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>
34 #include <trace/events/sched.h>
39 * Targeted preemption latency for CPU-bound tasks:
40 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
42 * NOTE: this latency value is not the same as the concept of
43 * 'timeslice length' - timeslices in CFS are of variable length
44 * and have no persistent notion like in traditional, time-slice
45 * based scheduling concepts.
47 * (to see the precise effective timeslice length of your workload,
48 * run vmstat and monitor the context-switches (cs) field)
50 unsigned int sysctl_sched_latency = 6000000ULL;
51 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
54 * The initial- and re-scaling of tunables is configurable
55 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
58 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
59 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
60 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
62 enum sched_tunable_scaling sysctl_sched_tunable_scaling
63 = SCHED_TUNABLESCALING_LOG;
66 * Minimal preemption granularity for CPU-bound tasks:
67 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
69 unsigned int sysctl_sched_min_granularity = 750000ULL;
70 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
73 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
75 static unsigned int sched_nr_latency = 8;
78 * After fork, child runs first. If set to 0 (default) then
79 * parent will (try to) run first.
81 unsigned int sysctl_sched_child_runs_first __read_mostly;
84 * SCHED_OTHER wake-up granularity.
85 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
87 * This option delays the preemption effects of decoupled workloads
88 * and reduces their over-scheduling. Synchronous workloads will still
89 * have immediate wakeup/sleep latencies.
91 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
92 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
94 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
97 * The exponential sliding window over which load is averaged for shares
101 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
103 #ifdef CONFIG_CFS_BANDWIDTH
105 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
106 * each time a cfs_rq requests quota.
108 * Note: in the case that the slice exceeds the runtime remaining (either due
109 * to consumption or the quota being specified to be smaller than the slice)
110 * we will always only issue the remaining available time.
112 * default: 5 msec, units: microseconds
114 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
117 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
123 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
129 static inline void update_load_set(struct load_weight *lw, unsigned long w)
136 * Increase the granularity value when there are more CPUs,
137 * because with more CPUs the 'effective latency' as visible
138 * to users decreases. But the relationship is not linear,
139 * so pick a second-best guess by going with the log2 of the
142 * This idea comes from the SD scheduler of Con Kolivas:
144 static unsigned int get_update_sysctl_factor(void)
146 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
149 switch (sysctl_sched_tunable_scaling) {
150 case SCHED_TUNABLESCALING_NONE:
153 case SCHED_TUNABLESCALING_LINEAR:
156 case SCHED_TUNABLESCALING_LOG:
158 factor = 1 + ilog2(cpus);
165 static void update_sysctl(void)
167 unsigned int factor = get_update_sysctl_factor();
169 #define SET_SYSCTL(name) \
170 (sysctl_##name = (factor) * normalized_sysctl_##name)
171 SET_SYSCTL(sched_min_granularity);
172 SET_SYSCTL(sched_latency);
173 SET_SYSCTL(sched_wakeup_granularity);
177 void sched_init_granularity(void)
182 #define WMULT_CONST (~0U)
183 #define WMULT_SHIFT 32
185 static void __update_inv_weight(struct load_weight *lw)
189 if (likely(lw->inv_weight))
192 w = scale_load_down(lw->weight);
194 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
196 else if (unlikely(!w))
197 lw->inv_weight = WMULT_CONST;
199 lw->inv_weight = WMULT_CONST / w;
203 * delta_exec * weight / lw.weight
205 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
207 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
208 * we're guaranteed shift stays positive because inv_weight is guaranteed to
209 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
211 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
212 * weight/lw.weight <= 1, and therefore our shift will also be positive.
214 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
216 u64 fact = scale_load_down(weight);
217 int shift = WMULT_SHIFT;
219 __update_inv_weight(lw);
221 if (unlikely(fact >> 32)) {
228 /* hint to use a 32x32->64 mul */
229 fact = (u64)(u32)fact * lw->inv_weight;
236 return mul_u64_u32_shr(delta_exec, fact, shift);
240 const struct sched_class fair_sched_class;
242 /**************************************************************
243 * CFS operations on generic schedulable entities:
246 #ifdef CONFIG_FAIR_GROUP_SCHED
248 /* cpu runqueue to which this cfs_rq is attached */
249 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
254 /* An entity is a task if it doesn't "own" a runqueue */
255 #define entity_is_task(se) (!se->my_q)
257 static inline struct task_struct *task_of(struct sched_entity *se)
259 #ifdef CONFIG_SCHED_DEBUG
260 WARN_ON_ONCE(!entity_is_task(se));
262 return container_of(se, struct task_struct, se);
265 /* Walk up scheduling entities hierarchy */
266 #define for_each_sched_entity(se) \
267 for (; se; se = se->parent)
269 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
274 /* runqueue on which this entity is (to be) queued */
275 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
280 /* runqueue "owned" by this group */
281 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
286 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
288 if (!cfs_rq->on_list) {
290 * Ensure we either appear before our parent (if already
291 * enqueued) or force our parent to appear after us when it is
292 * enqueued. The fact that we always enqueue bottom-up
293 * reduces this to two cases.
295 if (cfs_rq->tg->parent &&
296 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
297 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
298 &rq_of(cfs_rq)->leaf_cfs_rq_list);
300 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
301 &rq_of(cfs_rq)->leaf_cfs_rq_list);
308 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
310 if (cfs_rq->on_list) {
311 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
316 /* Iterate thr' all leaf cfs_rq's on a runqueue */
317 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
318 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
320 /* Do the two (enqueued) entities belong to the same group ? */
321 static inline struct cfs_rq *
322 is_same_group(struct sched_entity *se, struct sched_entity *pse)
324 if (se->cfs_rq == pse->cfs_rq)
330 static inline struct sched_entity *parent_entity(struct sched_entity *se)
336 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
338 int se_depth, pse_depth;
341 * preemption test can be made between sibling entities who are in the
342 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
343 * both tasks until we find their ancestors who are siblings of common
347 /* First walk up until both entities are at same depth */
348 se_depth = (*se)->depth;
349 pse_depth = (*pse)->depth;
351 while (se_depth > pse_depth) {
353 *se = parent_entity(*se);
356 while (pse_depth > se_depth) {
358 *pse = parent_entity(*pse);
361 while (!is_same_group(*se, *pse)) {
362 *se = parent_entity(*se);
363 *pse = parent_entity(*pse);
367 #else /* !CONFIG_FAIR_GROUP_SCHED */
369 static inline struct task_struct *task_of(struct sched_entity *se)
371 return container_of(se, struct task_struct, se);
374 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
376 return container_of(cfs_rq, struct rq, cfs);
379 #define entity_is_task(se) 1
381 #define for_each_sched_entity(se) \
382 for (; se; se = NULL)
384 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
386 return &task_rq(p)->cfs;
389 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
391 struct task_struct *p = task_of(se);
392 struct rq *rq = task_rq(p);
397 /* runqueue "owned" by this group */
398 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
403 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
407 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
411 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
412 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
414 static inline struct sched_entity *parent_entity(struct sched_entity *se)
420 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
424 #endif /* CONFIG_FAIR_GROUP_SCHED */
426 static __always_inline
427 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
429 /**************************************************************
430 * Scheduling class tree data structure manipulation methods:
433 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
435 s64 delta = (s64)(vruntime - max_vruntime);
437 max_vruntime = vruntime;
442 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
444 s64 delta = (s64)(vruntime - min_vruntime);
446 min_vruntime = vruntime;
451 static inline int entity_before(struct sched_entity *a,
452 struct sched_entity *b)
454 return (s64)(a->vruntime - b->vruntime) < 0;
457 static void update_min_vruntime(struct cfs_rq *cfs_rq)
459 u64 vruntime = cfs_rq->min_vruntime;
462 vruntime = cfs_rq->curr->vruntime;
464 if (cfs_rq->rb_leftmost) {
465 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
470 vruntime = se->vruntime;
472 vruntime = min_vruntime(vruntime, se->vruntime);
475 /* ensure we never gain time by being placed backwards. */
476 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
479 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
484 * Enqueue an entity into the rb-tree:
486 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
488 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
489 struct rb_node *parent = NULL;
490 struct sched_entity *entry;
494 * Find the right place in the rbtree:
498 entry = rb_entry(parent, struct sched_entity, run_node);
500 * We dont care about collisions. Nodes with
501 * the same key stay together.
503 if (entity_before(se, entry)) {
504 link = &parent->rb_left;
506 link = &parent->rb_right;
512 * Maintain a cache of leftmost tree entries (it is frequently
516 cfs_rq->rb_leftmost = &se->run_node;
518 rb_link_node(&se->run_node, parent, link);
519 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
522 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
524 if (cfs_rq->rb_leftmost == &se->run_node) {
525 struct rb_node *next_node;
527 next_node = rb_next(&se->run_node);
528 cfs_rq->rb_leftmost = next_node;
531 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
534 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
536 struct rb_node *left = cfs_rq->rb_leftmost;
541 return rb_entry(left, struct sched_entity, run_node);
544 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
546 struct rb_node *next = rb_next(&se->run_node);
551 return rb_entry(next, struct sched_entity, run_node);
554 #ifdef CONFIG_SCHED_DEBUG
555 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
557 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
562 return rb_entry(last, struct sched_entity, run_node);
565 /**************************************************************
566 * Scheduling class statistics methods:
569 int sched_proc_update_handler(struct ctl_table *table, int write,
570 void __user *buffer, size_t *lenp,
573 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
574 unsigned int factor = get_update_sysctl_factor();
579 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
580 sysctl_sched_min_granularity);
582 #define WRT_SYSCTL(name) \
583 (normalized_sysctl_##name = sysctl_##name / (factor))
584 WRT_SYSCTL(sched_min_granularity);
585 WRT_SYSCTL(sched_latency);
586 WRT_SYSCTL(sched_wakeup_granularity);
596 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
598 if (unlikely(se->load.weight != NICE_0_LOAD))
599 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
605 * The idea is to set a period in which each task runs once.
