stats *local, *busiest;
7932 local = &sds->local_stat;
7933 busiest = &sds->busiest_stat;
7935 if (busiest->group_type == group_imbalanced) {
7937 * In the group_imb case we cannot rely on group-wide averages
7938 * to ensure cpu-load equilibrium, look at wider averages. XXX
7940 busiest->load_per_task =
7941 min(busiest->load_per_task, sds->avg_load);
7945 * In the presence of smp nice balancing, certain scenarios can have
7946 * max load less than avg load(as we skip the groups at or below
7947 * its cpu_capacity, while calculating max_load..)
7949 if (busiest->avg_load <= sds->avg_load ||
7950 local->avg_load >= sds->avg_load) {
7951 /* Misfitting tasks should be migrated in any case */
7952 if (busiest->group_type == group_misfit_task) {
7953 env->imbalance = busiest->group_misfit_task;
7958 * Busiest group is overloaded, local is not, use the spare
7959 * cycles to maximize throughput
7961 if (busiest->group_type == group_overloaded &&
7962 local->group_type <= group_misfit_task) {
7963 env->imbalance = busiest->load_per_task;
7968 return fix_small_imbalance(env, sds);
7972 * If there aren't any idle cpus, avoid creating some.
7974 if (busiest->group_type == group_overloaded &&
7975 local->group_type == group_overloaded) {
7976 load_above_capacity = busiest->sum_nr_running *
7978 if (load_above_capacity > busiest->group_capacity)
7979 load_above_capacity -= busiest->group_capacity;
7981 load_above_capacity = ~0UL;
7985 * We're trying to get all the cpus to the average_load, so we don't
7986 * want to push ourselves above the average load, nor do we wish to
7987 * reduce the max loaded cpu below the average load. At the same time,
7988 * we also don't want to reduce the group load below the group capacity
7989 * (so that we can implement power-savings policies etc). Thus we look
7990 * for the minimum possible imbalance.
7992 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7994 /* How much load to actually move to equalise the imbalance */
7995 env->imbalance = min(
7996 max_pull * busiest->group_capacity,
7997 (sds->avg_load - local->avg_load) * local->group_capacity
7998 ) / SCHED_CAPACITY_SCALE;
8000 /* Boost imbalance to allow misfit task to be balanced. */
8001 if (busiest->group_type == group_misfit_task)
8002 env->imbalance = max_t(long, env->imbalance,
8003 busiest->group_misfit_task);
8006 * if *imbalance is less than the average load per runnable task
8007 * there is no guarantee that any tasks will be moved so we'll have
8008 * a think about bumping its value to force at least one task to be
8011 if (env->imbalance < busiest->load_per_task)
8012 return fix_small_imbalance(env, sds);
8015 /******* find_busiest_group() helpers end here *********************/
8018 * find_busiest_group - Returns the busiest group within the sched_domain
8019 * if there is an imbalance. If there isn't an imbalance, and
8020 * the user has opted for power-savings, it returns a group whose
8021 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
8022 * such a group exists.
8024 * Also calculates the amount of weighted load which should be moved
8025 * to restore balance.
8027 * @env: The load balancing environment.
8029 * Return: - The busiest group if imbalance exists.
8030 * - If no imbalance and user has opted for power-savings balance,
8031 * return the least loaded group whose CPUs can be
8032 * put to idle by rebalancing its tasks onto our group.
8034 static struct sched_group *find_busiest_group(struct lb_env *env)
8036 struct sg_lb_stats *local, *busiest;
8037 struct sd_lb_stats sds;
8039 init_sd_lb_stats(&sds);
8042 * Compute the various statistics relavent for load balancing at
8045 update_sd_lb_stats(env, &sds);
8047 if (energy_aware() && !env->dst_rq->rd->overutilized)
8050 local = &sds.local_stat;
8051 busiest = &sds.busiest_stat;
8053 /* ASYM feature bypasses nice load balance check */
8054 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
8055 check_asym_packing(env, &sds))
8058 /* There is no busy sibling group to pull tasks from */
8059 if (!sds.busiest || busiest->sum_nr_running == 0)
8062 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8063 / sds.total_capacity;
8066 * If the busiest group is imbalanced the below checks don't
8067 * work because they assume all things are equal, which typically
8068 * isn't true due to cpus_allowed constraints and the like.
8070 if (busiest->group_type == group_imbalanced)
8073 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
8074 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
8075 busiest->group_no_capacity)
8078 /* Misfitting tasks should be dealt with regardless of the avg load */
8079 if (busiest->group_type == group_misfit_task) {
8084 * If the local group is busier than the selected busiest group
8085 * don't try and pull any tasks.
8087 if (local->avg_load >= busiest->avg_load)
8091 * Don't pull any tasks if this group is already above the domain
8094 if (local->avg_load >= sds.avg_load)
8097 if (env->idle == CPU_IDLE) {
8099 * This cpu is idle. If the busiest group is not overloaded
8100 * and there is no imbalance between this and busiest group
8101 * wrt idle cpus, it is balanced. The imbalance becomes
8102 * significant if the diff is greater than 1 otherwise we
8103 * might end up to just move the imbalance on another group
8105 if ((busiest->group_type != group_overloaded) &&
8106 (local->idle_cpus <= (busiest->idle_cpus + 1)) &&
8107 !group_smaller_cpu_capacity(sds.busiest, sds.local))
8111 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8112 * imbalance_pct to be conservative.
