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) {
308 struct rq *rq = rq_of(cfs_rq);
309 int cpu = cpu_of(rq);
311 * Ensure we either appear before our parent (if already
312 * enqueued) or force our parent to appear after us when it is
313 * enqueued. The fact that we always enqueue bottom-up
314 * reduces this to two cases and a special case for the root
315 * cfs_rq. Furthermore, it also means that we will always reset
316 * tmp_alone_branch either when the branch is connected
317 * to a tree or when we reach the beg of the tree
319 if (cfs_rq->tg->parent &&
320 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
322 * If parent is already on the list, we add the child
323 * just before. Thanks to circular linked property of
324 * the list, this means to put the child at the tail
325 * of the list that starts by parent.
327 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
328 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
330 * The branch is now connected to its tree so we can
331 * reset tmp_alone_branch to the beginning of the
334 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
335 } else if (!cfs_rq->tg->parent) {
337 * cfs rq without parent should be put
338 * at the tail of the list.
340 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
341 &rq->leaf_cfs_rq_list);
343 * We have reach the beg of a tree so we can reset
344 * tmp_alone_branch to the beginning of the list.
346 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
349 * The parent has not already been added so we want to
350 * make sure that it will be put after us.
351 * tmp_alone_branch points to the beg of the branch
352 * where we will add parent.
354 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
355 rq->tmp_alone_branch);
357 * update tmp_alone_branch to points to the new beg
360 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
367 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
369 if (cfs_rq->on_list) {
370 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
375 /* Iterate thr' all leaf cfs_rq's on a runqueue */
376 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
377 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
379 /* Do the two (enqueued) entities belong to the same group ? */
380 static inline struct cfs_rq *
381 is_same_group(struct sched_entity *se, struct sched_entity *pse)
383 if (se->cfs_rq == pse->cfs_rq)
389 static inline struct sched_entity *parent_entity(struct sched_entity *se)
395 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
397 int se_depth, pse_depth;
400 * preemption test can be made between sibling entities who are in the
401 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
402 * both tasks until we find their ancestors who are siblings of common
406 /* First walk up until both entities are at same depth */
407 se_depth = (*se)->depth;
408 pse_depth = (*pse)->depth;
410 while (se_depth > pse_depth) {
412 *se = parent_entity(*se);
415 while (pse_depth > se_depth) {
417 *pse = parent_entity(*pse);
420 while (!is_same_group(*se, *pse)) {
421 *se = parent_entity(*se);
422 *pse = parent_entity(*pse);
426 #else /* !CONFIG_FAIR_GROUP_SCHED */
428 static inline struct task_struct *task_of(struct sched_entity *se)
430 return container_of(se, struct task_struct, se);
433 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
435 return container_of(cfs_rq, struct rq, cfs);
438 #define entity_is_task(se) 1
440 #define for_each_sched_entity(se) \
441 for (; se; se = NULL)
443 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
445 return &task_rq(p)->cfs;
448 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
450 struct task_struct *p = task_of(se);
451 struct rq *rq = task_rq(p);
456 /* runqueue "owned" by this group */
457 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
462 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
466 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
470 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
471 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
473 static inline struct sched_entity *parent_entity(struct sched_entity *se)
479 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
483 #endif /* CONFIG_FAIR_GROUP_SCHED */
485 static __always_inline
486 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
488 /**************************************************************
489 * Scheduling class tree data structure manipulation methods:
492 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
494 s64 delta = (s64)(vruntime - max_vruntime);
496 max_vruntime = vruntime;
501 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
503 s64 delta = (s64)(vruntime - min_vruntime);
505 min_vruntime = vruntime;
510 static inline int entity_before(struct sched_entity *a,
511 struct sched_entity *b)
513 return (s64)(a->vruntime - b->vruntime) < 0;
516 static void update_min_vruntime(struct cfs_rq *cfs_rq)
518 u64 vruntime = cfs_rq->min_vruntime;
521 vruntime = cfs_rq->curr->vruntime;
523 if (cfs_rq->rb_leftmost) {
524 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
529 vruntime = se->vruntime;
531 vruntime = min_vruntime(vruntime, se->vruntime);
534 /* ensure we never gain time by being placed backwards. */
535 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
538 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
543 * Enqueue an entity into the rb-tree:
545 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
547 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
548 struct rb_node *parent = NULL;
549 struct sched_entity *entry;
553 * Find the right place in the rbtree:
557 entry = rb_entry(parent, struct sched_entity, run_node);
559 * We dont care about collisions. Nodes with
560 * the same key stay together.
562 if (entity_before(se, entry)) {
563 link = &parent->rb_left;
565 link = &parent->rb_right;
571 * Maintain a cache of leftmost tree entries (it is frequently
575 cfs_rq->rb_leftmost = &se->run_node;
577 rb_link_node(&se->run_node, parent, link);
578 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
581 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
583 if (cfs_rq->rb_leftmost == &se->run_node) {
584 struct rb_node *next_node;
586 next_node = rb_next(&se->run_node);
587 cfs_rq->rb_leftmost = next_node;
590 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
593 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
595 struct rb_node *left = cfs_rq->rb_leftmost;
600 return rb_entry(left, struct sched_entity, run_node);
603 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
605 struct rb_node *next = rb_next(&se->run_node);
610 return rb_entry(next, struct sched_entity, run_node);
613 #ifdef CONFIG_SCHED_DEBUG
614 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
616 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
621 return rb_entry(last, struct sched_entity, run_node);
624 /**************************************************************
625 * Scheduling class statistics methods:
628 int sched_proc_update_handler(struct ctl_table *table, int write,
629 void __user *buffer, size_t *lenp,
632 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
633 unsigned int factor = get_update_sysctl_factor();
638 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
639 sysctl_sched_min_granularity);
641 #define WRT_SYSCTL(name) \
642 (normalized_sysctl_##name = sysctl_##name / (factor))
643 WRT_SYSCTL(sched_min_granularity);
644 WRT_SYSCTL(sched_latency);
645 WRT_SYSCTL(sched_wakeup_granularity);
655 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
657 if (unlikely(se->load.weight != NICE_0_LOAD))
658 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
664 * The idea is to set a period in which each task runs once.
666 * When there are too many tasks (sched_nr_latency) we have to stretch
667 * this period because otherwise the slices get too small.
669 * p = (nr <= nl) ? l : l*nr/nl
671 static u64 __sched_period(unsigned long nr_running)
673 if (unlikely(nr_running > sched_nr_latency))
674 return nr_running * sysctl_sched_min_granularity;
676 return sysctl_sched_latency;
680 * We calculate the wall-time slice from the period by taking a part
681 * proportional to the weight.
685 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
687 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
689 for_each_sched_entity(se) {
690 struct load_weight *load;
691 struct load_weight lw;
693 cfs_rq = cfs_rq_of(se);
694 load = &cfs_rq->load;
696 if (unlikely(!se->on_rq)) {
699 update_load_add(&lw, se->load.weight);
702 slice = __calc_delta(slice, se->load.weight, load);
708 * We calculate the vruntime slice of a to-be-inserted task.
712 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
714 return calc_delta_fair(sched_slice(cfs_rq, se), se);
718 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
719 static unsigned long task_h_load(struct task_struct *p);
722 * We choose a half-life close to 1 scheduling period.
723 * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are
724 * dependent on this value.
726 #define LOAD_AVG_PERIOD 32
727 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
728 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */
730 /* Give new sched_entity start runnable values to heavy its load in infant time */
731 void init_entity_runnable_average(struct sched_entity *se)
733 struct sched_avg *sa = &se->avg;
735 sa->last_update_time = 0;
737 * sched_avg's period_contrib should be strictly less then 1024, so
738 * we give it 1023 to make sure it is almost a period (1024us), and
739 * will definitely be update (after enqueue).
741 sa->period_contrib = 1023;
742 sa->load_avg = scale_load_down(se->load.weight);
743 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
745 * In previous Android versions, we used to have:
746 * sa->util_avg = sched_freq() ?
747 * sysctl_sched_initial_task_util :
748 * scale_load_down(SCHED_LOAD_SCALE);
749 * sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
750 * However, that functionality has been moved to enqueue.
751 * It is unclear if we should restore this in enqueue.
754 * At this point, util_avg won't be used in select_task_rq_fair anyway
758 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
761 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
762 static void attach_entity_cfs_rq(struct sched_entity *se);
765 * With new tasks being created, their initial util_avgs are extrapolated
766 * based on the cfs_rq's current util_avg:
768 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
770 * However, in many cases, the above util_avg does not give a desired
771 * value. Moreover, the sum of the util_avgs may be divergent, such
772 * as when the series is a harmonic series.
774 * To solve this problem, we also cap the util_avg of successive tasks to
775 * only 1/2 of the left utilization budget:
777 * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
779 * where n denotes the nth task.
781 * For example, a simplest series from the beginning would be like:
783 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
784 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
786 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
787 * if util_avg > util_avg_cap.
789 void post_init_entity_util_avg(struct sched_entity *se)
791 struct cfs_rq *cfs_rq = cfs_rq_of(se);
792 struct sched_avg *sa = &se->avg;
793 long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
796 if (cfs_rq->avg.util_avg != 0) {
797 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
798 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
800 if (sa->util_avg > cap)
806 * If we wish to restore tuning via setting initial util,
807 * this is where we should do it.
809 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
812 if (entity_is_task(se)) {
813 struct task_struct *p = task_of(se);
814 if (p->sched_class != &fair_sched_class) {
816 * For !fair tasks do:
818 update_cfs_rq_load_avg(now, cfs_rq, false);
819 attach_entity_load_avg(cfs_rq, se);
820 switched_from_fair(rq, p);
822 * such that the next switched_to_fair() has the
825 se->avg.last_update_time = cfs_rq_clock_task(cfs_rq);
830 attach_entity_cfs_rq(se);
833 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq);
834 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq);
836 void init_entity_runnable_average(struct sched_entity *se)
839 void post_init_entity_util_avg(struct sched_entity *se)
842 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
845 #endif /* CONFIG_SMP */
848 * Update the current task's runtime statistics.
850 static void update_curr(struct cfs_rq *cfs_rq)
852 struct sched_entity *curr = cfs_rq->curr;
853 u64 now = rq_clock_task(rq_of(cfs_rq));
859 delta_exec = now - curr->exec_start;
860 if (unlikely((s64)delta_exec <= 0))
863 curr->exec_start = now;
865 schedstat_set(curr->statistics.exec_max,
866 max(delta_exec, curr->statistics.exec_max));
868 curr->sum_exec_runtime += delta_exec;
869 schedstat_add(cfs_rq, exec_clock, delta_exec);
871 curr->vruntime += calc_delta_fair(delta_exec, curr);
872 update_min_vruntime(cfs_rq);
874 if (entity_is_task(curr)) {
875 struct task_struct *curtask = task_of(curr);
877 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
878 cpuacct_charge(curtask, delta_exec);
879 account_group_exec_runtime(curtask, delta_exec);
882 account_cfs_rq_runtime(cfs_rq, delta_exec);
885 static void update_curr_fair(struct rq *rq)
887 update_curr(cfs_rq_of(&rq->curr->se));
891 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
893 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
897 * Task is being enqueued - update stats:
899 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
902 * Are we enqueueing a waiting task? (for current tasks
903 * a dequeue/enqueue event is a NOP)
905 if (se != cfs_rq->curr)
906 update_stats_wait_start(cfs_rq, se);
910 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
912 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
913 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
914 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
915 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
916 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
917 #ifdef CONFIG_SCHEDSTATS
918 if (entity_is_task(se)) {
919 trace_sched_stat_wait(task_of(se),
920 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
923 schedstat_set(se->statistics.wait_start, 0);
927 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
930 * Mark the end of the wait period if dequeueing a
933 if (se != cfs_rq->curr)
934 update_stats_wait_end(cfs_rq, se);
938 * We are picking a new current task - update its stats:
941 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
944 * We are starting a new run period:
946 se->exec_start = rq_clock_task(rq_of(cfs_rq));
949 /**************************************************
950 * Scheduling class queueing methods:
953 #ifdef CONFIG_NUMA_BALANCING
955 * Approximate time to scan a full NUMA task in ms. The task scan period is
956 * calculated based on the tasks virtual memory size and
957 * numa_balancing_scan_size.
959 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
960 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
962 /* Portion of address space to scan in MB */
963 unsigned int sysctl_numa_balancing_scan_size = 256;
965 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
966 unsigned int sysctl_numa_balancing_scan_delay = 1000;
968 static unsigned int task_nr_scan_windows(struct task_struct *p)
970 unsigned long rss = 0;
971 unsigned long nr_scan_pages;
974 * Calculations based on RSS as non-present and empty pages are skipped
975 * by the PTE scanner and NUMA hinting faults should be trapped based
978 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
979 rss = get_mm_rss(p->mm);
983 rss = round_up(rss, nr_scan_pages);
984 return rss / nr_scan_pages;
987 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
988 #define MAX_SCAN_WINDOW 2560
990 static unsigned int task_scan_min(struct task_struct *p)
992 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
993 unsigned int scan, floor;
994 unsigned int windows = 1;
996 if (scan_size < MAX_SCAN_WINDOW)
997 windows = MAX_SCAN_WINDOW / scan_size;
998 floor = 1000 / windows;
1000 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1001 return max_t(unsigned int, floor, scan);
1004 static unsigned int task_scan_max(struct task_struct *p)
1006 unsigned int smin = task_scan_min(p);
1009 /* Watch for min being lower than max due to floor calculations */
1010 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1011 return max(smin, smax);
1014 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1016 rq->nr_numa_running += (p->numa_preferred_nid != -1);
1017 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1020 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1022 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1023 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1029 spinlock_t lock; /* nr_tasks, tasks */
1033 struct rcu_head rcu;
1034 nodemask_t active_nodes;
1035 unsigned long total_faults;
1037 * Faults_cpu is used to decide whether memory should move
1038 * towards the CPU. As a consequence, these stats are weighted
1039 * more by CPU use than by memory faults.
1041 unsigned long *faults_cpu;
1042 unsigned long faults[0];
1045 /* Shared or private faults. */
1046 #define NR_NUMA_HINT_FAULT_TYPES 2
1048 /* Memory and CPU locality */
1049 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1051 /* Averaged statistics, and temporary buffers. */
1052 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1054 pid_t task_numa_group_id(struct task_struct *p)
1056 return p->numa_group ? p->numa_group->gid : 0;
1060 * The averaged statistics, shared & private, memory & cpu,
1061 * occupy the first half of the array. The second half of the
1062 * array is for current counters, which are averaged into the
1063 * first set by task_numa_placement.
1065 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1067 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1070 static inline unsigned long task_faults(struct task_struct *p, int nid)
1072 if (!p->numa_faults)
1075 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1076 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1079 static inline unsigned long group_faults(struct task_struct *p, int nid)
1084 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1085 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1088 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1090 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1091 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1094 /* Handle placement on systems where not all nodes are directly connected. */
1095 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1096 int maxdist, bool task)
1098 unsigned long score = 0;
1102 * All nodes are directly connected, and the same distance
1103 * from each other. No need for fancy placement algorithms.
1105 if (sched_numa_topology_type == NUMA_DIRECT)
1109 * This code is called for each node, introducing N^2 complexity,
1110 * which should be ok given the number of nodes rarely exceeds 8.
1112 for_each_online_node(node) {
1113 unsigned long faults;
1114 int dist = node_distance(nid, node);
1117 * The furthest away nodes in the system are not interesting
1118 * for placement; nid was already counted.
1120 if (dist == sched_max_numa_distance || node == nid)
1124 * On systems with a backplane NUMA topology, compare groups
1125 * of nodes, and move tasks towards the group with the most
1126 * memory accesses. When comparing two nodes at distance
1127 * "hoplimit", only nodes closer by than "hoplimit" are part
1128 * of each group. Skip other nodes.
1130 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1134 /* Add up the faults from nearby nodes. */
1136 faults = task_faults(p, node);
1138 faults = group_faults(p, node);
1141 * On systems with a glueless mesh NUMA topology, there are
1142 * no fixed "groups of nodes". Instead, nodes that are not
1143 * directly connected bounce traffic through intermediate
1144 * nodes; a numa_group can occupy any set of nodes.
1145 * The further away a node is, the less the faults count.
1146 * This seems to result in good task placement.
1148 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1149 faults *= (sched_max_numa_distance - dist);
1150 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1160 * These return the fraction of accesses done by a particular task, or
1161 * task group, on a particular numa node. The group weight is given a
1162 * larger multiplier, in order to group tasks together that are almost
1163 * evenly spread out between numa nodes.
1165 static inline unsigned long task_weight(struct task_struct *p, int nid,
1168 unsigned long faults, total_faults;
1170 if (!p->numa_faults)
1173 total_faults = p->total_numa_faults;
1178 faults = task_faults(p, nid);
1179 faults += score_nearby_nodes(p, nid, dist, true);
1181 return 1000 * faults / total_faults;
1184 static inline unsigned long group_weight(struct task_struct *p, int nid,
1187 unsigned long faults, total_faults;
1192 total_faults = p->numa_group->total_faults;
1197 faults = group_faults(p, nid);
1198 faults += score_nearby_nodes(p, nid, dist, false);
1200 return 1000 * faults / total_faults;
1203 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1204 int src_nid, int dst_cpu)
1206 struct numa_group *ng = p->numa_group;
1207 int dst_nid = cpu_to_node(dst_cpu);
1208 int last_cpupid, this_cpupid;
1210 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1213 * Multi-stage node selection is used in conjunction with a periodic
1214 * migration fault to build a temporal task<->page relation. By using
1215 * a two-stage filter we remove short/unlikely relations.
1217 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1218 * a task's usage of a particular page (n_p) per total usage of this
1219 * page (n_t) (in a given time-span) to a probability.
1221 * Our periodic faults will sample this probability and getting the
1222 * same result twice in a row, given these samples are fully
1223 * independent, is then given by P(n)^2, provided our sample period
1224 * is sufficiently short compared to the usage pattern.
1226 * This quadric squishes small probabilities, making it less likely we
1227 * act on an unlikely task<->page relation.
1229 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1230 if (!cpupid_pid_unset(last_cpupid) &&
1231 cpupid_to_nid(last_cpupid) != dst_nid)
1234 /* Always allow migrate on private faults */
1235 if (cpupid_match_pid(p, last_cpupid))
1238 /* A shared fault, but p->numa_group has not been set up yet. */
1243 * Do not migrate if the destination is not a node that
1244 * is actively used by this numa group.
1246 if (!node_isset(dst_nid, ng->active_nodes))
1250 * Source is a node that is not actively used by this
1251 * numa group, while the destination is. Migrate.
1253 if (!node_isset(src_nid, ng->active_nodes))
1257 * Both source and destination are nodes in active
1258 * use by this numa group. Maximize memory bandwidth
1259 * by migrating from more heavily used groups, to less
1260 * heavily used ones, spreading the load around.
1261 * Use a 1/4 hysteresis to avoid spurious page movement.
1263 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1266 static unsigned long weighted_cpuload(const int cpu);
1267 static unsigned long source_load(int cpu, int type);
1268 static unsigned long target_load(int cpu, int type);
1269 static unsigned long capacity_of(int cpu);
1270 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1272 /* Cached statistics for all CPUs within a node */
1274 unsigned long nr_running;
1277 /* Total compute capacity of CPUs on a node */
1278 unsigned long compute_capacity;
1280 /* Approximate capacity in terms of runnable tasks on a node */
1281 unsigned long task_capacity;
1282 int has_free_capacity;
1286 * XXX borrowed from update_sg_lb_stats
1288 static void update_numa_stats(struct numa_stats *ns, int nid)
1290 int smt, cpu, cpus = 0;
1291 unsigned long capacity;
1293 memset(ns, 0, sizeof(*ns));
1294 for_each_cpu(cpu, cpumask_of_node(nid)) {
1295 struct rq *rq = cpu_rq(cpu);
1297 ns->nr_running += rq->nr_running;
1298 ns->load += weighted_cpuload(cpu);
1299 ns->compute_capacity += capacity_of(cpu);
1305 * If we raced with hotplug and there are no CPUs left in our mask
1306 * the @ns structure is NULL'ed and task_numa_compare() will
1307 * not find this node attractive.
1309 * We'll either bail at !has_free_capacity, or we'll detect a huge
1310 * imbalance and bail there.
1315 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1316 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1317 capacity = cpus / smt; /* cores */
1319 ns->task_capacity = min_t(unsigned, capacity,
1320 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1321 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1324 struct task_numa_env {
1325 struct task_struct *p;
1327 int src_cpu, src_nid;
1328 int dst_cpu, dst_nid;
1330 struct numa_stats src_stats, dst_stats;
1335 struct task_struct *best_task;
1340 static void task_numa_assign(struct task_numa_env *env,
1341 struct task_struct *p, long imp)
1344 put_task_struct(env->best_task);
1347 env->best_imp = imp;
1348 env->best_cpu = env->dst_cpu;
1351 static bool load_too_imbalanced(long src_load, long dst_load,
1352 struct task_numa_env *env)
1355 long orig_src_load, orig_dst_load;
1356 long src_capacity, dst_capacity;
1359 * The load is corrected for the CPU capacity available on each node.
1362 * ------------ vs ---------
1363 * src_capacity dst_capacity
1365 src_capacity = env->src_stats.compute_capacity;
1366 dst_capacity = env->dst_stats.compute_capacity;
1368 /* We care about the slope of the imbalance, not the direction. */
1369 if (dst_load < src_load)
1370 swap(dst_load, src_load);
1372 /* Is the difference below the threshold? */
1373 imb = dst_load * src_capacity * 100 -
1374 src_load * dst_capacity * env->imbalance_pct;
1379 * The imbalance is above the allowed threshold.
1380 * Compare it with the old imbalance.
1382 orig_src_load = env->src_stats.load;
1383 orig_dst_load = env->dst_stats.load;
1385 if (orig_dst_load < orig_src_load)
1386 swap(orig_dst_load, orig_src_load);
1388 old_imb = orig_dst_load * src_capacity * 100 -
1389 orig_src_load * dst_capacity * env->imbalance_pct;
1391 /* Would this change make things worse? */
1392 return (imb > old_imb);
1396 * This checks if the overall compute and NUMA accesses of the system would
1397 * be improved if the source tasks was migrated to the target dst_cpu taking
1398 * into account that it might be best if task running on the dst_cpu should
1399 * be exchanged with the source task
1401 static void task_numa_compare(struct task_numa_env *env,
1402 long taskimp, long groupimp)
1404 struct rq *src_rq = cpu_rq(env->src_cpu);
1405 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1406 struct task_struct *cur;
1407 long src_load, dst_load;
1409 long imp = env->p->numa_group ? groupimp : taskimp;
1411 int dist = env->dist;
1412 bool assigned = false;
1416 raw_spin_lock_irq(&dst_rq->lock);
1419 * No need to move the exiting task or idle task.
1421 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1425 * The task_struct must be protected here to protect the
1426 * p->numa_faults access in the task_weight since the
1427 * numa_faults could already be freed in the following path:
1428 * finish_task_switch()
1429 * --> put_task_struct()
1430 * --> __put_task_struct()
1431 * --> task_numa_free()
1433 get_task_struct(cur);
1436 raw_spin_unlock_irq(&dst_rq->lock);
1439 * Because we have preemption enabled we can get migrated around and
1440 * end try selecting ourselves (current == env->p) as a swap candidate.
1446 * "imp" is the fault differential for the source task between the
1447 * source and destination node. Calculate the total differential for
1448 * the source task and potential destination task. The more negative
1449 * the value is, the more rmeote accesses that would be expected to
1450 * be incurred if the tasks were swapped.
1453 /* Skip this swap candidate if cannot move to the source cpu */
1454 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1458 * If dst and source tasks are in the same NUMA group, or not
1459 * in any group then look only at task weights.
1461 if (cur->numa_group == env->p->numa_group) {
1462 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1463 task_weight(cur, env->dst_nid, dist);
1465 * Add some hysteresis to prevent swapping the
1466 * tasks within a group over tiny differences.
1468 if (cur->numa_group)
1472 * Compare the group weights. If a task is all by
1473 * itself (not part of a group), use the task weight
1476 if (cur->numa_group)
1477 imp += group_weight(cur, env->src_nid, dist) -
1478 group_weight(cur, env->dst_nid, dist);
1480 imp += task_weight(cur, env->src_nid, dist) -
1481 task_weight(cur, env->dst_nid, dist);
1485 if (imp <= env->best_imp && moveimp <= env->best_imp)
1489 /* Is there capacity at our destination? */
1490 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1491 !env->dst_stats.has_free_capacity)
1497 /* Balance doesn't matter much if we're running a task per cpu */
1498 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1499 dst_rq->nr_running == 1)
1503 * In the overloaded case, try and keep the load balanced.
1506 load = task_h_load(env->p);
1507 dst_load = env->dst_stats.load + load;
1508 src_load = env->src_stats.load - load;
1510 if (moveimp > imp && moveimp > env->best_imp) {
1512 * If the improvement from just moving env->p direction is
1513 * better than swapping tasks around, check if a move is
1514 * possible. Store a slightly smaller score than moveimp,
1515 * so an actually idle CPU will win.
1517 if (!load_too_imbalanced(src_load, dst_load, env)) {
1519 put_task_struct(cur);
1525 if (imp <= env->best_imp)
1529 load = task_h_load(cur);
1534 if (load_too_imbalanced(src_load, dst_load, env))
1538 * One idle CPU per node is evaluated for a task numa move.
1539 * Call select_idle_sibling to maybe find a better one.
1542 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1547 task_numa_assign(env, cur, imp);
1551 * The dst_rq->curr isn't assigned. The protection for task_struct is
1554 if (cur && !assigned)
1555 put_task_struct(cur);
1558 static void task_numa_find_cpu(struct task_numa_env *env,
1559 long taskimp, long groupimp)
1563 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1564 /* Skip this CPU if the source task cannot migrate */
1565 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1569 task_numa_compare(env, taskimp, groupimp);
1573 /* Only move tasks to a NUMA node less busy than the current node. */
1574 static bool numa_has_capacity(struct task_numa_env *env)
1576 struct numa_stats *src = &env->src_stats;
1577 struct numa_stats *dst = &env->dst_stats;
1579 if (src->has_free_capacity && !dst->has_free_capacity)
1583 * Only consider a task move if the source has a higher load
1584 * than the destination, corrected for CPU capacity on each node.
1586 * src->load dst->load
1587 * --------------------- vs ---------------------
1588 * src->compute_capacity dst->compute_capacity
1590 if (src->load * dst->compute_capacity * env->imbalance_pct >
1592 dst->load * src->compute_capacity * 100)
1598 static int task_numa_migrate(struct task_struct *p)
1600 struct task_numa_env env = {
1603 .src_cpu = task_cpu(p),
1604 .src_nid = task_node(p),
1606 .imbalance_pct = 112,
1612 struct sched_domain *sd;
1613 unsigned long taskweight, groupweight;
1615 long taskimp, groupimp;
1618 * Pick the lowest SD_NUMA domain, as that would have the smallest
1619 * imbalance and would be the first to start moving tasks about.
1621 * And we want to avoid any moving of tasks about, as that would create
1622 * random movement of tasks -- counter the numa conditions we're trying
1626 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1628 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1632 * Cpusets can break the scheduler domain tree into smaller
1633 * balance domains, some of which do not cross NUMA boundaries.
1634 * Tasks that are "trapped" in such domains cannot be migrated
1635 * elsewhere, so there is no point in (re)trying.
1637 if (unlikely(!sd)) {
1638 p->numa_preferred_nid = task_node(p);
1642 env.dst_nid = p->numa_preferred_nid;
1643 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1644 taskweight = task_weight(p, env.src_nid, dist);
1645 groupweight = group_weight(p, env.src_nid, dist);
1646 update_numa_stats(&env.src_stats, env.src_nid);
1647 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1648 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1649 update_numa_stats(&env.dst_stats, env.dst_nid);
1651 /* Try to find a spot on the preferred nid. */
1652 if (numa_has_capacity(&env))
1653 task_numa_find_cpu(&env, taskimp, groupimp);
1656 * Look at other nodes in these cases:
1657 * - there is no space available on the preferred_nid
1658 * - the task is part of a numa_group that is interleaved across
1659 * multiple NUMA nodes; in order to better consolidate the group,
1660 * we need to check other locations.