607 * When there are too many tasks (sched_nr_latency) we have to stretch
608 * this period because otherwise the slices get too small.
610 * p = (nr <= nl) ? l : l*nr/nl
612 static u64 __sched_period(unsigned long nr_running)
614 if (unlikely(nr_running > sched_nr_latency))
615 return nr_running * sysctl_sched_min_granularity;
617 return sysctl_sched_latency;
621 * We calculate the wall-time slice from the period by taking a part
622 * proportional to the weight.
626 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
628 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
630 for_each_sched_entity(se) {
631 struct load_weight *load;
632 struct load_weight lw;
634 cfs_rq = cfs_rq_of(se);
635 load = &cfs_rq->load;
637 if (unlikely(!se->on_rq)) {
640 update_load_add(&lw, se->load.weight);
643 slice = __calc_delta(slice, se->load.weight, load);
649 * We calculate the vruntime slice of a to-be-inserted task.
653 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
655 return calc_delta_fair(sched_slice(cfs_rq, se), se);
659 static int select_idle_sibling(struct task_struct *p, int cpu);
660 static unsigned long task_h_load(struct task_struct *p);
663 * We choose a half-life close to 1 scheduling period.
664 * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are
665 * dependent on this value.
667 #define LOAD_AVG_PERIOD 32
668 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
669 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */
671 /* Give new sched_entity start runnable values to heavy its load in infant time */
672 void init_entity_runnable_average(struct sched_entity *se)
674 struct sched_avg *sa = &se->avg;
676 sa->last_update_time = 0;
678 * sched_avg's period_contrib should be strictly less then 1024, so
679 * we give it 1023 to make sure it is almost a period (1024us), and
680 * will definitely be update (after enqueue).
682 sa->period_contrib = 1023;
683 sa->load_avg = scale_load_down(se->load.weight);
684 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
685 sa->util_avg = scale_load_down(SCHED_LOAD_SCALE);
686 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
687 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
691 void init_entity_runnable_average(struct sched_entity *se)
697 * Update the current task's runtime statistics.
699 static void update_curr(struct cfs_rq *cfs_rq)
701 struct sched_entity *curr = cfs_rq->curr;
702 u64 now = rq_clock_task(rq_of(cfs_rq));
708 delta_exec = now - curr->exec_start;
709 if (unlikely((s64)delta_exec <= 0))
712 curr->exec_start = now;
714 schedstat_set(curr->statistics.exec_max,
715 max(delta_exec, curr->statistics.exec_max));
717 curr->sum_exec_runtime += delta_exec;
718 schedstat_add(cfs_rq, exec_clock, delta_exec);
720 curr->vruntime += calc_delta_fair(delta_exec, curr);
721 update_min_vruntime(cfs_rq);
723 if (entity_is_task(curr)) {
724 struct task_struct *curtask = task_of(curr);
726 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
727 cpuacct_charge(curtask, delta_exec);
728 account_group_exec_runtime(curtask, delta_exec);
731 account_cfs_rq_runtime(cfs_rq, delta_exec);
734 static void update_curr_fair(struct rq *rq)
736 update_curr(cfs_rq_of(&rq->curr->se));
740 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
742 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
746 * Task is being enqueued - update stats:
748 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
751 * Are we enqueueing a waiting task? (for current tasks
752 * a dequeue/enqueue event is a NOP)
754 if (se != cfs_rq->curr)
755 update_stats_wait_start(cfs_rq, se);
759 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
761 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
762 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
763 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
764 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
765 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
766 #ifdef CONFIG_SCHEDSTATS
767 if (entity_is_task(se)) {
768 trace_sched_stat_wait(task_of(se),
769 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
772 schedstat_set(se->statistics.wait_start, 0);
776 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
779 * Mark the end of the wait period if dequeueing a
782 if (se != cfs_rq->curr)
783 update_stats_wait_end(cfs_rq, se);
787 * We are picking a new current task - update its stats:
790 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
793 * We are starting a new run period:
795 se->exec_start = rq_clock_task(rq_of(cfs_rq));
798 /**************************************************
799 * Scheduling class queueing methods:
802 #ifdef CONFIG_NUMA_BALANCING
804 * Approximate time to scan a full NUMA task in ms. The task scan period is
805 * calculated based on the tasks virtual memory size and
806 * numa_balancing_scan_size.
808 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
809 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
811 /* Portion of address space to scan in MB */
812 unsigned int sysctl_numa_balancing_scan_size = 256;
814 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
815 unsigned int sysctl_numa_balancing_scan_delay = 1000;
817 static unsigned int task_nr_scan_windows(struct task_struct *p)
819 unsigned long rss = 0;
820 unsigned long nr_scan_pages;
823 * Calculations based on RSS as non-present and empty pages are skipped
824 * by the PTE scanner and NUMA hinting faults should be trapped based
827 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
828 rss = get_mm_rss(p->mm);
832 rss = round_up(rss, nr_scan_pages);
833 return rss / nr_scan_pages;
836 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
837 #define MAX_SCAN_WINDOW 2560
839 static unsigned int task_scan_min(struct task_struct *p)
841 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
842 unsigned int scan, floor;
843 unsigned int windows = 1;
845 if (scan_size < MAX_SCAN_WINDOW)
846 windows = MAX_SCAN_WINDOW / scan_size;
847 floor = 1000 / windows;
849 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
850 return max_t(unsigned int, floor, scan);
853 static unsigned int task_scan_max(struct task_struct *p)
855 unsigned int smin = task_scan_min(p);
858 /* Watch for min being lower than max due to floor calculations */
859 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
860 return max(smin, smax);
863 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
865 rq->nr_numa_running += (p->numa_preferred_nid != -1);
866 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
869 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
871 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
872 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
878 spinlock_t lock; /* nr_tasks, tasks */
883 nodemask_t active_nodes;
884 unsigned long total_faults;
886 * Faults_cpu is used to decide whether memory should move
887 * towards the CPU. As a consequence, these stats are weighted
888 * more by CPU use than by memory faults.
890 unsigned long *faults_cpu;
891 unsigned long faults[0];
894 /* Shared or private faults. */
895 #define NR_NUMA_HINT_FAULT_TYPES 2
897 /* Memory and CPU locality */
898 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
900 /* Averaged statistics, and temporary buffers. */
901 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
903 pid_t task_numa_group_id(struct task_struct *p)
905 return p->numa_group ? p->numa_group->gid : 0;
909 * The averaged statistics, shared & private, memory & cpu,
910 * occupy the first half of the array. The second half of the
911 * array is for current counters, which are averaged into the
912 * first set by task_numa_placement.
914 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
916 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
919 static inline unsigned long task_faults(struct task_struct *p, int nid)
924 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
925 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
928 static inline unsigned long group_faults(struct task_struct *p, int nid)
933 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
934 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
937 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
939 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
940 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
943 /* Handle placement on systems where not all nodes are directly connected. */
944 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
945 int maxdist, bool task)
947 unsigned long score = 0;
951 * All nodes are directly connected, and the same distance
952 * from each other. No need for fancy placement algorithms.
954 if (sched_numa_topology_type == NUMA_DIRECT)
958 * This code is called for each node, introducing N^2 complexity,
959 * which should be ok given the number of nodes rarely exceeds 8.
961 for_each_online_node(node) {
962 unsigned long faults;
963 int dist = node_distance(nid, node);
966 * The furthest away nodes in the system are not interesting
967 * for placement; nid was already counted.
969 if (dist == sched_max_numa_distance || node == nid)
973 * On systems with a backplane NUMA topology, compare groups
974 * of nodes, and move tasks towards the group with the most
975 * memory accesses. When comparing two nodes at distance
976 * "hoplimit", only nodes closer by than "hoplimit" are part
977 * of each group. Skip other nodes.
979 if (sched_numa_topology_type == NUMA_BACKPLANE &&
983 /* Add up the faults from nearby nodes. */
985 faults = task_faults(p, node);
987 faults = group_faults(p, node);
990 * On systems with a glueless mesh NUMA topology, there are
991 * no fixed "groups of nodes". Instead, nodes that are not
992 * directly connected bounce traffic through intermediate
993 * nodes; a numa_group can occupy any set of nodes.
994 * The further away a node is, the less the faults count.
995 * This seems to result in good task placement.
997 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
998 faults *= (sched_max_numa_distance - dist);
999 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1009 * These return the fraction of accesses done by a particular task, or
1010 * task group, on a particular numa node. The group weight is given a
1011 * larger multiplier, in order to group tasks together that are almost
1012 * evenly spread out between numa nodes.
1014 static inline unsigned long task_weight(struct task_struct *p, int nid,
1017 unsigned long faults, total_faults;
1019 if (!p->numa_faults)
1022 total_faults = p->total_numa_faults;
1027 faults = task_faults(p, nid);
1028 faults += score_nearby_nodes(p, nid, dist, true);
1030 return 1000 * faults / total_faults;
1033 static inline unsigned long group_weight(struct task_struct *p, int nid,
1036 unsigned long faults, total_faults;
1041 total_faults = p->numa_group->total_faults;
1046 faults = group_faults(p, nid);
1047 faults += score_nearby_nodes(p, nid, dist, false);
1049 return 1000 * faults / total_faults;
1052 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1053 int src_nid, int dst_cpu)
1055 struct numa_group *ng = p->numa_group;
1056 int dst_nid = cpu_to_node(dst_cpu);
1057 int last_cpupid, this_cpupid;
1059 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1062 * Multi-stage node selection is used in conjunction with a periodic
1063 * migration fault to build a temporal task<->page relation. By using
1064 * a two-stage filter we remove short/unlikely relations.
1066 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1067 * a task's usage of a particular page (n_p) per total usage of this
1068 * page (n_t) (in a given time-span) to a probability.
1070 * Our periodic faults will sample this probability and getting the
1071 * same result twice in a row, given these samples are fully
1072 * independent, is then given by P(n)^2, provided our sample period
1073 * is sufficiently short compared to the usage pattern.