8114 if (100 * busiest->avg_load <=
8115 env->sd->imbalance_pct * local->avg_load)
8120 env->busiest_group_type = busiest->group_type;
8121 /* Looks like there is an imbalance. Compute it */
8122 calculate_imbalance(env, &sds);
8131 * find_busiest_queue - find the busiest runqueue among the cpus in group.
8133 static struct rq *find_busiest_queue(struct lb_env *env,
8134 struct sched_group *group)
8136 struct rq *busiest = NULL, *rq;
8137 unsigned long busiest_load = 0, busiest_capacity = 1;
8140 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
8141 unsigned long capacity, wl;
8145 rt = fbq_classify_rq(rq);
8148 * We classify groups/runqueues into three groups:
8149 * - regular: there are !numa tasks
8150 * - remote: there are numa tasks that run on the 'wrong' node
8151 * - all: there is no distinction
8153 * In order to avoid migrating ideally placed numa tasks,
8154 * ignore those when there's better options.
8156 * If we ignore the actual busiest queue to migrate another
8157 * task, the next balance pass can still reduce the busiest
8158 * queue by moving tasks around inside the node.
8160 * If we cannot move enough load due to this classification
8161 * the next pass will adjust the group classification and
8162 * allow migration of more tasks.
8164 * Both cases only affect the total convergence complexity.
8166 if (rt > env->fbq_type)
8169 capacity = capacity_of(i);
8171 wl = weighted_cpuload(i);
8174 * When comparing with imbalance, use weighted_cpuload()
8175 * which is not scaled with the cpu capacity.
8178 if (rq->nr_running == 1 && wl > env->imbalance &&
8179 !check_cpu_capacity(rq, env->sd) &&
8180 env->busiest_group_type != group_misfit_task)
8184 * For the load comparisons with the other cpu's, consider
8185 * the weighted_cpuload() scaled with the cpu capacity, so
8186 * that the load can be moved away from the cpu that is
8187 * potentially running at a lower capacity.
8189 * Thus we're looking for max(wl_i / capacity_i), crosswise
8190 * multiplication to rid ourselves of the division works out
8191 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8192 * our previous maximum.
8194 if (wl * busiest_capacity > busiest_load * capacity) {
8196 busiest_capacity = capacity;
8205 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8206 * so long as it is large enough.
8208 #define MAX_PINNED_INTERVAL 512
8210 /* Working cpumask for load_balance and load_balance_newidle. */
8211 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
8213 static int need_active_balance(struct lb_env *env)
8215 struct sched_domain *sd = env->sd;
8217 if (env->idle == CPU_NEWLY_IDLE) {
8220 * ASYM_PACKING needs to force migrate tasks from busy but
8221 * higher numbered CPUs in order to pack all tasks in the
8222 * lowest numbered CPUs.
8224 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
8229 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8230 * It's worth migrating the task if the src_cpu's capacity is reduced
8231 * because of other sched_class or IRQs if more capacity stays
8232 * available on dst_cpu.
8234 if ((env->idle != CPU_NOT_IDLE) &&
8235 (env->src_rq->cfs.h_nr_running == 1)) {
8236 if ((check_cpu_capacity(env->src_rq, sd)) &&
8237 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8241 if ((capacity_of(env->src_cpu) < capacity_of(env->dst_cpu)) &&
8242 env->src_rq->cfs.h_nr_running == 1 &&
8243 cpu_overutilized(env->src_cpu) &&
8244 !cpu_overutilized(env->dst_cpu)) {
8248 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8251 static int active_load_balance_cpu_stop(void *data);
8253 static int should_we_balance(struct lb_env *env)
8255 struct sched_group *sg = env->sd->groups;
8256 struct cpumask *sg_cpus, *sg_mask;
8257 int cpu, balance_cpu = -1;
8260 * In the newly idle case, we will allow all the cpu's
8261 * to do the newly idle load balance.
8263 if (env->idle == CPU_NEWLY_IDLE)
8266 sg_cpus = sched_group_cpus(sg);
8267 sg_mask = sched_group_mask(sg);
8268 /* Try to find first idle cpu */
8269 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
8270 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
8277 if (balance_cpu == -1)
8278 balance_cpu = group_balance_cpu(sg);
8281 * First idle cpu or the first cpu(busiest) in this sched group
8282 * is eligible for doing load balancing at this and above domains.
8284 return balance_cpu == env->dst_cpu;
8288 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8289 * tasks if there is an imbalance.
8291 static int load_balance(int this_cpu, struct rq *this_rq,
8292 struct sched_domain *sd, enum cpu_idle_type idle,
8293 int *continue_balancing)
8295 int ld_moved, cur_ld_moved, active_balance = 0;
8296 struct sched_domain *sd_parent = sd->parent;
8297 struct sched_group *group;
8299 unsigned long flags;
8300 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8302 struct lb_env env = {
8304 .dst_cpu = this_cpu,
8306 .dst_grpmask = sched_group_cpus(sd->groups),
8308 .loop_break = sched_nr_migrate_break,
8311 .tasks = LIST_HEAD_INIT(env.tasks),
8315 * For NEWLY_IDLE load_balancing, we don't need to consider
8316 * other cpus in our group
8318 if (idle == CPU_NEWLY_IDLE)
8319 env.dst_grpmask = NULL;
8321 cpumask_copy(cpus, cpu_active_mask);
8323 schedstat_inc(sd, lb_count[idle]);
8326 if (!should_we_balance(&env)) {
8327 *continue_balancing = 0;
8331 group = find_busiest_group(&env);
8333 schedstat_inc(sd, lb_nobusyg[idle]);
8337 busiest = find_busiest_queue(&env, group);
8339 schedstat_inc(sd, lb_nobusyq[idle]);
8343 BUG_ON(busiest == env.dst_rq);
8345 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
8347 env.src_cpu = busiest->cpu;
8348 env.src_rq = busiest;
8351 if (busiest->nr_running > 1) {
8353 * Attempt to move tasks. If find_busiest_group has found
8354 * an imbalance but busiest->nr_running <= 1, the group is
8355 * still unbalanced. ld_moved simply stays zero, so it is
8356 * correctly treated as an imbalance.