1662 if (env.best_cpu == -1 || (p->numa_group &&
1663 nodes_weight(p->numa_group->active_nodes) > 1)) {
1664 for_each_online_node(nid) {
1665 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1668 dist = node_distance(env.src_nid, env.dst_nid);
1669 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1671 taskweight = task_weight(p, env.src_nid, dist);
1672 groupweight = group_weight(p, env.src_nid, dist);
1675 /* Only consider nodes where both task and groups benefit */
1676 taskimp = task_weight(p, nid, dist) - taskweight;
1677 groupimp = group_weight(p, nid, dist) - groupweight;
1678 if (taskimp < 0 && groupimp < 0)
1683 update_numa_stats(&env.dst_stats, env.dst_nid);
1684 if (numa_has_capacity(&env))
1685 task_numa_find_cpu(&env, taskimp, groupimp);
1690 * If the task is part of a workload that spans multiple NUMA nodes,
1691 * and is migrating into one of the workload's active nodes, remember
1692 * this node as the task's preferred numa node, so the workload can
1694 * A task that migrated to a second choice node will be better off
1695 * trying for a better one later. Do not set the preferred node here.
1697 if (p->numa_group) {
1698 if (env.best_cpu == -1)
1703 if (node_isset(nid, p->numa_group->active_nodes))
1704 sched_setnuma(p, env.dst_nid);
1707 /* No better CPU than the current one was found. */
1708 if (env.best_cpu == -1)
1712 * Reset the scan period if the task is being rescheduled on an
1713 * alternative node to recheck if the tasks is now properly placed.
1715 p->numa_scan_period = task_scan_min(p);
1717 if (env.best_task == NULL) {
1718 ret = migrate_task_to(p, env.best_cpu);
1720 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1724 ret = migrate_swap(p, env.best_task);
1726 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1727 put_task_struct(env.best_task);
1731 /* Attempt to migrate a task to a CPU on the preferred node. */
1732 static void numa_migrate_preferred(struct task_struct *p)
1734 unsigned long interval = HZ;
1736 /* This task has no NUMA fault statistics yet */
1737 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1740 /* Periodically retry migrating the task to the preferred node */
1741 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1742 p->numa_migrate_retry = jiffies + interval;
1744 /* Success if task is already running on preferred CPU */
1745 if (task_node(p) == p->numa_preferred_nid)
1748 /* Otherwise, try migrate to a CPU on the preferred node */
1749 task_numa_migrate(p);
1753 * Find the nodes on which the workload is actively running. We do this by
1754 * tracking the nodes from which NUMA hinting faults are triggered. This can
1755 * be different from the set of nodes where the workload's memory is currently
1758 * The bitmask is used to make smarter decisions on when to do NUMA page
1759 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1760 * are added when they cause over 6/16 of the maximum number of faults, but
1761 * only removed when they drop below 3/16.
1763 static void update_numa_active_node_mask(struct numa_group *numa_group)
1765 unsigned long faults, max_faults = 0;
1768 for_each_online_node(nid) {
1769 faults = group_faults_cpu(numa_group, nid);
1770 if (faults > max_faults)
1771 max_faults = faults;
1774 for_each_online_node(nid) {
1775 faults = group_faults_cpu(numa_group, nid);
1776 if (!node_isset(nid, numa_group->active_nodes)) {
1777 if (faults > max_faults * 6 / 16)
1778 node_set(nid, numa_group->active_nodes);
1779 } else if (faults < max_faults * 3 / 16)
1780 node_clear(nid, numa_group->active_nodes);
1785 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1786 * increments. The more local the fault statistics are, the higher the scan
1787 * period will be for the next scan window. If local/(local+remote) ratio is
1788 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1789 * the scan period will decrease. Aim for 70% local accesses.
1791 #define NUMA_PERIOD_SLOTS 10
1792 #define NUMA_PERIOD_THRESHOLD 7
1795 * Increase the scan period (slow down scanning) if the majority of
1796 * our memory is already on our local node, or if the majority of
1797 * the page accesses are shared with other processes.
1798 * Otherwise, decrease the scan period.
1800 static void update_task_scan_period(struct task_struct *p,
1801 unsigned long shared, unsigned long private)
1803 unsigned int period_slot;
1807 unsigned long remote = p->numa_faults_locality[0];
1808 unsigned long local = p->numa_faults_locality[1];
1811 * If there were no record hinting faults then either the task is
1812 * completely idle or all activity is areas that are not of interest
1813 * to automatic numa balancing. Related to that, if there were failed
1814 * migration then it implies we are migrating too quickly or the local
1815 * node is overloaded. In either case, scan slower
1817 if (local + shared == 0 || p->numa_faults_locality[2]) {
1818 p->numa_scan_period = min(p->numa_scan_period_max,
1819 p->numa_scan_period << 1);
1821 p->mm->numa_next_scan = jiffies +
1822 msecs_to_jiffies(p->numa_scan_period);
1828 * Prepare to scale scan period relative to the current period.
1829 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1830 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1831 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1833 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1834 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1835 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1836 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1839 diff = slot * period_slot;
1841 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1844 * Scale scan rate increases based on sharing. There is an
1845 * inverse relationship between the degree of sharing and
1846 * the adjustment made to the scanning period. Broadly
1847 * speaking the intent is that there is little point
1848 * scanning faster if shared accesses dominate as it may
1849 * simply bounce migrations uselessly
1851 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1852 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1855 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1856 task_scan_min(p), task_scan_max(p));
1857 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1861 * Get the fraction of time the task has been running since the last
1862 * NUMA placement cycle. The scheduler keeps similar statistics, but
1863 * decays those on a 32ms period, which is orders of magnitude off
1864 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1865 * stats only if the task is so new there are no NUMA statistics yet.
1867 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1869 u64 runtime, delta, now;
1870 /* Use the start of this time slice to avoid calculations. */
1871 now = p->se.exec_start;
1872 runtime = p->se.sum_exec_runtime;
1874 if (p->last_task_numa_placement) {
1875 delta = runtime - p->last_sum_exec_runtime;
1876 *period = now - p->last_task_numa_placement;
1878 delta = p->se.avg.load_sum / p->se.load.weight;
1879 *period = LOAD_AVG_MAX;
1882 p->last_sum_exec_runtime = runtime;
1883 p->last_task_numa_placement = now;
1889 * Determine the preferred nid for a task in a numa_group. This needs to
1890 * be done in a way that produces consistent results with group_weight,
1891 * otherwise workloads might not converge.
1893 static int preferred_group_nid(struct task_struct *p, int nid)
1898 /* Direct connections between all NUMA nodes. */
1899 if (sched_numa_topology_type == NUMA_DIRECT)
1903 * On a system with glueless mesh NUMA topology, group_weight
1904 * scores nodes according to the number of NUMA hinting faults on
1905 * both the node itself, and on nearby nodes.
1907 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1908 unsigned long score, max_score = 0;
1909 int node, max_node = nid;
1911 dist = sched_max_numa_distance;
1913 for_each_online_node(node) {
1914 score = group_weight(p, node, dist);
1915 if (score > max_score) {
1924 * Finding the preferred nid in a system with NUMA backplane
1925 * interconnect topology is more involved. The goal is to locate
1926 * tasks from numa_groups near each other in the system, and
1927 * untangle workloads from different sides of the system. This requires
1928 * searching down the hierarchy of node groups, recursively searching
1929 * inside the highest scoring group of nodes. The nodemask tricks
1930 * keep the complexity of the search down.
1932 nodes = node_online_map;
1933 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1934 unsigned long max_faults = 0;
1935 nodemask_t max_group = NODE_MASK_NONE;
1938 /* Are there nodes at this distance from each other? */
1939 if (!find_numa_distance(dist))
1942 for_each_node_mask(a, nodes) {
1943 unsigned long faults = 0;
1944 nodemask_t this_group;
1945 nodes_clear(this_group);
1947 /* Sum group's NUMA faults; includes a==b case. */
1948 for_each_node_mask(b, nodes) {
1949 if (node_distance(a, b) < dist) {
1950 faults += group_faults(p, b);
1951 node_set(b, this_group);
1952 node_clear(b, nodes);
1956 /* Remember the top group. */
1957 if (faults > max_faults) {
1958 max_faults = faults;
1959 max_group = this_group;
1961 * subtle: at the smallest distance there is
1962 * just one node left in each "group", the
1963 * winner is the preferred nid.
1968 /* Next round, evaluate the nodes within max_group. */
1976 static void task_numa_placement(struct task_struct *p)
1978 int seq, nid, max_nid = -1, max_group_nid = -1;
1979 unsigned long max_faults = 0, max_group_faults = 0;
1980 unsigned long fault_types[2] = { 0, 0 };
1981 unsigned long total_faults;
1982 u64 runtime, period;
1983 spinlock_t *group_lock = NULL;
1986 * The p->mm->numa_scan_seq field gets updated without
1987 * exclusive access. Use READ_ONCE() here to ensure
1988 * that the field is read in a single access:
1990 seq = READ_ONCE(p->mm->numa_scan_seq);
1991 if (p->numa_scan_seq == seq)
1993 p->numa_scan_seq = seq;
1994 p->numa_scan_period_max = task_scan_max(p);
1996 total_faults = p->numa_faults_locality[0] +
1997 p->numa_faults_locality[1];
1998 runtime = numa_get_avg_runtime(p, &period);
2000 /* If the task is part of a group prevent parallel updates to group stats */
2001 if (p->numa_group) {
2002 group_lock = &p->numa_group->lock;
2003 spin_lock_irq(group_lock);
2006 /* Find the node with the highest number of faults */
2007 for_each_online_node(nid) {
2008 /* Keep track of the offsets in numa_faults array */
2009 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2010 unsigned long faults = 0, group_faults = 0;
2013 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2014 long diff, f_diff, f_weight;
2016 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2017 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2018 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2019 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2021 /* Decay existing window, copy faults since last scan */
2022 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2023 fault_types[priv] += p->numa_faults[membuf_idx];
2024 p->numa_faults[membuf_idx] = 0;
2027 * Normalize the faults_from, so all tasks in a group
2028 * count according to CPU use, instead of by the raw
2029 * number of faults. Tasks with little runtime have
2030 * little over-all impact on throughput, and thus their
2031 * faults are less important.
2033 f_weight = div64_u64(runtime << 16, period + 1);
2034 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2036 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2037 p->numa_faults[cpubuf_idx] = 0;
2039 p->numa_faults[mem_idx] += diff;
2040 p->numa_faults[cpu_idx] += f_diff;
2041 faults += p->numa_faults[mem_idx];
2042 p->total_numa_faults += diff;
2043 if (p->numa_group) {
2045 * safe because we can only change our own group
2047 * mem_idx represents the offset for a given
2048 * nid and priv in a specific region because it
2049 * is at the beginning of the numa_faults array.
2051 p->numa_group->faults[mem_idx] += diff;
2052 p->numa_group->faults_cpu[mem_idx] += f_diff;
2053 p->numa_group->total_faults += diff;
2054 group_faults += p->numa_group->faults[mem_idx];
2058 if (faults > max_faults) {
2059 max_faults = faults;
2063 if (group_faults > max_group_faults) {
2064 max_group_faults = group_faults;
2065 max_group_nid = nid;
2069 update_task_scan_period(p, fault_types[0], fault_types[1]);
2071 if (p->numa_group) {
2072 update_numa_active_node_mask(p->numa_group);
2073 spin_unlock_irq(group_lock);
2074 max_nid = preferred_group_nid(p, max_group_nid);
2078 /* Set the new preferred node */
2079 if (max_nid != p->numa_preferred_nid)
2080 sched_setnuma(p, max_nid);
2082 if (task_node(p) != p->numa_preferred_nid)
2083 numa_migrate_preferred(p);
2087 static inline int get_numa_group(struct numa_group *grp)
2089 return atomic_inc_not_zero(&grp->refcount);
2092 static inline void put_numa_group(struct numa_group *grp)
2094 if (atomic_dec_and_test(&grp->refcount))
2095 kfree_rcu(grp, rcu);
2098 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2101 struct numa_group *grp, *my_grp;
2102 struct task_struct *tsk;
2104 int cpu = cpupid_to_cpu(cpupid);
2107 if (unlikely(!p->numa_group)) {
2108 unsigned int size = sizeof(struct numa_group) +
2109 4*nr_node_ids*sizeof(unsigned long);
2111 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2115 atomic_set(&grp->refcount, 1);
2116 spin_lock_init(&grp->lock);
2118 /* Second half of the array tracks nids where faults happen */
2119 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2122 node_set(task_node(current), grp->active_nodes);
2124 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2125 grp->faults[i] = p->numa_faults[i];
2127 grp->total_faults = p->total_numa_faults;
2130 rcu_assign_pointer(p->numa_group, grp);
2134 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2136 if (!cpupid_match_pid(tsk, cpupid))
2139 grp = rcu_dereference(tsk->numa_group);
2143 my_grp = p->numa_group;
2148 * Only join the other group if its bigger; if we're the bigger group,
2149 * the other task will join us.
2151 if (my_grp->nr_tasks > grp->nr_tasks)
2155 * Tie-break on the grp address.
2157 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2160 /* Always join threads in the same process. */
2161 if (tsk->mm == current->mm)
2164 /* Simple filter to avoid false positives due to PID collisions */
2165 if (flags & TNF_SHARED)
2168 /* Update priv based on whether false sharing was detected */
2171 if (join && !get_numa_group(grp))
2179 BUG_ON(irqs_disabled());
2180 double_lock_irq(&my_grp->lock, &grp->lock);
2182 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2183 my_grp->faults[i] -= p->numa_faults[i];
2184 grp->faults[i] += p->numa_faults[i];
2186 my_grp->total_faults -= p->total_numa_faults;
2187 grp->total_faults += p->total_numa_faults;
2192 spin_unlock(&my_grp->lock);
2193 spin_unlock_irq(&grp->lock);
2195 rcu_assign_pointer(p->numa_group, grp);
2197 put_numa_group(my_grp);
2205 void task_numa_free(struct task_struct *p)
2207 struct numa_group *grp = p->numa_group;
2208 void *numa_faults = p->numa_faults;
2209 unsigned long flags;
2213 spin_lock_irqsave(&grp->lock, flags);
2214 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2215 grp->faults[i] -= p->numa_faults[i];
2216 grp->total_faults -= p->total_numa_faults;
2219 spin_unlock_irqrestore(&grp->lock, flags);
2220 RCU_INIT_POINTER(p->numa_group, NULL);
2221 put_numa_group(grp);
2224 p->numa_faults = NULL;
2229 * Got a PROT_NONE fault for a page on @node.
2231 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2233 struct task_struct *p = current;
2234 bool migrated = flags & TNF_MIGRATED;
2235 int cpu_node = task_node(current);
2236 int local = !!(flags & TNF_FAULT_LOCAL);
2239 if (!static_branch_likely(&sched_numa_balancing))
2242 /* for example, ksmd faulting in a user's mm */
2246 /* Allocate buffer to track faults on a per-node basis */
2247 if (unlikely(!p->numa_faults)) {
2248 int size = sizeof(*p->numa_faults) *
2249 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2251 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2252 if (!p->numa_faults)
2255 p->total_numa_faults = 0;
2256 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2260 * First accesses are treated as private, otherwise consider accesses
2261 * to be private if the accessing pid has not changed
2263 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2266 priv = cpupid_match_pid(p, last_cpupid);
2267 if (!priv && !(flags & TNF_NO_GROUP))
2268 task_numa_group(p, last_cpupid, flags, &priv);
2272 * If a workload spans multiple NUMA nodes, a shared fault that
2273 * occurs wholly within the set of nodes that the workload is
2274 * actively using should be counted as local. This allows the
2275 * scan rate to slow down when a workload has settled down.
2277 if (!priv && !local && p->numa_group &&
2278 node_isset(cpu_node, p->numa_group->active_nodes) &&
2279 node_isset(mem_node, p->numa_group->active_nodes))
2282 task_numa_placement(p);
2285 * Retry task to preferred node migration periodically, in case it
2286 * case it previously failed, or the scheduler moved us.
2288 if (time_after(jiffies, p->numa_migrate_retry))
2289 numa_migrate_preferred(p);
2292 p->numa_pages_migrated += pages;
2293 if (flags & TNF_MIGRATE_FAIL)
2294 p->numa_faults_locality[2] += pages;
2296 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2297 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2298 p->numa_faults_locality[local] += pages;
2301 static void reset_ptenuma_scan(struct task_struct *p)
2304 * We only did a read acquisition of the mmap sem, so
2305 * p->mm->numa_scan_seq is written to without exclusive access
2306 * and the update is not guaranteed to be atomic. That's not
2307 * much of an issue though, since this is just used for
2308 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2309 * expensive, to avoid any form of compiler optimizations:
2311 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2312 p->mm->numa_scan_offset = 0;
2316 * The expensive part of numa migration is done from task_work context.
2317 * Triggered from task_tick_numa().
2319 void task_numa_work(struct callback_head *work)
2321 unsigned long migrate, next_scan, now = jiffies;
2322 struct task_struct *p = current;
2323 struct mm_struct *mm = p->mm;
2324 struct vm_area_struct *vma;
2325 unsigned long start, end;
2326 unsigned long nr_pte_updates = 0;
2327 long pages, virtpages;
2329 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2331 work->next = work; /* protect against double add */
2333 * Who cares about NUMA placement when they're dying.
2335 * NOTE: make sure not to dereference p->mm before this check,
2336 * exit_task_work() happens _after_ exit_mm() so we could be called
2337 * without p->mm even though we still had it when we enqueued this
2340 if (p->flags & PF_EXITING)
2343 if (!mm->numa_next_scan) {
2344 mm->numa_next_scan = now +
2345 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2349 * Enforce maximal scan/migration frequency..
2351 migrate = mm->numa_next_scan;
2352 if (time_before(now, migrate))
2355 if (p->numa_scan_period == 0) {
2356 p->numa_scan_period_max = task_scan_max(p);
2357 p->numa_scan_period = task_scan_min(p);
2360 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2361 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2365 * Delay this task enough that another task of this mm will likely win
2366 * the next time around.
2368 p->node_stamp += 2 * TICK_NSEC;
2370 start = mm->numa_scan_offset;
2371 pages = sysctl_numa_balancing_scan_size;
2372 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2373 virtpages = pages * 8; /* Scan up to this much virtual space */
2378 down_read(&mm->mmap_sem);
2379 vma = find_vma(mm, start);
2381 reset_ptenuma_scan(p);
2385 for (; vma; vma = vma->vm_next) {
2386 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2387 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2392 * Shared library pages mapped by multiple processes are not
2393 * migrated as it is expected they are cache replicated. Avoid
2394 * hinting faults in read-only file-backed mappings or the vdso
2395 * as migrating the pages will be of marginal benefit.
2398 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2402 * Skip inaccessible VMAs to avoid any confusion between
2403 * PROT_NONE and NUMA hinting ptes
2405 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2409 start = max(start, vma->vm_start);
2410 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2411 end = min(end, vma->vm_end);
2412 nr_pte_updates = change_prot_numa(vma, start, end);
2415 * Try to scan sysctl_numa_balancing_size worth of
2416 * hpages that have at least one present PTE that
2417 * is not already pte-numa. If the VMA contains
2418 * areas that are unused or already full of prot_numa
2419 * PTEs, scan up to virtpages, to skip through those
2423 pages -= (end - start) >> PAGE_SHIFT;
2424 virtpages -= (end - start) >> PAGE_SHIFT;
2427 if (pages <= 0 || virtpages <= 0)
2431 } while (end != vma->vm_end);
2436 * It is possible to reach the end of the VMA list but the last few
2437 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2438 * would find the !migratable VMA on the next scan but not reset the
2439 * scanner to the start so check it now.
2442 mm->numa_scan_offset = start;
2444 reset_ptenuma_scan(p);
2445 up_read(&mm->mmap_sem);
2449 * Drive the periodic memory faults..
2451 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2453 struct callback_head *work = &curr->numa_work;
2457 * We don't care about NUMA placement if we don't have memory.
2459 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2463 * Using runtime rather than walltime has the dual advantage that
2464 * we (mostly) drive the selection from busy threads and that the
2465 * task needs to have done some actual work before we bother with
2468 now = curr->se.sum_exec_runtime;
2469 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2471 if (now > curr->node_stamp + period) {
2472 if (!curr->node_stamp)
2473 curr->numa_scan_period = task_scan_min(curr);
2474 curr->node_stamp += period;
2476 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2477 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2478 task_work_add(curr, work, true);
2483 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2487 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2491 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2494 #endif /* CONFIG_NUMA_BALANCING */
2497 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2499 update_load_add(&cfs_rq->load, se->load.weight);
2500 if (!parent_entity(se))
2501 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2503 if (entity_is_task(se)) {
2504 struct rq *rq = rq_of(cfs_rq);
2506 account_numa_enqueue(rq, task_of(se));
2507 list_add(&se->group_node, &rq->cfs_tasks);
2510 cfs_rq->nr_running++;
2514 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2516 update_load_sub(&cfs_rq->load, se->load.weight);
2517 if (!parent_entity(se))
2518 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2519 if (entity_is_task(se)) {
2520 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2521 list_del_init(&se->group_node);
2523 cfs_rq->nr_running--;
2526 #ifdef CONFIG_FAIR_GROUP_SCHED
2528 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2533 * Use this CPU's real-time load instead of the last load contribution
2534 * as the updating of the contribution is delayed, and we will use the
2535 * the real-time load to calc the share. See update_tg_load_avg().
2537 tg_weight = atomic_long_read(&tg->load_avg);
2538 tg_weight -= cfs_rq->tg_load_avg_contrib;
2539 tg_weight += cfs_rq->load.weight;
2544 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2546 long tg_weight, load, shares;
2548 tg_weight = calc_tg_weight(tg, cfs_rq);
2549 load = cfs_rq->load.weight;
2551 shares = (tg->shares * load);
2553 shares /= tg_weight;
2555 if (shares < MIN_SHARES)
2556 shares = MIN_SHARES;
2557 if (shares > tg->shares)
2558 shares = tg->shares;
2562 # else /* CONFIG_SMP */
2563 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2567 # endif /* CONFIG_SMP */
2568 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2569 unsigned long weight)
2572 /* commit outstanding execution time */
2573 if (cfs_rq->curr == se)
2574 update_curr(cfs_rq);
2575 account_entity_dequeue(cfs_rq, se);
2578 update_load_set(&se->load, weight);
2581 account_entity_enqueue(cfs_rq, se);
2584 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2586 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2588 struct task_group *tg;
2589 struct sched_entity *se;
2593 se = tg->se[cpu_of(rq_of(cfs_rq))];
2594 if (!se || throttled_hierarchy(cfs_rq))
2597 if (likely(se->load.weight == tg->shares))
2600 shares = calc_cfs_shares(cfs_rq, tg);
2602 reweight_entity(cfs_rq_of(se), se, shares);
2604 #else /* CONFIG_FAIR_GROUP_SCHED */
2605 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2608 #endif /* CONFIG_FAIR_GROUP_SCHED */
2611 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2612 static const u32 runnable_avg_yN_inv[] = {
2613 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2614 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2615 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2616 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2617 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2618 0x85aac367, 0x82cd8698,
2622 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2623 * over-estimates when re-combining.
2625 static const u32 runnable_avg_yN_sum[] = {
2626 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2627 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2628 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2633 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2635 static __always_inline u64 decay_load(u64 val, u64 n)
2637 unsigned int local_n;
2641 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2644 /* after bounds checking we can collapse to 32-bit */
2648 * As y^PERIOD = 1/2, we can combine
2649 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2650 * With a look-up table which covers y^n (n<PERIOD)
2652 * To achieve constant time decay_load.
2654 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2655 val >>= local_n / LOAD_AVG_PERIOD;
2656 local_n %= LOAD_AVG_PERIOD;
2659 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2664 * For updates fully spanning n periods, the contribution to runnable
2665 * average will be: \Sum 1024*y^n
2667 * We can compute this reasonably efficiently by combining:
2668 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2670 static u32 __compute_runnable_contrib(u64 n)
2674 if (likely(n <= LOAD_AVG_PERIOD))
2675 return runnable_avg_yN_sum[n];
2676 else if (unlikely(n >= LOAD_AVG_MAX_N))
2677 return LOAD_AVG_MAX;
2679 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2681 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2682 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2684 n -= LOAD_AVG_PERIOD;
2685 } while (n > LOAD_AVG_PERIOD);
2687 contrib = decay_load(contrib, n);
2688 return contrib + runnable_avg_yN_sum[n];
2691 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2692 #error "load tracking assumes 2^10 as unit"
2695 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2698 * We can represent the historical contribution to runnable average as the
2699 * coefficients of a geometric series. To do this we sub-divide our runnable
2700 * history into segments of approximately 1ms (1024us); label the segment that
2701 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2703 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2705 * (now) (~1ms ago) (~2ms ago)
2707 * Let u_i denote the fraction of p_i that the entity was runnable.
2709 * We then designate the fractions u_i as our co-efficients, yielding the
2710 * following representation of historical load:
2711 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2713 * We choose y based on the with of a reasonably scheduling period, fixing:
2716 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2717 * approximately half as much as the contribution to load within the last ms
2720 * When a period "rolls over" and we have new u_0`, multiplying the previous
2721 * sum again by y is sufficient to update:
2722 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2723 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2725 static __always_inline int
2726 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2727 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2729 u64 delta, scaled_delta, periods;
2731 unsigned int delta_w, scaled_delta_w, decayed = 0;
2732 unsigned long scale_freq, scale_cpu;
2734 delta = now - sa->last_update_time;
2736 * This should only happen when time goes backwards, which it
2737 * unfortunately does during sched clock init when we swap over to TSC.
2739 if ((s64)delta < 0) {
2740 sa->last_update_time = now;
2745 * Use 1024ns as the unit of measurement since it's a reasonable
2746 * approximation of 1us and fast to compute.
2751 sa->last_update_time = now;
2753 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2754 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2755 trace_sched_contrib_scale_f(cpu, scale_freq, scale_cpu);
2757 /* delta_w is the amount already accumulated against our next period */
2758 delta_w = sa->period_contrib;
2759 if (delta + delta_w >= 1024) {
2762 /* how much left for next period will start over, we don't know yet */
2763 sa->period_contrib = 0;
2766 * Now that we know we're crossing a period boundary, figure
2767 * out how much from delta we need to complete the current
2768 * period and accrue it.
2770 delta_w = 1024 - delta_w;
2771 scaled_delta_w = cap_scale(delta_w, scale_freq);
2773 sa->load_sum += weight * scaled_delta_w;
2775 cfs_rq->runnable_load_sum +=
2776 weight * scaled_delta_w;
2780 sa->util_sum += scaled_delta_w * scale_cpu;
2784 /* Figure out how many additional periods this update spans */
2785 periods = delta / 1024;
2788 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2790 cfs_rq->runnable_load_sum =
2791 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2793 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2795 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2796 contrib = __compute_runnable_contrib(periods);
2797 contrib = cap_scale(contrib, scale_freq);
2799 sa->load_sum += weight * contrib;
2801 cfs_rq->runnable_load_sum += weight * contrib;
2804 sa->util_sum += contrib * scale_cpu;
2807 /* Remainder of delta accrued against u_0` */
2808 scaled_delta = cap_scale(delta, scale_freq);
2810 sa->load_sum += weight * scaled_delta;
2812 cfs_rq->runnable_load_sum += weight * scaled_delta;
2815 sa->util_sum += scaled_delta * scale_cpu;
2817 sa->period_contrib += delta;
2820 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2822 cfs_rq->runnable_load_avg =
2823 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2825 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2832 * Signed add and clamp on underflow.
2834 * Explicitly do a load-store to ensure the intermediate value never hits
2835 * memory. This allows lockless observations without ever seeing the negative
2838 #define add_positive(_ptr, _val) do { \
2839 typeof(_ptr) ptr = (_ptr); \
2840 typeof(_val) val = (_val); \
2841 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2845 if (val < 0 && res > var) \
2848 WRITE_ONCE(*ptr, res); \
2851 #ifdef CONFIG_FAIR_GROUP_SCHED
2853 * update_tg_load_avg - update the tg's load avg
2854 * @cfs_rq: the cfs_rq whose avg changed
2855 * @force: update regardless of how small the difference
2857 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
2858 * However, because tg->load_avg is a global value there are performance
2861 * In order to avoid having to look at the other cfs_rq's, we use a
2862 * differential update where we store the last value we propagated. This in
2863 * turn allows skipping updates if the differential is 'small'.
2865 * Updating tg's load_avg is necessary before update_cfs_share() (which is
2866 * done) and effective_load() (which is not done because it is too costly).
2868 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2870 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2873 * No need to update load_avg for root_task_group as it is not used.