1075 * This quadric squishes small probabilities, making it less likely we
1076 * act on an unlikely task<->page relation.
1078 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1079 if (!cpupid_pid_unset(last_cpupid) &&
1080 cpupid_to_nid(last_cpupid) != dst_nid)
1083 /* Always allow migrate on private faults */
1084 if (cpupid_match_pid(p, last_cpupid))
1087 /* A shared fault, but p->numa_group has not been set up yet. */
1092 * Do not migrate if the destination is not a node that
1093 * is actively used by this numa group.
1095 if (!node_isset(dst_nid, ng->active_nodes))
1099 * Source is a node that is not actively used by this
1100 * numa group, while the destination is. Migrate.
1102 if (!node_isset(src_nid, ng->active_nodes))
1106 * Both source and destination are nodes in active
1107 * use by this numa group. Maximize memory bandwidth
1108 * by migrating from more heavily used groups, to less
1109 * heavily used ones, spreading the load around.
1110 * Use a 1/4 hysteresis to avoid spurious page movement.
1112 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1115 static unsigned long weighted_cpuload(const int cpu);
1116 static unsigned long source_load(int cpu, int type);
1117 static unsigned long target_load(int cpu, int type);
1118 static unsigned long capacity_of(int cpu);
1119 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1121 /* Cached statistics for all CPUs within a node */
1123 unsigned long nr_running;
1126 /* Total compute capacity of CPUs on a node */
1127 unsigned long compute_capacity;
1129 /* Approximate capacity in terms of runnable tasks on a node */
1130 unsigned long task_capacity;
1131 int has_free_capacity;
1135 * XXX borrowed from update_sg_lb_stats
1137 static void update_numa_stats(struct numa_stats *ns, int nid)
1139 int smt, cpu, cpus = 0;
1140 unsigned long capacity;
1142 memset(ns, 0, sizeof(*ns));
1143 for_each_cpu(cpu, cpumask_of_node(nid)) {
1144 struct rq *rq = cpu_rq(cpu);
1146 ns->nr_running += rq->nr_running;
1147 ns->load += weighted_cpuload(cpu);
1148 ns->compute_capacity += capacity_of(cpu);
1154 * If we raced with hotplug and there are no CPUs left in our mask
1155 * the @ns structure is NULL'ed and task_numa_compare() will
1156 * not find this node attractive.
1158 * We'll either bail at !has_free_capacity, or we'll detect a huge
1159 * imbalance and bail there.
1164 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1165 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1166 capacity = cpus / smt; /* cores */
1168 ns->task_capacity = min_t(unsigned, capacity,
1169 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1170 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1173 struct task_numa_env {
1174 struct task_struct *p;
1176 int src_cpu, src_nid;
1177 int dst_cpu, dst_nid;
1179 struct numa_stats src_stats, dst_stats;
1184 struct task_struct *best_task;
1189 static void task_numa_assign(struct task_numa_env *env,
1190 struct task_struct *p, long imp)
1193 put_task_struct(env->best_task);
1196 env->best_imp = imp;
1197 env->best_cpu = env->dst_cpu;
1200 static bool load_too_imbalanced(long src_load, long dst_load,
1201 struct task_numa_env *env)
1204 long orig_src_load, orig_dst_load;
1205 long src_capacity, dst_capacity;
1208 * The load is corrected for the CPU capacity available on each node.
1211 * ------------ vs ---------
1212 * src_capacity dst_capacity
1214 src_capacity = env->src_stats.compute_capacity;
1215 dst_capacity = env->dst_stats.compute_capacity;
1217 /* We care about the slope of the imbalance, not the direction. */
1218 if (dst_load < src_load)
1219 swap(dst_load, src_load);
1221 /* Is the difference below the threshold? */
1222 imb = dst_load * src_capacity * 100 -
1223 src_load * dst_capacity * env->imbalance_pct;
1228 * The imbalance is above the allowed threshold.
1229 * Compare it with the old imbalance.
1231 orig_src_load = env->src_stats.load;
1232 orig_dst_load = env->dst_stats.load;
1234 if (orig_dst_load < orig_src_load)
1235 swap(orig_dst_load, orig_src_load);
1237 old_imb = orig_dst_load * src_capacity * 100 -
1238 orig_src_load * dst_capacity * env->imbalance_pct;
1240 /* Would this change make things worse? */
1241 return (imb > old_imb);
1245 * This checks if the overall compute and NUMA accesses of the system would
1246 * be improved if the source tasks was migrated to the target dst_cpu taking
1247 * into account that it might be best if task running on the dst_cpu should
1248 * be exchanged with the source task
1250 static void task_numa_compare(struct task_numa_env *env,
1251 long taskimp, long groupimp)
1253 struct rq *src_rq = cpu_rq(env->src_cpu);
1254 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1255 struct task_struct *cur;
1256 long src_load, dst_load;
1258 long imp = env->p->numa_group ? groupimp : taskimp;
1260 int dist = env->dist;
1261 bool assigned = false;
1265 raw_spin_lock_irq(&dst_rq->lock);
1268 * No need to move the exiting task or idle task.
1270 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1274 * The task_struct must be protected here to protect the
1275 * p->numa_faults access in the task_weight since the
1276 * numa_faults could already be freed in the following path:
1277 * finish_task_switch()
1278 * --> put_task_struct()
1279 * --> __put_task_struct()
1280 * --> task_numa_free()
1282 get_task_struct(cur);
1285 raw_spin_unlock_irq(&dst_rq->lock);
1288 * Because we have preemption enabled we can get migrated around and
1289 * end try selecting ourselves (current == env->p) as a swap candidate.
1295 * "imp" is the fault differential for the source task between the
1296 * source and destination node. Calculate the total differential for
1297 * the source task and potential destination task. The more negative
1298 * the value is, the more rmeote accesses that would be expected to
1299 * be incurred if the tasks were swapped.
1302 /* Skip this swap candidate if cannot move to the source cpu */
1303 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1307 * If dst and source tasks are in the same NUMA group, or not
1308 * in any group then look only at task weights.
1310 if (cur->numa_group == env->p->numa_group) {
1311 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1312 task_weight(cur, env->dst_nid, dist);
1314 * Add some hysteresis to prevent swapping the
1315 * tasks within a group over tiny differences.
1317 if (cur->numa_group)
1321 * Compare the group weights. If a task is all by
1322 * itself (not part of a group), use the task weight
1325 if (cur->numa_group)
1326 imp += group_weight(cur, env->src_nid, dist) -
1327 group_weight(cur, env->dst_nid, dist);
1329 imp += task_weight(cur, env->src_nid, dist) -
1330 task_weight(cur, env->dst_nid, dist);
1334 if (imp <= env->best_imp && moveimp <= env->best_imp)
1338 /* Is there capacity at our destination? */
1339 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1340 !env->dst_stats.has_free_capacity)
1346 /* Balance doesn't matter much if we're running a task per cpu */
1347 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1348 dst_rq->nr_running == 1)
1352 * In the overloaded case, try and keep the load balanced.
1355 load = task_h_load(env->p);
1356 dst_load = env->dst_stats.load + load;
1357 src_load = env->src_stats.load - load;
1359 if (moveimp > imp && moveimp > env->best_imp) {
1361 * If the improvement from just moving env->p direction is
1362 * better than swapping tasks around, check if a move is
1363 * possible. Store a slightly smaller score than moveimp,
1364 * so an actually idle CPU will win.
1366 if (!load_too_imbalanced(src_load, dst_load, env)) {
1368 put_task_struct(cur);
1374 if (imp <= env->best_imp)
1378 load = task_h_load(cur);
1383 if (load_too_imbalanced(src_load, dst_load, env))
1387 * One idle CPU per node is evaluated for a task numa move.
1388 * Call select_idle_sibling to maybe find a better one.
1391 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1395 task_numa_assign(env, cur, imp);
1399 * The dst_rq->curr isn't assigned. The protection for task_struct is
1402 if (cur && !assigned)
1403 put_task_struct(cur);
1406 static void task_numa_find_cpu(struct task_numa_env *env,
1407 long taskimp, long groupimp)
1411 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1412 /* Skip this CPU if the source task cannot migrate */
1413 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1417 task_numa_compare(env, taskimp, groupimp);
1421 /* Only move tasks to a NUMA node less busy than the current node. */
1422 static bool numa_has_capacity(struct task_numa_env *env)
1424 struct numa_stats *src = &env->src_stats;
1425 struct numa_stats *dst = &env->dst_stats;
1427 if (src->has_free_capacity && !dst->has_free_capacity)
1431 * Only consider a task move if the source has a higher load
1432 * than the destination, corrected for CPU capacity on each node.
1434 * src->load dst->load
1435 * --------------------- vs ---------------------
1436 * src->compute_capacity dst->compute_capacity
1438 if (src->load * dst->compute_capacity * env->imbalance_pct >
1440 dst->load * src->compute_capacity * 100)
1446 static int task_numa_migrate(struct task_struct *p)
1448 struct task_numa_env env = {
1451 .src_cpu = task_cpu(p),
1452 .src_nid = task_node(p),
1454 .imbalance_pct = 112,
1460 struct sched_domain *sd;
1461 unsigned long taskweight, groupweight;
1463 long taskimp, groupimp;
1466 * Pick the lowest SD_NUMA domain, as that would have the smallest
1467 * imbalance and would be the first to start moving tasks about.
1469 * And we want to avoid any moving of tasks about, as that would create
1470 * random movement of tasks -- counter the numa conditions we're trying
1474 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1476 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1480 * Cpusets can break the scheduler domain tree into smaller
1481 * balance domains, some of which do not cross NUMA boundaries.
1482 * Tasks that are "trapped" in such domains cannot be migrated
1483 * elsewhere, so there is no point in (re)trying.