8358 env.flags |= LBF_ALL_PINNED;
8359 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8362 raw_spin_lock_irqsave(&busiest->lock, flags);
8365 * cur_ld_moved - load moved in current iteration
8366 * ld_moved - cumulative load moved across iterations
8368 cur_ld_moved = detach_tasks(&env);
8370 * We want to potentially lower env.src_cpu's OPP.
8373 update_capacity_of(env.src_cpu);
8376 * We've detached some tasks from busiest_rq. Every
8377 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8378 * unlock busiest->lock, and we are able to be sure
8379 * that nobody can manipulate the tasks in parallel.
8380 * See task_rq_lock() family for the details.
8383 raw_spin_unlock(&busiest->lock);
8387 ld_moved += cur_ld_moved;
8390 local_irq_restore(flags);
8392 if (env.flags & LBF_NEED_BREAK) {
8393 env.flags &= ~LBF_NEED_BREAK;
8398 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8399 * us and move them to an alternate dst_cpu in our sched_group
8400 * where they can run. The upper limit on how many times we
8401 * iterate on same src_cpu is dependent on number of cpus in our
8404 * This changes load balance semantics a bit on who can move
8405 * load to a given_cpu. In addition to the given_cpu itself
8406 * (or a ilb_cpu acting on its behalf where given_cpu is
8407 * nohz-idle), we now have balance_cpu in a position to move
8408 * load to given_cpu. In rare situations, this may cause
8409 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8410 * _independently_ and at _same_ time to move some load to
8411 * given_cpu) causing exceess load to be moved to given_cpu.
8412 * This however should not happen so much in practice and
8413 * moreover subsequent load balance cycles should correct the
8414 * excess load moved.
8416 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8418 /* Prevent to re-select dst_cpu via env's cpus */
8419 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8421 env.dst_rq = cpu_rq(env.new_dst_cpu);
8422 env.dst_cpu = env.new_dst_cpu;
8423 env.flags &= ~LBF_DST_PINNED;
8425 env.loop_break = sched_nr_migrate_break;
8428 * Go back to "more_balance" rather than "redo" since we
8429 * need to continue with same src_cpu.
8435 * We failed to reach balance because of affinity.
8438 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8440 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8441 *group_imbalance = 1;
8444 /* All tasks on this runqueue were pinned by CPU affinity */
8445 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8446 cpumask_clear_cpu(cpu_of(busiest), cpus);
8447 if (!cpumask_empty(cpus)) {
8449 env.loop_break = sched_nr_migrate_break;
8452 goto out_all_pinned;
8457 schedstat_inc(sd, lb_failed[idle]);
8459 * Increment the failure counter only on periodic balance.
8460 * We do not want newidle balance, which can be very
8461 * frequent, pollute the failure counter causing
8462 * excessive cache_hot migrations and active balances.
8464 if (idle != CPU_NEWLY_IDLE)
8465 if (env.src_grp_nr_running > 1)
8466 sd->nr_balance_failed++;
8468 if (need_active_balance(&env)) {
8469 raw_spin_lock_irqsave(&busiest->lock, flags);
8471 /* don't kick the active_load_balance_cpu_stop,
8472 * if the curr task on busiest cpu can't be
8475 if (!cpumask_test_cpu(this_cpu,
8476 tsk_cpus_allowed(busiest->curr))) {
8477 raw_spin_unlock_irqrestore(&busiest->lock,
8479 env.flags |= LBF_ALL_PINNED;
8480 goto out_one_pinned;
8484 * ->active_balance synchronizes accesses to
8485 * ->active_balance_work. Once set, it's cleared
8486 * only after active load balance is finished.
8488 if (!busiest->active_balance) {
8489 busiest->active_balance = 1;
8490 busiest->push_cpu = this_cpu;
8493 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8495 if (active_balance) {
8496 stop_one_cpu_nowait(cpu_of(busiest),
8497 active_load_balance_cpu_stop, busiest,
8498 &busiest->active_balance_work);
8502 * We've kicked active balancing, reset the failure
8505 sd->nr_balance_failed = sd->cache_nice_tries+1;
8508 sd->nr_balance_failed = 0;
8510 if (likely(!active_balance)) {
8511 /* We were unbalanced, so reset the balancing interval */
8512 sd->balance_interval = sd->min_interval;
8515 * If we've begun active balancing, start to back off. This
8516 * case may not be covered by the all_pinned logic if there
8517 * is only 1 task on the busy runqueue (because we don't call
8520 if (sd->balance_interval < sd->max_interval)
8521 sd->balance_interval *= 2;
8528 * We reach balance although we may have faced some affinity
8529 * constraints. Clear the imbalance flag if it was set.
8532 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8534 if (*group_imbalance)
8535 *group_imbalance = 0;
8540 * We reach balance because all tasks are pinned at this level so
8541 * we can't migrate them. Let the imbalance flag set so parent level
8542 * can try to migrate them.