2875 if (cfs_rq->tg == &root_task_group)
2878 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2879 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2880 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2885 * Called within set_task_rq() right before setting a task's cpu. The
2886 * caller only guarantees p->pi_lock is held; no other assumptions,
2887 * including the state of rq->lock, should be made.
2889 void set_task_rq_fair(struct sched_entity *se,
2890 struct cfs_rq *prev, struct cfs_rq *next)
2892 if (!sched_feat(ATTACH_AGE_LOAD))
2896 * We are supposed to update the task to "current" time, then its up to
2897 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
2898 * getting what current time is, so simply throw away the out-of-date
2899 * time. This will result in the wakee task is less decayed, but giving
2900 * the wakee more load sounds not bad.
2902 if (se->avg.last_update_time && prev) {
2903 u64 p_last_update_time;
2904 u64 n_last_update_time;
2906 #ifndef CONFIG_64BIT
2907 u64 p_last_update_time_copy;
2908 u64 n_last_update_time_copy;
2911 p_last_update_time_copy = prev->load_last_update_time_copy;
2912 n_last_update_time_copy = next->load_last_update_time_copy;
2916 p_last_update_time = prev->avg.last_update_time;
2917 n_last_update_time = next->avg.last_update_time;
2919 } while (p_last_update_time != p_last_update_time_copy ||
2920 n_last_update_time != n_last_update_time_copy);
2922 p_last_update_time = prev->avg.last_update_time;
2923 n_last_update_time = next->avg.last_update_time;
2925 __update_load_avg(p_last_update_time, cpu_of(rq_of(prev)),
2926 &se->avg, 0, 0, NULL);
2927 se->avg.last_update_time = n_last_update_time;
2931 /* Take into account change of utilization of a child task group */
2933 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se)
2935 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
2936 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
2938 /* Nothing to update */
2942 /* Set new sched_entity's utilization */
2943 se->avg.util_avg = gcfs_rq->avg.util_avg;
2944 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
2946 /* Update parent cfs_rq utilization */
2947 add_positive(&cfs_rq->avg.util_avg, delta);
2948 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
2951 /* Take into account change of load of a child task group */
2953 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se)
2955 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
2956 long delta, load = gcfs_rq->avg.load_avg;
2959 * If the load of group cfs_rq is null, the load of the
2960 * sched_entity will also be null so we can skip the formula
2965 /* Get tg's load and ensure tg_load > 0 */
2966 tg_load = atomic_long_read(&gcfs_rq->tg->load_avg) + 1;
2968 /* Ensure tg_load >= load and updated with current load*/
2969 tg_load -= gcfs_rq->tg_load_avg_contrib;
2973 * We need to compute a correction term in the case that the
2974 * task group is consuming more CPU than a task of equal
2975 * weight. A task with a weight equals to tg->shares will have
2976 * a load less or equal to scale_load_down(tg->shares).
2977 * Similarly, the sched_entities that represent the task group
2978 * at parent level, can't have a load higher than
2979 * scale_load_down(tg->shares). And the Sum of sched_entities'
2980 * load must be <= scale_load_down(tg->shares).
2982 if (tg_load > scale_load_down(gcfs_rq->tg->shares)) {
2983 /* scale gcfs_rq's load into tg's shares*/
2984 load *= scale_load_down(gcfs_rq->tg->shares);
2989 delta = load - se->avg.load_avg;
2991 /* Nothing to update */
2995 /* Set new sched_entity's load */
2996 se->avg.load_avg = load;
2997 se->avg.load_sum = se->avg.load_avg * LOAD_AVG_MAX;
2999 /* Update parent cfs_rq load */
3000 add_positive(&cfs_rq->avg.load_avg, delta);
3001 cfs_rq->avg.load_sum = cfs_rq->avg.load_avg * LOAD_AVG_MAX;
3004 * If the sched_entity is already enqueued, we also have to update the
3005 * runnable load avg.
3008 /* Update parent cfs_rq runnable_load_avg */
3009 add_positive(&cfs_rq->runnable_load_avg, delta);
3010 cfs_rq->runnable_load_sum = cfs_rq->runnable_load_avg * LOAD_AVG_MAX;
3014 static inline void set_tg_cfs_propagate(struct cfs_rq *cfs_rq)
3016 cfs_rq->propagate_avg = 1;
3019 static inline int test_and_clear_tg_cfs_propagate(struct sched_entity *se)
3021 struct cfs_rq *cfs_rq = group_cfs_rq(se);
3023 if (!cfs_rq->propagate_avg)
3026 cfs_rq->propagate_avg = 0;
3030 /* Update task and its cfs_rq load average */
3031 static inline int propagate_entity_load_avg(struct sched_entity *se)
3033 struct cfs_rq *cfs_rq;
3035 if (entity_is_task(se))
3038 if (!test_and_clear_tg_cfs_propagate(se))
3041 cfs_rq = cfs_rq_of(se);
3043 set_tg_cfs_propagate(cfs_rq);
3045 update_tg_cfs_util(cfs_rq, se);
3046 update_tg_cfs_load(cfs_rq, se);
3051 #else /* CONFIG_FAIR_GROUP_SCHED */
3053 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3055 static inline int propagate_entity_load_avg(struct sched_entity *se)
3060 static inline void set_tg_cfs_propagate(struct cfs_rq *cfs_rq) {}
3062 #endif /* CONFIG_FAIR_GROUP_SCHED */
3064 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
3066 if (&this_rq()->cfs == cfs_rq) {
3068 * There are a few boundary cases this might miss but it should
3069 * get called often enough that that should (hopefully) not be
3070 * a real problem -- added to that it only calls on the local
3071 * CPU, so if we enqueue remotely we'll miss an update, but
3072 * the next tick/schedule should update.
3074 * It will not get called when we go idle, because the idle
3075 * thread is a different class (!fair), nor will the utilization
3076 * number include things like RT tasks.
3078 * As is, the util number is not freq-invariant (we'd have to
3079 * implement arch_scale_freq_capacity() for that).
3083 cpufreq_update_util(rq_of(cfs_rq), 0);
3087 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
3090 * Unsigned subtract and clamp on underflow.
3092 * Explicitly do a load-store to ensure the intermediate value never hits
3093 * memory. This allows lockless observations without ever seeing the negative
3096 #define sub_positive(_ptr, _val) do { \
3097 typeof(_ptr) ptr = (_ptr); \
3098 typeof(*ptr) val = (_val); \
3099 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3103 WRITE_ONCE(*ptr, res); \
3107 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3108 * @now: current time, as per cfs_rq_clock_task()
3109 * @cfs_rq: cfs_rq to update
3110 * @update_freq: should we call cfs_rq_util_change() or will the call do so
3112 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3113 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3114 * post_init_entity_util_avg().
3116 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3118 * Returns true if the load decayed or we removed load.
3120 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3121 * call update_tg_load_avg() when this function returns true.
3124 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
3126 struct sched_avg *sa = &cfs_rq->avg;
3127 int decayed, removed = 0, removed_util = 0;
3129 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
3130 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
3131 sub_positive(&sa->load_avg, r);
3132 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
3134 set_tg_cfs_propagate(cfs_rq);
3137 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
3138 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
3139 sub_positive(&sa->util_avg, r);
3140 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
3142 set_tg_cfs_propagate(cfs_rq);
3145 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
3146 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
3148 #ifndef CONFIG_64BIT
3150 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3153 /* Trace CPU load, unless cfs_rq belongs to a non-root task_group */
3154 if (cfs_rq == &rq_of(cfs_rq)->cfs)
3155 trace_sched_load_avg_cpu(cpu_of(rq_of(cfs_rq)), cfs_rq);
3157 if (update_freq && (decayed || removed_util))
3158 cfs_rq_util_change(cfs_rq);
3160 return decayed || removed;
3164 * Optional action to be done while updating the load average
3166 #define UPDATE_TG 0x1
3167 #define SKIP_AGE_LOAD 0x2
3169 /* Update task and its cfs_rq load average */
3170 static inline void update_load_avg(struct sched_entity *se, int flags)
3172 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3173 u64 now = cfs_rq_clock_task(cfs_rq);
3174 int cpu = cpu_of(rq_of(cfs_rq));
3178 * Track task load average for carrying it to new CPU after migrated, and
3179 * track group sched_entity load average for task_h_load calc in migration
3181 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD)) {
3182 __update_load_avg(now, cpu, &se->avg,
3183 se->on_rq * scale_load_down(se->load.weight),
3184 cfs_rq->curr == se, NULL);
3187 decayed = update_cfs_rq_load_avg(now, cfs_rq, true);
3188 decayed |= propagate_entity_load_avg(se);
3190 if (decayed && (flags & UPDATE_TG))
3191 update_tg_load_avg(cfs_rq, 0);
3193 if (entity_is_task(se))
3194 trace_sched_load_avg_task(task_of(se), &se->avg);
3198 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3199 * @cfs_rq: cfs_rq to attach to
3200 * @se: sched_entity to attach
3202 * Must call update_cfs_rq_load_avg() before this, since we rely on
3203 * cfs_rq->avg.last_update_time being current.
3205 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3207 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3208 cfs_rq->avg.load_avg += se->avg.load_avg;
3209 cfs_rq->avg.load_sum += se->avg.load_sum;
3210 cfs_rq->avg.util_avg += se->avg.util_avg;
3211 cfs_rq->avg.util_sum += se->avg.util_sum;
3212 set_tg_cfs_propagate(cfs_rq);
3214 cfs_rq_util_change(cfs_rq);
3218 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3219 * @cfs_rq: cfs_rq to detach from
3220 * @se: sched_entity to detach
3222 * Must call update_cfs_rq_load_avg() before this, since we rely on
3223 * cfs_rq->avg.last_update_time being current.
3225 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3228 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3229 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
3230 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3231 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3232 set_tg_cfs_propagate(cfs_rq);
3234 cfs_rq_util_change(cfs_rq);
3237 /* Add the load generated by se into cfs_rq's load average */
3239 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3241 struct sched_avg *sa = &se->avg;
3243 cfs_rq->runnable_load_avg += sa->load_avg;
3244 cfs_rq->runnable_load_sum += sa->load_sum;
3246 if (!sa->last_update_time) {
3247 attach_entity_load_avg(cfs_rq, se);
3248 update_tg_load_avg(cfs_rq, 0);
3252 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3254 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3256 cfs_rq->runnable_load_avg =
3257 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
3258 cfs_rq->runnable_load_sum =
3259 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
3262 #ifndef CONFIG_64BIT
3263 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3265 u64 last_update_time_copy;
3266 u64 last_update_time;
3269 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3271 last_update_time = cfs_rq->avg.last_update_time;
3272 } while (last_update_time != last_update_time_copy);
3274 return last_update_time;
3277 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3279 return cfs_rq->avg.last_update_time;
3284 * Synchronize entity load avg of dequeued entity without locking
3287 void sync_entity_load_avg(struct sched_entity *se)
3289 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3290 u64 last_update_time;
3292 last_update_time = cfs_rq_last_update_time(cfs_rq);
3293 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
3297 * Task first catches up with cfs_rq, and then subtract
3298 * itself from the cfs_rq (task must be off the queue now).
3300 void remove_entity_load_avg(struct sched_entity *se)
3302 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3305 * Newly created task or never used group entity should not be removed
3306 * from its (source) cfs_rq
3308 if (se->avg.last_update_time == 0)
3311 sync_entity_load_avg(se);
3312 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3313 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3317 * Update the rq's load with the elapsed running time before entering
3318 * idle. if the last scheduled task is not a CFS task, idle_enter will
3319 * be the only way to update the runnable statistic.
3321 void idle_enter_fair(struct rq *this_rq)
3326 * Update the rq's load with the elapsed idle time before a task is
3327 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
3328 * be the only way to update the runnable statistic.
3330 void idle_exit_fair(struct rq *this_rq)
3334 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3336 return cfs_rq->runnable_load_avg;
3339 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3341 return cfs_rq->avg.load_avg;
3344 static int idle_balance(struct rq *this_rq);
3346 #else /* CONFIG_SMP */
3349 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
3354 #define UPDATE_TG 0x0
3355 #define SKIP_AGE_LOAD 0x0
3357 static inline void update_load_avg(struct sched_entity *se, int not_used1){}
3359 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3361 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3362 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3365 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3367 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3369 static inline int idle_balance(struct rq *rq)
3374 #endif /* CONFIG_SMP */
3376 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
3378 #ifdef CONFIG_SCHEDSTATS
3379 struct task_struct *tsk = NULL;
3381 if (entity_is_task(se))
3384 if (se->statistics.sleep_start) {
3385 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
3390 if (unlikely(delta > se->statistics.sleep_max))
3391 se->statistics.sleep_max = delta;
3393 se->statistics.sleep_start = 0;
3394 se->statistics.sum_sleep_runtime += delta;
3397 account_scheduler_latency(tsk, delta >> 10, 1);
3398 trace_sched_stat_sleep(tsk, delta);
3401 if (se->statistics.block_start) {
3402 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
3407 if (unlikely(delta > se->statistics.block_max))
3408 se->statistics.block_max = delta;
3410 se->statistics.block_start = 0;
3411 se->statistics.sum_sleep_runtime += delta;
3414 if (tsk->in_iowait) {
3415 se->statistics.iowait_sum += delta;
3416 se->statistics.iowait_count++;
3417 trace_sched_stat_iowait(tsk, delta);
3420 trace_sched_stat_blocked(tsk, delta);
3421 trace_sched_blocked_reason(tsk);
3424 * Blocking time is in units of nanosecs, so shift by
3425 * 20 to get a milliseconds-range estimation of the
3426 * amount of time that the task spent sleeping:
3428 if (unlikely(prof_on == SLEEP_PROFILING)) {
3429 profile_hits(SLEEP_PROFILING,
3430 (void *)get_wchan(tsk),
3433 account_scheduler_latency(tsk, delta >> 10, 0);
3439 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3441 #ifdef CONFIG_SCHED_DEBUG
3442 s64 d = se->vruntime - cfs_rq->min_vruntime;
3447 if (d > 3*sysctl_sched_latency)
3448 schedstat_inc(cfs_rq, nr_spread_over);
3453 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3455 u64 vruntime = cfs_rq->min_vruntime;
3458 * The 'current' period is already promised to the current tasks,
3459 * however the extra weight of the new task will slow them down a
3460 * little, place the new task so that it fits in the slot that
3461 * stays open at the end.
3463 if (initial && sched_feat(START_DEBIT))
3464 vruntime += sched_vslice(cfs_rq, se);
3466 /* sleeps up to a single latency don't count. */
3468 unsigned long thresh = sysctl_sched_latency;
3471 * Halve their sleep time's effect, to allow
3472 * for a gentler effect of sleepers:
3474 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3480 /* ensure we never gain time by being placed backwards. */
3481 se->vruntime = max_vruntime(se->vruntime, vruntime);
3484 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3487 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3490 * Update the normalized vruntime before updating min_vruntime
3491 * through calling update_curr().
3493 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3494 se->vruntime += cfs_rq->min_vruntime;
3497 * Update run-time statistics of the 'current'.
3499 update_curr(cfs_rq);
3500 update_load_avg(se, UPDATE_TG);
3501 enqueue_entity_load_avg(cfs_rq, se);
3502 account_entity_enqueue(cfs_rq, se);
3503 update_cfs_shares(cfs_rq);
3505 if (flags & ENQUEUE_WAKEUP) {
3506 place_entity(cfs_rq, se, 0);
3507 enqueue_sleeper(cfs_rq, se);
3510 update_stats_enqueue(cfs_rq, se);
3511 check_spread(cfs_rq, se);
3512 if (se != cfs_rq->curr)
3513 __enqueue_entity(cfs_rq, se);
3516 if (cfs_rq->nr_running == 1) {
3517 list_add_leaf_cfs_rq(cfs_rq);
3518 check_enqueue_throttle(cfs_rq);
3522 static void __clear_buddies_last(struct sched_entity *se)
3524 for_each_sched_entity(se) {
3525 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3526 if (cfs_rq->last != se)
3529 cfs_rq->last = NULL;
3533 static void __clear_buddies_next(struct sched_entity *se)
3535 for_each_sched_entity(se) {
3536 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3537 if (cfs_rq->next != se)
3540 cfs_rq->next = NULL;
3544 static void __clear_buddies_skip(struct sched_entity *se)
3546 for_each_sched_entity(se) {
3547 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3548 if (cfs_rq->skip != se)
3551 cfs_rq->skip = NULL;
3555 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3557 if (cfs_rq->last == se)
3558 __clear_buddies_last(se);
3560 if (cfs_rq->next == se)
3561 __clear_buddies_next(se);
3563 if (cfs_rq->skip == se)
3564 __clear_buddies_skip(se);
3567 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3570 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3573 * Update run-time statistics of the 'current'.
3575 update_curr(cfs_rq);
3576 update_load_avg(se, UPDATE_TG);
3577 dequeue_entity_load_avg(cfs_rq, se);
3579 update_stats_dequeue(cfs_rq, se);
3580 if (flags & DEQUEUE_SLEEP) {
3581 #ifdef CONFIG_SCHEDSTATS
3582 if (entity_is_task(se)) {
3583 struct task_struct *tsk = task_of(se);
3585 if (tsk->state & TASK_INTERRUPTIBLE)
3586 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3587 if (tsk->state & TASK_UNINTERRUPTIBLE)
3588 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3593 clear_buddies(cfs_rq, se);
3595 if (se != cfs_rq->curr)
3596 __dequeue_entity(cfs_rq, se);
3598 account_entity_dequeue(cfs_rq, se);
3601 * Normalize the entity after updating the min_vruntime because the
3602 * update can refer to the ->curr item and we need to reflect this
3603 * movement in our normalized position.
3605 if (!(flags & DEQUEUE_SLEEP))
3606 se->vruntime -= cfs_rq->min_vruntime;
3608 /* return excess runtime on last dequeue */
3609 return_cfs_rq_runtime(cfs_rq);
3611 update_min_vruntime(cfs_rq);
3612 update_cfs_shares(cfs_rq);
3616 * Preempt the current task with a newly woken task if needed:
3619 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3621 unsigned long ideal_runtime, delta_exec;
3622 struct sched_entity *se;
3625 ideal_runtime = sched_slice(cfs_rq, curr);
3626 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3627 if (delta_exec > ideal_runtime) {
3628 resched_curr(rq_of(cfs_rq));
3630 * The current task ran long enough, ensure it doesn't get
3631 * re-elected due to buddy favours.
3633 clear_buddies(cfs_rq, curr);
3638 * Ensure that a task that missed wakeup preemption by a
3639 * narrow margin doesn't have to wait for a full slice.
3640 * This also mitigates buddy induced latencies under load.
3642 if (delta_exec < sysctl_sched_min_granularity)
3645 se = __pick_first_entity(cfs_rq);
3646 delta = curr->vruntime - se->vruntime;
3651 if (delta > ideal_runtime)
3652 resched_curr(rq_of(cfs_rq));
3656 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3658 /* 'current' is not kept within the tree. */
3661 * Any task has to be enqueued before it get to execute on
3662 * a CPU. So account for the time it spent waiting on the
3665 update_stats_wait_end(cfs_rq, se);
3666 __dequeue_entity(cfs_rq, se);
3667 update_load_avg(se, UPDATE_TG);
3670 update_stats_curr_start(cfs_rq, se);
3672 #ifdef CONFIG_SCHEDSTATS
3674 * Track our maximum slice length, if the CPU's load is at
3675 * least twice that of our own weight (i.e. dont track it
3676 * when there are only lesser-weight tasks around):
3678 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3679 se->statistics.slice_max = max(se->statistics.slice_max,
3680 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3683 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3687 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3690 * Pick the next process, keeping these things in mind, in this order:
3691 * 1) keep things fair between processes/task groups
3692 * 2) pick the "next" process, since someone really wants that to run
3693 * 3) pick the "last" process, for cache locality
3694 * 4) do not run the "skip" process, if something else is available
3696 static struct sched_entity *
3697 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3699 struct sched_entity *left = __pick_first_entity(cfs_rq);
3700 struct sched_entity *se;
3703 * If curr is set we have to see if its left of the leftmost entity
3704 * still in the tree, provided there was anything in the tree at all.
3706 if (!left || (curr && entity_before(curr, left)))
3709 se = left; /* ideally we run the leftmost entity */
3712 * Avoid running the skip buddy, if running something else can
3713 * be done without getting too unfair.
3715 if (cfs_rq->skip == se) {
3716 struct sched_entity *second;
3719 second = __pick_first_entity(cfs_rq);
3721 second = __pick_next_entity(se);
3722 if (!second || (curr && entity_before(curr, second)))
3726 if (second && wakeup_preempt_entity(second, left) < 1)
3731 * Prefer last buddy, try to return the CPU to a preempted task.
3733 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3737 * Someone really wants this to run. If it's not unfair, run it.
3739 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3742 clear_buddies(cfs_rq, se);
3747 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3749 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3752 * If still on the runqueue then deactivate_task()
3753 * was not called and update_curr() has to be done:
3756 update_curr(cfs_rq);
3758 /* throttle cfs_rqs exceeding runtime */
3759 check_cfs_rq_runtime(cfs_rq);
3761 check_spread(cfs_rq, prev);
3763 update_stats_wait_start(cfs_rq, prev);
3764 /* Put 'current' back into the tree. */
3765 __enqueue_entity(cfs_rq, prev);
3766 /* in !on_rq case, update occurred at dequeue */
3767 update_load_avg(prev, 0);
3769 cfs_rq->curr = NULL;
3773 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3776 * Update run-time statistics of the 'current'.
3778 update_curr(cfs_rq);
3781 * Ensure that runnable average is periodically updated.
3783 update_load_avg(curr, UPDATE_TG);
3784 update_cfs_shares(cfs_rq);
3786 #ifdef CONFIG_SCHED_HRTICK
3788 * queued ticks are scheduled to match the slice, so don't bother
3789 * validating it and just reschedule.
3792 resched_curr(rq_of(cfs_rq));
3796 * don't let the period tick interfere with the hrtick preemption
3798 if (!sched_feat(DOUBLE_TICK) &&
3799 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3803 if (cfs_rq->nr_running > 1)
3804 check_preempt_tick(cfs_rq, curr);
3808 /**************************************************
3809 * CFS bandwidth control machinery
3812 #ifdef CONFIG_CFS_BANDWIDTH
3814 #ifdef HAVE_JUMP_LABEL
3815 static struct static_key __cfs_bandwidth_used;
3817 static inline bool cfs_bandwidth_used(void)
3819 return static_key_false(&__cfs_bandwidth_used);
3822 void cfs_bandwidth_usage_inc(void)
3824 static_key_slow_inc(&__cfs_bandwidth_used);
3827 void cfs_bandwidth_usage_dec(void)
3829 static_key_slow_dec(&__cfs_bandwidth_used);
3831 #else /* HAVE_JUMP_LABEL */
3832 static bool cfs_bandwidth_used(void)
3837 void cfs_bandwidth_usage_inc(void) {}
3838 void cfs_bandwidth_usage_dec(void) {}
3839 #endif /* HAVE_JUMP_LABEL */
3842 * default period for cfs group bandwidth.
3843 * default: 0.1s, units: nanoseconds
3845 static inline u64 default_cfs_period(void)
3847 return 100000000ULL;
3850 static inline u64 sched_cfs_bandwidth_slice(void)
3852 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3856 * Replenish runtime according to assigned quota and update expiration time.
3857 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3858 * additional synchronization around rq->lock.
3860 * requires cfs_b->lock
3862 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3866 if (cfs_b->quota == RUNTIME_INF)
3869 now = sched_clock_cpu(smp_processor_id());
3870 cfs_b->runtime = cfs_b->quota;
3871 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3874 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3876 return &tg->cfs_bandwidth;
3879 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3880 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3882 if (unlikely(cfs_rq->throttle_count))
3883 return cfs_rq->throttled_clock_task;
3885 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3888 /* returns 0 on failure to allocate runtime */
3889 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3891 struct task_group *tg = cfs_rq->tg;
3892 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3893 u64 amount = 0, min_amount, expires;
3895 /* note: this is a positive sum as runtime_remaining <= 0 */
3896 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3898 raw_spin_lock(&cfs_b->lock);
3899 if (cfs_b->quota == RUNTIME_INF)
3900 amount = min_amount;
3902 start_cfs_bandwidth(cfs_b);
3904 if (cfs_b->runtime > 0) {
3905 amount = min(cfs_b->runtime, min_amount);
3906 cfs_b->runtime -= amount;
3910 expires = cfs_b->runtime_expires;
3911 raw_spin_unlock(&cfs_b->lock);
3913 cfs_rq->runtime_remaining += amount;
3915 * we may have advanced our local expiration to account for allowed
3916 * spread between our sched_clock and the one on which runtime was
3919 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3920 cfs_rq->runtime_expires = expires;
3922 return cfs_rq->runtime_remaining > 0;
3926 * Note: This depends on the synchronization provided by sched_clock and the
3927 * fact that rq->clock snapshots this value.
3929 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3931 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3933 /* if the deadline is ahead of our clock, nothing to do */
3934 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3937 if (cfs_rq->runtime_remaining < 0)
3941 * If the local deadline has passed we have to consider the
3942 * possibility that our sched_clock is 'fast' and the global deadline
3943 * has not truly expired.
3945 * Fortunately we can check determine whether this the case by checking
3946 * whether the global deadline has advanced. It is valid to compare
3947 * cfs_b->runtime_expires without any locks since we only care about
3948 * exact equality, so a partial write will still work.
3951 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3952 /* extend local deadline, drift is bounded above by 2 ticks */
3953 cfs_rq->runtime_expires += TICK_NSEC;
3955 /* global deadline is ahead, expiration has passed */
3956 cfs_rq->runtime_remaining = 0;
3960 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3962 /* dock delta_exec before expiring quota (as it could span periods) */
3963 cfs_rq->runtime_remaining -= delta_exec;
3964 expire_cfs_rq_runtime(cfs_rq);
3966 if (likely(cfs_rq->runtime_remaining > 0))
3970 * if we're unable to extend our runtime we resched so that the active
3971 * hierarchy can be throttled
3973 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3974 resched_curr(rq_of(cfs_rq));
3977 static __always_inline
3978 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3980 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3983 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3986 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3988 return cfs_bandwidth_used() && cfs_rq->throttled;
3991 /* check whether cfs_rq, or any parent, is throttled */
3992 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3994 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3998 * Ensure that neither of the group entities corresponding to src_cpu or
3999 * dest_cpu are members of a throttled hierarchy when performing group
4000 * load-balance operations.