1485 if (unlikely(!sd)) {
1486 p->numa_preferred_nid = task_node(p);
1490 env.dst_nid = p->numa_preferred_nid;
1491 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1492 taskweight = task_weight(p, env.src_nid, dist);
1493 groupweight = group_weight(p, env.src_nid, dist);
1494 update_numa_stats(&env.src_stats, env.src_nid);
1495 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1496 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1497 update_numa_stats(&env.dst_stats, env.dst_nid);
1499 /* Try to find a spot on the preferred nid. */
1500 if (numa_has_capacity(&env))
1501 task_numa_find_cpu(&env, taskimp, groupimp);
1504 * Look at other nodes in these cases:
1505 * - there is no space available on the preferred_nid
1506 * - the task is part of a numa_group that is interleaved across
1507 * multiple NUMA nodes; in order to better consolidate the group,
1508 * we need to check other locations.
1510 if (env.best_cpu == -1 || (p->numa_group &&
1511 nodes_weight(p->numa_group->active_nodes) > 1)) {
1512 for_each_online_node(nid) {
1513 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1516 dist = node_distance(env.src_nid, env.dst_nid);
1517 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1519 taskweight = task_weight(p, env.src_nid, dist);
1520 groupweight = group_weight(p, env.src_nid, dist);
1523 /* Only consider nodes where both task and groups benefit */
1524 taskimp = task_weight(p, nid, dist) - taskweight;
1525 groupimp = group_weight(p, nid, dist) - groupweight;
1526 if (taskimp < 0 && groupimp < 0)
1531 update_numa_stats(&env.dst_stats, env.dst_nid);
1532 if (numa_has_capacity(&env))
1533 task_numa_find_cpu(&env, taskimp, groupimp);
1538 * If the task is part of a workload that spans multiple NUMA nodes,
1539 * and is migrating into one of the workload's active nodes, remember
1540 * this node as the task's preferred numa node, so the workload can
1542 * A task that migrated to a second choice node will be better off
1543 * trying for a better one later. Do not set the preferred node here.
1545 if (p->numa_group) {
1546 if (env.best_cpu == -1)
1551 if (node_isset(nid, p->numa_group->active_nodes))
1552 sched_setnuma(p, env.dst_nid);
1555 /* No better CPU than the current one was found. */
1556 if (env.best_cpu == -1)
1560 * Reset the scan period if the task is being rescheduled on an
1561 * alternative node to recheck if the tasks is now properly placed.
1563 p->numa_scan_period = task_scan_min(p);
1565 if (env.best_task == NULL) {
1566 ret = migrate_task_to(p, env.best_cpu);
1568 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1572 ret = migrate_swap(p, env.best_task);
1574 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1575 put_task_struct(env.best_task);
1579 /* Attempt to migrate a task to a CPU on the preferred node. */
1580 static void numa_migrate_preferred(struct task_struct *p)
1582 unsigned long interval = HZ;
1584 /* This task has no NUMA fault statistics yet */
1585 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1588 /* Periodically retry migrating the task to the preferred node */
1589 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1590 p->numa_migrate_retry = jiffies + interval;
1592 /* Success if task is already running on preferred CPU */
1593 if (task_node(p) == p->numa_preferred_nid)
1596 /* Otherwise, try migrate to a CPU on the preferred node */
1597 task_numa_migrate(p);
1601 * Find the nodes on which the workload is actively running. We do this by
1602 * tracking the nodes from which NUMA hinting faults are triggered. This can
1603 * be different from the set of nodes where the workload's memory is currently
1606 * The bitmask is used to make smarter decisions on when to do NUMA page
1607 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1608 * are added when they cause over 6/16 of the maximum number of faults, but
1609 * only removed when they drop below 3/16.
1611 static void update_numa_active_node_mask(struct numa_group *numa_group)
1613 unsigned long faults, max_faults = 0;
1616 for_each_online_node(nid) {
1617 faults = group_faults_cpu(numa_group, nid);
1618 if (faults > max_faults)
1619 max_faults = faults;
1622 for_each_online_node(nid) {
1623 faults = group_faults_cpu(numa_group, nid);
1624 if (!node_isset(nid, numa_group->active_nodes)) {
1625 if (faults > max_faults * 6 / 16)
1626 node_set(nid, numa_group->active_nodes);
1627 } else if (faults < max_faults * 3 / 16)
1628 node_clear(nid, numa_group->active_nodes);
1633 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1634 * increments. The more local the fault statistics are, the higher the scan
1635 * period will be for the next scan window. If local/(local+remote) ratio is
1636 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1637 * the scan period will decrease. Aim for 70% local accesses.
1639 #define NUMA_PERIOD_SLOTS 10
1640 #define NUMA_PERIOD_THRESHOLD 7
1643 * Increase the scan period (slow down scanning) if the majority of
1644 * our memory is already on our local node, or if the majority of
1645 * the page accesses are shared with other processes.
1646 * Otherwise, decrease the scan period.
1648 static void update_task_scan_period(struct task_struct *p,
1649 unsigned long shared, unsigned long private)
1651 unsigned int period_slot;
1655 unsigned long remote = p->numa_faults_locality[0];
1656 unsigned long local = p->numa_faults_locality[1];
1659 * If there were no record hinting faults then either the task is
1660 * completely idle or all activity is areas that are not of interest
1661 * to automatic numa balancing. Related to that, if there were failed
1662 * migration then it implies we are migrating too quickly or the local
1663 * node is overloaded. In either case, scan slower
1665 if (local + shared == 0 || p->numa_faults_locality[2]) {
1666 p->numa_scan_period = min(p->numa_scan_period_max,
1667 p->numa_scan_period << 1);
1669 p->mm->numa_next_scan = jiffies +
1670 msecs_to_jiffies(p->numa_scan_period);
1676 * Prepare to scale scan period relative to the current period.
1677 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1678 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1679 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1681 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1682 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1683 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1684 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1687 diff = slot * period_slot;
1689 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1692 * Scale scan rate increases based on sharing. There is an
1693 * inverse relationship between the degree of sharing and
1694 * the adjustment made to the scanning period. Broadly
1695 * speaking the intent is that there is little point
1696 * scanning faster if shared accesses dominate as it may
1697 * simply bounce migrations uselessly
1699 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1700 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1703 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1704 task_scan_min(p), task_scan_max(p));
1705 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1709 * Get the fraction of time the task has been running since the last
1710 * NUMA placement cycle. The scheduler keeps similar statistics, but
1711 * decays those on a 32ms period, which is orders of magnitude off
1712 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1713 * stats only if the task is so new there are no NUMA statistics yet.
1715 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1717 u64 runtime, delta, now;
1718 /* Use the start of this time slice to avoid calculations. */
1719 now = p->se.exec_start;
1720 runtime = p->se.sum_exec_runtime;
1722 if (p->last_task_numa_placement) {
1723 delta = runtime - p->last_sum_exec_runtime;
1724 *period = now - p->last_task_numa_placement;
1726 /* Avoid time going backwards, prevent potential divide error: */
1727 if (unlikely((s64)*period < 0))
1730 delta = p->se.avg.load_sum / p->se.load.weight;
1731 *period = LOAD_AVG_MAX;
1734 p->last_sum_exec_runtime = runtime;
1735 p->last_task_numa_placement = now;
1741 * Determine the preferred nid for a task in a numa_group. This needs to
1742 * be done in a way that produces consistent results with group_weight,
1743 * otherwise workloads might not converge.
1745 static int preferred_group_nid(struct task_struct *p, int nid)
1750 /* Direct connections between all NUMA nodes. */
1751 if (sched_numa_topology_type == NUMA_DIRECT)
1755 * On a system with glueless mesh NUMA topology, group_weight
1756 * scores nodes according to the number of NUMA hinting faults on
1757 * both the node itself, and on nearby nodes.
1759 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1760 unsigned long score, max_score = 0;
1761 int node, max_node = nid;
1763 dist = sched_max_numa_distance;
1765 for_each_online_node(node) {
1766 score = group_weight(p, node, dist);
1767 if (score > max_score) {
1776 * Finding the preferred nid in a system with NUMA backplane
1777 * interconnect topology is more involved. The goal is to locate
1778 * tasks from numa_groups near each other in the system, and
1779 * untangle workloads from different sides of the system. This requires
1780 * searching down the hierarchy of node groups, recursively searching
1781 * inside the highest scoring group of nodes. The nodemask tricks
1782 * keep the complexity of the search down.
1784 nodes = node_online_map;
1785 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1786 unsigned long max_faults = 0;
1787 nodemask_t max_group = NODE_MASK_NONE;
1790 /* Are there nodes at this distance from each other? */
1791 if (!find_numa_distance(dist))
1794 for_each_node_mask(a, nodes) {
1795 unsigned long faults = 0;
1796 nodemask_t this_group;
1797 nodes_clear(this_group);
1799 /* Sum group's NUMA faults; includes a==b case. */
1800 for_each_node_mask(b, nodes) {
1801 if (node_distance(a, b) < dist) {
1802 faults += group_faults(p, b);
1803 node_set(b, this_group);
1804 node_clear(b, nodes);
1808 /* Remember the top group. */
1809 if (faults > max_faults) {
1810 max_faults = faults;
1811 max_group = this_group;
1813 * subtle: at the smallest distance there is
1814 * just one node left in each "group", the
1815 * winner is the preferred nid.