8544 schedstat_inc(sd, lb_balanced[idle]);
8546 sd->nr_balance_failed = 0;
8549 /* tune up the balancing interval */
8550 if (((env.flags & LBF_ALL_PINNED) &&
8551 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8552 (sd->balance_interval < sd->max_interval))
8553 sd->balance_interval *= 2;
8560 static inline unsigned long
8561 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8563 unsigned long interval = sd->balance_interval;
8566 interval *= sd->busy_factor;
8568 /* scale ms to jiffies */
8569 interval = msecs_to_jiffies(interval);
8570 interval = clamp(interval, 1UL, max_load_balance_interval);
8576 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
8578 unsigned long interval, next;
8580 interval = get_sd_balance_interval(sd, cpu_busy);
8581 next = sd->last_balance + interval;
8583 if (time_after(*next_balance, next))
8584 *next_balance = next;
8588 * idle_balance is called by schedule() if this_cpu is about to become
8589 * idle. Attempts to pull tasks from other CPUs.
8591 static int idle_balance(struct rq *this_rq)
8593 unsigned long next_balance = jiffies + HZ;
8594 int this_cpu = this_rq->cpu;
8595 struct sched_domain *sd;
8596 int pulled_task = 0;
8598 long removed_util=0;
8600 idle_enter_fair(this_rq);
8603 * We must set idle_stamp _before_ calling idle_balance(), such that we
8604 * measure the duration of idle_balance() as idle time.
8606 this_rq->idle_stamp = rq_clock(this_rq);
8608 if (!energy_aware() &&
8609 (this_rq->avg_idle < sysctl_sched_migration_cost ||
8610 !this_rq->rd->overload)) {
8612 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8614 update_next_balance(sd, 0, &next_balance);
8620 raw_spin_unlock(&this_rq->lock);
8623 * If removed_util_avg is !0 we most probably migrated some task away
8624 * from this_cpu. In this case we might be willing to trigger an OPP
8625 * update, but we want to do so if we don't find anybody else to pull
8626 * here (we will trigger an OPP update with the pulled task's enqueue
8629 * Record removed_util before calling update_blocked_averages, and use
8630 * it below (before returning) to see if an OPP update is required.
8632 removed_util = atomic_long_read(&(this_rq->cfs).removed_util_avg);
8633 update_blocked_averages(this_cpu);
8635 for_each_domain(this_cpu, sd) {
8636 int continue_balancing = 1;
8637 u64 t0, domain_cost;
8639 if (!(sd->flags & SD_LOAD_BALANCE))
8642 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8643 update_next_balance(sd, 0, &next_balance);
8647 if (sd->flags & SD_BALANCE_NEWIDLE) {
8648 t0 = sched_clock_cpu(this_cpu);
8650 pulled_task = load_balance(this_cpu, this_rq,
8652 &continue_balancing);
8654 domain_cost = sched_clock_cpu(this_cpu) - t0;
8655 if (domain_cost > sd->max_newidle_lb_cost)
8656 sd->max_newidle_lb_cost = domain_cost;
8658 curr_cost += domain_cost;
8661 update_next_balance(sd, 0, &next_balance);
8664 * Stop searching for tasks to pull if there are
8665 * now runnable tasks on this rq.
8667 if (pulled_task || this_rq->nr_running > 0)
8672 raw_spin_lock(&this_rq->lock);
8674 if (curr_cost > this_rq->max_idle_balance_cost)
8675 this_rq->max_idle_balance_cost = curr_cost;
8678 * While browsing the domains, we released the rq lock, a task could
8679 * have been enqueued in the meantime. Since we're not going idle,
8680 * pretend we pulled a task.
8682 if (this_rq->cfs.h_nr_running && !pulled_task)
8686 /* Move the next balance forward */
8687 if (time_after(this_rq->next_balance, next_balance))
8688 this_rq->next_balance = next_balance;
8690 /* Is there a task of a high priority class? */
8691 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8695 idle_exit_fair(this_rq);
8696 this_rq->idle_stamp = 0;
8697 } else if (removed_util) {
8699 * No task pulled and someone has been migrated away.
8700 * Good case to trigger an OPP update.
8702 update_capacity_of(this_cpu);
8709 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8710 * running tasks off the busiest CPU onto idle CPUs. It requires at
8711 * least 1 task to be running on each physical CPU where possible, and
8712 * avoids physical / logical imbalances.
8714 static int active_load_balance_cpu_stop(void *data)
8716 struct rq *busiest_rq = data;
8717 int busiest_cpu = cpu_of(busiest_rq);
8718 int target_cpu = busiest_rq->push_cpu;
8719 struct rq *target_rq = cpu_rq(target_cpu);
8720 struct sched_domain *sd;
8721 struct task_struct *p = NULL;
8723 raw_spin_lock_irq(&busiest_rq->lock);
8725 /* make sure the requested cpu hasn't gone down in the meantime */
8726 if (unlikely(busiest_cpu != smp_processor_id() ||
8727 !busiest_rq->active_balance))
8730 /* Is there any task to move? */
8731 if (busiest_rq->nr_running <= 1)
8735 * This condition is "impossible", if it occurs
8736 * we need to fix it. Originally reported by
8737 * Bjorn Helgaas on a 128-cpu setup.
8739 BUG_ON(busiest_rq == target_rq);
8741 /* Search for an sd spanning us and the target CPU. */
8743 for_each_domain(target_cpu, sd) {
8744 if ((sd->flags & SD_LOAD_BALANCE) &&
8745 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8750 struct lb_env env = {
8752 .dst_cpu = target_cpu,
8753 .dst_rq = target_rq,
8754 .src_cpu = busiest_rq->cpu,
8755 .src_rq = busiest_rq,
8759 schedstat_inc(sd, alb_count);
8761 p = detach_one_task(&env);
8763 schedstat_inc(sd, alb_pushed);
8765 * We want to potentially lower env.src_cpu's OPP.