4002 static inline int throttled_lb_pair(struct task_group *tg,
4003 int src_cpu, int dest_cpu)
4005 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4007 src_cfs_rq = tg->cfs_rq[src_cpu];
4008 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4010 return throttled_hierarchy(src_cfs_rq) ||
4011 throttled_hierarchy(dest_cfs_rq);
4014 /* updated child weight may affect parent so we have to do this bottom up */
4015 static int tg_unthrottle_up(struct task_group *tg, void *data)
4017 struct rq *rq = data;
4018 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4020 cfs_rq->throttle_count--;
4022 if (!cfs_rq->throttle_count) {
4023 /* adjust cfs_rq_clock_task() */
4024 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4025 cfs_rq->throttled_clock_task;
4032 static int tg_throttle_down(struct task_group *tg, void *data)
4034 struct rq *rq = data;
4035 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4037 /* group is entering throttled state, stop time */
4038 if (!cfs_rq->throttle_count)
4039 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4040 cfs_rq->throttle_count++;
4045 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4047 struct rq *rq = rq_of(cfs_rq);
4048 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4049 struct sched_entity *se;
4050 long task_delta, dequeue = 1;
4053 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4055 /* freeze hierarchy runnable averages while throttled */
4057 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4060 task_delta = cfs_rq->h_nr_running;
4061 for_each_sched_entity(se) {
4062 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4063 /* throttled entity or throttle-on-deactivate */
4068 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4069 qcfs_rq->h_nr_running -= task_delta;
4071 if (qcfs_rq->load.weight)
4076 sub_nr_running(rq, task_delta);
4078 cfs_rq->throttled = 1;
4079 cfs_rq->throttled_clock = rq_clock(rq);
4080 raw_spin_lock(&cfs_b->lock);
4081 empty = list_empty(&cfs_b->throttled_cfs_rq);
4084 * Add to the _head_ of the list, so that an already-started
4085 * distribute_cfs_runtime will not see us
4087 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4090 * If we're the first throttled task, make sure the bandwidth
4094 start_cfs_bandwidth(cfs_b);
4096 raw_spin_unlock(&cfs_b->lock);
4099 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4101 struct rq *rq = rq_of(cfs_rq);
4102 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4103 struct sched_entity *se;
4107 se = cfs_rq->tg->se[cpu_of(rq)];
4109 cfs_rq->throttled = 0;
4111 update_rq_clock(rq);
4113 raw_spin_lock(&cfs_b->lock);
4114 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4115 list_del_rcu(&cfs_rq->throttled_list);
4116 raw_spin_unlock(&cfs_b->lock);
4118 /* update hierarchical throttle state */
4119 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4121 if (!cfs_rq->load.weight)
4124 task_delta = cfs_rq->h_nr_running;
4125 for_each_sched_entity(se) {
4129 cfs_rq = cfs_rq_of(se);
4131 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4132 cfs_rq->h_nr_running += task_delta;
4134 if (cfs_rq_throttled(cfs_rq))
4139 add_nr_running(rq, task_delta);
4141 /* determine whether we need to wake up potentially idle cpu */
4142 if (rq->curr == rq->idle && rq->cfs.nr_running)
4146 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4147 u64 remaining, u64 expires)
4149 struct cfs_rq *cfs_rq;
4151 u64 starting_runtime = remaining;
4154 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4156 struct rq *rq = rq_of(cfs_rq);
4158 raw_spin_lock(&rq->lock);
4159 if (!cfs_rq_throttled(cfs_rq))
4162 runtime = -cfs_rq->runtime_remaining + 1;
4163 if (runtime > remaining)
4164 runtime = remaining;
4165 remaining -= runtime;
4167 cfs_rq->runtime_remaining += runtime;
4168 cfs_rq->runtime_expires = expires;
4170 /* we check whether we're throttled above */
4171 if (cfs_rq->runtime_remaining > 0)
4172 unthrottle_cfs_rq(cfs_rq);
4175 raw_spin_unlock(&rq->lock);
4182 return starting_runtime - remaining;
4186 * Responsible for refilling a task_group's bandwidth and unthrottling its
4187 * cfs_rqs as appropriate. If there has been no activity within the last
4188 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4189 * used to track this state.
4191 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4193 u64 runtime, runtime_expires;
4196 /* no need to continue the timer with no bandwidth constraint */
4197 if (cfs_b->quota == RUNTIME_INF)
4198 goto out_deactivate;
4200 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4201 cfs_b->nr_periods += overrun;
4204 * idle depends on !throttled (for the case of a large deficit), and if
4205 * we're going inactive then everything else can be deferred
4207 if (cfs_b->idle && !throttled)
4208 goto out_deactivate;
4210 __refill_cfs_bandwidth_runtime(cfs_b);
4213 /* mark as potentially idle for the upcoming period */
4218 /* account preceding periods in which throttling occurred */
4219 cfs_b->nr_throttled += overrun;
4221 runtime_expires = cfs_b->runtime_expires;
4224 * This check is repeated as we are holding onto the new bandwidth while
4225 * we unthrottle. This can potentially race with an unthrottled group
4226 * trying to acquire new bandwidth from the global pool. This can result
4227 * in us over-using our runtime if it is all used during this loop, but
4228 * only by limited amounts in that extreme case.
4230 while (throttled && cfs_b->runtime > 0) {
4231 runtime = cfs_b->runtime;
4232 raw_spin_unlock(&cfs_b->lock);
4233 /* we can't nest cfs_b->lock while distributing bandwidth */
4234 runtime = distribute_cfs_runtime(cfs_b, runtime,
4236 raw_spin_lock(&cfs_b->lock);
4238 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4240 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4244 * While we are ensured activity in the period following an
4245 * unthrottle, this also covers the case in which the new bandwidth is
4246 * insufficient to cover the existing bandwidth deficit. (Forcing the
4247 * timer to remain active while there are any throttled entities.)
4257 /* a cfs_rq won't donate quota below this amount */
4258 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4259 /* minimum remaining period time to redistribute slack quota */
4260 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4261 /* how long we wait to gather additional slack before distributing */
4262 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4265 * Are we near the end of the current quota period?
4267 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4268 * hrtimer base being cleared by hrtimer_start. In the case of
4269 * migrate_hrtimers, base is never cleared, so we are fine.
4271 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4273 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4276 /* if the call-back is running a quota refresh is already occurring */
4277 if (hrtimer_callback_running(refresh_timer))
4280 /* is a quota refresh about to occur? */
4281 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4282 if (remaining < min_expire)
4288 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4290 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4292 /* if there's a quota refresh soon don't bother with slack */
4293 if (runtime_refresh_within(cfs_b, min_left))
4296 hrtimer_start(&cfs_b->slack_timer,
4297 ns_to_ktime(cfs_bandwidth_slack_period),
4301 /* we know any runtime found here is valid as update_curr() precedes return */
4302 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4304 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4305 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4307 if (slack_runtime <= 0)
4310 raw_spin_lock(&cfs_b->lock);
4311 if (cfs_b->quota != RUNTIME_INF &&
4312 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4313 cfs_b->runtime += slack_runtime;
4315 /* we are under rq->lock, defer unthrottling using a timer */
4316 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4317 !list_empty(&cfs_b->throttled_cfs_rq))
4318 start_cfs_slack_bandwidth(cfs_b);
4320 raw_spin_unlock(&cfs_b->lock);
4322 /* even if it's not valid for return we don't want to try again */
4323 cfs_rq->runtime_remaining -= slack_runtime;
4326 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4328 if (!cfs_bandwidth_used())
4331 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4334 __return_cfs_rq_runtime(cfs_rq);
4338 * This is done with a timer (instead of inline with bandwidth return) since
4339 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4341 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4343 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4346 /* confirm we're still not at a refresh boundary */
4347 raw_spin_lock(&cfs_b->lock);
4348 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4349 raw_spin_unlock(&cfs_b->lock);
4353 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4354 runtime = cfs_b->runtime;
4356 expires = cfs_b->runtime_expires;
4357 raw_spin_unlock(&cfs_b->lock);
4362 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4364 raw_spin_lock(&cfs_b->lock);
4365 if (expires == cfs_b->runtime_expires)
4366 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4367 raw_spin_unlock(&cfs_b->lock);
4371 * When a group wakes up we want to make sure that its quota is not already
4372 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4373 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4375 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4377 if (!cfs_bandwidth_used())
4380 /* Synchronize hierarchical throttle counter: */
4381 if (unlikely(!cfs_rq->throttle_uptodate)) {
4382 struct rq *rq = rq_of(cfs_rq);
4383 struct cfs_rq *pcfs_rq;
4384 struct task_group *tg;
4386 cfs_rq->throttle_uptodate = 1;
4388 /* Get closest up-to-date node, because leaves go first: */
4389 for (tg = cfs_rq->tg->parent; tg; tg = tg->parent) {
4390 pcfs_rq = tg->cfs_rq[cpu_of(rq)];
4391 if (pcfs_rq->throttle_uptodate)
4395 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4396 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4400 /* an active group must be handled by the update_curr()->put() path */
4401 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4404 /* ensure the group is not already throttled */
4405 if (cfs_rq_throttled(cfs_rq))
4408 /* update runtime allocation */
4409 account_cfs_rq_runtime(cfs_rq, 0);
4410 if (cfs_rq->runtime_remaining <= 0)
4411 throttle_cfs_rq(cfs_rq);
4414 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4415 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4417 if (!cfs_bandwidth_used())
4420 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4424 * it's possible for a throttled entity to be forced into a running
4425 * state (e.g. set_curr_task), in this case we're finished.
4427 if (cfs_rq_throttled(cfs_rq))
4430 throttle_cfs_rq(cfs_rq);
4434 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4436 struct cfs_bandwidth *cfs_b =
4437 container_of(timer, struct cfs_bandwidth, slack_timer);
4439 do_sched_cfs_slack_timer(cfs_b);
4441 return HRTIMER_NORESTART;
4444 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4446 struct cfs_bandwidth *cfs_b =
4447 container_of(timer, struct cfs_bandwidth, period_timer);
4451 raw_spin_lock(&cfs_b->lock);
4453 overrun = hrtimer_forward_now(timer, cfs_b->period);
4457 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4460 cfs_b->period_active = 0;
4461 raw_spin_unlock(&cfs_b->lock);
4463 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4466 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4468 raw_spin_lock_init(&cfs_b->lock);
4470 cfs_b->quota = RUNTIME_INF;
4471 cfs_b->period = ns_to_ktime(default_cfs_period());
4473 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4474 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4475 cfs_b->period_timer.function = sched_cfs_period_timer;
4476 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4477 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4480 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4482 cfs_rq->runtime_enabled = 0;
4483 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4486 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4488 lockdep_assert_held(&cfs_b->lock);
4490 if (!cfs_b->period_active) {
4491 cfs_b->period_active = 1;
4492 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4493 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4497 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4499 /* init_cfs_bandwidth() was not called */
4500 if (!cfs_b->throttled_cfs_rq.next)
4503 hrtimer_cancel(&cfs_b->period_timer);
4504 hrtimer_cancel(&cfs_b->slack_timer);
4507 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4509 struct cfs_rq *cfs_rq;
4511 for_each_leaf_cfs_rq(rq, cfs_rq) {
4512 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4514 raw_spin_lock(&cfs_b->lock);
4515 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4516 raw_spin_unlock(&cfs_b->lock);
4520 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4522 struct cfs_rq *cfs_rq;
4524 for_each_leaf_cfs_rq(rq, cfs_rq) {
4525 if (!cfs_rq->runtime_enabled)
4529 * clock_task is not advancing so we just need to make sure
4530 * there's some valid quota amount
4532 cfs_rq->runtime_remaining = 1;
4534 * Offline rq is schedulable till cpu is completely disabled
4535 * in take_cpu_down(), so we prevent new cfs throttling here.
4537 cfs_rq->runtime_enabled = 0;
4539 if (cfs_rq_throttled(cfs_rq))
4540 unthrottle_cfs_rq(cfs_rq);
4544 #else /* CONFIG_CFS_BANDWIDTH */
4545 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4547 return rq_clock_task(rq_of(cfs_rq));
4550 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4551 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4552 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4553 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4555 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4560 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4565 static inline int throttled_lb_pair(struct task_group *tg,
4566 int src_cpu, int dest_cpu)
4571 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4573 #ifdef CONFIG_FAIR_GROUP_SCHED
4574 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4577 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4581 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4582 static inline void update_runtime_enabled(struct rq *rq) {}
4583 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4585 #endif /* CONFIG_CFS_BANDWIDTH */
4587 /**************************************************
4588 * CFS operations on tasks:
4591 #ifdef CONFIG_SCHED_HRTICK
4592 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4594 struct sched_entity *se = &p->se;
4595 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4597 WARN_ON(task_rq(p) != rq);
4599 if (cfs_rq->nr_running > 1) {
4600 u64 slice = sched_slice(cfs_rq, se);
4601 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4602 s64 delta = slice - ran;
4609 hrtick_start(rq, delta);
4614 * called from enqueue/dequeue and updates the hrtick when the
4615 * current task is from our class and nr_running is low enough
4618 static void hrtick_update(struct rq *rq)
4620 struct task_struct *curr = rq->curr;
4622 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4625 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4626 hrtick_start_fair(rq, curr);
4628 #else /* !CONFIG_SCHED_HRTICK */
4630 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4634 static inline void hrtick_update(struct rq *rq)
4640 static bool cpu_overutilized(int cpu);
4641 unsigned long boosted_cpu_util(int cpu);
4643 #define boosted_cpu_util(cpu) cpu_util(cpu)
4647 static void update_capacity_of(int cpu)
4649 unsigned long req_cap;
4654 /* Convert scale-invariant capacity to cpu. */
4655 req_cap = boosted_cpu_util(cpu);
4656 req_cap = req_cap * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
4657 set_cfs_cpu_capacity(cpu, true, req_cap);
4662 * The enqueue_task method is called before nr_running is
4663 * increased. Here we update the fair scheduling stats and
4664 * then put the task into the rbtree:
4667 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4669 struct cfs_rq *cfs_rq;
4670 struct sched_entity *se = &p->se;
4672 int task_new = flags & ENQUEUE_WAKEUP_NEW;
4673 int task_wakeup = flags & ENQUEUE_WAKEUP;
4677 * If in_iowait is set, the code below may not trigger any cpufreq
4678 * utilization updates, so do it here explicitly with the IOWAIT flag
4682 cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_IOWAIT);
4684 for_each_sched_entity(se) {
4687 cfs_rq = cfs_rq_of(se);
4688 enqueue_entity(cfs_rq, se, flags);
4691 * end evaluation on encountering a throttled cfs_rq
4693 * note: in the case of encountering a throttled cfs_rq we will
4694 * post the final h_nr_running increment below.
4696 if (cfs_rq_throttled(cfs_rq))
4698 cfs_rq->h_nr_running++;
4699 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4701 flags = ENQUEUE_WAKEUP;
4704 for_each_sched_entity(se) {
4705 cfs_rq = cfs_rq_of(se);
4706 cfs_rq->h_nr_running++;
4707 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4709 if (cfs_rq_throttled(cfs_rq))
4712 update_load_avg(se, UPDATE_TG);
4713 update_cfs_shares(cfs_rq);
4717 add_nr_running(rq, 1);
4722 * Update SchedTune accounting.
4724 * We do it before updating the CPU capacity to ensure the
4725 * boost value of the current task is accounted for in the
4726 * selection of the OPP.
4728 * We do it also in the case where we enqueue a throttled task;
4729 * we could argue that a throttled task should not boost a CPU,
4731 * a) properly implementing CPU boosting considering throttled
4732 * tasks will increase a lot the complexity of the solution
4733 * b) it's not easy to quantify the benefits introduced by
4734 * such a more complex solution.
4735 * Thus, for the time being we go for the simple solution and boost
4736 * also for throttled RQs.
4738 schedtune_enqueue_task(p, cpu_of(rq));
4741 walt_inc_cumulative_runnable_avg(rq, p);
4742 if (!task_new && !rq->rd->overutilized &&
4743 cpu_overutilized(rq->cpu)) {
4744 rq->rd->overutilized = true;
4745 trace_sched_overutilized(true);
4749 * We want to potentially trigger a freq switch
4750 * request only for tasks that are waking up; this is
4751 * because we get here also during load balancing, but
4752 * in these cases it seems wise to trigger as single
4753 * request after load balancing is done.
4755 if (task_new || task_wakeup)
4756 update_capacity_of(cpu_of(rq));
4759 #endif /* CONFIG_SMP */
4763 static void set_next_buddy(struct sched_entity *se);
4766 * The dequeue_task method is called before nr_running is
4767 * decreased. We remove the task from the rbtree and
4768 * update the fair scheduling stats:
4770 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4772 struct cfs_rq *cfs_rq;
4773 struct sched_entity *se = &p->se;
4774 int task_sleep = flags & DEQUEUE_SLEEP;
4776 for_each_sched_entity(se) {
4777 cfs_rq = cfs_rq_of(se);
4778 dequeue_entity(cfs_rq, se, flags);
4781 * end evaluation on encountering a throttled cfs_rq
4783 * note: in the case of encountering a throttled cfs_rq we will
4784 * post the final h_nr_running decrement below.
4786 if (cfs_rq_throttled(cfs_rq))
4788 cfs_rq->h_nr_running--;
4789 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4791 /* Don't dequeue parent if it has other entities besides us */
4792 if (cfs_rq->load.weight) {
4793 /* Avoid re-evaluating load for this entity: */
4794 se = parent_entity(se);
4796 * Bias pick_next to pick a task from this cfs_rq, as
4797 * p is sleeping when it is within its sched_slice.
4799 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4803 flags |= DEQUEUE_SLEEP;
4806 for_each_sched_entity(se) {
4807 cfs_rq = cfs_rq_of(se);
4808 cfs_rq->h_nr_running--;
4809 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4811 if (cfs_rq_throttled(cfs_rq))
4814 update_load_avg(se, UPDATE_TG);
4815 update_cfs_shares(cfs_rq);
4819 sub_nr_running(rq, 1);
4824 * Update SchedTune accounting
4826 * We do it before updating the CPU capacity to ensure the
4827 * boost value of the current task is accounted for in the
4828 * selection of the OPP.
4830 schedtune_dequeue_task(p, cpu_of(rq));
4833 walt_dec_cumulative_runnable_avg(rq, p);
4836 * We want to potentially trigger a freq switch
4837 * request only for tasks that are going to sleep;
4838 * this is because we get here also during load
4839 * balancing, but in these cases it seems wise to
4840 * trigger as single request after load balancing is
4844 if (rq->cfs.nr_running)
4845 update_capacity_of(cpu_of(rq));
4846 else if (sched_freq())
4847 set_cfs_cpu_capacity(cpu_of(rq), false, 0);
4851 #endif /* CONFIG_SMP */
4859 * per rq 'load' arrray crap; XXX kill this.
4863 * The exact cpuload at various idx values, calculated at every tick would be
4864 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4866 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4867 * on nth tick when cpu may be busy, then we have:
4868 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4869 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4871 * decay_load_missed() below does efficient calculation of
4872 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4873 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4875 * The calculation is approximated on a 128 point scale.
4876 * degrade_zero_ticks is the number of ticks after which load at any
4877 * particular idx is approximated to be zero.
4878 * degrade_factor is a precomputed table, a row for each load idx.
4879 * Each column corresponds to degradation factor for a power of two ticks,
4880 * based on 128 point scale.
4882 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4883 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4885 * With this power of 2 load factors, we can degrade the load n times
4886 * by looking at 1 bits in n and doing as many mult/shift instead of
4887 * n mult/shifts needed by the exact degradation.
4889 #define DEGRADE_SHIFT 7
4890 static const unsigned char
4891 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4892 static const unsigned char
4893 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4894 {0, 0, 0, 0, 0, 0, 0, 0},
4895 {64, 32, 8, 0, 0, 0, 0, 0},
4896 {96, 72, 40, 12, 1, 0, 0},
4897 {112, 98, 75, 43, 15, 1, 0},
4898 {120, 112, 98, 76, 45, 16, 2} };
4901 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4902 * would be when CPU is idle and so we just decay the old load without
4903 * adding any new load.
4905 static unsigned long
4906 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4910 if (!missed_updates)
4913 if (missed_updates >= degrade_zero_ticks[idx])
4917 return load >> missed_updates;
4919 while (missed_updates) {
4920 if (missed_updates % 2)
4921 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4923 missed_updates >>= 1;
4930 * Update rq->cpu_load[] statistics. This function is usually called every
4931 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4932 * every tick. We fix it up based on jiffies.
4934 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4935 unsigned long pending_updates)
4939 this_rq->nr_load_updates++;
4941 /* Update our load: */
4942 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4943 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4944 unsigned long old_load, new_load;
4946 /* scale is effectively 1 << i now, and >> i divides by scale */
4948 old_load = this_rq->cpu_load[i];
4949 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4950 new_load = this_load;
4952 * Round up the averaging division if load is increasing. This
4953 * prevents us from getting stuck on 9 if the load is 10, for
4956 if (new_load > old_load)
4957 new_load += scale - 1;
4959 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4962 sched_avg_update(this_rq);
4965 /* Used instead of source_load when we know the type == 0 */
4966 static unsigned long weighted_cpuload(const int cpu)
4968 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4971 #ifdef CONFIG_NO_HZ_COMMON
4973 * There is no sane way to deal with nohz on smp when using jiffies because the
4974 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4975 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4977 * Therefore we cannot use the delta approach from the regular tick since that
4978 * would seriously skew the load calculation. However we'll make do for those
4979 * updates happening while idle (nohz_idle_balance) or coming out of idle
4980 * (tick_nohz_idle_exit).
4982 * This means we might still be one tick off for nohz periods.
4986 * Called from nohz_idle_balance() to update the load ratings before doing the
4989 static void update_idle_cpu_load(struct rq *this_rq)
4991 unsigned long curr_jiffies = READ_ONCE(jiffies);
4992 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4993 unsigned long pending_updates;
4996 * bail if there's load or we're actually up-to-date.
4998 if (load || curr_jiffies == this_rq->last_load_update_tick)
5001 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5002 this_rq->last_load_update_tick = curr_jiffies;
5004 __update_cpu_load(this_rq, load, pending_updates);
5008 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
5010 void update_cpu_load_nohz(void)
5012 struct rq *this_rq = this_rq();
5013 unsigned long curr_jiffies = READ_ONCE(jiffies);
5014 unsigned long pending_updates;
5016 if (curr_jiffies == this_rq->last_load_update_tick)
5019 raw_spin_lock(&this_rq->lock);
5020 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5021 if (pending_updates) {
5022 this_rq->last_load_update_tick = curr_jiffies;
5024 * We were idle, this means load 0, the current load might be
5025 * !0 due to remote wakeups and the sort.
5027 __update_cpu_load(this_rq, 0, pending_updates);
5029 raw_spin_unlock(&this_rq->lock);
5031 #endif /* CONFIG_NO_HZ */
5034 * Called from scheduler_tick()
5036 void update_cpu_load_active(struct rq *this_rq)
5038 unsigned long load = weighted_cpuload(cpu_of(this_rq));
5040 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
5042 this_rq->last_load_update_tick = jiffies;
5043 __update_cpu_load(this_rq, load, 1);
5047 * Return a low guess at the load of a migration-source cpu weighted
5048 * according to the scheduling class and "nice" value.
5050 * We want to under-estimate the load of migration sources, to
5051 * balance conservatively.
5053 static unsigned long source_load(int cpu, int type)
5055 struct rq *rq = cpu_rq(cpu);
5056 unsigned long total = weighted_cpuload(cpu);
5058 if (type == 0 || !sched_feat(LB_BIAS))
5061 return min(rq->cpu_load[type-1], total);
5065 * Return a high guess at the load of a migration-target cpu weighted
5066 * according to the scheduling class and "nice" value.
5068 static unsigned long target_load(int cpu, int type)
5070 struct rq *rq = cpu_rq(cpu);
5071 unsigned long total = weighted_cpuload(cpu);
5073 if (type == 0 || !sched_feat(LB_BIAS))
5076 return max(rq->cpu_load[type-1], total);
5080 static unsigned long cpu_avg_load_per_task(int cpu)
5082 struct rq *rq = cpu_rq(cpu);
5083 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5084 unsigned long load_avg = weighted_cpuload(cpu);
5087 return load_avg / nr_running;
5092 static void record_wakee(struct task_struct *p)
5095 * Rough decay (wiping) for cost saving, don't worry
5096 * about the boundary, really active task won't care
5099 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5100 current->wakee_flips >>= 1;
5101 current->wakee_flip_decay_ts = jiffies;
5104 if (current->last_wakee != p) {
5105 current->last_wakee = p;
5106 current->wakee_flips++;
5110 static void task_waking_fair(struct task_struct *p)
5112 struct sched_entity *se = &p->se;
5113 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5116 #ifndef CONFIG_64BIT
5117 u64 min_vruntime_copy;
5120 min_vruntime_copy = cfs_rq->min_vruntime_copy;
5122 min_vruntime = cfs_rq->min_vruntime;
5123 } while (min_vruntime != min_vruntime_copy);
5125 min_vruntime = cfs_rq->min_vruntime;
5128 se->vruntime -= min_vruntime;
5132 #ifdef CONFIG_FAIR_GROUP_SCHED
5134 * effective_load() calculates the load change as seen from the root_task_group
5136 * Adding load to a group doesn't make a group heavier, but can cause movement
5137 * of group shares between cpus. Assuming the shares were perfectly aligned one
5138 * can calculate the shift in shares.
5140 * Calculate the effective load difference if @wl is added (subtracted) to @tg
5141 * on this @cpu and results in a total addition (subtraction) of @wg to the
5142 * total group weight.
5144 * Given a runqueue weight distribution (rw_i) we can compute a shares
5145 * distribution (s_i) using:
5147 * s_i = rw_i / \Sum rw_j (1)
5149 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
5150 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
5151 * shares distribution (s_i):
5153 * rw_i = { 2, 4, 1, 0 }
5154 * s_i = { 2/7, 4/7, 1/7, 0 }
5156 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
5157 * task used to run on and the CPU the waker is running on), we need to
5158 * compute the effect of waking a task on either CPU and, in case of a sync
5159 * wakeup, compute the effect of the current task going to sleep.
5161 * So for a change of @wl to the local @cpu with an overall group weight change
5162 * of @wl we can compute the new shares distribution (s'_i) using:
5164 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
5166 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
5167 * differences in waking a task to CPU 0. The additional task changes the
5168 * weight and shares distributions like:
5170 * rw'_i = { 3, 4, 1, 0 }
5171 * s'_i = { 3/8, 4/8, 1/8, 0 }
5173 * We can then compute the difference in effective weight by using:
5175 * dw_i = S * (s'_i - s_i) (3)
5177 * Where 'S' is the group weight as seen by its parent.
5179 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
5180 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
5181 * 4/7) times the weight of the group.
5183 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
5185 struct sched_entity *se = tg->se[cpu];
5187 if (!tg->parent) /* the trivial, non-cgroup case */
5190 for_each_sched_entity(se) {
5191 struct cfs_rq *cfs_rq = se->my_q;
5192 long W, w = cfs_rq_load_avg(cfs_rq);
5197 * W = @wg + \Sum rw_j
5199 W = wg + atomic_long_read(&tg->load_avg);
5201 /* Ensure \Sum rw_j >= rw_i */
5202 W -= cfs_rq->tg_load_avg_contrib;
5211 * wl = S * s'_i; see (2)
5214 wl = (w * (long)tg->shares) / W;
5219 * Per the above, wl is the new se->load.weight value; since
5220 * those are clipped to [MIN_SHARES, ...) do so now. See
5221 * calc_cfs_shares().
5223 if (wl < MIN_SHARES)
5227 * wl = dw_i = S * (s'_i - s_i); see (3)
5229 wl -= se->avg.load_avg;
5232 * Recursively apply this logic to all parent groups to compute
5233 * the final effective load change on the root group. Since
5234 * only the @tg group gets extra weight, all parent groups can
5235 * only redistribute existing shares. @wl is the shift in shares
5236 * resulting from this level per the above.
5245 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
5253 * Returns the current capacity of cpu after applying both
5254 * cpu and freq scaling.
5256 unsigned long capacity_curr_of(int cpu)
5258 return cpu_rq(cpu)->cpu_capacity_orig *
5259 arch_scale_freq_capacity(NULL, cpu)
5260 >> SCHED_CAPACITY_SHIFT;
5263 static inline bool energy_aware(void)
5265 return sched_feat(ENERGY_AWARE);
5269 struct sched_group *sg_top;
5270 struct sched_group *sg_cap;
5277 struct task_struct *task;
5292 * __cpu_norm_util() returns the cpu util relative to a specific capacity,
5293 * i.e. it's busy ratio, in the range [0..SCHED_LOAD_SCALE] which is useful for
5294 * energy calculations. Using the scale-invariant util returned by
5295 * cpu_util() and approximating scale-invariant util by:
5297 * util ~ (curr_freq/max_freq)*1024 * capacity_orig/1024 * running_time/time
5299 * the normalized util can be found using the specific capacity.
5301 * capacity = capacity_orig * curr_freq/max_freq
5303 * norm_util = running_time/time ~ util/capacity
5305 static unsigned long __cpu_norm_util(int cpu, unsigned long capacity, int delta)
5307 int util = __cpu_util(cpu, delta);
5309 if (util >= capacity)
5310 return SCHED_CAPACITY_SCALE;
5312 return (util << SCHED_CAPACITY_SHIFT)/capacity;
5315 static int calc_util_delta(struct energy_env *eenv, int cpu)
5317 if (cpu == eenv->src_cpu)
5318 return -eenv->util_delta;
5319 if (cpu == eenv->dst_cpu)
5320 return eenv->util_delta;
5325 unsigned long group_max_util(struct energy_env *eenv)
5328 unsigned long max_util = 0;
5330 for_each_cpu(i, sched_group_cpus(eenv->sg_cap)) {
5331 delta = calc_util_delta(eenv, i);
5332 max_util = max(max_util, __cpu_util(i, delta));
5339 * group_norm_util() returns the approximated group util relative to it's
5340 * current capacity (busy ratio) in the range [0..SCHED_LOAD_SCALE] for use in
5341 * energy calculations. Since task executions may or may not overlap in time in
5342 * the group the true normalized util is between max(cpu_norm_util(i)) and
5343 * sum(cpu_norm_util(i)) when iterating over all cpus in the group, i. The
5344 * latter is used as the estimate as it leads to a more pessimistic energy
5345 * estimate (more busy).