1820 /* Next round, evaluate the nodes within max_group. */
1828 static void task_numa_placement(struct task_struct *p)
1830 int seq, nid, max_nid = -1, max_group_nid = -1;
1831 unsigned long max_faults = 0, max_group_faults = 0;
1832 unsigned long fault_types[2] = { 0, 0 };
1833 unsigned long total_faults;
1834 u64 runtime, period;
1835 spinlock_t *group_lock = NULL;
1838 * The p->mm->numa_scan_seq field gets updated without
1839 * exclusive access. Use READ_ONCE() here to ensure
1840 * that the field is read in a single access:
1842 seq = READ_ONCE(p->mm->numa_scan_seq);
1843 if (p->numa_scan_seq == seq)
1845 p->numa_scan_seq = seq;
1846 p->numa_scan_period_max = task_scan_max(p);
1848 total_faults = p->numa_faults_locality[0] +
1849 p->numa_faults_locality[1];
1850 runtime = numa_get_avg_runtime(p, &period);
1852 /* If the task is part of a group prevent parallel updates to group stats */
1853 if (p->numa_group) {
1854 group_lock = &p->numa_group->lock;
1855 spin_lock_irq(group_lock);
1858 /* Find the node with the highest number of faults */
1859 for_each_online_node(nid) {
1860 /* Keep track of the offsets in numa_faults array */
1861 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1862 unsigned long faults = 0, group_faults = 0;
1865 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1866 long diff, f_diff, f_weight;
1868 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1869 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1870 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1871 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1873 /* Decay existing window, copy faults since last scan */
1874 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1875 fault_types[priv] += p->numa_faults[membuf_idx];
1876 p->numa_faults[membuf_idx] = 0;
1879 * Normalize the faults_from, so all tasks in a group
1880 * count according to CPU use, instead of by the raw
1881 * number of faults. Tasks with little runtime have
1882 * little over-all impact on throughput, and thus their
1883 * faults are less important.
1885 f_weight = div64_u64(runtime << 16, period + 1);
1886 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1888 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1889 p->numa_faults[cpubuf_idx] = 0;
1891 p->numa_faults[mem_idx] += diff;
1892 p->numa_faults[cpu_idx] += f_diff;
1893 faults += p->numa_faults[mem_idx];
1894 p->total_numa_faults += diff;
1895 if (p->numa_group) {
1897 * safe because we can only change our own group
1899 * mem_idx represents the offset for a given
1900 * nid and priv in a specific region because it
1901 * is at the beginning of the numa_faults array.
1903 p->numa_group->faults[mem_idx] += diff;
1904 p->numa_group->faults_cpu[mem_idx] += f_diff;
1905 p->numa_group->total_faults += diff;
1906 group_faults += p->numa_group->faults[mem_idx];
1910 if (faults > max_faults) {
1911 max_faults = faults;
1915 if (group_faults > max_group_faults) {
1916 max_group_faults = group_faults;
1917 max_group_nid = nid;
1921 update_task_scan_period(p, fault_types[0], fault_types[1]);
1923 if (p->numa_group) {
1924 update_numa_active_node_mask(p->numa_group);
1925 spin_unlock_irq(group_lock);
1926 max_nid = preferred_group_nid(p, max_group_nid);
1930 /* Set the new preferred node */
1931 if (max_nid != p->numa_preferred_nid)
1932 sched_setnuma(p, max_nid);
1934 if (task_node(p) != p->numa_preferred_nid)
1935 numa_migrate_preferred(p);
1939 static inline int get_numa_group(struct numa_group *grp)
1941 return atomic_inc_not_zero(&grp->refcount);
1944 static inline void put_numa_group(struct numa_group *grp)
1946 if (atomic_dec_and_test(&grp->refcount))
1947 kfree_rcu(grp, rcu);
1950 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1953 struct numa_group *grp, *my_grp;
1954 struct task_struct *tsk;
1956 int cpu = cpupid_to_cpu(cpupid);
1959 if (unlikely(!p->numa_group)) {
1960 unsigned int size = sizeof(struct numa_group) +
1961 4*nr_node_ids*sizeof(unsigned long);
1963 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1967 atomic_set(&grp->refcount, 1);
1968 spin_lock_init(&grp->lock);
1970 /* Second half of the array tracks nids where faults happen */
1971 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1974 node_set(task_node(current), grp->active_nodes);
1976 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1977 grp->faults[i] = p->numa_faults[i];
1979 grp->total_faults = p->total_numa_faults;
1982 rcu_assign_pointer(p->numa_group, grp);
1986 tsk = READ_ONCE(cpu_rq(cpu)->curr);
1988 if (!cpupid_match_pid(tsk, cpupid))
1991 grp = rcu_dereference(tsk->numa_group);
1995 my_grp = p->numa_group;
2000 * Only join the other group if its bigger; if we're the bigger group,
2001 * the other task will join us.
2003 if (my_grp->nr_tasks > grp->nr_tasks)
2007 * Tie-break on the grp address.
2009 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2012 /* Always join threads in the same process. */
2013 if (tsk->mm == current->mm)
2016 /* Simple filter to avoid false positives due to PID collisions */
2017 if (flags & TNF_SHARED)
2020 /* Update priv based on whether false sharing was detected */
2023 if (join && !get_numa_group(grp))
2031 BUG_ON(irqs_disabled());
2032 double_lock_irq(&my_grp->lock, &grp->lock);
2034 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2035 my_grp->faults[i] -= p->numa_faults[i];
2036 grp->faults[i] += p->numa_faults[i];
2038 my_grp->total_faults -= p->total_numa_faults;
2039 grp->total_faults += p->total_numa_faults;
2044 spin_unlock(&my_grp->lock);
2045 spin_unlock_irq(&grp->lock);
2047 rcu_assign_pointer(p->numa_group, grp);
2049 put_numa_group(my_grp);
2057 void task_numa_free(struct task_struct *p)
2059 struct numa_group *grp = p->numa_group;
2060 void *numa_faults = p->numa_faults;
2061 unsigned long flags;
2065 spin_lock_irqsave(&grp->lock, flags);
2066 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2067 grp->faults[i] -= p->numa_faults[i];
2068 grp->total_faults -= p->total_numa_faults;
2071 spin_unlock_irqrestore(&grp->lock, flags);
2072 RCU_INIT_POINTER(p->numa_group, NULL);
2073 put_numa_group(grp);
2076 p->numa_faults = NULL;
2081 * Got a PROT_NONE fault for a page on @node.
2083 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2085 struct task_struct *p = current;
2086 bool migrated = flags & TNF_MIGRATED;
2087 int cpu_node = task_node(current);
2088 int local = !!(flags & TNF_FAULT_LOCAL);
2091 if (!static_branch_likely(&sched_numa_balancing))
2094 /* for example, ksmd faulting in a user's mm */
2098 /* Allocate buffer to track faults on a per-node basis */
2099 if (unlikely(!p->numa_faults)) {
2100 int size = sizeof(*p->numa_faults) *
2101 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2103 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2104 if (!p->numa_faults)
2107 p->total_numa_faults = 0;
2108 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2112 * First accesses are treated as private, otherwise consider accesses
2113 * to be private if the accessing pid has not changed
2115 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2118 priv = cpupid_match_pid(p, last_cpupid);
2119 if (!priv && !(flags & TNF_NO_GROUP))
2120 task_numa_group(p, last_cpupid, flags, &priv);
2124 * If a workload spans multiple NUMA nodes, a shared fault that
2125 * occurs wholly within the set of nodes that the workload is
2126 * actively using should be counted as local. This allows the
2127 * scan rate to slow down when a workload has settled down.
2129 if (!priv && !local && p->numa_group &&
2130 node_isset(cpu_node, p->numa_group->active_nodes) &&
2131 node_isset(mem_node, p->numa_group->active_nodes))
2134 task_numa_placement(p);
2137 * Retry task to preferred node migration periodically, in case it
2138 * case it previously failed, or the scheduler moved us.
2140 if (time_after(jiffies, p->numa_migrate_retry))
2141 numa_migrate_preferred(p);
2144 p->numa_pages_migrated += pages;
2145 if (flags & TNF_MIGRATE_FAIL)
2146 p->numa_faults_locality[2] += pages;
2148 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2149 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2150 p->numa_faults_locality[local] += pages;
2153 static void reset_ptenuma_scan(struct task_struct *p)
2156 * We only did a read acquisition of the mmap sem, so
2157 * p->mm->numa_scan_seq is written to without exclusive access
2158 * and the update is not guaranteed to be atomic. That's not
2159 * much of an issue though, since this is just used for
2160 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2161 * expensive, to avoid any form of compiler optimizations:
2163 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2164 p->mm->numa_scan_offset = 0;
2168 * The expensive part of numa migration is done from task_work context.
2169 * Triggered from task_tick_numa().
2171 void task_numa_work(struct callback_head *work)
2173 unsigned long migrate, next_scan, now = jiffies;
2174 struct task_struct *p = current;
2175 struct mm_struct *mm = p->mm;
2176 struct vm_area_struct *vma;
2177 unsigned long start, end;
2178 unsigned long nr_pte_updates = 0;
2179 long pages, virtpages;
2181 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2183 work->next = work; /* protect against double add */
2185 * Who cares about NUMA placement when they're dying.
2187 * NOTE: make sure not to dereference p->mm before this check,
2188 * exit_task_work() happens _after_ exit_mm() so we could be called
2189 * without p->mm even though we still had it when we enqueued this
2192 if (p->flags & PF_EXITING)
2195 if (!mm->numa_next_scan) {
2196 mm->numa_next_scan = now +
2197 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2201 * Enforce maximal scan/migration frequency..
2203 migrate = mm->numa_next_scan;
2204 if (time_before(now, migrate))
2207 if (p->numa_scan_period == 0) {
2208 p->numa_scan_period_max = task_scan_max(p);
2209 p->numa_scan_period = task_scan_min(p);
2212 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2213 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2217 * Delay this task enough that another task of this mm will likely win
2218 * the next time around.
2220 p->node_stamp += 2 * TICK_NSEC;
2222 start = mm->numa_scan_offset;
2223 pages = sysctl_numa_balancing_scan_size;
2224 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2225 virtpages = pages * 8; /* Scan up to this much virtual space */
2230 if (!down_read_trylock(&mm->mmap_sem))
2232 vma = find_vma(mm, start);
2234 reset_ptenuma_scan(p);
2238 for (; vma; vma = vma->vm_next) {
2239 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2240 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2245 * Shared library pages mapped by multiple processes are not
2246 * migrated as it is expected they are cache replicated. Avoid
2247 * hinting faults in read-only file-backed mappings or the vdso
2248 * as migrating the pages will be of marginal benefit.