8767 update_capacity_of(env.src_cpu);
8770 schedstat_inc(sd, alb_failed);
8774 busiest_rq->active_balance = 0;
8775 raw_spin_unlock(&busiest_rq->lock);
8778 attach_one_task(target_rq, p);
8785 static inline int on_null_domain(struct rq *rq)
8787 return unlikely(!rcu_dereference_sched(rq->sd));
8790 #ifdef CONFIG_NO_HZ_COMMON
8792 * idle load balancing details
8793 * - When one of the busy CPUs notice that there may be an idle rebalancing
8794 * needed, they will kick the idle load balancer, which then does idle
8795 * load balancing for all the idle CPUs.
8798 cpumask_var_t idle_cpus_mask;
8800 unsigned long next_balance; /* in jiffy units */
8801 } nohz ____cacheline_aligned;
8803 static inline int find_new_ilb(void)
8805 int ilb = cpumask_first(nohz.idle_cpus_mask);
8807 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8814 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8815 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8816 * CPU (if there is one).
8818 static void nohz_balancer_kick(void)
8822 nohz.next_balance++;
8824 ilb_cpu = find_new_ilb();
8826 if (ilb_cpu >= nr_cpu_ids)
8829 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8832 * Use smp_send_reschedule() instead of resched_cpu().
8833 * This way we generate a sched IPI on the target cpu which
8834 * is idle. And the softirq performing nohz idle load balance
8835 * will be run before returning from the IPI.
8837 smp_send_reschedule(ilb_cpu);
8841 static inline void nohz_balance_exit_idle(int cpu)
8843 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8845 * Completely isolated CPUs don't ever set, so we must test.
8847 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8848 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8849 atomic_dec(&nohz.nr_cpus);
8851 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8855 static inline void set_cpu_sd_state_busy(void)
8857 struct sched_domain *sd;
8858 int cpu = smp_processor_id();
8861 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8863 if (!sd || !sd->nohz_idle)
8867 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
8872 void set_cpu_sd_state_idle(void)
8874 struct sched_domain *sd;
8875 int cpu = smp_processor_id();
8878 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8880 if (!sd || sd->nohz_idle)
8884 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
8890 * This routine will record that the cpu is going idle with tick stopped.
8891 * This info will be used in performing idle load balancing in the future.
8893 void nohz_balance_enter_idle(int cpu)
8896 * If this cpu is going down, then nothing needs to be done.
8898 if (!cpu_active(cpu))
8901 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8905 * If we're a completely isolated CPU, we don't play.
8907 if (on_null_domain(cpu_rq(cpu)))
8910 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8911 atomic_inc(&nohz.nr_cpus);
8912 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8915 static int sched_ilb_notifier(struct notifier_block *nfb,
8916 unsigned long action, void *hcpu)
8918 switch (action & ~CPU_TASKS_FROZEN) {
8920 nohz_balance_exit_idle(smp_processor_id());
8928 static DEFINE_SPINLOCK(balancing);
8931 * Scale the max load_balance interval with the number of CPUs in the system.
8932 * This trades load-balance latency on larger machines for less cross talk.
8934 void update_max_interval(void)
8936 max_load_balance_interval = HZ*num_online_cpus()/10;
8940 * It checks each scheduling domain to see if it is due to be balanced,
8941 * and initiates a balancing operation if so.
8943 * Balancing parameters are set up in init_sched_domains.
8945 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8947 int continue_balancing = 1;
8949 unsigned long interval;
8950 struct sched_domain *sd;
8951 /* Earliest time when we have to do rebalance again */
8952 unsigned long next_balance = jiffies + 60*HZ;
8953 int update_next_balance = 0;
8954 int need_serialize, need_decay = 0;
8957 update_blocked_averages(cpu);
8960 for_each_domain(cpu, sd) {
8962 * Decay the newidle max times here because this is a regular
8963 * visit to all the domains. Decay ~1% per second.
8965 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8966 sd->max_newidle_lb_cost =
8967 (sd->max_newidle_lb_cost * 253) / 256;
8968 sd->next_decay_max_lb_cost = jiffies + HZ;
8971 max_cost += sd->max_newidle_lb_cost;
8973 if (!(sd->flags & SD_LOAD_BALANCE))
8977 * Stop the load balance at this level. There is another
8978 * CPU in our sched group which is doing load balancing more
8981 if (!continue_balancing) {
8987 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8989 need_serialize = sd->flags & SD_SERIALIZE;
8990 if (need_serialize) {
8991 if (!spin_trylock(&balancing))
8995 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8996 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8998 * The LBF_DST_PINNED logic could have changed
8999 * env->dst_cpu, so we can't know our idle
9000 * state even if we migrated tasks. Update it.
9002 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9004 sd->last_balance = jiffies;
9005 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9008 spin_unlock(&balancing);
9010 if (time_after(next_balance, sd->last_balance + interval)) {
9011 next_balance = sd->last_balance + interval;
9012 update_next_balance = 1;
9017 * Ensure the rq-wide value also decays but keep it at a
9018 * reasonable floor to avoid funnies with rq->avg_idle.
9020 rq->max_idle_balance_cost =
9021 max((u64)sysctl_sched_migration_cost, max_cost);
9026 * next_balance will be updated only when there is a need.