5348 long group_norm_util(struct energy_env *eenv, struct sched_group *sg)
5351 unsigned long util_sum = 0;
5352 unsigned long capacity = sg->sge->cap_states[eenv->cap_idx].cap;
5354 for_each_cpu(i, sched_group_cpus(sg)) {
5355 delta = calc_util_delta(eenv, i);
5356 util_sum += __cpu_norm_util(i, capacity, delta);
5359 if (util_sum > SCHED_CAPACITY_SCALE)
5360 return SCHED_CAPACITY_SCALE;
5364 static int find_new_capacity(struct energy_env *eenv,
5365 const struct sched_group_energy * const sge)
5368 unsigned long util = group_max_util(eenv);
5370 for (idx = 0; idx < sge->nr_cap_states; idx++) {
5371 if (sge->cap_states[idx].cap >= util)
5375 eenv->cap_idx = idx;
5380 static int group_idle_state(struct sched_group *sg)
5382 int i, state = INT_MAX;
5384 /* Find the shallowest idle state in the sched group. */
5385 for_each_cpu(i, sched_group_cpus(sg))
5386 state = min(state, idle_get_state_idx(cpu_rq(i)));
5388 /* Take non-cpuidle idling into account (active idle/arch_cpu_idle()) */
5395 * sched_group_energy(): Computes the absolute energy consumption of cpus
5396 * belonging to the sched_group including shared resources shared only by
5397 * members of the group. Iterates over all cpus in the hierarchy below the
5398 * sched_group starting from the bottom working it's way up before going to
5399 * the next cpu until all cpus are covered at all levels. The current
5400 * implementation is likely to gather the same util statistics multiple times.
5401 * This can probably be done in a faster but more complex way.
5402 * Note: sched_group_energy() may fail when racing with sched_domain updates.
5404 static int sched_group_energy(struct energy_env *eenv)
5406 struct sched_domain *sd;
5407 int cpu, total_energy = 0;
5408 struct cpumask visit_cpus;
5409 struct sched_group *sg;
5411 WARN_ON(!eenv->sg_top->sge);
5413 cpumask_copy(&visit_cpus, sched_group_cpus(eenv->sg_top));
5415 while (!cpumask_empty(&visit_cpus)) {
5416 struct sched_group *sg_shared_cap = NULL;
5418 cpu = cpumask_first(&visit_cpus);
5421 * Is the group utilization affected by cpus outside this
5424 sd = rcu_dereference(per_cpu(sd_scs, cpu));
5428 * We most probably raced with hotplug; returning a
5429 * wrong energy estimation is better than entering an
5435 sg_shared_cap = sd->parent->groups;
5437 for_each_domain(cpu, sd) {
5440 /* Has this sched_domain already been visited? */
5441 if (sd->child && group_first_cpu(sg) != cpu)
5445 unsigned long group_util;
5446 int sg_busy_energy, sg_idle_energy;
5447 int cap_idx, idle_idx;
5449 if (sg_shared_cap && sg_shared_cap->group_weight >= sg->group_weight)
5450 eenv->sg_cap = sg_shared_cap;
5454 cap_idx = find_new_capacity(eenv, sg->sge);
5456 if (sg->group_weight == 1) {
5457 /* Remove capacity of src CPU (before task move) */
5458 if (eenv->util_delta == 0 &&
5459 cpumask_test_cpu(eenv->src_cpu, sched_group_cpus(sg))) {
5460 eenv->cap.before = sg->sge->cap_states[cap_idx].cap;
5461 eenv->cap.delta -= eenv->cap.before;
5463 /* Add capacity of dst CPU (after task move) */
5464 if (eenv->util_delta != 0 &&
5465 cpumask_test_cpu(eenv->dst_cpu, sched_group_cpus(sg))) {
5466 eenv->cap.after = sg->sge->cap_states[cap_idx].cap;
5467 eenv->cap.delta += eenv->cap.after;
5471 idle_idx = group_idle_state(sg);
5472 group_util = group_norm_util(eenv, sg);
5473 sg_busy_energy = (group_util * sg->sge->cap_states[cap_idx].power)
5474 >> SCHED_CAPACITY_SHIFT;
5475 sg_idle_energy = ((SCHED_LOAD_SCALE-group_util)
5476 * sg->sge->idle_states[idle_idx].power)
5477 >> SCHED_CAPACITY_SHIFT;
5479 total_energy += sg_busy_energy + sg_idle_energy;
5482 cpumask_xor(&visit_cpus, &visit_cpus, sched_group_cpus(sg));
5484 if (cpumask_equal(sched_group_cpus(sg), sched_group_cpus(eenv->sg_top)))
5487 } while (sg = sg->next, sg != sd->groups);
5490 cpumask_clear_cpu(cpu, &visit_cpus);
5494 eenv->energy = total_energy;
5498 static inline bool cpu_in_sg(struct sched_group *sg, int cpu)
5500 return cpu != -1 && cpumask_test_cpu(cpu, sched_group_cpus(sg));
5504 * energy_diff(): Estimate the energy impact of changing the utilization
5505 * distribution. eenv specifies the change: utilisation amount, source, and
5506 * destination cpu. Source or destination cpu may be -1 in which case the
5507 * utilization is removed from or added to the system (e.g. task wake-up). If
5508 * both are specified, the utilization is migrated.
5510 static inline int __energy_diff(struct energy_env *eenv)
5512 struct sched_domain *sd;
5513 struct sched_group *sg;
5514 int sd_cpu = -1, energy_before = 0, energy_after = 0;
5517 struct energy_env eenv_before = {
5519 .src_cpu = eenv->src_cpu,
5520 .dst_cpu = eenv->dst_cpu,
5521 .nrg = { 0, 0, 0, 0},
5525 if (eenv->src_cpu == eenv->dst_cpu)
5528 sd_cpu = (eenv->src_cpu != -1) ? eenv->src_cpu : eenv->dst_cpu;
5529 sd = rcu_dereference(per_cpu(sd_ea, sd_cpu));
5532 return 0; /* Error */
5537 if (cpu_in_sg(sg, eenv->src_cpu) || cpu_in_sg(sg, eenv->dst_cpu)) {
5538 eenv_before.sg_top = eenv->sg_top = sg;
5540 if (sched_group_energy(&eenv_before))
5541 return 0; /* Invalid result abort */
5542 energy_before += eenv_before.energy;
5544 /* Keep track of SRC cpu (before) capacity */
5545 eenv->cap.before = eenv_before.cap.before;
5546 eenv->cap.delta = eenv_before.cap.delta;
5548 if (sched_group_energy(eenv))
5549 return 0; /* Invalid result abort */
5550 energy_after += eenv->energy;
5552 } while (sg = sg->next, sg != sd->groups);
5554 eenv->nrg.before = energy_before;
5555 eenv->nrg.after = energy_after;
5556 eenv->nrg.diff = eenv->nrg.after - eenv->nrg.before;
5559 trace_sched_energy_diff(eenv->task,
5560 eenv->src_cpu, eenv->dst_cpu, eenv->util_delta,
5561 eenv->nrg.before, eenv->nrg.after, eenv->nrg.diff,
5562 eenv->cap.before, eenv->cap.after, eenv->cap.delta,
5563 eenv->nrg.delta, eenv->payoff);
5566 * Dead-zone margin preventing too many migrations.
5569 margin = eenv->nrg.before >> 6; /* ~1.56% */
5571 diff = eenv->nrg.after - eenv->nrg.before;
5573 eenv->nrg.diff = (abs(diff) < margin) ? 0 : eenv->nrg.diff;
5575 return eenv->nrg.diff;
5578 #ifdef CONFIG_SCHED_TUNE
5580 struct target_nrg schedtune_target_nrg;
5583 * System energy normalization
5584 * Returns the normalized value, in the range [0..SCHED_LOAD_SCALE],
5585 * corresponding to the specified energy variation.
5588 normalize_energy(int energy_diff)
5591 #ifdef CONFIG_SCHED_DEBUG
5594 /* Check for boundaries */
5595 max_delta = schedtune_target_nrg.max_power;
5596 max_delta -= schedtune_target_nrg.min_power;
5597 WARN_ON(abs(energy_diff) >= max_delta);
5600 /* Do scaling using positive numbers to increase the range */
5601 normalized_nrg = (energy_diff < 0) ? -energy_diff : energy_diff;
5603 /* Scale by energy magnitude */
5604 normalized_nrg <<= SCHED_LOAD_SHIFT;
5606 /* Normalize on max energy for target platform */
5607 normalized_nrg = reciprocal_divide(
5608 normalized_nrg, schedtune_target_nrg.rdiv);
5610 return (energy_diff < 0) ? -normalized_nrg : normalized_nrg;
5614 energy_diff(struct energy_env *eenv)
5616 int boost = schedtune_task_boost(eenv->task);
5619 /* Conpute "absolute" energy diff */
5620 __energy_diff(eenv);
5622 /* Return energy diff when boost margin is 0 */
5624 return eenv->nrg.diff;
5626 /* Compute normalized energy diff */
5627 nrg_delta = normalize_energy(eenv->nrg.diff);
5628 eenv->nrg.delta = nrg_delta;
5630 eenv->payoff = schedtune_accept_deltas(
5636 * When SchedTune is enabled, the energy_diff() function will return
5637 * the computed energy payoff value. Since the energy_diff() return
5638 * value is expected to be negative by its callers, this evaluation
5639 * function return a negative value each time the evaluation return a
5640 * positive payoff, which is the condition for the acceptance of
5641 * a scheduling decision
5643 return -eenv->payoff;
5645 #else /* CONFIG_SCHED_TUNE */
5646 #define energy_diff(eenv) __energy_diff(eenv)
5650 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5651 * A waker of many should wake a different task than the one last awakened
5652 * at a frequency roughly N times higher than one of its wakees. In order
5653 * to determine whether we should let the load spread vs consolodating to
5654 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
5655 * partner, and a factor of lls_size higher frequency in the other. With
5656 * both conditions met, we can be relatively sure that the relationship is
5657 * non-monogamous, with partner count exceeding socket size. Waker/wakee
5658 * being client/server, worker/dispatcher, interrupt source or whatever is
5659 * irrelevant, spread criteria is apparent partner count exceeds socket size.
5661 static int wake_wide(struct task_struct *p)
5663 unsigned int master = current->wakee_flips;
5664 unsigned int slave = p->wakee_flips;
5665 int factor = this_cpu_read(sd_llc_size);
5668 swap(master, slave);
5669 if (slave < factor || master < slave * factor)
5674 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5675 int prev_cpu, int sync)
5677 s64 this_load, load;
5678 s64 this_eff_load, prev_eff_load;
5680 struct task_group *tg;
5681 unsigned long weight;
5685 this_cpu = smp_processor_id();
5686 load = source_load(prev_cpu, idx);
5687 this_load = target_load(this_cpu, idx);
5690 * If sync wakeup then subtract the (maximum possible)
5691 * effect of the currently running task from the load
5692 * of the current CPU:
5695 tg = task_group(current);
5696 weight = current->se.avg.load_avg;
5698 this_load += effective_load(tg, this_cpu, -weight, -weight);
5699 load += effective_load(tg, prev_cpu, 0, -weight);
5703 weight = p->se.avg.load_avg;
5706 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5707 * due to the sync cause above having dropped this_load to 0, we'll
5708 * always have an imbalance, but there's really nothing you can do
5709 * about that, so that's good too.
5711 * Otherwise check if either cpus are near enough in load to allow this
5712 * task to be woken on this_cpu.
5714 this_eff_load = 100;
5715 this_eff_load *= capacity_of(prev_cpu);
5717 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5718 prev_eff_load *= capacity_of(this_cpu);
5720 if (this_load > 0) {
5721 this_eff_load *= this_load +
5722 effective_load(tg, this_cpu, weight, weight);
5724 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5727 balanced = this_eff_load <= prev_eff_load;
5729 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5734 schedstat_inc(sd, ttwu_move_affine);
5735 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5740 static inline unsigned long task_util(struct task_struct *p)
5742 #ifdef CONFIG_SCHED_WALT
5743 if (!walt_disabled && sysctl_sched_use_walt_task_util) {
5744 unsigned long demand = p->ravg.demand;
5745 return (demand << 10) / walt_ravg_window;
5748 return p->se.avg.util_avg;
5751 static inline unsigned long boosted_task_util(struct task_struct *task);
5753 static inline bool __task_fits(struct task_struct *p, int cpu, int util)
5755 unsigned long capacity = capacity_of(cpu);
5757 util += boosted_task_util(p);
5759 return (capacity * 1024) > (util * capacity_margin);
5762 static inline bool task_fits_max(struct task_struct *p, int cpu)
5764 unsigned long capacity = capacity_of(cpu);
5765 unsigned long max_capacity = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5767 if (capacity == max_capacity)
5770 if (capacity * capacity_margin > max_capacity * 1024)
5773 return __task_fits(p, cpu, 0);
5776 static bool cpu_overutilized(int cpu)
5778 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5781 #ifdef CONFIG_SCHED_TUNE
5784 schedtune_margin(unsigned long signal, long boost)
5786 long long margin = 0;
5789 * Signal proportional compensation (SPC)
5791 * The Boost (B) value is used to compute a Margin (M) which is
5792 * proportional to the complement of the original Signal (S):
5793 * M = B * (SCHED_LOAD_SCALE - S), if B is positive
5794 * M = B * S, if B is negative
5795 * The obtained M could be used by the caller to "boost" S.
5798 margin = SCHED_LOAD_SCALE - signal;
5801 margin = -signal * boost;
5803 * Fast integer division by constant:
5804 * Constant : (C) = 100
5805 * Precision : 0.1% (P) = 0.1
5806 * Reference : C * 100 / P (R) = 100000
5809 * Shift bits : ceil(log(R,2)) (S) = 17
5810 * Mult const : round(2^S/C) (M) = 1311
5823 schedtune_cpu_margin(unsigned long util, int cpu)
5825 int boost = schedtune_cpu_boost(cpu);
5830 return schedtune_margin(util, boost);
5834 schedtune_task_margin(struct task_struct *task)
5836 int boost = schedtune_task_boost(task);
5843 util = task_util(task);
5844 margin = schedtune_margin(util, boost);
5849 #else /* CONFIG_SCHED_TUNE */
5852 schedtune_cpu_margin(unsigned long util, int cpu)
5858 schedtune_task_margin(struct task_struct *task)
5863 #endif /* CONFIG_SCHED_TUNE */
5866 boosted_cpu_util(int cpu)
5868 unsigned long util = cpu_util(cpu);
5869 long margin = schedtune_cpu_margin(util, cpu);
5871 trace_sched_boost_cpu(cpu, util, margin);
5873 return util + margin;
5876 static inline unsigned long
5877 boosted_task_util(struct task_struct *task)
5879 unsigned long util = task_util(task);
5880 long margin = schedtune_task_margin(task);
5882 trace_sched_boost_task(task, util, margin);
5884 return util + margin;
5887 static int cpu_util_wake(int cpu, struct task_struct *p);
5889 static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
5891 return capacity_orig_of(cpu) - cpu_util_wake(cpu, p);
5895 * find_idlest_group finds and returns the least busy CPU group within the
5898 static struct sched_group *
5899 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5900 int this_cpu, int sd_flag)
5902 struct sched_group *idlest = NULL, *group = sd->groups;
5903 struct sched_group *most_spare_sg = NULL;
5904 unsigned long min_load = ULONG_MAX, this_load = 0;
5905 unsigned long most_spare = 0, this_spare = 0;
5906 int load_idx = sd->forkexec_idx;
5907 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5909 if (sd_flag & SD_BALANCE_WAKE)
5910 load_idx = sd->wake_idx;
5913 unsigned long load, avg_load, spare_cap, max_spare_cap;
5917 /* Skip over this group if it has no CPUs allowed */
5918 if (!cpumask_intersects(sched_group_cpus(group),
5919 tsk_cpus_allowed(p)))
5922 local_group = cpumask_test_cpu(this_cpu,
5923 sched_group_cpus(group));
5926 * Tally up the load of all CPUs in the group and find
5927 * the group containing the CPU with most spare capacity.
5932 for_each_cpu(i, sched_group_cpus(group)) {
5933 /* Bias balancing toward cpus of our domain */
5935 load = source_load(i, load_idx);
5937 load = target_load(i, load_idx);
5941 spare_cap = capacity_spare_wake(i, p);
5943 if (spare_cap > max_spare_cap)
5944 max_spare_cap = spare_cap;
5947 /* Adjust by relative CPU capacity of the group */
5948 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5951 this_load = avg_load;
5952 this_spare = max_spare_cap;
5954 if (avg_load < min_load) {
5955 min_load = avg_load;
5959 if (most_spare < max_spare_cap) {
5960 most_spare = max_spare_cap;
5961 most_spare_sg = group;
5964 } while (group = group->next, group != sd->groups);
5967 * The cross-over point between using spare capacity or least load
5968 * is too conservative for high utilization tasks on partially
5969 * utilized systems if we require spare_capacity > task_util(p),
5970 * so we allow for some task stuffing by using
5971 * spare_capacity > task_util(p)/2.
5973 if (this_spare > task_util(p) / 2 &&
5974 imbalance*this_spare > 100*most_spare)
5976 else if (most_spare > task_util(p) / 2)
5977 return most_spare_sg;
5979 if (!idlest || 100*this_load < imbalance*min_load)
5985 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5988 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5990 unsigned long load, min_load = ULONG_MAX;
5991 unsigned int min_exit_latency = UINT_MAX;
5992 u64 latest_idle_timestamp = 0;
5993 int least_loaded_cpu = this_cpu;
5994 int shallowest_idle_cpu = -1;
5997 /* Check if we have any choice: */
5998 if (group->group_weight == 1)
5999 return cpumask_first(sched_group_cpus(group));
6001 /* Traverse only the allowed CPUs */
6002 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
6004 struct rq *rq = cpu_rq(i);
6005 struct cpuidle_state *idle = idle_get_state(rq);
6006 if (idle && idle->exit_latency < min_exit_latency) {
6008 * We give priority to a CPU whose idle state
6009 * has the smallest exit latency irrespective
6010 * of any idle timestamp.
6012 min_exit_latency = idle->exit_latency;
6013 latest_idle_timestamp = rq->idle_stamp;
6014 shallowest_idle_cpu = i;
6015 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
6016 rq->idle_stamp > latest_idle_timestamp) {
6018 * If equal or no active idle state, then
6019 * the most recently idled CPU might have
6022 latest_idle_timestamp = rq->idle_stamp;
6023 shallowest_idle_cpu = i;
6025 } else if (shallowest_idle_cpu == -1) {
6026 load = weighted_cpuload(i);
6027 if (load < min_load || (load == min_load && i == this_cpu)) {
6029 least_loaded_cpu = i;
6034 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
6038 * Try and locate an idle CPU in the sched_domain.
6040 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6042 struct sched_domain *sd;
6043 struct sched_group *sg;
6044 int best_idle_cpu = -1;
6045 int best_idle_cstate = INT_MAX;
6046 unsigned long best_idle_capacity = ULONG_MAX;
6048 if (!sysctl_sched_cstate_aware) {
6049 if (idle_cpu(target))
6053 * If the prevous cpu is cache affine and idle, don't be stupid.
6055 if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev))
6060 * Otherwise, iterate the domains and find an elegible idle cpu.
6062 sd = rcu_dereference(per_cpu(sd_llc, target));
6063 for_each_lower_domain(sd) {
6067 if (!cpumask_intersects(sched_group_cpus(sg),
6068 tsk_cpus_allowed(p)))
6071 if (sysctl_sched_cstate_aware) {
6072 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg)) {
6073 int idle_idx = idle_get_state_idx(cpu_rq(i));
6074 unsigned long new_usage = boosted_task_util(p);
6075 unsigned long capacity_orig = capacity_orig_of(i);
6077 if (new_usage > capacity_orig || !idle_cpu(i))
6080 if (i == target && new_usage <= capacity_curr_of(target))
6083 if (idle_idx < best_idle_cstate &&
6084 capacity_orig <= best_idle_capacity) {
6086 best_idle_cstate = idle_idx;
6087 best_idle_capacity = capacity_orig;
6091 for_each_cpu(i, sched_group_cpus(sg)) {
6092 if (i == target || !idle_cpu(i))
6096 target = cpumask_first_and(sched_group_cpus(sg),
6097 tsk_cpus_allowed(p));
6102 } while (sg != sd->groups);
6105 if (best_idle_cpu >= 0)
6106 target = best_idle_cpu;
6112 static int start_cpu(bool boosted)
6114 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
6116 RCU_LOCKDEP_WARN(rcu_read_lock_sched_held(),
6117 "sched RCU must be held");
6119 return boosted ? rd->max_cap_orig_cpu : rd->min_cap_orig_cpu;
6122 static inline int find_best_target(struct task_struct *p, bool boosted, bool prefer_idle)
6124 int target_cpu = -1;
6125 unsigned long target_util = prefer_idle ? ULONG_MAX : 0;
6126 unsigned long backup_capacity = ULONG_MAX;
6127 int best_idle_cpu = -1;
6128 int best_idle_cstate = INT_MAX;
6129 int backup_cpu = -1;
6130 unsigned long min_util = boosted_task_util(p);
6131 struct sched_domain *sd;
6132 struct sched_group *sg;
6133 int cpu = start_cpu(boosted);
6138 sd = rcu_dereference(per_cpu(sd_ea, cpu));
6148 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg)) {
6149 unsigned long cur_capacity, new_util;
6155 * p's blocked utilization is still accounted for on prev_cpu
6156 * so prev_cpu will receive a negative bias due to the double
6157 * accounting. However, the blocked utilization may be zero.
6159 new_util = cpu_util(i) + task_util(p);
6162 * Ensure minimum capacity to grant the required boost.
6163 * The target CPU can be already at a capacity level higher
6164 * than the one required to boost the task.
6166 new_util = max(min_util, new_util);
6168 if (new_util > capacity_orig_of(i))
6171 #ifdef CONFIG_SCHED_WALT
6172 if (walt_cpu_high_irqload(i))
6177 * Unconditionally favoring tasks that prefer idle cpus to
6180 if (idle_cpu(i) && prefer_idle)
6183 cur_capacity = capacity_curr_of(i);
6185 if (new_util < cur_capacity) {
6186 if (cpu_rq(i)->nr_running) {
6188 * Find a target cpu with the lowest/highest
6189 * utilization if prefer_idle/!prefer_idle.
6191 if ((prefer_idle && target_util > new_util) ||
6192 (!prefer_idle && target_util < new_util)) {
6193 target_util = new_util;
6196 } else if (!prefer_idle) {
6197 int idle_idx = idle_get_state_idx(cpu_rq(i));
6199 if (best_idle_cpu < 0 ||
6200 (sysctl_sched_cstate_aware &&
6201 best_idle_cstate > idle_idx)) {
6202 best_idle_cstate = idle_idx;
6206 } else if (backup_capacity > cur_capacity) {
6207 /* Find a backup cpu with least capacity. */
6208 backup_capacity = cur_capacity;
6212 } while (sg = sg->next, sg != sd->groups);
6215 target_cpu = best_idle_cpu >= 0 ? best_idle_cpu : backup_cpu;
6221 * cpu_util_wake: Compute cpu utilization with any contributions from
6222 * the waking task p removed.
6224 static int cpu_util_wake(int cpu, struct task_struct *p)
6226 unsigned long util, capacity;
6228 /* Task has no contribution or is new */
6229 if (cpu != task_cpu(p) || !p->se.avg.last_update_time)
6230 return cpu_util(cpu);
6232 capacity = capacity_orig_of(cpu);
6233 util = max_t(long, cpu_rq(cpu)->cfs.avg.util_avg - task_util(p), 0);
6235 return (util >= capacity) ? capacity : util;
6239 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6240 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6242 * In that case WAKE_AFFINE doesn't make sense and we'll let
6243 * BALANCE_WAKE sort things out.
6245 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6247 long min_cap, max_cap;
6249 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6250 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity.val;
6252 /* Minimum capacity is close to max, no need to abort wake_affine */
6253 if (max_cap - min_cap < max_cap >> 3)
6256 /* Bring task utilization in sync with prev_cpu */
6257 sync_entity_load_avg(&p->se);
6259 return min_cap * 1024 < task_util(p) * capacity_margin;
6262 static int select_energy_cpu_brute(struct task_struct *p, int prev_cpu, int sync)
6264 struct sched_domain *sd;
6265 int target_cpu = prev_cpu, tmp_target;
6266 bool boosted, prefer_idle;
6268 if (sysctl_sched_sync_hint_enable && sync) {
6269 int cpu = smp_processor_id();
6271 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
6276 #ifdef CONFIG_CGROUP_SCHEDTUNE
6277 boosted = schedtune_task_boost(p) > 0;
6278 prefer_idle = schedtune_prefer_idle(p) > 0;
6280 boosted = get_sysctl_sched_cfs_boost() > 0;
6284 sd = rcu_dereference(per_cpu(sd_ea, prev_cpu));
6285 /* Find a cpu with sufficient capacity */
6286 tmp_target = find_best_target(p, boosted, prefer_idle);
6290 if (tmp_target >= 0) {
6291 target_cpu = tmp_target;
6292 if ((boosted || prefer_idle) && idle_cpu(target_cpu))
6296 if (target_cpu != prev_cpu) {
6297 struct energy_env eenv = {
6298 .util_delta = task_util(p),
6299 .src_cpu = prev_cpu,
6300 .dst_cpu = target_cpu,
6304 /* Not enough spare capacity on previous cpu */
6305 if (cpu_overutilized(prev_cpu))
6308 if (energy_diff(&eenv) >= 0)
6309 target_cpu = prev_cpu;
6318 * select_task_rq_fair: Select target runqueue for the waking task in domains
6319 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6320 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6322 * Balances load by selecting the idlest cpu in the idlest group, or under
6323 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
6325 * Returns the target cpu number.
6327 * preempt must be disabled.
6330 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6332 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
6333 int cpu = smp_processor_id();
6334 int new_cpu = prev_cpu;
6335 int want_affine = 0;
6336 int sync = wake_flags & WF_SYNC;
6338 if (sd_flag & SD_BALANCE_WAKE)
6339 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
6340 && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
6342 if (energy_aware() && !(cpu_rq(prev_cpu)->rd->overutilized))
6343 return select_energy_cpu_brute(p, prev_cpu, sync);
6346 for_each_domain(cpu, tmp) {
6347 if (!(tmp->flags & SD_LOAD_BALANCE))
6351 * If both cpu and prev_cpu are part of this domain,
6352 * cpu is a valid SD_WAKE_AFFINE target.
6354 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6355 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6360 if (tmp->flags & sd_flag)
6362 else if (!want_affine)
6367 sd = NULL; /* Prefer wake_affine over balance flags */
6368 if (cpu != prev_cpu && wake_affine(affine_sd, p, prev_cpu, sync))
6373 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
6374 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6377 struct sched_group *group;
6380 if (!(sd->flags & sd_flag)) {
6385 group = find_idlest_group(sd, p, cpu, sd_flag);
6391 new_cpu = find_idlest_cpu(group, p, cpu);
6392 if (new_cpu == -1 || new_cpu == cpu) {
6393 /* Now try balancing at a lower domain level of cpu */
6398 /* Now try balancing at a lower domain level of new_cpu */
6400 weight = sd->span_weight;
6402 for_each_domain(cpu, tmp) {
6403 if (weight <= tmp->span_weight)
6405 if (tmp->flags & sd_flag)
6408 /* while loop will break here if sd == NULL */
6416 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
6417 * cfs_rq_of(p) references at time of call are still valid and identify the
6418 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
6419 * other assumptions, including the state of rq->lock, should be made.
6421 static void migrate_task_rq_fair(struct task_struct *p)
6424 * We are supposed to update the task to "current" time, then its up to date
6425 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
6426 * what current time is, so simply throw away the out-of-date time. This
6427 * will result in the wakee task is less decayed, but giving the wakee more
6428 * load sounds not bad.
6430 remove_entity_load_avg(&p->se);
6432 /* Tell new CPU we are migrated */
6433 p->se.avg.last_update_time = 0;
6435 /* We have migrated, no longer consider this task hot */
6436 p->se.exec_start = 0;
6439 static void task_dead_fair(struct task_struct *p)
6441 remove_entity_load_avg(&p->se);
6444 #define task_fits_max(p, cpu) true
6445 #endif /* CONFIG_SMP */
6447 static unsigned long
6448 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
6450 unsigned long gran = sysctl_sched_wakeup_granularity;
6453 * Since its curr running now, convert the gran from real-time
6454 * to virtual-time in his units.