2251 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2255 * Skip inaccessible VMAs to avoid any confusion between
2256 * PROT_NONE and NUMA hinting ptes
2258 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2262 start = max(start, vma->vm_start);
2263 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2264 end = min(end, vma->vm_end);
2265 nr_pte_updates = change_prot_numa(vma, start, end);
2268 * Try to scan sysctl_numa_balancing_size worth of
2269 * hpages that have at least one present PTE that
2270 * is not already pte-numa. If the VMA contains
2271 * areas that are unused or already full of prot_numa
2272 * PTEs, scan up to virtpages, to skip through those
2276 pages -= (end - start) >> PAGE_SHIFT;
2277 virtpages -= (end - start) >> PAGE_SHIFT;
2280 if (pages <= 0 || virtpages <= 0)
2284 } while (end != vma->vm_end);
2289 * It is possible to reach the end of the VMA list but the last few
2290 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2291 * would find the !migratable VMA on the next scan but not reset the
2292 * scanner to the start so check it now.
2295 mm->numa_scan_offset = start;
2297 reset_ptenuma_scan(p);
2298 up_read(&mm->mmap_sem);
2302 * Drive the periodic memory faults..
2304 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2306 struct callback_head *work = &curr->numa_work;
2310 * We don't care about NUMA placement if we don't have memory.
2312 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2316 * Using runtime rather than walltime has the dual advantage that
2317 * we (mostly) drive the selection from busy threads and that the
2318 * task needs to have done some actual work before we bother with
2321 now = curr->se.sum_exec_runtime;
2322 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2324 if (now > curr->node_stamp + period) {
2325 if (!curr->node_stamp)
2326 curr->numa_scan_period = task_scan_min(curr);
2327 curr->node_stamp += period;
2329 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2330 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2331 task_work_add(curr, work, true);
2336 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2340 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2344 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2347 #endif /* CONFIG_NUMA_BALANCING */
2350 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2352 update_load_add(&cfs_rq->load, se->load.weight);
2353 if (!parent_entity(se))
2354 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2356 if (entity_is_task(se)) {
2357 struct rq *rq = rq_of(cfs_rq);
2359 account_numa_enqueue(rq, task_of(se));
2360 list_add(&se->group_node, &rq->cfs_tasks);
2363 cfs_rq->nr_running++;
2367 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2369 update_load_sub(&cfs_rq->load, se->load.weight);
2370 if (!parent_entity(se))
2371 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2372 if (entity_is_task(se)) {
2373 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2374 list_del_init(&se->group_node);
2376 cfs_rq->nr_running--;
2379 #ifdef CONFIG_FAIR_GROUP_SCHED
2381 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2386 * Use this CPU's real-time load instead of the last load contribution
2387 * as the updating of the contribution is delayed, and we will use the
2388 * the real-time load to calc the share. See update_tg_load_avg().
2390 tg_weight = atomic_long_read(&tg->load_avg);
2391 tg_weight -= cfs_rq->tg_load_avg_contrib;
2392 tg_weight += cfs_rq->load.weight;
2397 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2399 long tg_weight, load, shares;
2401 tg_weight = calc_tg_weight(tg, cfs_rq);
2402 load = cfs_rq->load.weight;
2404 shares = (tg->shares * load);
2406 shares /= tg_weight;
2408 if (shares < MIN_SHARES)
2409 shares = MIN_SHARES;
2410 if (shares > tg->shares)
2411 shares = tg->shares;
2415 # else /* CONFIG_SMP */
2416 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2420 # endif /* CONFIG_SMP */
2421 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2422 unsigned long weight)
2425 /* commit outstanding execution time */
2426 if (cfs_rq->curr == se)
2427 update_curr(cfs_rq);
2428 account_entity_dequeue(cfs_rq, se);
2431 update_load_set(&se->load, weight);
2434 account_entity_enqueue(cfs_rq, se);
2437 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2439 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2441 struct task_group *tg;
2442 struct sched_entity *se;
2446 se = tg->se[cpu_of(rq_of(cfs_rq))];
2447 if (!se || throttled_hierarchy(cfs_rq))
2450 if (likely(se->load.weight == tg->shares))
2453 shares = calc_cfs_shares(cfs_rq, tg);
2455 reweight_entity(cfs_rq_of(se), se, shares);
2457 #else /* CONFIG_FAIR_GROUP_SCHED */
2458 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2461 #endif /* CONFIG_FAIR_GROUP_SCHED */
2464 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2465 static const u32 runnable_avg_yN_inv[] = {
2466 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2467 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2468 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2469 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2470 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2471 0x85aac367, 0x82cd8698,
2475 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2476 * over-estimates when re-combining.
2478 static const u32 runnable_avg_yN_sum[] = {
2479 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2480 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2481 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2486 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2488 static __always_inline u64 decay_load(u64 val, u64 n)
2490 unsigned int local_n;
2494 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2497 /* after bounds checking we can collapse to 32-bit */
2501 * As y^PERIOD = 1/2, we can combine
2502 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2503 * With a look-up table which covers y^n (n<PERIOD)
2505 * To achieve constant time decay_load.
2507 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2508 val >>= local_n / LOAD_AVG_PERIOD;
2509 local_n %= LOAD_AVG_PERIOD;
2512 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2517 * For updates fully spanning n periods, the contribution to runnable
2518 * average will be: \Sum 1024*y^n
2520 * We can compute this reasonably efficiently by combining:
2521 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2523 static u32 __compute_runnable_contrib(u64 n)
2527 if (likely(n <= LOAD_AVG_PERIOD))
2528 return runnable_avg_yN_sum[n];
2529 else if (unlikely(n >= LOAD_AVG_MAX_N))
2530 return LOAD_AVG_MAX;
2532 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2534 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2535 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2537 n -= LOAD_AVG_PERIOD;
2538 } while (n > LOAD_AVG_PERIOD);
2540 contrib = decay_load(contrib, n);
2541 return contrib + runnable_avg_yN_sum[n];
2544 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2545 #error "load tracking assumes 2^10 as unit"
2548 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2551 * We can represent the historical contribution to runnable average as the
2552 * coefficients of a geometric series. To do this we sub-divide our runnable
2553 * history into segments of approximately 1ms (1024us); label the segment that
2554 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2556 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2558 * (now) (~1ms ago) (~2ms ago)
2560 * Let u_i denote the fraction of p_i that the entity was runnable.
2562 * We then designate the fractions u_i as our co-efficients, yielding the
2563 * following representation of historical load:
2564 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2566 * We choose y based on the with of a reasonably scheduling period, fixing:
2569 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2570 * approximately half as much as the contribution to load within the last ms
2573 * When a period "rolls over" and we have new u_0`, multiplying the previous
2574 * sum again by y is sufficient to update:
2575 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2576 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2578 static __always_inline int
2579 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2580 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2582 u64 delta, scaled_delta, periods;
2584 unsigned int delta_w, scaled_delta_w, decayed = 0;
2585 unsigned long scale_freq, scale_cpu;
2587 delta = now - sa->last_update_time;
2589 * This should only happen when time goes backwards, which it
2590 * unfortunately does during sched clock init when we swap over to TSC.
2592 if ((s64)delta < 0) {
2593 sa->last_update_time = now;
2598 * Use 1024ns as the unit of measurement since it's a reasonable
2599 * approximation of 1us and fast to compute.
2604 sa->last_update_time = now;
2606 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2607 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2609 /* delta_w is the amount already accumulated against our next period */
2610 delta_w = sa->period_contrib;
2611 if (delta + delta_w >= 1024) {
2614 /* how much left for next period will start over, we don't know yet */
2615 sa->period_contrib = 0;
2618 * Now that we know we're crossing a period boundary, figure
2619 * out how much from delta we need to complete the current
2620 * period and accrue it.
2622 delta_w = 1024 - delta_w;
2623 scaled_delta_w = cap_scale(delta_w, scale_freq);
2625 sa->load_sum += weight * scaled_delta_w;
2627 cfs_rq->runnable_load_sum +=
2628 weight * scaled_delta_w;
2632 sa->util_sum += scaled_delta_w * scale_cpu;
2636 /* Figure out how many additional periods this update spans */
2637 periods = delta / 1024;
2640 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2642 cfs_rq->runnable_load_sum =
2643 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2645 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2647 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2648 contrib = __compute_runnable_contrib(periods);
2649 contrib = cap_scale(contrib, scale_freq);
2651 sa->load_sum += weight * contrib;
2653 cfs_rq->runnable_load_sum += weight * contrib;
2656 sa->util_sum += contrib * scale_cpu;
2659 /* Remainder of delta accrued against u_0` */
2660 scaled_delta = cap_scale(delta, scale_freq);
2662 sa->load_sum += weight * scaled_delta;
2664 cfs_rq->runnable_load_sum += weight * scaled_delta;
2667 sa->util_sum += scaled_delta * scale_cpu;
2669 sa->period_contrib += delta;
2672 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2674 cfs_rq->runnable_load_avg =
2675 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2677 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2683 #ifdef CONFIG_FAIR_GROUP_SCHED
2685 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2686 * and effective_load (which is not done because it is too costly).
2688 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2690 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2692 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2693 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2694 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2698 #else /* CONFIG_FAIR_GROUP_SCHED */
2699 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2700 #endif /* CONFIG_FAIR_GROUP_SCHED */
2702 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2705 * Unsigned subtract and clamp on underflow.