9027 * When the cpu is attached to null domain for ex, it will not be
9030 if (likely(update_next_balance)) {
9031 rq->next_balance = next_balance;
9033 #ifdef CONFIG_NO_HZ_COMMON
9035 * If this CPU has been elected to perform the nohz idle
9036 * balance. Other idle CPUs have already rebalanced with
9037 * nohz_idle_balance() and nohz.next_balance has been
9038 * updated accordingly. This CPU is now running the idle load
9039 * balance for itself and we need to update the
9040 * nohz.next_balance accordingly.
9042 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9043 nohz.next_balance = rq->next_balance;
9048 #ifdef CONFIG_NO_HZ_COMMON
9050 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9051 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9053 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9055 int this_cpu = this_rq->cpu;
9058 /* Earliest time when we have to do rebalance again */
9059 unsigned long next_balance = jiffies + 60*HZ;
9060 int update_next_balance = 0;
9062 if (idle != CPU_IDLE ||
9063 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
9066 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9067 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9071 * If this cpu gets work to do, stop the load balancing
9072 * work being done for other cpus. Next load
9073 * balancing owner will pick it up.
9078 rq = cpu_rq(balance_cpu);
9081 * If time for next balance is due,
9084 if (time_after_eq(jiffies, rq->next_balance)) {
9085 raw_spin_lock_irq(&rq->lock);
9086 update_rq_clock(rq);
9087 update_idle_cpu_load(rq);
9088 raw_spin_unlock_irq(&rq->lock);
9089 rebalance_domains(rq, CPU_IDLE);
9092 if (time_after(next_balance, rq->next_balance)) {
9093 next_balance = rq->next_balance;
9094 update_next_balance = 1;
9099 * next_balance will be updated only when there is a need.
9100 * When the CPU is attached to null domain for ex, it will not be
9103 if (likely(update_next_balance))
9104 nohz.next_balance = next_balance;
9106 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
9110 * Current heuristic for kicking the idle load balancer in the presence
9111 * of an idle cpu in the system.
9112 * - This rq has more than one task.
9113 * - This rq has at least one CFS task and the capacity of the CPU is
9114 * significantly reduced because of RT tasks or IRQs.
9115 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
9116 * multiple busy cpu.
9117 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
9118 * domain span are idle.
9120 static inline bool nohz_kick_needed(struct rq *rq)
9122 unsigned long now = jiffies;
9123 struct sched_domain *sd;
9124 struct sched_group_capacity *sgc;
9125 int nr_busy, cpu = rq->cpu;
9128 if (unlikely(rq->idle_balance))
9132 * We may be recently in ticked or tickless idle mode. At the first
9133 * busy tick after returning from idle, we will update the busy stats.
9135 set_cpu_sd_state_busy();
9136 nohz_balance_exit_idle(cpu);
9139 * None are in tickless mode and hence no need for NOHZ idle load
9142 if (likely(!atomic_read(&nohz.nr_cpus)))
9145 if (time_before(now, nohz.next_balance))
9148 if (rq->nr_running >= 2 &&
9149 (!energy_aware() || cpu_overutilized(cpu)))
9153 sd = rcu_dereference(per_cpu(sd_busy, cpu));
9154 if (sd && !energy_aware()) {
9155 sgc = sd->groups->sgc;
9156 nr_busy = atomic_read(&sgc->nr_busy_cpus);
9165 sd = rcu_dereference(rq->sd);
9167 if ((rq->cfs.h_nr_running >= 1) &&
9168 check_cpu_capacity(rq, sd)) {
9174 sd = rcu_dereference(per_cpu(sd_asym, cpu));
9175 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
9176 sched_domain_span(sd)) < cpu)) {
9186 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
9190 * run_rebalance_domains is triggered when needed from the scheduler tick.
9191 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9193 static void run_rebalance_domains(struct softirq_action *h)
9195 struct rq *this_rq = this_rq();
9196 enum cpu_idle_type idle = this_rq->idle_balance ?
9197 CPU_IDLE : CPU_NOT_IDLE;
9200 * If this cpu has a pending nohz_balance_kick, then do the
9201 * balancing on behalf of the other idle cpus whose ticks are
9202 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9203 * give the idle cpus a chance to load balance. Else we may
9204 * load balance only within the local sched_domain hierarchy
9205 * and abort nohz_idle_balance altogether if we pull some load.
9207 nohz_idle_balance(this_rq, idle);
9208 rebalance_domains(this_rq, idle);
9212 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9214 void trigger_load_balance(struct rq *rq)
9216 /* Don't need to rebalance while attached to NULL domain */
9217 if (unlikely(on_null_domain(rq)))
9220 if (time_after_eq(jiffies, rq->next_balance))
9221 raise_softirq(SCHED_SOFTIRQ);
9222 #ifdef CONFIG_NO_HZ_COMMON
9223 if (nohz_kick_needed(rq))
9224 nohz_balancer_kick();
9228 static void rq_online_fair(struct rq *rq)
9232 update_runtime_enabled(rq);
9235 static void rq_offline_fair(struct rq *rq)
9239 /* Ensure any throttled groups are reachable by pick_next_task */
9240 unthrottle_offline_cfs_rqs(rq);
9243 #endif /* CONFIG_SMP */
9246 * scheduler tick hitting a task of our scheduling class:
9248 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9250 struct cfs_rq *cfs_rq;
9251 struct sched_entity *se = &curr->se;
9253 for_each_sched_entity(se) {
9254 cfs_rq = cfs_rq_of(se);
9255 entity_tick(cfs_rq, se, queued);
9258 if (static_branch_unlikely(&sched_numa_balancing))
9259 task_tick_numa(rq, curr);
9262 if (!rq->rd->overutilized && cpu_overutilized(task_cpu(curr))) {
9263 rq->rd->overutilized = true;
9264 trace_sched_overutilized(true);
9267 rq->misfit_task = !task_fits_max(curr, rq->cpu);
9273 * called on fork with the child task as argument from the parent's context
9274 * - child not yet on the tasklist
9275 * - preemption disabled
9277 static void task_fork_fair(struct task_struct *p)
9279 struct cfs_rq *cfs_rq;
9280 struct sched_entity *se = &p->se, *curr;
9281 int this_cpu = smp_processor_id();
9282 struct rq *rq = this_rq();
9283 unsigned long flags;
9285 raw_spin_lock_irqsave(&rq->lock, flags);
9287 update_rq_clock(rq);
9289 cfs_rq = task_cfs_rq(current);
9290 curr = cfs_rq->curr;
9293 * Not only the cpu but also the task_group of the parent might have
9294 * been changed after parent->se.parent,cfs_rq were copied to
9295 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
9296 * of child point to valid ones.