6456 * By using 'se' instead of 'curr' we penalize light tasks, so
6457 * they get preempted easier. That is, if 'se' < 'curr' then
6458 * the resulting gran will be larger, therefore penalizing the
6459 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6460 * be smaller, again penalizing the lighter task.
6462 * This is especially important for buddies when the leftmost
6463 * task is higher priority than the buddy.
6465 return calc_delta_fair(gran, se);
6469 * Should 'se' preempt 'curr'.
6483 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6485 s64 gran, vdiff = curr->vruntime - se->vruntime;
6490 gran = wakeup_gran(curr, se);
6497 static void set_last_buddy(struct sched_entity *se)
6499 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6502 for_each_sched_entity(se)
6503 cfs_rq_of(se)->last = se;
6506 static void set_next_buddy(struct sched_entity *se)
6508 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6511 for_each_sched_entity(se)
6512 cfs_rq_of(se)->next = se;
6515 static void set_skip_buddy(struct sched_entity *se)
6517 for_each_sched_entity(se)
6518 cfs_rq_of(se)->skip = se;
6522 * Preempt the current task with a newly woken task if needed:
6524 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6526 struct task_struct *curr = rq->curr;
6527 struct sched_entity *se = &curr->se, *pse = &p->se;
6528 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6529 int scale = cfs_rq->nr_running >= sched_nr_latency;
6530 int next_buddy_marked = 0;
6532 if (unlikely(se == pse))
6536 * This is possible from callers such as attach_tasks(), in which we
6537 * unconditionally check_prempt_curr() after an enqueue (which may have
6538 * lead to a throttle). This both saves work and prevents false
6539 * next-buddy nomination below.
6541 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6544 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6545 set_next_buddy(pse);
6546 next_buddy_marked = 1;
6550 * We can come here with TIF_NEED_RESCHED already set from new task
6553 * Note: this also catches the edge-case of curr being in a throttled
6554 * group (e.g. via set_curr_task), since update_curr() (in the
6555 * enqueue of curr) will have resulted in resched being set. This
6556 * prevents us from potentially nominating it as a false LAST_BUDDY
6559 if (test_tsk_need_resched(curr))
6562 /* Idle tasks are by definition preempted by non-idle tasks. */
6563 if (unlikely(curr->policy == SCHED_IDLE) &&
6564 likely(p->policy != SCHED_IDLE))
6568 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6569 * is driven by the tick):
6571 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6574 find_matching_se(&se, &pse);
6575 update_curr(cfs_rq_of(se));
6577 if (wakeup_preempt_entity(se, pse) == 1) {
6579 * Bias pick_next to pick the sched entity that is
6580 * triggering this preemption.
6582 if (!next_buddy_marked)
6583 set_next_buddy(pse);
6592 * Only set the backward buddy when the current task is still
6593 * on the rq. This can happen when a wakeup gets interleaved
6594 * with schedule on the ->pre_schedule() or idle_balance()
6595 * point, either of which can * drop the rq lock.
6597 * Also, during early boot the idle thread is in the fair class,
6598 * for obvious reasons its a bad idea to schedule back to it.
6600 if (unlikely(!se->on_rq || curr == rq->idle))
6603 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6607 static struct task_struct *
6608 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
6610 struct cfs_rq *cfs_rq = &rq->cfs;
6611 struct sched_entity *se;
6612 struct task_struct *p;
6616 #ifdef CONFIG_FAIR_GROUP_SCHED
6617 if (!cfs_rq->nr_running)
6620 if (prev->sched_class != &fair_sched_class)
6624 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6625 * likely that a next task is from the same cgroup as the current.
6627 * Therefore attempt to avoid putting and setting the entire cgroup
6628 * hierarchy, only change the part that actually changes.
6632 struct sched_entity *curr = cfs_rq->curr;
6635 * Since we got here without doing put_prev_entity() we also
6636 * have to consider cfs_rq->curr. If it is still a runnable
6637 * entity, update_curr() will update its vruntime, otherwise
6638 * forget we've ever seen it.
6642 update_curr(cfs_rq);
6647 * This call to check_cfs_rq_runtime() will do the
6648 * throttle and dequeue its entity in the parent(s).
6649 * Therefore the 'simple' nr_running test will indeed
6652 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
6656 se = pick_next_entity(cfs_rq, curr);
6657 cfs_rq = group_cfs_rq(se);
6663 * Since we haven't yet done put_prev_entity and if the selected task
6664 * is a different task than we started out with, try and touch the
6665 * least amount of cfs_rqs.
6668 struct sched_entity *pse = &prev->se;
6670 while (!(cfs_rq = is_same_group(se, pse))) {
6671 int se_depth = se->depth;
6672 int pse_depth = pse->depth;
6674 if (se_depth <= pse_depth) {
6675 put_prev_entity(cfs_rq_of(pse), pse);
6676 pse = parent_entity(pse);
6678 if (se_depth >= pse_depth) {
6679 set_next_entity(cfs_rq_of(se), se);
6680 se = parent_entity(se);
6684 put_prev_entity(cfs_rq, pse);
6685 set_next_entity(cfs_rq, se);
6688 if (hrtick_enabled(rq))
6689 hrtick_start_fair(rq, p);
6691 rq->misfit_task = !task_fits_max(p, rq->cpu);
6698 if (!cfs_rq->nr_running)
6701 put_prev_task(rq, prev);
6704 se = pick_next_entity(cfs_rq, NULL);
6705 set_next_entity(cfs_rq, se);
6706 cfs_rq = group_cfs_rq(se);
6711 if (hrtick_enabled(rq))
6712 hrtick_start_fair(rq, p);
6714 rq->misfit_task = !task_fits_max(p, rq->cpu);
6719 rq->misfit_task = 0;
6721 * This is OK, because current is on_cpu, which avoids it being picked
6722 * for load-balance and preemption/IRQs are still disabled avoiding
6723 * further scheduler activity on it and we're being very careful to
6724 * re-start the picking loop.
6726 lockdep_unpin_lock(&rq->lock);
6727 new_tasks = idle_balance(rq);
6728 lockdep_pin_lock(&rq->lock);
6730 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6731 * possible for any higher priority task to appear. In that case we
6732 * must re-start the pick_next_entity() loop.
6744 * Account for a descheduled task:
6746 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6748 struct sched_entity *se = &prev->se;
6749 struct cfs_rq *cfs_rq;
6751 for_each_sched_entity(se) {
6752 cfs_rq = cfs_rq_of(se);
6753 put_prev_entity(cfs_rq, se);
6758 * sched_yield() is very simple
6760 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6762 static void yield_task_fair(struct rq *rq)
6764 struct task_struct *curr = rq->curr;
6765 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6766 struct sched_entity *se = &curr->se;
6769 * Are we the only task in the tree?
6771 if (unlikely(rq->nr_running == 1))
6774 clear_buddies(cfs_rq, se);
6776 if (curr->policy != SCHED_BATCH) {
6777 update_rq_clock(rq);
6779 * Update run-time statistics of the 'current'.
6781 update_curr(cfs_rq);
6783 * Tell update_rq_clock() that we've just updated,
6784 * so we don't do microscopic update in schedule()
6785 * and double the fastpath cost.
6787 rq_clock_skip_update(rq, true);
6793 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6795 struct sched_entity *se = &p->se;
6797 /* throttled hierarchies are not runnable */
6798 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6801 /* Tell the scheduler that we'd really like pse to run next. */
6804 yield_task_fair(rq);
6810 /**************************************************
6811 * Fair scheduling class load-balancing methods.
6815 * The purpose of load-balancing is to achieve the same basic fairness the
6816 * per-cpu scheduler provides, namely provide a proportional amount of compute
6817 * time to each task. This is expressed in the following equation:
6819 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6821 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6822 * W_i,0 is defined as:
6824 * W_i,0 = \Sum_j w_i,j (2)
6826 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6827 * is derived from the nice value as per prio_to_weight[].
6829 * The weight average is an exponential decay average of the instantaneous
6832 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6834 * C_i is the compute capacity of cpu i, typically it is the
6835 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6836 * can also include other factors [XXX].
6838 * To achieve this balance we define a measure of imbalance which follows
6839 * directly from (1):
6841 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6843 * We them move tasks around to minimize the imbalance. In the continuous
6844 * function space it is obvious this converges, in the discrete case we get
6845 * a few fun cases generally called infeasible weight scenarios.
6848 * - infeasible weights;
6849 * - local vs global optima in the discrete case. ]
6854 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6855 * for all i,j solution, we create a tree of cpus that follows the hardware
6856 * topology where each level pairs two lower groups (or better). This results
6857 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6858 * tree to only the first of the previous level and we decrease the frequency
6859 * of load-balance at each level inv. proportional to the number of cpus in
6865 * \Sum { --- * --- * 2^i } = O(n) (5)
6867 * `- size of each group
6868 * | | `- number of cpus doing load-balance
6870 * `- sum over all levels
6872 * Coupled with a limit on how many tasks we can migrate every balance pass,
6873 * this makes (5) the runtime complexity of the balancer.
6875 * An important property here is that each CPU is still (indirectly) connected
6876 * to every other cpu in at most O(log n) steps:
6878 * The adjacency matrix of the resulting graph is given by:
6881 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6884 * And you'll find that:
6886 * A^(log_2 n)_i,j != 0 for all i,j (7)
6888 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6889 * The task movement gives a factor of O(m), giving a convergence complexity
6892 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6897 * In order to avoid CPUs going idle while there's still work to do, new idle
6898 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6899 * tree itself instead of relying on other CPUs to bring it work.
6901 * This adds some complexity to both (5) and (8) but it reduces the total idle
6909 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6912 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6917 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6919 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6921 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6924 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6925 * rewrite all of this once again.]
6928 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6930 enum fbq_type { regular, remote, all };
6939 #define LBF_ALL_PINNED 0x01
6940 #define LBF_NEED_BREAK 0x02
6941 #define LBF_DST_PINNED 0x04
6942 #define LBF_SOME_PINNED 0x08
6945 struct sched_domain *sd;
6953 struct cpumask *dst_grpmask;
6955 enum cpu_idle_type idle;
6957 unsigned int src_grp_nr_running;
6958 /* The set of CPUs under consideration for load-balancing */
6959 struct cpumask *cpus;
6964 unsigned int loop_break;
6965 unsigned int loop_max;
6967 enum fbq_type fbq_type;
6968 enum group_type busiest_group_type;
6969 struct list_head tasks;
6973 * Is this task likely cache-hot:
6975 static int task_hot(struct task_struct *p, struct lb_env *env)
6979 lockdep_assert_held(&env->src_rq->lock);
6981 if (p->sched_class != &fair_sched_class)
6984 if (unlikely(p->policy == SCHED_IDLE))
6988 * Buddy candidates are cache hot:
6990 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6991 (&p->se == cfs_rq_of(&p->se)->next ||
6992 &p->se == cfs_rq_of(&p->se)->last))
6995 if (sysctl_sched_migration_cost == -1)
6997 if (sysctl_sched_migration_cost == 0)
7000 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7002 return delta < (s64)sysctl_sched_migration_cost;
7005 #ifdef CONFIG_NUMA_BALANCING
7007 * Returns 1, if task migration degrades locality
7008 * Returns 0, if task migration improves locality i.e migration preferred.
7009 * Returns -1, if task migration is not affected by locality.
7011 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7013 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7014 unsigned long src_faults, dst_faults;
7015 int src_nid, dst_nid;
7017 if (!static_branch_likely(&sched_numa_balancing))
7020 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7023 src_nid = cpu_to_node(env->src_cpu);
7024 dst_nid = cpu_to_node(env->dst_cpu);
7026 if (src_nid == dst_nid)
7029 /* Migrating away from the preferred node is always bad. */
7030 if (src_nid == p->numa_preferred_nid) {
7031 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7037 /* Encourage migration to the preferred node. */
7038 if (dst_nid == p->numa_preferred_nid)
7042 src_faults = group_faults(p, src_nid);
7043 dst_faults = group_faults(p, dst_nid);
7045 src_faults = task_faults(p, src_nid);
7046 dst_faults = task_faults(p, dst_nid);
7049 return dst_faults < src_faults;
7053 static inline int migrate_degrades_locality(struct task_struct *p,
7061 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7064 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7068 lockdep_assert_held(&env->src_rq->lock);
7071 * We do not migrate tasks that are:
7072 * 1) throttled_lb_pair, or
7073 * 2) cannot be migrated to this CPU due to cpus_allowed, or
7074 * 3) running (obviously), or
7075 * 4) are cache-hot on their current CPU.
7077 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7080 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
7083 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
7085 env->flags |= LBF_SOME_PINNED;
7088 * Remember if this task can be migrated to any other cpu in
7089 * our sched_group. We may want to revisit it if we couldn't
7090 * meet load balance goals by pulling other tasks on src_cpu.
7092 * Also avoid computing new_dst_cpu if we have already computed
7093 * one in current iteration.
7095 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
7098 /* Prevent to re-select dst_cpu via env's cpus */
7099 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7100 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
7101 env->flags |= LBF_DST_PINNED;
7102 env->new_dst_cpu = cpu;
7110 /* Record that we found atleast one task that could run on dst_cpu */
7111 env->flags &= ~LBF_ALL_PINNED;
7113 if (task_running(env->src_rq, p)) {
7114 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
7119 * Aggressive migration if:
7120 * 1) destination numa is preferred
7121 * 2) task is cache cold, or
7122 * 3) too many balance attempts have failed.
7124 tsk_cache_hot = migrate_degrades_locality(p, env);
7125 if (tsk_cache_hot == -1)
7126 tsk_cache_hot = task_hot(p, env);
7128 if (tsk_cache_hot <= 0 ||
7129 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7130 if (tsk_cache_hot == 1) {
7131 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
7132 schedstat_inc(p, se.statistics.nr_forced_migrations);
7137 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
7142 * detach_task() -- detach the task for the migration specified in env
7144 static void detach_task(struct task_struct *p, struct lb_env *env)
7146 lockdep_assert_held(&env->src_rq->lock);
7148 deactivate_task(env->src_rq, p, 0);
7149 p->on_rq = TASK_ON_RQ_MIGRATING;
7150 double_lock_balance(env->src_rq, env->dst_rq);
7151 set_task_cpu(p, env->dst_cpu);
7152 double_unlock_balance(env->src_rq, env->dst_rq);
7156 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7157 * part of active balancing operations within "domain".
7159 * Returns a task if successful and NULL otherwise.
7161 static struct task_struct *detach_one_task(struct lb_env *env)
7163 struct task_struct *p, *n;
7165 lockdep_assert_held(&env->src_rq->lock);
7167 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
7168 if (!can_migrate_task(p, env))
7171 detach_task(p, env);
7174 * Right now, this is only the second place where
7175 * lb_gained[env->idle] is updated (other is detach_tasks)
7176 * so we can safely collect stats here rather than
7177 * inside detach_tasks().
7179 schedstat_inc(env->sd, lb_gained[env->idle]);
7185 static const unsigned int sched_nr_migrate_break = 32;
7188 * detach_tasks() -- tries to detach up to imbalance weighted load from
7189 * busiest_rq, as part of a balancing operation within domain "sd".
7191 * Returns number of detached tasks if successful and 0 otherwise.
7193 static int detach_tasks(struct lb_env *env)
7195 struct list_head *tasks = &env->src_rq->cfs_tasks;
7196 struct task_struct *p;
7200 lockdep_assert_held(&env->src_rq->lock);
7202 if (env->imbalance <= 0)
7205 while (!list_empty(tasks)) {
7207 * We don't want to steal all, otherwise we may be treated likewise,
7208 * which could at worst lead to a livelock crash.
7210 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7213 p = list_first_entry(tasks, struct task_struct, se.group_node);
7216 /* We've more or less seen every task there is, call it quits */
7217 if (env->loop > env->loop_max)
7220 /* take a breather every nr_migrate tasks */
7221 if (env->loop > env->loop_break) {
7222 env->loop_break += sched_nr_migrate_break;
7223 env->flags |= LBF_NEED_BREAK;
7227 if (!can_migrate_task(p, env))
7230 load = task_h_load(p);
7232 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7235 if ((load / 2) > env->imbalance)
7238 detach_task(p, env);
7239 list_add(&p->se.group_node, &env->tasks);
7242 env->imbalance -= load;
7244 #ifdef CONFIG_PREEMPT
7246 * NEWIDLE balancing is a source of latency, so preemptible
7247 * kernels will stop after the first task is detached to minimize
7248 * the critical section.
7250 if (env->idle == CPU_NEWLY_IDLE)
7255 * We only want to steal up to the prescribed amount of
7258 if (env->imbalance <= 0)
7263 list_move_tail(&p->se.group_node, tasks);
7267 * Right now, this is one of only two places we collect this stat
7268 * so we can safely collect detach_one_task() stats here rather
7269 * than inside detach_one_task().
7271 schedstat_add(env->sd, lb_gained[env->idle], detached);
7277 * attach_task() -- attach the task detached by detach_task() to its new rq.
7279 static void attach_task(struct rq *rq, struct task_struct *p)
7281 lockdep_assert_held(&rq->lock);
7283 BUG_ON(task_rq(p) != rq);
7284 p->on_rq = TASK_ON_RQ_QUEUED;
7285 activate_task(rq, p, 0);
7286 check_preempt_curr(rq, p, 0);
7290 * attach_one_task() -- attaches the task returned from detach_one_task() to
7293 static void attach_one_task(struct rq *rq, struct task_struct *p)
7295 raw_spin_lock(&rq->lock);
7298 * We want to potentially raise target_cpu's OPP.
7300 update_capacity_of(cpu_of(rq));
7301 raw_spin_unlock(&rq->lock);
7305 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7308 static void attach_tasks(struct lb_env *env)
7310 struct list_head *tasks = &env->tasks;
7311 struct task_struct *p;
7313 raw_spin_lock(&env->dst_rq->lock);
7315 while (!list_empty(tasks)) {
7316 p = list_first_entry(tasks, struct task_struct, se.group_node);
7317 list_del_init(&p->se.group_node);
7319 attach_task(env->dst_rq, p);
7323 * We want to potentially raise env.dst_cpu's OPP.
7325 update_capacity_of(env->dst_cpu);
7327 raw_spin_unlock(&env->dst_rq->lock);
7330 #ifdef CONFIG_FAIR_GROUP_SCHED
7331 static void update_blocked_averages(int cpu)
7333 struct rq *rq = cpu_rq(cpu);
7334 struct cfs_rq *cfs_rq;
7335 unsigned long flags;
7337 raw_spin_lock_irqsave(&rq->lock, flags);
7338 update_rq_clock(rq);
7341 * Iterates the task_group tree in a bottom up fashion, see
7342 * list_add_leaf_cfs_rq() for details.
7344 for_each_leaf_cfs_rq(rq, cfs_rq) {
7345 /* throttled entities do not contribute to load */
7346 if (throttled_hierarchy(cfs_rq))
7349 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq,
7351 update_tg_load_avg(cfs_rq, 0);
7353 /* Propagate pending load changes to the parent */
7354 if (cfs_rq->tg->se[cpu])
7355 update_load_avg(cfs_rq->tg->se[cpu], 0);
7357 raw_spin_unlock_irqrestore(&rq->lock, flags);
7361 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7362 * This needs to be done in a top-down fashion because the load of a child
7363 * group is a fraction of its parents load.
7365 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7367 struct rq *rq = rq_of(cfs_rq);
7368 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7369 unsigned long now = jiffies;
7372 if (cfs_rq->last_h_load_update == now)
7375 cfs_rq->h_load_next = NULL;
7376 for_each_sched_entity(se) {
7377 cfs_rq = cfs_rq_of(se);
7378 cfs_rq->h_load_next = se;
7379 if (cfs_rq->last_h_load_update == now)
7384 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7385 cfs_rq->last_h_load_update = now;
7388 while ((se = cfs_rq->h_load_next) != NULL) {
7389 load = cfs_rq->h_load;
7390 load = div64_ul(load * se->avg.load_avg,
7391 cfs_rq_load_avg(cfs_rq) + 1);
7392 cfs_rq = group_cfs_rq(se);
7393 cfs_rq->h_load = load;
7394 cfs_rq->last_h_load_update = now;
7398 static unsigned long task_h_load(struct task_struct *p)
7400 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7402 update_cfs_rq_h_load(cfs_rq);
7403 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7404 cfs_rq_load_avg(cfs_rq) + 1);
7407 static inline void update_blocked_averages(int cpu)
7409 struct rq *rq = cpu_rq(cpu);
7410 struct cfs_rq *cfs_rq = &rq->cfs;
7411 unsigned long flags;
7413 raw_spin_lock_irqsave(&rq->lock, flags);
7414 update_rq_clock(rq);
7415 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true);
7416 raw_spin_unlock_irqrestore(&rq->lock, flags);
7419 static unsigned long task_h_load(struct task_struct *p)
7421 return p->se.avg.load_avg;
7425 /********** Helpers for find_busiest_group ************************/
7428 * sg_lb_stats - stats of a sched_group required for load_balancing
7430 struct sg_lb_stats {
7431 unsigned long avg_load; /*Avg load across the CPUs of the group */
7432 unsigned long group_load; /* Total load over the CPUs of the group */
7433 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7434 unsigned long load_per_task;
7435 unsigned long group_capacity;
7436 unsigned long group_util; /* Total utilization of the group */
7437 unsigned int sum_nr_running; /* Nr tasks running in the group */
7438 unsigned int idle_cpus;
7439 unsigned int group_weight;
7440 enum group_type group_type;
7441 int group_no_capacity;
7442 int group_misfit_task; /* A cpu has a task too big for its capacity */
7443 #ifdef CONFIG_NUMA_BALANCING
7444 unsigned int nr_numa_running;
7445 unsigned int nr_preferred_running;
7450 * sd_lb_stats - Structure to store the statistics of a sched_domain
7451 * during load balancing.
7453 struct sd_lb_stats {
7454 struct sched_group *busiest; /* Busiest group in this sd */
7455 struct sched_group *local; /* Local group in this sd */
7456 unsigned long total_load; /* Total load of all groups in sd */
7457 unsigned long total_capacity; /* Total capacity of all groups in sd */
7458 unsigned long avg_load; /* Average load across all groups in sd */
7460 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7461 struct sg_lb_stats local_stat; /* Statistics of the local group */
7464 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7467 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7468 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7469 * We must however clear busiest_stat::avg_load because
7470 * update_sd_pick_busiest() reads this before assignment.
7472 *sds = (struct sd_lb_stats){
7476 .total_capacity = 0UL,
7479 .sum_nr_running = 0,
7480 .group_type = group_other,
7486 * get_sd_load_idx - Obtain the load index for a given sched domain.
7487 * @sd: The sched_domain whose load_idx is to be obtained.
7488 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7490 * Return: The load index.
7492 static inline int get_sd_load_idx(struct sched_domain *sd,
7493 enum cpu_idle_type idle)
7499 load_idx = sd->busy_idx;
7502 case CPU_NEWLY_IDLE:
7503 load_idx = sd->newidle_idx;
7506 load_idx = sd->idle_idx;
7513 static unsigned long scale_rt_capacity(int cpu)
7515 struct rq *rq = cpu_rq(cpu);
7516 u64 total, used, age_stamp, avg;
7520 * Since we're reading these variables without serialization make sure
7521 * we read them once before doing sanity checks on them.
7523 age_stamp = READ_ONCE(rq->age_stamp);
7524 avg = READ_ONCE(rq->rt_avg);
7525 delta = __rq_clock_broken(rq) - age_stamp;
7527 if (unlikely(delta < 0))
7530 total = sched_avg_period() + delta;
7532 used = div_u64(avg, total);
7535 * deadline bandwidth is defined at system level so we must
7536 * weight this bandwidth with the max capacity of the system.
7537 * As a reminder, avg_bw is 20bits width and
7538 * scale_cpu_capacity is 10 bits width
7540 used += div_u64(rq->dl.avg_bw, arch_scale_cpu_capacity(NULL, cpu));
7542 if (likely(used < SCHED_CAPACITY_SCALE))
7543 return SCHED_CAPACITY_SCALE - used;
7548 void init_max_cpu_capacity(struct max_cpu_capacity *mcc)
7550 raw_spin_lock_init(&mcc->lock);
7555 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7557 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7558 struct sched_group *sdg = sd->groups;
7559 struct max_cpu_capacity *mcc;
7560 unsigned long max_capacity;
7562 unsigned long flags;
7564 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7566 mcc = &cpu_rq(cpu)->rd->max_cpu_capacity;
7568 raw_spin_lock_irqsave(&mcc->lock, flags);
7569 max_capacity = mcc->val;
7570 max_cap_cpu = mcc->cpu;
7572 if ((max_capacity > capacity && max_cap_cpu == cpu) ||
7573 (max_capacity < capacity)) {
7574 mcc->val = capacity;
7576 #ifdef CONFIG_SCHED_DEBUG
7577 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7578 printk_deferred(KERN_INFO "CPU%d: update max cpu_capacity %lu\n",
7583 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7585 skip_unlock: __attribute__ ((unused));
7586 capacity *= scale_rt_capacity(cpu);
7587 capacity >>= SCHED_CAPACITY_SHIFT;
7592 cpu_rq(cpu)->cpu_capacity = capacity;
7593 sdg->sgc->capacity = capacity;
7594 sdg->sgc->max_capacity = capacity;
7595 sdg->sgc->min_capacity = capacity;
7598 void update_group_capacity(struct sched_domain *sd, int cpu)
7600 struct sched_domain *child = sd->child;
7601 struct sched_group *group, *sdg = sd->groups;
7602 unsigned long capacity, max_capacity, min_capacity;
7603 unsigned long interval;
7605 interval = msecs_to_jiffies(sd->balance_interval);
7606 interval = clamp(interval, 1UL, max_load_balance_interval);
7607 sdg->sgc->next_update = jiffies + interval;
7610 update_cpu_capacity(sd, cpu);
7616 min_capacity = ULONG_MAX;
7618 if (child->flags & SD_OVERLAP) {
7620 * SD_OVERLAP domains cannot assume that child groups
7621 * span the current group.
7624 for_each_cpu(cpu, sched_group_cpus(sdg)) {
7625 struct sched_group_capacity *sgc;
7626 struct rq *rq = cpu_rq(cpu);
7629 * build_sched_domains() -> init_sched_groups_capacity()
7630 * gets here before we've attached the domains to the
7633 * Use capacity_of(), which is set irrespective of domains
7634 * in update_cpu_capacity().
7636 * This avoids capacity from being 0 and
7637 * causing divide-by-zero issues on boot.
7639 if (unlikely(!rq->sd)) {
7640 capacity += capacity_of(cpu);
7642 sgc = rq->sd->groups->sgc;
7643 capacity += sgc->capacity;
7646 max_capacity = max(capacity, max_capacity);
7647 min_capacity = min(capacity, min_capacity);
7651 * !SD_OVERLAP domains can assume that child groups
7652 * span the current group.
7655 group = child->groups;
7657 struct sched_group_capacity *sgc = group->sgc;
7659 capacity += sgc->capacity;
7660 max_capacity = max(sgc->max_capacity, max_capacity);
7661 min_capacity = min(sgc->min_capacity, min_capacity);
7662 group = group->next;
7663 } while (group != child->groups);
7666 sdg->sgc->capacity = capacity;
7667 sdg->sgc->max_capacity = max_capacity;
7668 sdg->sgc->min_capacity = min_capacity;
7672 * Check whether the capacity of the rq has been noticeably reduced by side
7673 * activity. The imbalance_pct is used for the threshold.
7674 * Return true is the capacity is reduced
7677 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7679 return ((rq->cpu_capacity * sd->imbalance_pct) <
7680 (rq->cpu_capacity_orig * 100));
7684 * Group imbalance indicates (and tries to solve) the problem where balancing
7685 * groups is inadequate due to tsk_cpus_allowed() constraints.
7687 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7688 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7691 * { 0 1 2 3 } { 4 5 6 7 }
7694 * If we were to balance group-wise we'd place two tasks in the first group and
7695 * two tasks in the second group. Clearly this is undesired as it will overload
7696 * cpu 3 and leave one of the cpus in the second group unused.
7698 * The current solution to this issue is detecting the skew in the first group
7699 * by noticing the lower domain failed to reach balance and had difficulty
7700 * moving tasks due to affinity constraints.