2707 * Explicitly do a load-store to ensure the intermediate value never hits
2708 * memory. This allows lockless observations without ever seeing the negative
2711 #define sub_positive(_ptr, _val) do { \
2712 typeof(_ptr) ptr = (_ptr); \
2713 typeof(*ptr) val = (_val); \
2714 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2718 WRITE_ONCE(*ptr, res); \
2721 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2722 static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
2724 struct sched_avg *sa = &cfs_rq->avg;
2725 int decayed, removed = 0;
2727 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2728 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2729 sub_positive(&sa->load_avg, r);
2730 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
2734 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2735 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2736 sub_positive(&sa->util_avg, r);
2737 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
2740 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2741 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2743 #ifndef CONFIG_64BIT
2745 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2748 return decayed || removed;
2751 /* Update task and its cfs_rq load average */
2752 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2754 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2755 u64 now = cfs_rq_clock_task(cfs_rq);
2756 int cpu = cpu_of(rq_of(cfs_rq));
2759 * Track task load average for carrying it to new CPU after migrated, and
2760 * track group sched_entity load average for task_h_load calc in migration
2762 __update_load_avg(now, cpu, &se->avg,
2763 se->on_rq * scale_load_down(se->load.weight),
2764 cfs_rq->curr == se, NULL);
2766 if (update_cfs_rq_load_avg(now, cfs_rq) && update_tg)
2767 update_tg_load_avg(cfs_rq, 0);
2770 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2772 if (!sched_feat(ATTACH_AGE_LOAD))
2776 * If we got migrated (either between CPUs or between cgroups) we'll
2777 * have aged the average right before clearing @last_update_time.
2779 if (se->avg.last_update_time) {
2780 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2781 &se->avg, 0, 0, NULL);
2784 * XXX: we could have just aged the entire load away if we've been
2785 * absent from the fair class for too long.
2790 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2791 cfs_rq->avg.load_avg += se->avg.load_avg;
2792 cfs_rq->avg.load_sum += se->avg.load_sum;
2793 cfs_rq->avg.util_avg += se->avg.util_avg;
2794 cfs_rq->avg.util_sum += se->avg.util_sum;
2797 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2799 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2800 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2801 cfs_rq->curr == se, NULL);
2803 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2804 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
2805 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
2806 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
2809 /* Add the load generated by se into cfs_rq's load average */
2811 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2813 struct sched_avg *sa = &se->avg;
2814 u64 now = cfs_rq_clock_task(cfs_rq);
2815 int migrated, decayed;
2817 migrated = !sa->last_update_time;
2819 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2820 se->on_rq * scale_load_down(se->load.weight),
2821 cfs_rq->curr == se, NULL);
2824 decayed = update_cfs_rq_load_avg(now, cfs_rq);
2826 cfs_rq->runnable_load_avg += sa->load_avg;
2827 cfs_rq->runnable_load_sum += sa->load_sum;
2830 attach_entity_load_avg(cfs_rq, se);
2832 if (decayed || migrated)
2833 update_tg_load_avg(cfs_rq, 0);
2836 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2838 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2840 update_load_avg(se, 1);
2842 cfs_rq->runnable_load_avg =
2843 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2844 cfs_rq->runnable_load_sum =
2845 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2848 #ifndef CONFIG_64BIT
2849 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2851 u64 last_update_time_copy;
2852 u64 last_update_time;
2855 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2857 last_update_time = cfs_rq->avg.last_update_time;
2858 } while (last_update_time != last_update_time_copy);
2860 return last_update_time;
2863 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2865 return cfs_rq->avg.last_update_time;
2870 * Task first catches up with cfs_rq, and then subtract
2871 * itself from the cfs_rq (task must be off the queue now).
2873 void remove_entity_load_avg(struct sched_entity *se)
2875 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2876 u64 last_update_time;
2879 * Newly created task or never used group entity should not be removed
2880 * from its (source) cfs_rq
2882 if (se->avg.last_update_time == 0)
2885 last_update_time = cfs_rq_last_update_time(cfs_rq);
2887 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
2888 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2889 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2893 * Update the rq's load with the elapsed running time before entering
2894 * idle. if the last scheduled task is not a CFS task, idle_enter will
2895 * be the only way to update the runnable statistic.
2897 void idle_enter_fair(struct rq *this_rq)
2902 * Update the rq's load with the elapsed idle time before a task is
2903 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2904 * be the only way to update the runnable statistic.
2906 void idle_exit_fair(struct rq *this_rq)
2910 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
2912 return cfs_rq->runnable_load_avg;
2915 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
2917 return cfs_rq->avg.load_avg;
2920 static int idle_balance(struct rq *this_rq);
2922 #else /* CONFIG_SMP */
2924 static inline void update_load_avg(struct sched_entity *se, int update_tg) {}
2926 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2928 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2929 static inline void remove_entity_load_avg(struct sched_entity *se) {}
2932 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2934 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2936 static inline int idle_balance(struct rq *rq)
2941 #endif /* CONFIG_SMP */
2943 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2945 #ifdef CONFIG_SCHEDSTATS
2946 struct task_struct *tsk = NULL;
2948 if (entity_is_task(se))
2951 if (se->statistics.sleep_start) {
2952 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2957 if (unlikely(delta > se->statistics.sleep_max))
2958 se->statistics.sleep_max = delta;
2960 se->statistics.sleep_start = 0;
2961 se->statistics.sum_sleep_runtime += delta;
2964 account_scheduler_latency(tsk, delta >> 10, 1);
2965 trace_sched_stat_sleep(tsk, delta);
2968 if (se->statistics.block_start) {
2969 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2974 if (unlikely(delta > se->statistics.block_max))
2975 se->statistics.block_max = delta;
2977 se->statistics.block_start = 0;
2978 se->statistics.sum_sleep_runtime += delta;
2981 if (tsk->in_iowait) {
2982 se->statistics.iowait_sum += delta;
2983 se->statistics.iowait_count++;
2984 trace_sched_stat_iowait(tsk, delta);
2987 trace_sched_stat_blocked(tsk, delta);
2990 * Blocking time is in units of nanosecs, so shift by
2991 * 20 to get a milliseconds-range estimation of the
2992 * amount of time that the task spent sleeping:
2994 if (unlikely(prof_on == SLEEP_PROFILING)) {
2995 profile_hits(SLEEP_PROFILING,
2996 (void *)get_wchan(tsk),
2999 account_scheduler_latency(tsk, delta >> 10, 0);
3005 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3007 #ifdef CONFIG_SCHED_DEBUG
3008 s64 d = se->vruntime - cfs_rq->min_vruntime;
3013 if (d > 3*sysctl_sched_latency)
3014 schedstat_inc(cfs_rq, nr_spread_over);
3019 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3021 u64 vruntime = cfs_rq->min_vruntime;
3024 * The 'current' period is already promised to the current tasks,
3025 * however the extra weight of the new task will slow them down a
3026 * little, place the new task so that it fits in the slot that
3027 * stays open at the end.
3029 if (initial && sched_feat(START_DEBIT))
3030 vruntime += sched_vslice(cfs_rq, se);
3032 /* sleeps up to a single latency don't count. */
3034 unsigned long thresh = sysctl_sched_latency;
3037 * Halve their sleep time's effect, to allow
3038 * for a gentler effect of sleepers:
3040 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3046 /* ensure we never gain time by being placed backwards. */
3047 se->vruntime = max_vruntime(se->vruntime, vruntime);
3050 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3053 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3056 * Update the normalized vruntime before updating min_vruntime
3057 * through calling update_curr().
3059 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3060 se->vruntime += cfs_rq->min_vruntime;
3063 * Update run-time statistics of the 'current'.
3065 update_curr(cfs_rq);
3066 enqueue_entity_load_avg(cfs_rq, se);
3067 account_entity_enqueue(cfs_rq, se);
3068 update_cfs_shares(cfs_rq);
3070 if (flags & ENQUEUE_WAKEUP) {
3071 place_entity(cfs_rq, se, 0);
3072 enqueue_sleeper(cfs_rq, se);
3075 update_stats_enqueue(cfs_rq, se);
3076 check_spread(cfs_rq, se);
3077 if (se != cfs_rq->curr)
3078 __enqueue_entity(cfs_rq, se);
3081 if (cfs_rq->nr_running == 1) {
3082 list_add_leaf_cfs_rq(cfs_rq);
3083 check_enqueue_throttle(cfs_rq);
3087 static void __clear_buddies_last(struct sched_entity *se)
3089 for_each_sched_entity(se) {
3090 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3091 if (cfs_rq->last != se)
3094 cfs_rq->last = NULL;
3098 static void __clear_buddies_next(struct sched_entity *se)
3100 for_each_sched_entity(se) {
3101 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3102 if (cfs_rq->next != se)
3105 cfs_rq->next = NULL;
3109 static void __clear_buddies_skip(struct sched_entity *se)
3111 for_each_sched_entity(se) {
3112 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3113 if (cfs_rq->skip != se)
3116 cfs_rq->skip = NULL;
3120 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3122 if (cfs_rq->last == se)
3123 __clear_buddies_last(se);
3125 if (cfs_rq->next == se)
3126 __clear_buddies_next(se);
3128 if (cfs_rq->skip == se)
3129 __clear_buddies_skip(se);
3132 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3135 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3138 * Update run-time statistics of the 'current'.
3140 update_curr(cfs_rq);
3141 dequeue_entity_load_avg(cfs_rq, se);
3143 update_stats_dequeue(cfs_rq, se);
3144 if (flags & DEQUEUE_SLEEP) {
3145 #ifdef CONFIG_SCHEDSTATS
3146 if (entity_is_task(se)) {
3147 struct task_struct *tsk = task_of(se);
3149 if (tsk->state & TASK_INTERRUPTIBLE)
3150 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3151 if (tsk->state & TASK_UNINTERRUPTIBLE)
3152 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3157 clear_buddies(cfs_rq, se);
3159 if (se != cfs_rq->curr)
3160 __dequeue_entity(cfs_rq, se);
3162 account_entity_dequeue(cfs_rq, se);
3165 * Normalize the entity after updating the min_vruntime because the
3166 * update can refer to the ->curr item and we need to reflect this
3167 * movement in our normalized position.