9299 __set_task_cpu(p, this_cpu);
9302 update_curr(cfs_rq);
9305 se->vruntime = curr->vruntime;
9306 place_entity(cfs_rq, se, 1);
9308 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9310 * Upon rescheduling, sched_class::put_prev_task() will place
9311 * 'current' within the tree based on its new key value.
9313 swap(curr->vruntime, se->vruntime);
9317 se->vruntime -= cfs_rq->min_vruntime;
9319 raw_spin_unlock_irqrestore(&rq->lock, flags);
9323 * Priority of the task has changed. Check to see if we preempt
9327 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9329 if (!task_on_rq_queued(p))
9333 * Reschedule if we are currently running on this runqueue and
9334 * our priority decreased, or if we are not currently running on
9335 * this runqueue and our priority is higher than the current's
9337 if (rq->curr == p) {
9338 if (p->prio > oldprio)
9341 check_preempt_curr(rq, p, 0);
9344 static inline bool vruntime_normalized(struct task_struct *p)
9346 struct sched_entity *se = &p->se;
9349 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9350 * the dequeue_entity(.flags=0) will already have normalized the
9357 * When !on_rq, vruntime of the task has usually NOT been normalized.
9358 * But there are some cases where it has already been normalized:
9360 * - A forked child which is waiting for being woken up by
9361 * wake_up_new_task().
9362 * - A task which has been woken up by try_to_wake_up() and
9363 * waiting for actually being woken up by sched_ttwu_pending().
9365 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
9371 static void detach_entity_cfs_rq(struct sched_entity *se)
9373 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9375 /* Catch up with the cfs_rq and remove our load when we leave */
9376 update_load_avg(se, 0);
9377 detach_entity_load_avg(cfs_rq, se);
9378 update_tg_load_avg(cfs_rq, false);
9381 static void attach_entity_cfs_rq(struct sched_entity *se)
9383 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9385 #ifdef CONFIG_FAIR_GROUP_SCHED
9387 * Since the real-depth could have been changed (only FAIR
9388 * class maintain depth value), reset depth properly.
9390 se->depth = se->parent ? se->parent->depth + 1 : 0;
9393 /* Synchronize entity with its cfs_rq */
9394 update_load_avg(se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
9395 attach_entity_load_avg(cfs_rq, se);
9396 update_tg_load_avg(cfs_rq, false);
9399 static void detach_task_cfs_rq(struct task_struct *p)
9401 struct sched_entity *se = &p->se;
9402 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9404 if (!vruntime_normalized(p)) {
9406 * Fix up our vruntime so that the current sleep doesn't
9407 * cause 'unlimited' sleep bonus.
9409 place_entity(cfs_rq, se, 0);
9410 se->vruntime -= cfs_rq->min_vruntime;
9413 detach_entity_cfs_rq(se);
9416 static void attach_task_cfs_rq(struct task_struct *p)
9418 struct sched_entity *se = &p->se;
9419 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9421 attach_entity_cfs_rq(se);
9423 if (!vruntime_normalized(p))
9424 se->vruntime += cfs_rq->min_vruntime;
9427 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9429 detach_task_cfs_rq(p);
9432 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9434 attach_task_cfs_rq(p);
9436 if (task_on_rq_queued(p)) {
9438 * We were most likely switched from sched_rt, so
9439 * kick off the schedule if running, otherwise just see
9440 * if we can still preempt the current task.
9445 check_preempt_curr(rq, p, 0);
9449 /* Account for a task changing its policy or group.
9451 * This routine is mostly called to set cfs_rq->curr field when a task
9452 * migrates between groups/classes.