7702 * When this is so detected; this group becomes a candidate for busiest; see
7703 * update_sd_pick_busiest(). And calculate_imbalance() and
7704 * find_busiest_group() avoid some of the usual balance conditions to allow it
7705 * to create an effective group imbalance.
7707 * This is a somewhat tricky proposition since the next run might not find the
7708 * group imbalance and decide the groups need to be balanced again. A most
7709 * subtle and fragile situation.
7712 static inline int sg_imbalanced(struct sched_group *group)
7714 return group->sgc->imbalance;
7718 * group_has_capacity returns true if the group has spare capacity that could
7719 * be used by some tasks.
7720 * We consider that a group has spare capacity if the * number of task is
7721 * smaller than the number of CPUs or if the utilization is lower than the
7722 * available capacity for CFS tasks.
7723 * For the latter, we use a threshold to stabilize the state, to take into
7724 * account the variance of the tasks' load and to return true if the available
7725 * capacity in meaningful for the load balancer.
7726 * As an example, an available capacity of 1% can appear but it doesn't make
7727 * any benefit for the load balance.
7730 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7732 if (sgs->sum_nr_running < sgs->group_weight)
7735 if ((sgs->group_capacity * 100) >
7736 (sgs->group_util * env->sd->imbalance_pct))
7743 * group_is_overloaded returns true if the group has more tasks than it can
7745 * group_is_overloaded is not equals to !group_has_capacity because a group
7746 * with the exact right number of tasks, has no more spare capacity but is not
7747 * overloaded so both group_has_capacity and group_is_overloaded return
7751 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7753 if (sgs->sum_nr_running <= sgs->group_weight)
7756 if ((sgs->group_capacity * 100) <
7757 (sgs->group_util * env->sd->imbalance_pct))
7765 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7766 * per-cpu capacity than sched_group ref.
7769 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7771 return sg->sgc->max_capacity + capacity_margin - SCHED_LOAD_SCALE <
7772 ref->sgc->max_capacity;
7776 group_type group_classify(struct sched_group *group,
7777 struct sg_lb_stats *sgs)
7779 if (sgs->group_no_capacity)
7780 return group_overloaded;
7782 if (sg_imbalanced(group))
7783 return group_imbalanced;
7785 if (sgs->group_misfit_task)
7786 return group_misfit_task;
7792 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7793 * @env: The load balancing environment.
7794 * @group: sched_group whose statistics are to be updated.
7795 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7796 * @local_group: Does group contain this_cpu.
7797 * @sgs: variable to hold the statistics for this group.
7798 * @overload: Indicate more than one runnable task for any CPU.
7799 * @overutilized: Indicate overutilization for any CPU.
7801 static inline void update_sg_lb_stats(struct lb_env *env,
7802 struct sched_group *group, int load_idx,
7803 int local_group, struct sg_lb_stats *sgs,
7804 bool *overload, bool *overutilized)
7809 memset(sgs, 0, sizeof(*sgs));
7811 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7812 struct rq *rq = cpu_rq(i);
7814 /* Bias balancing toward cpus of our domain */
7816 load = target_load(i, load_idx);
7818 load = source_load(i, load_idx);
7820 sgs->group_load += load;
7821 sgs->group_util += cpu_util(i);
7822 sgs->sum_nr_running += rq->cfs.h_nr_running;
7824 nr_running = rq->nr_running;
7828 #ifdef CONFIG_NUMA_BALANCING
7829 sgs->nr_numa_running += rq->nr_numa_running;
7830 sgs->nr_preferred_running += rq->nr_preferred_running;
7832 sgs->sum_weighted_load += weighted_cpuload(i);
7834 * No need to call idle_cpu() if nr_running is not 0
7836 if (!nr_running && idle_cpu(i))
7839 if (cpu_overutilized(i)) {
7840 *overutilized = true;
7841 if (!sgs->group_misfit_task && rq->misfit_task)
7842 sgs->group_misfit_task = capacity_of(i);
7846 /* Adjust by relative CPU capacity of the group */
7847 sgs->group_capacity = group->sgc->capacity;
7848 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7850 if (sgs->sum_nr_running)
7851 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7853 sgs->group_weight = group->group_weight;
7855 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7856 sgs->group_type = group_classify(group, sgs);
7860 * update_sd_pick_busiest - return 1 on busiest group
7861 * @env: The load balancing environment.
7862 * @sds: sched_domain statistics
7863 * @sg: sched_group candidate to be checked for being the busiest
7864 * @sgs: sched_group statistics
7866 * Determine if @sg is a busier group than the previously selected
7869 * Return: %true if @sg is a busier group than the previously selected
7870 * busiest group. %false otherwise.
7872 static bool update_sd_pick_busiest(struct lb_env *env,
7873 struct sd_lb_stats *sds,
7874 struct sched_group *sg,
7875 struct sg_lb_stats *sgs)
7877 struct sg_lb_stats *busiest = &sds->busiest_stat;
7879 if (sgs->group_type > busiest->group_type)
7882 if (sgs->group_type < busiest->group_type)
7886 * Candidate sg doesn't face any serious load-balance problems
7887 * so don't pick it if the local sg is already filled up.
7889 if (sgs->group_type == group_other &&
7890 !group_has_capacity(env, &sds->local_stat))
7893 if (sgs->avg_load <= busiest->avg_load)
7896 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
7900 * Candidate sg has no more than one task per CPU and
7901 * has higher per-CPU capacity. Migrating tasks to less
7902 * capable CPUs may harm throughput. Maximize throughput,
7903 * power/energy consequences are not considered.
7905 if (sgs->sum_nr_running <= sgs->group_weight &&
7906 group_smaller_cpu_capacity(sds->local, sg))
7910 /* This is the busiest node in its class. */
7911 if (!(env->sd->flags & SD_ASYM_PACKING))
7915 * ASYM_PACKING needs to move all the work to the lowest
7916 * numbered CPUs in the group, therefore mark all groups
7917 * higher than ourself as busy.
7919 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
7923 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
7930 #ifdef CONFIG_NUMA_BALANCING
7931 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7933 if (sgs->sum_nr_running > sgs->nr_numa_running)
7935 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7940 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7942 if (rq->nr_running > rq->nr_numa_running)
7944 if (rq->nr_running > rq->nr_preferred_running)
7949 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7954 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7958 #endif /* CONFIG_NUMA_BALANCING */
7961 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7962 * @env: The load balancing environment.
7963 * @sds: variable to hold the statistics for this sched_domain.
7965 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7967 struct sched_domain *child = env->sd->child;
7968 struct sched_group *sg = env->sd->groups;
7969 struct sg_lb_stats tmp_sgs;
7970 int load_idx, prefer_sibling = 0;
7971 bool overload = false, overutilized = false;
7973 if (child && child->flags & SD_PREFER_SIBLING)
7976 load_idx = get_sd_load_idx(env->sd, env->idle);
7979 struct sg_lb_stats *sgs = &tmp_sgs;
7982 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
7985 sgs = &sds->local_stat;
7987 if (env->idle != CPU_NEWLY_IDLE ||
7988 time_after_eq(jiffies, sg->sgc->next_update))
7989 update_group_capacity(env->sd, env->dst_cpu);
7992 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7993 &overload, &overutilized);
7999 * In case the child domain prefers tasks go to siblings
8000 * first, lower the sg capacity so that we'll try
8001 * and move all the excess tasks away. We lower the capacity
8002 * of a group only if the local group has the capacity to fit
8003 * these excess tasks. The extra check prevents the case where
8004 * you always pull from the heaviest group when it is already
8005 * under-utilized (possible with a large weight task outweighs
8006 * the tasks on the system).
8008 if (prefer_sibling && sds->local &&
8009 group_has_capacity(env, &sds->local_stat) &&
8010 (sgs->sum_nr_running > 1)) {
8011 sgs->group_no_capacity = 1;
8012 sgs->group_type = group_classify(sg, sgs);
8016 * Ignore task groups with misfit tasks if local group has no
8017 * capacity or if per-cpu capacity isn't higher.
8019 if (sgs->group_type == group_misfit_task &&
8020 (!group_has_capacity(env, &sds->local_stat) ||
8021 !group_smaller_cpu_capacity(sg, sds->local)))
8022 sgs->group_type = group_other;
8024 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
8026 sds->busiest_stat = *sgs;
8030 /* Now, start updating sd_lb_stats */
8031 sds->total_load += sgs->group_load;
8032 sds->total_capacity += sgs->group_capacity;
8035 } while (sg != env->sd->groups);
8037 if (env->sd->flags & SD_NUMA)
8038 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8040 env->src_grp_nr_running = sds->busiest_stat.sum_nr_running;
8042 if (!env->sd->parent) {
8043 /* update overload indicator if we are at root domain */
8044 if (env->dst_rq->rd->overload != overload)
8045 env->dst_rq->rd->overload = overload;
8047 /* Update over-utilization (tipping point, U >= 0) indicator */
8048 if (env->dst_rq->rd->overutilized != overutilized) {
8049 env->dst_rq->rd->overutilized = overutilized;
8050 trace_sched_overutilized(overutilized);
8053 if (!env->dst_rq->rd->overutilized && overutilized) {
8054 env->dst_rq->rd->overutilized = true;
8055 trace_sched_overutilized(true);
8062 * check_asym_packing - Check to see if the group is packed into the
8065 * This is primarily intended to used at the sibling level. Some
8066 * cores like POWER7 prefer to use lower numbered SMT threads. In the
8067 * case of POWER7, it can move to lower SMT modes only when higher
8068 * threads are idle. When in lower SMT modes, the threads will
8069 * perform better since they share less core resources. Hence when we
8070 * have idle threads, we want them to be the higher ones.
8072 * This packing function is run on idle threads. It checks to see if
8073 * the busiest CPU in this domain (core in the P7 case) has a higher
8074 * CPU number than the packing function is being run on. Here we are
8075 * assuming lower CPU number will be equivalent to lower a SMT thread
8078 * Return: 1 when packing is required and a task should be moved to
8079 * this CPU. The amount of the imbalance is returned in *imbalance.
8081 * @env: The load balancing environment.
8082 * @sds: Statistics of the sched_domain which is to be packed
8084 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
8088 if (!(env->sd->flags & SD_ASYM_PACKING))
8094 busiest_cpu = group_first_cpu(sds->busiest);
8095 if (env->dst_cpu > busiest_cpu)
8098 env->imbalance = DIV_ROUND_CLOSEST(
8099 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
8100 SCHED_CAPACITY_SCALE);
8106 * fix_small_imbalance - Calculate the minor imbalance that exists
8107 * amongst the groups of a sched_domain, during
8109 * @env: The load balancing environment.
8110 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
8113 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8115 unsigned long tmp, capa_now = 0, capa_move = 0;
8116 unsigned int imbn = 2;
8117 unsigned long scaled_busy_load_per_task;
8118 struct sg_lb_stats *local, *busiest;
8120 local = &sds->local_stat;
8121 busiest = &sds->busiest_stat;
8123 if (!local->sum_nr_running)
8124 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
8125 else if (busiest->load_per_task > local->load_per_task)
8128 scaled_busy_load_per_task =
8129 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8130 busiest->group_capacity;
8132 if (busiest->avg_load + scaled_busy_load_per_task >=
8133 local->avg_load + (scaled_busy_load_per_task * imbn)) {
8134 env->imbalance = busiest->load_per_task;
8139 * OK, we don't have enough imbalance to justify moving tasks,
8140 * however we may be able to increase total CPU capacity used by
8144 capa_now += busiest->group_capacity *
8145 min(busiest->load_per_task, busiest->avg_load);
8146 capa_now += local->group_capacity *
8147 min(local->load_per_task, local->avg_load);
8148 capa_now /= SCHED_CAPACITY_SCALE;
8150 /* Amount of load we'd subtract */
8151 if (busiest->avg_load > scaled_busy_load_per_task) {
8152 capa_move += busiest->group_capacity *
8153 min(busiest->load_per_task,
8154 busiest->avg_load - scaled_busy_load_per_task);
8157 /* Amount of load we'd add */
8158 if (busiest->avg_load * busiest->group_capacity <
8159 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
8160 tmp = (busiest->avg_load * busiest->group_capacity) /
8161 local->group_capacity;
8163 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8164 local->group_capacity;
8166 capa_move += local->group_capacity *
8167 min(local->load_per_task, local->avg_load + tmp);
8168 capa_move /= SCHED_CAPACITY_SCALE;
8170 /* Move if we gain throughput */
8171 if (capa_move > capa_now)
8172 env->imbalance = busiest->load_per_task;
8176 * calculate_imbalance - Calculate the amount of imbalance present within the
8177 * groups of a given sched_domain during load balance.
8178 * @env: load balance environment
8179 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8181 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8183 unsigned long max_pull, load_above_capacity = ~0UL;
8184 struct sg_lb_stats *local, *busiest;
8186 local = &sds->local_stat;
8187 busiest = &sds->busiest_stat;
8189 if (busiest->group_type == group_imbalanced) {
8191 * In the group_imb case we cannot rely on group-wide averages
8192 * to ensure cpu-load equilibrium, look at wider averages. XXX
8194 busiest->load_per_task =
8195 min(busiest->load_per_task, sds->avg_load);
8199 * In the presence of smp nice balancing, certain scenarios can have
8200 * max load less than avg load(as we skip the groups at or below
8201 * its cpu_capacity, while calculating max_load..)
8203 if (busiest->avg_load <= sds->avg_load ||
8204 local->avg_load >= sds->avg_load) {
8205 /* Misfitting tasks should be migrated in any case */
8206 if (busiest->group_type == group_misfit_task) {
8207 env->imbalance = busiest->group_misfit_task;
8212 * Busiest group is overloaded, local is not, use the spare
8213 * cycles to maximize throughput
8215 if (busiest->group_type == group_overloaded &&
8216 local->group_type <= group_misfit_task) {
8217 env->imbalance = busiest->load_per_task;
8222 return fix_small_imbalance(env, sds);
8226 * If there aren't any idle cpus, avoid creating some.
8228 if (busiest->group_type == group_overloaded &&
8229 local->group_type == group_overloaded) {
8230 load_above_capacity = busiest->sum_nr_running *
8232 if (load_above_capacity > busiest->group_capacity)
8233 load_above_capacity -= busiest->group_capacity;
8235 load_above_capacity = ~0UL;
8239 * We're trying to get all the cpus to the average_load, so we don't
8240 * want to push ourselves above the average load, nor do we wish to
8241 * reduce the max loaded cpu below the average load. At the same time,
8242 * we also don't want to reduce the group load below the group capacity
8243 * (so that we can implement power-savings policies etc). Thus we look
8244 * for the minimum possible imbalance.
8246 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8248 /* How much load to actually move to equalise the imbalance */
8249 env->imbalance = min(
8250 max_pull * busiest->group_capacity,
8251 (sds->avg_load - local->avg_load) * local->group_capacity
8252 ) / SCHED_CAPACITY_SCALE;
8254 /* Boost imbalance to allow misfit task to be balanced. */
8255 if (busiest->group_type == group_misfit_task)
8256 env->imbalance = max_t(long, env->imbalance,
8257 busiest->group_misfit_task);
8260 * if *imbalance is less than the average load per runnable task
8261 * there is no guarantee that any tasks will be moved so we'll have
8262 * a think about bumping its value to force at least one task to be
8265 if (env->imbalance < busiest->load_per_task)
8266 return fix_small_imbalance(env, sds);
8269 /******* find_busiest_group() helpers end here *********************/
8272 * find_busiest_group - Returns the busiest group within the sched_domain
8273 * if there is an imbalance. If there isn't an imbalance, and
8274 * the user has opted for power-savings, it returns a group whose
8275 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
8276 * such a group exists.
8278 * Also calculates the amount of weighted load which should be moved
8279 * to restore balance.
8281 * @env: The load balancing environment.
8283 * Return: - The busiest group if imbalance exists.
8284 * - If no imbalance and user has opted for power-savings balance,
8285 * return the least loaded group whose CPUs can be
8286 * put to idle by rebalancing its tasks onto our group.
8288 static struct sched_group *find_busiest_group(struct lb_env *env)
8290 struct sg_lb_stats *local, *busiest;
8291 struct sd_lb_stats sds;
8293 init_sd_lb_stats(&sds);
8296 * Compute the various statistics relavent for load balancing at
8299 update_sd_lb_stats(env, &sds);
8301 if (energy_aware() && !env->dst_rq->rd->overutilized)
8304 local = &sds.local_stat;
8305 busiest = &sds.busiest_stat;
8307 /* ASYM feature bypasses nice load balance check */
8308 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
8309 check_asym_packing(env, &sds))
8312 /* There is no busy sibling group to pull tasks from */
8313 if (!sds.busiest || busiest->sum_nr_running == 0)
8316 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8317 / sds.total_capacity;
8320 * If the busiest group is imbalanced the below checks don't
8321 * work because they assume all things are equal, which typically
8322 * isn't true due to cpus_allowed constraints and the like.
8324 if (busiest->group_type == group_imbalanced)
8327 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
8328 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
8329 busiest->group_no_capacity)
8332 /* Misfitting tasks should be dealt with regardless of the avg load */
8333 if (busiest->group_type == group_misfit_task) {
8338 * If the local group is busier than the selected busiest group
8339 * don't try and pull any tasks.
8341 if (local->avg_load >= busiest->avg_load)
8345 * Don't pull any tasks if this group is already above the domain
8348 if (local->avg_load >= sds.avg_load)
8351 if (env->idle == CPU_IDLE) {
8353 * This cpu is idle. If the busiest group is not overloaded
8354 * and there is no imbalance between this and busiest group
8355 * wrt idle cpus, it is balanced. The imbalance becomes
8356 * significant if the diff is greater than 1 otherwise we
8357 * might end up to just move the imbalance on another group
8359 if ((busiest->group_type != group_overloaded) &&
8360 (local->idle_cpus <= (busiest->idle_cpus + 1)) &&
8361 !group_smaller_cpu_capacity(sds.busiest, sds.local))
8365 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8366 * imbalance_pct to be conservative.
8368 if (100 * busiest->avg_load <=
8369 env->sd->imbalance_pct * local->avg_load)
8374 env->busiest_group_type = busiest->group_type;
8375 /* Looks like there is an imbalance. Compute it */
8376 calculate_imbalance(env, &sds);
8385 * find_busiest_queue - find the busiest runqueue among the cpus in group.
8387 static struct rq *find_busiest_queue(struct lb_env *env,
8388 struct sched_group *group)
8390 struct rq *busiest = NULL, *rq;
8391 unsigned long busiest_load = 0, busiest_capacity = 1;
8394 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
8395 unsigned long capacity, wl;
8399 rt = fbq_classify_rq(rq);
8402 * We classify groups/runqueues into three groups:
8403 * - regular: there are !numa tasks
8404 * - remote: there are numa tasks that run on the 'wrong' node
8405 * - all: there is no distinction
8407 * In order to avoid migrating ideally placed numa tasks,
8408 * ignore those when there's better options.
8410 * If we ignore the actual busiest queue to migrate another
8411 * task, the next balance pass can still reduce the busiest
8412 * queue by moving tasks around inside the node.
8414 * If we cannot move enough load due to this classification
8415 * the next pass will adjust the group classification and
8416 * allow migration of more tasks.
8418 * Both cases only affect the total convergence complexity.
8420 if (rt > env->fbq_type)
8423 capacity = capacity_of(i);
8425 wl = weighted_cpuload(i);
8428 * When comparing with imbalance, use weighted_cpuload()
8429 * which is not scaled with the cpu capacity.
8432 if (rq->nr_running == 1 && wl > env->imbalance &&
8433 !check_cpu_capacity(rq, env->sd) &&
8434 env->busiest_group_type != group_misfit_task)
8438 * For the load comparisons with the other cpu's, consider
8439 * the weighted_cpuload() scaled with the cpu capacity, so
8440 * that the load can be moved away from the cpu that is
8441 * potentially running at a lower capacity.
8443 * Thus we're looking for max(wl_i / capacity_i), crosswise
8444 * multiplication to rid ourselves of the division works out
8445 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8446 * our previous maximum.
8448 if (wl * busiest_capacity > busiest_load * capacity) {
8450 busiest_capacity = capacity;
8459 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8460 * so long as it is large enough.
8462 #define MAX_PINNED_INTERVAL 512
8464 /* Working cpumask for load_balance and load_balance_newidle. */
8465 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
8467 static int need_active_balance(struct lb_env *env)
8469 struct sched_domain *sd = env->sd;
8471 if (env->idle == CPU_NEWLY_IDLE) {
8474 * ASYM_PACKING needs to force migrate tasks from busy but
8475 * higher numbered CPUs in order to pack all tasks in the
8476 * lowest numbered CPUs.
8478 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
8483 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8484 * It's worth migrating the task if the src_cpu's capacity is reduced
8485 * because of other sched_class or IRQs if more capacity stays
8486 * available on dst_cpu.
8488 if ((env->idle != CPU_NOT_IDLE) &&
8489 (env->src_rq->cfs.h_nr_running == 1)) {
8490 if ((check_cpu_capacity(env->src_rq, sd)) &&
8491 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8495 if ((capacity_of(env->src_cpu) < capacity_of(env->dst_cpu)) &&
8496 env->src_rq->cfs.h_nr_running == 1 &&
8497 cpu_overutilized(env->src_cpu) &&
8498 !cpu_overutilized(env->dst_cpu)) {
8502 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8505 static int active_load_balance_cpu_stop(void *data);
8507 static int should_we_balance(struct lb_env *env)
8509 struct sched_group *sg = env->sd->groups;
8510 struct cpumask *sg_cpus, *sg_mask;
8511 int cpu, balance_cpu = -1;
8514 * In the newly idle case, we will allow all the cpu's
8515 * to do the newly idle load balance.
8517 if (env->idle == CPU_NEWLY_IDLE)
8520 sg_cpus = sched_group_cpus(sg);
8521 sg_mask = sched_group_mask(sg);
8522 /* Try to find first idle cpu */
8523 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
8524 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
8531 if (balance_cpu == -1)
8532 balance_cpu = group_balance_cpu(sg);
8535 * First idle cpu or the first cpu(busiest) in this sched group
8536 * is eligible for doing load balancing at this and above domains.
8538 return balance_cpu == env->dst_cpu;
8542 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8543 * tasks if there is an imbalance.
8545 static int load_balance(int this_cpu, struct rq *this_rq,
8546 struct sched_domain *sd, enum cpu_idle_type idle,
8547 int *continue_balancing)
8549 int ld_moved, cur_ld_moved, active_balance = 0;
8550 struct sched_domain *sd_parent = sd->parent;
8551 struct sched_group *group;
8553 unsigned long flags;
8554 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8556 struct lb_env env = {
8558 .dst_cpu = this_cpu,
8560 .dst_grpmask = sched_group_cpus(sd->groups),
8562 .loop_break = sched_nr_migrate_break,
8565 .tasks = LIST_HEAD_INIT(env.tasks),
8569 * For NEWLY_IDLE load_balancing, we don't need to consider
8570 * other cpus in our group
8572 if (idle == CPU_NEWLY_IDLE)
8573 env.dst_grpmask = NULL;
8575 cpumask_copy(cpus, cpu_active_mask);
8577 schedstat_inc(sd, lb_count[idle]);
8580 if (!should_we_balance(&env)) {
8581 *continue_balancing = 0;
8585 group = find_busiest_group(&env);
8587 schedstat_inc(sd, lb_nobusyg[idle]);
8591 busiest = find_busiest_queue(&env, group);
8593 schedstat_inc(sd, lb_nobusyq[idle]);
8597 BUG_ON(busiest == env.dst_rq);
8599 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
8601 env.src_cpu = busiest->cpu;
8602 env.src_rq = busiest;
8605 if (busiest->nr_running > 1) {
8607 * Attempt to move tasks. If find_busiest_group has found
8608 * an imbalance but busiest->nr_running <= 1, the group is
8609 * still unbalanced. ld_moved simply stays zero, so it is
8610 * correctly treated as an imbalance.
8612 env.flags |= LBF_ALL_PINNED;
8613 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8616 raw_spin_lock_irqsave(&busiest->lock, flags);
8619 * cur_ld_moved - load moved in current iteration
8620 * ld_moved - cumulative load moved across iterations
8622 cur_ld_moved = detach_tasks(&env);
8624 * We want to potentially lower env.src_cpu's OPP.
8627 update_capacity_of(env.src_cpu);
8630 * We've detached some tasks from busiest_rq. Every
8631 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8632 * unlock busiest->lock, and we are able to be sure
8633 * that nobody can manipulate the tasks in parallel.
8634 * See task_rq_lock() family for the details.
8637 raw_spin_unlock(&busiest->lock);
8641 ld_moved += cur_ld_moved;
8644 local_irq_restore(flags);
8646 if (env.flags & LBF_NEED_BREAK) {
8647 env.flags &= ~LBF_NEED_BREAK;
8652 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8653 * us and move them to an alternate dst_cpu in our sched_group
8654 * where they can run. The upper limit on how many times we
8655 * iterate on same src_cpu is dependent on number of cpus in our
8658 * This changes load balance semantics a bit on who can move
8659 * load to a given_cpu. In addition to the given_cpu itself
8660 * (or a ilb_cpu acting on its behalf where given_cpu is
8661 * nohz-idle), we now have balance_cpu in a position to move
8662 * load to given_cpu. In rare situations, this may cause
8663 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8664 * _independently_ and at _same_ time to move some load to
8665 * given_cpu) causing exceess load to be moved to given_cpu.
8666 * This however should not happen so much in practice and
8667 * moreover subsequent load balance cycles should correct the
8668 * excess load moved.
8670 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8672 /* Prevent to re-select dst_cpu via env's cpus */
8673 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8675 env.dst_rq = cpu_rq(env.new_dst_cpu);
8676 env.dst_cpu = env.new_dst_cpu;
8677 env.flags &= ~LBF_DST_PINNED;
8679 env.loop_break = sched_nr_migrate_break;
8682 * Go back to "more_balance" rather than "redo" since we
8683 * need to continue with same src_cpu.
8689 * We failed to reach balance because of affinity.
8692 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8694 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8695 *group_imbalance = 1;
8698 /* All tasks on this runqueue were pinned by CPU affinity */
8699 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8700 cpumask_clear_cpu(cpu_of(busiest), cpus);
8701 if (!cpumask_empty(cpus)) {
8703 env.loop_break = sched_nr_migrate_break;
8706 goto out_all_pinned;
8711 schedstat_inc(sd, lb_failed[idle]);
8713 * Increment the failure counter only on periodic balance.
8714 * We do not want newidle balance, which can be very
8715 * frequent, pollute the failure counter causing
8716 * excessive cache_hot migrations and active balances.
8718 if (idle != CPU_NEWLY_IDLE)
8719 if (env.src_grp_nr_running > 1)
8720 sd->nr_balance_failed++;
8722 if (need_active_balance(&env)) {
8723 raw_spin_lock_irqsave(&busiest->lock, flags);
8725 /* don't kick the active_load_balance_cpu_stop,
8726 * if the curr task on busiest cpu can't be
8729 if (!cpumask_test_cpu(this_cpu,
8730 tsk_cpus_allowed(busiest->curr))) {
8731 raw_spin_unlock_irqrestore(&busiest->lock,
8733 env.flags |= LBF_ALL_PINNED;
8734 goto out_one_pinned;
8738 * ->active_balance synchronizes accesses to
8739 * ->active_balance_work. Once set, it's cleared
8740 * only after active load balance is finished.
8742 if (!busiest->active_balance) {
8743 busiest->active_balance = 1;
8744 busiest->push_cpu = this_cpu;
8747 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8749 if (active_balance) {
8750 stop_one_cpu_nowait(cpu_of(busiest),
8751 active_load_balance_cpu_stop, busiest,
8752 &busiest->active_balance_work);
8756 * We've kicked active balancing, reset the failure
8759 sd->nr_balance_failed = sd->cache_nice_tries+1;
8762 sd->nr_balance_failed = 0;
8764 if (likely(!active_balance)) {
8765 /* We were unbalanced, so reset the balancing interval */
8766 sd->balance_interval = sd->min_interval;
8769 * If we've begun active balancing, start to back off. This
8770 * case may not be covered by the all_pinned logic if there
8771 * is only 1 task on the busy runqueue (because we don't call
8774 if (sd->balance_interval < sd->max_interval)
8775 sd->balance_interval *= 2;
8782 * We reach balance although we may have faced some affinity
8783 * constraints. Clear the imbalance flag if it was set.
8786 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8788 if (*group_imbalance)
8789 *group_imbalance = 0;
8794 * We reach balance because all tasks are pinned at this level so
8795 * we can't migrate them. Let the imbalance flag set so parent level
8796 * can try to migrate them.
8798 schedstat_inc(sd, lb_balanced[idle]);
8800 sd->nr_balance_failed = 0;
8803 /* tune up the balancing interval */
8804 if (((env.flags & LBF_ALL_PINNED) &&
8805 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8806 (sd->balance_interval < sd->max_interval))
8807 sd->balance_interval *= 2;
8814 static inline unsigned long
8815 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8817 unsigned long interval = sd->balance_interval;
8820 interval *= sd->busy_factor;
8822 /* scale ms to jiffies */
8823 interval = msecs_to_jiffies(interval);
8824 interval = clamp(interval, 1UL, max_load_balance_interval);
8830 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
8832 unsigned long interval, next;
8834 interval = get_sd_balance_interval(sd, cpu_busy);
8835 next = sd->last_balance + interval;
8837 if (time_after(*next_balance, next))
8838 *next_balance = next;
8842 * idle_balance is called by schedule() if this_cpu is about to become
8843 * idle. Attempts to pull tasks from other CPUs.
8845 static int idle_balance(struct rq *this_rq)
8847 unsigned long next_balance = jiffies + HZ;
8848 int this_cpu = this_rq->cpu;
8849 struct sched_domain *sd;
8850 int pulled_task = 0;
8852 long removed_util=0;
8854 idle_enter_fair(this_rq);
8857 * We must set idle_stamp _before_ calling idle_balance(), such that we
8858 * measure the duration of idle_balance() as idle time.
8860 this_rq->idle_stamp = rq_clock(this_rq);
8862 if (!energy_aware() &&
8863 (this_rq->avg_idle < sysctl_sched_migration_cost ||
8864 !this_rq->rd->overload)) {
8866 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8868 update_next_balance(sd, 0, &next_balance);
8874 raw_spin_unlock(&this_rq->lock);
8877 * If removed_util_avg is !0 we most probably migrated some task away
8878 * from this_cpu. In this case we might be willing to trigger an OPP
8879 * update, but we want to do so if we don't find anybody else to pull
8880 * here (we will trigger an OPP update with the pulled task's enqueue
8883 * Record removed_util before calling update_blocked_averages, and use
8884 * it below (before returning) to see if an OPP update is required.
8886 removed_util = atomic_long_read(&(this_rq->cfs).removed_util_avg);
8887 update_blocked_averages(this_cpu);
8889 for_each_domain(this_cpu, sd) {
8890 int continue_balancing = 1;
8891 u64 t0, domain_cost;
8893 if (!(sd->flags & SD_LOAD_BALANCE))
8896 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8897 update_next_balance(sd, 0, &next_balance);
8901 if (sd->flags & SD_BALANCE_NEWIDLE) {
8902 t0 = sched_clock_cpu(this_cpu);
8904 pulled_task = load_balance(this_cpu, this_rq,
8906 &continue_balancing);
8908 domain_cost = sched_clock_cpu(this_cpu) - t0;
8909 if (domain_cost > sd->max_newidle_lb_cost)
8910 sd->max_newidle_lb_cost = domain_cost;
8912 curr_cost += domain_cost;
8915 update_next_balance(sd, 0, &next_balance);
8918 * Stop searching for tasks to pull if there are
8919 * now runnable tasks on this rq.
8921 if (pulled_task || this_rq->nr_running > 0)
8926 raw_spin_lock(&this_rq->lock);
8928 if (curr_cost > this_rq->max_idle_balance_cost)
8929 this_rq->max_idle_balance_cost = curr_cost;
8932 * While browsing the domains, we released the rq lock, a task could
8933 * have been enqueued in the meantime. Since we're not going idle,
8934 * pretend we pulled a task.
8936 if (this_rq->cfs.h_nr_running && !pulled_task)
8940 /* Move the next balance forward */
8941 if (time_after(this_rq->next_balance, next_balance))
8942 this_rq->next_balance = next_balance;
8944 /* Is there a task of a high priority class? */
8945 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8949 idle_exit_fair(this_rq);
8950 this_rq->idle_stamp = 0;
8951 } else if (removed_util) {
8953 * No task pulled and someone has been migrated away.
8954 * Good case to trigger an OPP update.
8956 update_capacity_of(this_cpu);
8963 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8964 * running tasks off the busiest CPU onto idle CPUs. It requires at
8965 * least 1 task to be running on each physical CPU where possible, and
8966 * avoids physical / logical imbalances.
8968 static int active_load_balance_cpu_stop(void *data)
8970 struct rq *busiest_rq = data;
8971 int busiest_cpu = cpu_of(busiest_rq);
8972 int target_cpu = busiest_rq->push_cpu;
8973 struct rq *target_rq = cpu_rq(target_cpu);
8974 struct sched_domain *sd;
8975 struct task_struct *p = NULL;
8977 raw_spin_lock_irq(&busiest_rq->lock);
8979 /* make sure the requested cpu hasn't gone down in the meantime */
8980 if (unlikely(busiest_cpu != smp_processor_id() ||
8981 !busiest_rq->active_balance))
8984 /* Is there any task to move? */
8985 if (busiest_rq->nr_running <= 1)
8989 * This condition is "impossible", if it occurs
8990 * we need to fix it. Originally reported by
8991 * Bjorn Helgaas on a 128-cpu setup.
8993 BUG_ON(busiest_rq == target_rq);
8995 /* Search for an sd spanning us and the target CPU. */
8997 for_each_domain(target_cpu, sd) {
8998 if ((sd->flags & SD_LOAD_BALANCE) &&
8999 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
9004 struct lb_env env = {
9006 .dst_cpu = target_cpu,
9007 .dst_rq = target_rq,
9008 .src_cpu = busiest_rq->cpu,
9009 .src_rq = busiest_rq,
9013 schedstat_inc(sd, alb_count);
9015 p = detach_one_task(&env);
9017 schedstat_inc(sd, alb_pushed);
9019 * We want to potentially lower env.src_cpu's OPP.
9021 update_capacity_of(env.src_cpu);
9024 schedstat_inc(sd, alb_failed);
9028 busiest_rq->active_balance = 0;
9029 raw_spin_unlock(&busiest_rq->lock);
9032 attach_one_task(target_rq, p);
9039 static inline int on_null_domain(struct rq *rq)
9041 return unlikely(!rcu_dereference_sched(rq->sd));
9044 #ifdef CONFIG_NO_HZ_COMMON
9046 * idle load balancing details
9047 * - When one of the busy CPUs notice that there may be an idle rebalancing
9048 * needed, they will kick the idle load balancer, which then does idle
9049 * load balancing for all the idle CPUs.
9052 cpumask_var_t idle_cpus_mask;
9054 unsigned long next_balance; /* in jiffy units */
9055 } nohz ____cacheline_aligned;
9057 static inline int find_new_ilb(void)
9059 int ilb = cpumask_first(nohz.idle_cpus_mask);
9061 if (ilb < nr_cpu_ids && idle_cpu(ilb))
9068 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
9069 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
9070 * CPU (if there is one).
9072 static void nohz_balancer_kick(void)
9076 nohz.next_balance++;
9078 ilb_cpu = find_new_ilb();
9080 if (ilb_cpu >= nr_cpu_ids)
9083 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
9086 * Use smp_send_reschedule() instead of resched_cpu().
9087 * This way we generate a sched IPI on the target cpu which
9088 * is idle. And the softirq performing nohz idle load balance
9089 * will be run before returning from the IPI.
9091 smp_send_reschedule(ilb_cpu);
9095 static inline void nohz_balance_exit_idle(int cpu)
9097 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
9099 * Completely isolated CPUs don't ever set, so we must test.
9101 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
9102 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
9103 atomic_dec(&nohz.nr_cpus);
9105 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
9109 static inline void set_cpu_sd_state_busy(void)
9111 struct sched_domain *sd;
9112 int cpu = smp_processor_id();
9115 sd = rcu_dereference(per_cpu(sd_busy, cpu));
9117 if (!sd || !sd->nohz_idle)
9121 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
9126 void set_cpu_sd_state_idle(void)
9128 struct sched_domain *sd;
9129 int cpu = smp_processor_id();
9132 sd = rcu_dereference(per_cpu(sd_busy, cpu));
9134 if (!sd || sd->nohz_idle)
9138 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
9144 * This routine will record that the cpu is going idle with tick stopped.
9145 * This info will be used in performing idle load balancing in the future.
9147 void nohz_balance_enter_idle(int cpu)
9150 * If this cpu is going down, then nothing needs to be done.
9152 if (!cpu_active(cpu))
9155 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
9159 * If we're a completely isolated CPU, we don't play.
9161 if (on_null_domain(cpu_rq(cpu)))
9164 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9165 atomic_inc(&nohz.nr_cpus);
9166 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
9169 static int sched_ilb_notifier(struct notifier_block *nfb,
9170 unsigned long action, void *hcpu)
9172 switch (action & ~CPU_TASKS_FROZEN) {
9174 nohz_balance_exit_idle(smp_processor_id());
9182 static DEFINE_SPINLOCK(balancing);
9185 * Scale the max load_balance interval with the number of CPUs in the system.
9186 * This trades load-balance latency on larger machines for less cross talk.
9188 void update_max_interval(void)
9190 max_load_balance_interval = HZ*num_online_cpus()/10;
9194 * It checks each scheduling domain to see if it is due to be balanced,
9195 * and initiates a balancing operation if so.
9197 * Balancing parameters are set up in init_sched_domains.
9199 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
9201 int continue_balancing = 1;
9203 unsigned long interval;
9204 struct sched_domain *sd;
9205 /* Earliest time when we have to do rebalance again */
9206 unsigned long next_balance = jiffies + 60*HZ;
9207 int update_next_balance = 0;
9208 int need_serialize, need_decay = 0;
9211 update_blocked_averages(cpu);
9214 for_each_domain(cpu, sd) {
9216 * Decay the newidle max times here because this is a regular
9217 * visit to all the domains. Decay ~1% per second.
9219 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
9220 sd->max_newidle_lb_cost =
9221 (sd->max_newidle_lb_cost * 253) / 256;
9222 sd->next_decay_max_lb_cost = jiffies + HZ;
9225 max_cost += sd->max_newidle_lb_cost;
9227 if (!(sd->flags & SD_LOAD_BALANCE))
9231 * Stop the load balance at this level. There is another
9232 * CPU in our sched group which is doing load balancing more
9235 if (!continue_balancing) {
9241 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9243 need_serialize = sd->flags & SD_SERIALIZE;
9244 if (need_serialize) {
9245 if (!spin_trylock(&balancing))
9249 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9250 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9252 * The LBF_DST_PINNED logic could have changed
9253 * env->dst_cpu, so we can't know our idle
9254 * state even if we migrated tasks. Update it.
9256 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9258 sd->last_balance = jiffies;
9259 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9262 spin_unlock(&balancing);
9264 if (time_after(next_balance, sd->last_balance + interval)) {
9265 next_balance = sd->last_balance + interval;
9266 update_next_balance = 1;
9271 * Ensure the rq-wide value also decays but keep it at a
9272 * reasonable floor to avoid funnies with rq->avg_idle.
9274 rq->max_idle_balance_cost =
9275 max((u64)sysctl_sched_migration_cost, max_cost);
9280 * next_balance will be updated only when there is a need.
9281 * When the cpu is attached to null domain for ex, it will not be
9284 if (likely(update_next_balance)) {
9285 rq->next_balance = next_balance;
9287 #ifdef CONFIG_NO_HZ_COMMON
9289 * If this CPU has been elected to perform the nohz idle
9290 * balance. Other idle CPUs have already rebalanced with
9291 * nohz_idle_balance() and nohz.next_balance has been
9292 * updated accordingly. This CPU is now running the idle load
9293 * balance for itself and we need to update the
9294 * nohz.next_balance accordingly.
9296 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9297 nohz.next_balance = rq->next_balance;
9302 #ifdef CONFIG_NO_HZ_COMMON
9304 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9305 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9307 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9309 int this_cpu = this_rq->cpu;
9312 /* Earliest time when we have to do rebalance again */
9313 unsigned long next_balance = jiffies + 60*HZ;
9314 int update_next_balance = 0;
9316 if (idle != CPU_IDLE ||
9317 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
9320 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9321 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9325 * If this cpu gets work to do, stop the load balancing
9326 * work being done for other cpus. Next load
9327 * balancing owner will pick it up.
9332 rq = cpu_rq(balance_cpu);
9335 * If time for next balance is due,
9338 if (time_after_eq(jiffies, rq->next_balance)) {
9339 raw_spin_lock_irq(&rq->lock);
9340 update_rq_clock(rq);
9341 update_idle_cpu_load(rq);
9342 raw_spin_unlock_irq(&rq->lock);
9343 rebalance_domains(rq, CPU_IDLE);
9346 if (time_after(next_balance, rq->next_balance)) {
9347 next_balance = rq->next_balance;
9348 update_next_balance = 1;
9353 * next_balance will be updated only when there is a need.
9354 * When the CPU is attached to null domain for ex, it will not be
9357 if (likely(update_next_balance))
9358 nohz.next_balance = next_balance;
9360 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
9364 * Current heuristic for kicking the idle load balancer in the presence
9365 * of an idle cpu in the system.
9366 * - This rq has more than one task.
9367 * - This rq has at least one CFS task and the capacity of the CPU is
9368 * significantly reduced because of RT tasks or IRQs.
9369 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
9370 * multiple busy cpu.
9371 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
9372 * domain span are idle.
9374 static inline bool nohz_kick_needed(struct rq *rq)
9376 unsigned long now = jiffies;
9377 struct sched_domain *sd;
9378 struct sched_group_capacity *sgc;
9379 int nr_busy, cpu = rq->cpu;
9382 if (unlikely(rq->idle_balance))
9386 * We may be recently in ticked or tickless idle mode. At the first
9387 * busy tick after returning from idle, we will update the busy stats.
9389 set_cpu_sd_state_busy();
9390 nohz_balance_exit_idle(cpu);
9393 * None are in tickless mode and hence no need for NOHZ idle load
9396 if (likely(!atomic_read(&nohz.nr_cpus)))
9399 if (time_before(now, nohz.next_balance))
9402 if (rq->nr_running >= 2 &&
9403 (!energy_aware() || cpu_overutilized(cpu)))
9407 sd = rcu_dereference(per_cpu(sd_busy, cpu));
9408 if (sd && !energy_aware()) {
9409 sgc = sd->groups->sgc;
9410 nr_busy = atomic_read(&sgc->nr_busy_cpus);
9419 sd = rcu_dereference(rq->sd);
9421 if ((rq->cfs.h_nr_running >= 1) &&
9422 check_cpu_capacity(rq, sd)) {
9428 sd = rcu_dereference(per_cpu(sd_asym, cpu));
9429 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
9430 sched_domain_span(sd)) < cpu)) {
9440 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
9444 * run_rebalance_domains is triggered when needed from the scheduler tick.
9445 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9447 static void run_rebalance_domains(struct softirq_action *h)
9449 struct rq *this_rq = this_rq();
9450 enum cpu_idle_type idle = this_rq->idle_balance ?
9451 CPU_IDLE : CPU_NOT_IDLE;
9454 * If this cpu has a pending nohz_balance_kick, then do the
9455 * balancing on behalf of the other idle cpus whose ticks are
9456 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9457 * give the idle cpus a chance to load balance. Else we may
9458 * load balance only within the local sched_domain hierarchy
9459 * and abort nohz_idle_balance altogether if we pull some load.
9461 nohz_idle_balance(this_rq, idle);
9462 rebalance_domains(this_rq, idle);
9466 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9468 void trigger_load_balance(struct rq *rq)
9470 /* Don't need to rebalance while attached to NULL domain */
9471 if (unlikely(on_null_domain(rq)))
9474 if (time_after_eq(jiffies, rq->next_balance))
9475 raise_softirq(SCHED_SOFTIRQ);
9476 #ifdef CONFIG_NO_HZ_COMMON
9477 if (nohz_kick_needed(rq))
9478 nohz_balancer_kick();
9482 static void rq_online_fair(struct rq *rq)
9486 update_runtime_enabled(rq);
9489 static void rq_offline_fair(struct rq *rq)
9493 /* Ensure any throttled groups are reachable by pick_next_task */
9494 unthrottle_offline_cfs_rqs(rq);
9497 #endif /* CONFIG_SMP */
9500 * scheduler tick hitting a task of our scheduling class:
9502 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9504 struct cfs_rq *cfs_rq;
9505 struct sched_entity *se = &curr->se;
9507 for_each_sched_entity(se) {
9508 cfs_rq = cfs_rq_of(se);
9509 entity_tick(cfs_rq, se, queued);
9512 if (static_branch_unlikely(&sched_numa_balancing))
9513 task_tick_numa(rq, curr);
9516 if (!rq->rd->overutilized && cpu_overutilized(task_cpu(curr))) {
9517 rq->rd->overutilized = true;
9518 trace_sched_overutilized(true);
9521 rq->misfit_task = !task_fits_max(curr, rq->cpu);
9527 * called on fork with the child task as argument from the parent's context
9528 * - child not yet on the tasklist
9529 * - preemption disabled
9531 static void task_fork_fair(struct task_struct *p)
9533 struct cfs_rq *cfs_rq;
9534 struct sched_entity *se = &p->se, *curr;
9535 int this_cpu = smp_processor_id();
9536 struct rq *rq = this_rq();
9537 unsigned long flags;
9539 raw_spin_lock_irqsave(&rq->lock, flags);
9541 update_rq_clock(rq);
9543 cfs_rq = task_cfs_rq(current);
9544 curr = cfs_rq->curr;
9547 * Not only the cpu but also the task_group of the parent might have
9548 * been changed after parent->se.parent,cfs_rq were copied to
9549 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
9550 * of child point to valid ones.
9553 __set_task_cpu(p, this_cpu);
9556 update_curr(cfs_rq);
9559 se->vruntime = curr->vruntime;
9560 place_entity(cfs_rq, se, 1);
9562 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9564 * Upon rescheduling, sched_class::put_prev_task() will place
9565 * 'current' within the tree based on its new key value.
9567 swap(curr->vruntime, se->vruntime);
9571 se->vruntime -= cfs_rq->min_vruntime;
9573 raw_spin_unlock_irqrestore(&rq->lock, flags);
9577 * Priority of the task has changed. Check to see if we preempt
9581 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9583 if (!task_on_rq_queued(p))
9587 * Reschedule if we are currently running on this runqueue and
9588 * our priority decreased, or if we are not currently running on
9589 * this runqueue and our priority is higher than the current's
9591 if (rq->curr == p) {
9592 if (p->prio > oldprio)
9595 check_preempt_curr(rq, p, 0);
9598 static inline bool vruntime_normalized(struct task_struct *p)
9600 struct sched_entity *se = &p->se;
9603 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9604 * the dequeue_entity(.flags=0) will already have normalized the
9611 * When !on_rq, vruntime of the task has usually NOT been normalized.
9612 * But there are some cases where it has already been normalized:
9614 * - A forked child which is waiting for being woken up by
9615 * wake_up_new_task().
9616 * - A task which has been woken up by try_to_wake_up() and
9617 * waiting for actually being woken up by sched_ttwu_pending().
9619 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
9625 #ifdef CONFIG_FAIR_GROUP_SCHED
9627 * Propagate the changes of the sched_entity across the tg tree to make it
9628 * visible to the root
9630 static void propagate_entity_cfs_rq(struct sched_entity *se)
9632 struct cfs_rq *cfs_rq;
9634 /* Start to propagate at parent */
9637 for_each_sched_entity(se) {
9638 cfs_rq = cfs_rq_of(se);
9640 if (cfs_rq_throttled(cfs_rq))
9643 update_load_avg(se, UPDATE_TG);
9647 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
9650 static void detach_entity_cfs_rq(struct sched_entity *se)
9652 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9654 /* Catch up with the cfs_rq and remove our load when we leave */
9655 update_load_avg(se, 0);
9656 detach_entity_load_avg(cfs_rq, se);
9657 update_tg_load_avg(cfs_rq, false);
9658 propagate_entity_cfs_rq(se);
9661 static void attach_entity_cfs_rq(struct sched_entity *se)
9663 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9665 #ifdef CONFIG_FAIR_GROUP_SCHED
9667 * Since the real-depth could have been changed (only FAIR
9668 * class maintain depth value), reset depth properly.
9670 se->depth = se->parent ? se->parent->depth + 1 : 0;
9673 /* Synchronize entity with its cfs_rq */
9674 update_load_avg(se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
9675 attach_entity_load_avg(cfs_rq, se);
9676 update_tg_load_avg(cfs_rq, false);
9677 propagate_entity_cfs_rq(se);
9680 static void detach_task_cfs_rq(struct task_struct *p)
9682 struct sched_entity *se = &p->se;
9683 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9685 if (!vruntime_normalized(p)) {
9687 * Fix up our vruntime so that the current sleep doesn't
9688 * cause 'unlimited' sleep bonus.
9690 place_entity(cfs_rq, se, 0);
9691 se->vruntime -= cfs_rq->min_vruntime;
9694 detach_entity_cfs_rq(se);
9697 static void attach_task_cfs_rq(struct task_struct *p)
9699 struct sched_entity *se = &p->se;
9700 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9702 attach_entity_cfs_rq(se);
9704 if (!vruntime_normalized(p))
9705 se->vruntime += cfs_rq->min_vruntime;
9708 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9710 detach_task_cfs_rq(p);
9713 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9715 attach_task_cfs_rq(p);
9717 if (task_on_rq_queued(p)) {
9719 * We were most likely switched from sched_rt, so
9720 * kick off the schedule if running, otherwise just see
9721 * if we can still preempt the current task.
9726 check_preempt_curr(rq, p, 0);
9730 /* Account for a task changing its policy or group.
9732 * This routine is mostly called to set cfs_rq->curr field when a task
9733 * migrates between groups/classes.
9735 static void set_curr_task_fair(struct rq *rq)
9737 struct sched_entity *se = &rq->curr->se;
9739 for_each_sched_entity(se) {
9740 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9742 set_next_entity(cfs_rq, se);
9743 /* ensure bandwidth has been allocated on our new cfs_rq */
9744 account_cfs_rq_runtime(cfs_rq, 0);
9748 void init_cfs_rq(struct cfs_rq *cfs_rq)
9750 cfs_rq->tasks_timeline = RB_ROOT;
9751 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9752 #ifndef CONFIG_64BIT
9753 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9756 #ifdef CONFIG_FAIR_GROUP_SCHED
9757 cfs_rq->propagate_avg = 0;
9759 atomic_long_set(&cfs_rq->removed_load_avg, 0);
9760 atomic_long_set(&cfs_rq->removed_util_avg, 0);
9764 #ifdef CONFIG_FAIR_GROUP_SCHED
9765 static void task_move_group_fair(struct task_struct *p)
9767 detach_task_cfs_rq(p);
9768 set_task_rq(p, task_cpu(p));
9771 /* Tell se's cfs_rq has been changed -- migrated */
9772 p->se.avg.last_update_time = 0;
9774 attach_task_cfs_rq(p);
9777 void free_fair_sched_group(struct task_group *tg)
9781 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9783 for_each_possible_cpu(i) {
9785 kfree(tg->cfs_rq[i]);
9788 remove_entity_load_avg(tg->se[i]);
9797 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9799 struct sched_entity *se;
9800 struct cfs_rq *cfs_rq;
9804 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9807 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9811 tg->shares = NICE_0_LOAD;
9813 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9815 for_each_possible_cpu(i) {
9818 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9819 GFP_KERNEL, cpu_to_node(i));
9823 se = kzalloc_node(sizeof(struct sched_entity),
9824 GFP_KERNEL, cpu_to_node(i));
9828 init_cfs_rq(cfs_rq);
9829 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9830 init_entity_runnable_average(se);
9832 raw_spin_lock_irq(&rq->lock);
9833 post_init_entity_util_avg(se);
9834 raw_spin_unlock_irq(&rq->lock);
9845 void unregister_fair_sched_group(struct task_group *tg, int cpu)
9847 struct rq *rq = cpu_rq(cpu);
9848 unsigned long flags;
9851 * Only empty task groups can be destroyed; so we can speculatively
9852 * check on_list without danger of it being re-added.
9854 if (!tg->cfs_rq[cpu]->on_list)
9857 raw_spin_lock_irqsave(&rq->lock, flags);
9858 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9859 raw_spin_unlock_irqrestore(&rq->lock, flags);
9862 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9863 struct sched_entity *se, int cpu,
9864 struct sched_entity *parent)
9866 struct rq *rq = cpu_rq(cpu);
9870 init_cfs_rq_runtime(cfs_rq);
9872 tg->cfs_rq[cpu] = cfs_rq;
9875 /* se could be NULL for root_task_group */
9880 se->cfs_rq = &rq->cfs;
9883 se->cfs_rq = parent->my_q;
9884 se->depth = parent->depth + 1;
9888 /* guarantee group entities always have weight */
9889 update_load_set(&se->load, NICE_0_LOAD);
9890 se->parent = parent;
9893 static DEFINE_MUTEX(shares_mutex);
9895 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9898 unsigned long flags;
9901 * We can't change the weight of the root cgroup.
9906 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9908 mutex_lock(&shares_mutex);
9909 if (tg->shares == shares)
9912 tg->shares = shares;
9913 for_each_possible_cpu(i) {
9914 struct rq *rq = cpu_rq(i);
9915 struct sched_entity *se;
9918 /* Propagate contribution to hierarchy */
9919 raw_spin_lock_irqsave(&rq->lock, flags);
9921 /* Possible calls to update_curr() need rq clock */
9922 update_rq_clock(rq);
9923 for_each_sched_entity(se)
9924 update_cfs_shares(group_cfs_rq(se));
9925 raw_spin_unlock_irqrestore(&rq->lock, flags);
9929 mutex_unlock(&shares_mutex);
9932 #else /* CONFIG_FAIR_GROUP_SCHED */
9934 void free_fair_sched_group(struct task_group *tg) { }
9936 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9941 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
9943 #endif /* CONFIG_FAIR_GROUP_SCHED */
9946 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9948 struct sched_entity *se = &task->se;
9949 unsigned int rr_interval = 0;
9952 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9955 if (rq->cfs.load.weight)
9956 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9962 * All the scheduling class methods:
9964 const struct sched_class fair_sched_class = {
9965 .next = &idle_sched_class,
9966 .enqueue_task = enqueue_task_fair,
9967 .dequeue_task = dequeue_task_fair,
9968 .yield_task = yield_task_fair,
9969 .yield_to_task = yield_to_task_fair,
9971 .check_preempt_curr = check_preempt_wakeup,
9973 .pick_next_task = pick_next_task_fair,
9974 .put_prev_task = put_prev_task_fair,
9977 .select_task_rq = select_task_rq_fair,
9978 .migrate_task_rq = migrate_task_rq_fair,
9980 .rq_online = rq_online_fair,
9981 .rq_offline = rq_offline_fair,
9983 .task_waking = task_waking_fair,
9984 .task_dead = task_dead_fair,
9985 .set_cpus_allowed = set_cpus_allowed_common,
9988 .set_curr_task = set_curr_task_fair,
9989 .task_tick = task_tick_fair,
9990 .task_fork = task_fork_fair,
9992 .prio_changed = prio_changed_fair,
9993 .switched_from = switched_from_fair,
9994 .switched_to = switched_to_fair,
9996 .get_rr_interval = get_rr_interval_fair,
9998 .update_curr = update_curr_fair,
10000 #ifdef CONFIG_FAIR_GROUP_SCHED
10001 .task_move_group = task_move_group_fair,
10005 #ifdef CONFIG_SCHED_DEBUG
10006 void print_cfs_stats(struct seq_file *m, int cpu)
10008 struct cfs_rq *cfs_rq;
10011 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
10012 print_cfs_rq(m, cpu, cfs_rq);
10016 #ifdef CONFIG_NUMA_BALANCING
10017 void show_numa_stats(struct task_struct *p, struct seq_file *m)
10020 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
10022 for_each_online_node(node) {
10023 if (p->numa_faults) {
10024 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
10025 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
10027 if (p->numa_group) {
10028 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
10029 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
10031 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
10034 #endif /* CONFIG_NUMA_BALANCING */
10035 #endif /* CONFIG_SCHED_DEBUG */
10037 __init void init_sched_fair_class(void)
10040 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
10042 #ifdef CONFIG_NO_HZ_COMMON
10043 nohz.next_balance = jiffies;
10044 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
10045 cpu_notifier(sched_ilb_notifier, 0);