9454 static void set_curr_task_fair(struct rq *rq)
9456 struct sched_entity *se = &rq->curr->se;
9458 for_each_sched_entity(se) {
9459 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9461 set_next_entity(cfs_rq, se);
9462 /* ensure bandwidth has been allocated on our new cfs_rq */
9463 account_cfs_rq_runtime(cfs_rq, 0);
9467 void init_cfs_rq(struct cfs_rq *cfs_rq)
9469 cfs_rq->tasks_timeline = RB_ROOT;
9470 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9471 #ifndef CONFIG_64BIT
9472 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9475 atomic_long_set(&cfs_rq->removed_load_avg, 0);
9476 atomic_long_set(&cfs_rq->removed_util_avg, 0);
9480 #ifdef CONFIG_FAIR_GROUP_SCHED
9481 static void task_move_group_fair(struct task_struct *p)
9483 detach_task_cfs_rq(p);
9484 set_task_rq(p, task_cpu(p));
9487 /* Tell se's cfs_rq has been changed -- migrated */
9488 p->se.avg.last_update_time = 0;
9490 attach_task_cfs_rq(p);
9493 void free_fair_sched_group(struct task_group *tg)
9497 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9499 for_each_possible_cpu(i) {
9501 kfree(tg->cfs_rq[i]);
9504 remove_entity_load_avg(tg->se[i]);
9513 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9515 struct sched_entity *se;
9516 struct cfs_rq *cfs_rq;
9520 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9523 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9527 tg->shares = NICE_0_LOAD;
9529 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9531 for_each_possible_cpu(i) {
9534 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9535 GFP_KERNEL, cpu_to_node(i));
9539 se = kzalloc_node(sizeof(struct sched_entity),
9540 GFP_KERNEL, cpu_to_node(i));
9544 init_cfs_rq(cfs_rq);
9545 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9546 init_entity_runnable_average(se);
9548 raw_spin_lock_irq(&rq->lock);
9549 post_init_entity_util_avg(se);
9550 raw_spin_unlock_irq(&rq->lock);
9561 void unregister_fair_sched_group(struct task_group *tg, int cpu)
9563 struct rq *rq = cpu_rq(cpu);
9564 unsigned long flags;
9567 * Only empty task groups can be destroyed; so we can speculatively
9568 * check on_list without danger of it being re-added.
9570 if (!tg->cfs_rq[cpu]->on_list)
9573 raw_spin_lock_irqsave(&rq->lock, flags);
9574 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9575 raw_spin_unlock_irqrestore(&rq->lock, flags);
9578 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9579 struct sched_entity *se, int cpu,
9580 struct sched_entity *parent)
9582 struct rq *rq = cpu_rq(cpu);
9586 init_cfs_rq_runtime(cfs_rq);
9588 tg->cfs_rq[cpu] = cfs_rq;
9591 /* se could be NULL for root_task_group */
9596 se->cfs_rq = &rq->cfs;
9599 se->cfs_rq = parent->my_q;
9600 se->depth = parent->depth + 1;
9604 /* guarantee group entities always have weight */
9605 update_load_set(&se->load, NICE_0_LOAD);
9606 se->parent = parent;
9609 static DEFINE_MUTEX(shares_mutex);
9611 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9614 unsigned long flags;
9617 * We can't change the weight of the root cgroup.
9622 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9624 mutex_lock(&shares_mutex);
9625 if (tg->shares == shares)
9628 tg->shares = shares;
9629 for_each_possible_cpu(i) {
9630 struct rq *rq = cpu_rq(i);
9631 struct sched_entity *se;
9634 /* Propagate contribution to hierarchy */
9635 raw_spin_lock_irqsave(&rq->lock, flags);
9637 /* Possible calls to update_curr() need rq clock */
9638 update_rq_clock(rq);
9639 for_each_sched_entity(se)
9640 update_cfs_shares(group_cfs_rq(se));
9641 raw_spin_unlock_irqrestore(&rq->lock, flags);
9645 mutex_unlock(&shares_mutex);
9648 #else /* CONFIG_FAIR_GROUP_SCHED */
9650 void free_fair_sched_group(struct task_group *tg) { }
9652 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9657 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
9659 #endif /* CONFIG_FAIR_GROUP_SCHED */
9662 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9664 struct sched_entity *se = &task->se;
9665 unsigned int rr_interval = 0;
9668 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9671 if (rq->cfs.load.weight)
9672 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9678 * All the scheduling class methods:
9680 const struct sched_class fair_sched_class = {
9681 .next = &idle_sched_class,
9682 .enqueue_task = enqueue_task_fair,
9683 .dequeue_task = dequeue_task_fair,
9684 .yield_task = yield_task_fair,
9685 .yield_to_task = yield_to_task_fair,
9687 .check_preempt_curr = check_preempt_wakeup,
9689 .pick_next_task = pick_next_task_fair,
9690 .put_prev_task = put_prev_task_fair,
9693 .select_task_rq = select_task_rq_fair,
9694 .migrate_task_rq = migrate_task_rq_fair,
9696 .rq_online = rq_online_fair,
9697 .rq_offline = rq_offline_fair,
9699 .task_waking = task_waking_fair,
9700 .task_dead = task_dead_fair,
9701 .set_cpus_allowed = set_cpus_allowed_common,
9704 .set_curr_task = set_curr_task_fair,
9705 .task_tick = task_tick_fair,
9706 .task_fork = task_fork_fair,
9708 .prio_changed = prio_changed_fair,
9709 .switched_from = switched_from_fair,
9710 .switched_to = switched_to_fair,
9712 .get_rr_interval = get_rr_interval_fair,
9714 .update_curr = update_curr_fair,
9716 #ifdef CONFIG_FAIR_GROUP_SCHED
9717 .task_move_group = task_move_group_fair,
9721 #ifdef CONFIG_SCHED_DEBUG
9722 void print_cfs_stats(struct seq_file *m, int cpu)
9724 struct cfs_rq *cfs_rq;
9727 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
9728 print_cfs_rq(m, cpu, cfs_rq);
9732 #ifdef CONFIG_NUMA_BALANCING
9733 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9736 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9738 for_each_online_node(node) {
9739 if (p->numa_faults) {
9740 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9741 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9743 if (p->numa_group) {
9744 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9745 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9747 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9750 #endif /* CONFIG_NUMA_BALANCING */
9751 #endif /* CONFIG_SCHED_DEBUG */
9753 __init void init_sched_fair_class(void)
9756 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9758 #ifdef CONFIG_NO_HZ_COMMON
9759 nohz.next_balance = jiffies;
9760 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
9761 cpu_notifier(sched_ilb_notifier, 0);