3169 if (!(flags & DEQUEUE_SLEEP))
3170 se->vruntime -= cfs_rq->min_vruntime;
3172 /* return excess runtime on last dequeue */
3173 return_cfs_rq_runtime(cfs_rq);
3175 update_min_vruntime(cfs_rq);
3176 update_cfs_shares(cfs_rq);
3180 * Preempt the current task with a newly woken task if needed:
3183 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3185 unsigned long ideal_runtime, delta_exec;
3186 struct sched_entity *se;
3189 ideal_runtime = sched_slice(cfs_rq, curr);
3190 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3191 if (delta_exec > ideal_runtime) {
3192 resched_curr(rq_of(cfs_rq));
3194 * The current task ran long enough, ensure it doesn't get
3195 * re-elected due to buddy favours.
3197 clear_buddies(cfs_rq, curr);
3202 * Ensure that a task that missed wakeup preemption by a
3203 * narrow margin doesn't have to wait for a full slice.
3204 * This also mitigates buddy induced latencies under load.
3206 if (delta_exec < sysctl_sched_min_granularity)
3209 se = __pick_first_entity(cfs_rq);
3210 delta = curr->vruntime - se->vruntime;
3215 if (delta > ideal_runtime)
3216 resched_curr(rq_of(cfs_rq));
3220 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3222 /* 'current' is not kept within the tree. */
3225 * Any task has to be enqueued before it get to execute on
3226 * a CPU. So account for the time it spent waiting on the
3229 update_stats_wait_end(cfs_rq, se);
3230 __dequeue_entity(cfs_rq, se);
3231 update_load_avg(se, 1);
3234 update_stats_curr_start(cfs_rq, se);
3236 #ifdef CONFIG_SCHEDSTATS
3238 * Track our maximum slice length, if the CPU's load is at
3239 * least twice that of our own weight (i.e. dont track it
3240 * when there are only lesser-weight tasks around):
3242 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3243 se->statistics.slice_max = max(se->statistics.slice_max,
3244 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3247 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3251 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3254 * Pick the next process, keeping these things in mind, in this order:
3255 * 1) keep things fair between processes/task groups
3256 * 2) pick the "next" process, since someone really wants that to run
3257 * 3) pick the "last" process, for cache locality
3258 * 4) do not run the "skip" process, if something else is available
3260 static struct sched_entity *
3261 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3263 struct sched_entity *left = __pick_first_entity(cfs_rq);
3264 struct sched_entity *se;
3267 * If curr is set we have to see if its left of the leftmost entity
3268 * still in the tree, provided there was anything in the tree at all.
3270 if (!left || (curr && entity_before(curr, left)))
3273 se = left; /* ideally we run the leftmost entity */
3276 * Avoid running the skip buddy, if running something else can
3277 * be done without getting too unfair.
3279 if (cfs_rq->skip == se) {
3280 struct sched_entity *second;
3283 second = __pick_first_entity(cfs_rq);
3285 second = __pick_next_entity(se);
3286 if (!second || (curr && entity_before(curr, second)))
3290 if (second && wakeup_preempt_entity(second, left) < 1)
3295 * Prefer last buddy, try to return the CPU to a preempted task.
3297 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3301 * Someone really wants this to run. If it's not unfair, run it.
3303 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3306 clear_buddies(cfs_rq, se);
3311 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3313 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3316 * If still on the runqueue then deactivate_task()
3317 * was not called and update_curr() has to be done:
3320 update_curr(cfs_rq);
3322 /* throttle cfs_rqs exceeding runtime */
3323 check_cfs_rq_runtime(cfs_rq);
3325 check_spread(cfs_rq, prev);
3327 update_stats_wait_start(cfs_rq, prev);
3328 /* Put 'current' back into the tree. */
3329 __enqueue_entity(cfs_rq, prev);
3330 /* in !on_rq case, update occurred at dequeue */
3331 update_load_avg(prev, 0);
3333 cfs_rq->curr = NULL;
3337 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3340 * Update run-time statistics of the 'current'.
3342 update_curr(cfs_rq);
3345 * Ensure that runnable average is periodically updated.
3347 update_load_avg(curr, 1);
3348 update_cfs_shares(cfs_rq);
3350 #ifdef CONFIG_SCHED_HRTICK
3352 * queued ticks are scheduled to match the slice, so don't bother
3353 * validating it and just reschedule.
3356 resched_curr(rq_of(cfs_rq));
3360 * don't let the period tick interfere with the hrtick preemption
3362 if (!sched_feat(DOUBLE_TICK) &&
3363 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3367 if (cfs_rq->nr_running > 1)
3368 check_preempt_tick(cfs_rq, curr);
3372 /**************************************************
3373 * CFS bandwidth control machinery
3376 #ifdef CONFIG_CFS_BANDWIDTH
3378 #ifdef HAVE_JUMP_LABEL
3379 static struct static_key __cfs_bandwidth_used;
3381 static inline bool cfs_bandwidth_used(void)
3383 return static_key_false(&__cfs_bandwidth_used);
3386 void cfs_bandwidth_usage_inc(void)
3388 static_key_slow_inc(&__cfs_bandwidth_used);
3391 void cfs_bandwidth_usage_dec(void)
3393 static_key_slow_dec(&__cfs_bandwidth_used);
3395 #else /* HAVE_JUMP_LABEL */
3396 static bool cfs_bandwidth_used(void)
3401 void cfs_bandwidth_usage_inc(void) {}
3402 void cfs_bandwidth_usage_dec(void) {}
3403 #endif /* HAVE_JUMP_LABEL */
3406 * default period for cfs group bandwidth.
3407 * default: 0.1s, units: nanoseconds
3409 static inline u64 default_cfs_period(void)
3411 return 100000000ULL;
3414 static inline u64 sched_cfs_bandwidth_slice(void)
3416 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3420 * Replenish runtime according to assigned quota and update expiration time.
3421 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3422 * additional synchronization around rq->lock.
3424 * requires cfs_b->lock
3426 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3430 if (cfs_b->quota == RUNTIME_INF)
3433 now = sched_clock_cpu(smp_processor_id());
3434 cfs_b->runtime = cfs_b->quota;
3435 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3438 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3440 return &tg->cfs_bandwidth;
3443 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3444 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3446 if (unlikely(cfs_rq->throttle_count))
3447 return cfs_rq->throttled_clock_task;
3449 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3452 /* returns 0 on failure to allocate runtime */
3453 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3455 struct task_group *tg = cfs_rq->tg;
3456 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3457 u64 amount = 0, min_amount, expires;
3459 /* note: this is a positive sum as runtime_remaining <= 0 */
3460 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3462 raw_spin_lock(&cfs_b->lock);
3463 if (cfs_b->quota == RUNTIME_INF)
3464 amount = min_amount;
3466 start_cfs_bandwidth(cfs_b);
3468 if (cfs_b->runtime > 0) {
3469 amount = min(cfs_b->runtime, min_amount);
3470 cfs_b->runtime -= amount;
3474 expires = cfs_b->runtime_expires;
3475 raw_spin_unlock(&cfs_b->lock);
3477 cfs_rq->runtime_remaining += amount;
3479 * we may have advanced our local expiration to account for allowed
3480 * spread between our sched_clock and the one on which runtime was
3483 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3484 cfs_rq->runtime_expires = expires;
3486 return cfs_rq->runtime_remaining > 0;
3490 * Note: This depends on the synchronization provided by sched_clock and the
3491 * fact that rq->clock snapshots this value.
3493 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3495 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3497 /* if the deadline is ahead of our clock, nothing to do */
3498 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3501 if (cfs_rq->runtime_remaining < 0)
3505 * If the local deadline has passed we have to consider the
3506 * possibility that our sched_clock is 'fast' and the global deadline
3507 * has not truly expired.
3509 * Fortunately we can check determine whether this the case by checking
3510 * whether the global deadline has advanced. It is valid to compare
3511 * cfs_b->runtime_expires without any locks since we only care about
3512 * exact equality, so a partial write will still work.
3515 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3516 /* extend local deadline, drift is bounded above by 2 ticks */
3517 cfs_rq->runtime_expires += TICK_NSEC;
3519 /* global deadline is ahead, expiration has passed */
3520 cfs_rq->runtime_remaining = 0;
3524 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3526 /* dock delta_exec before expiring quota (as it could span periods) */
3527 cfs_rq->runtime_remaining -= delta_exec;
3528 expire_cfs_rq_runtime(cfs_rq);
3530 if (likely(cfs_rq->runtime_remaining > 0))
3534 * if we're unable to extend our runtime we resched so that the active
3535 * hierarchy can be throttled
3537 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3538 resched_curr(rq_of(cfs_rq));
3541 static __always_inline
3542 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3544 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3547 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3550 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3552 return cfs_bandwidth_used() && cfs_rq->throttled;
3555 /* check whether cfs_rq, or any parent, is throttled */
3556 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3558 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3562 * Ensure that neither of the group entities corresponding to src_cpu or
3563 * dest_cpu are members of a throttled hierarchy when performing group
3564 * load-balance operations.
3566 static inline int throttled_lb_pair(struct task_group *tg,
3567 int src_cpu, int dest_cpu)
3569 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3571 src_cfs_rq = tg->cfs_rq[src_cpu];
3572 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3574 return throttled_hierarchy(src_cfs_rq) ||
3575 throttled_hierarchy(dest_cfs_rq);
3578 /* updated child weight may affect parent so we have to do this bottom up */
3579 static int tg_unthrottle_up(struct task_group *tg, void *data)
3581 struct rq *rq = data;
3582 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3584 cfs_rq->throttle_count--;
3586 if (!cfs_rq->throttle_count) {
3587 /* adjust cfs_rq_clock_task() */
3588 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3589 cfs_rq->throttled_clock_task;
3596 static int tg_throttle_down(struct task_group *tg, void *data)
3598 struct rq *rq = data;
3599 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3601 /* group is entering throttled state, stop time */
3602 if (!cfs_rq->throttle_count)
3603 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3604 cfs_rq->throttle_count++;
3609 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3611 struct rq *rq = rq_of(cfs_rq);
3612 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3613 struct sched_entity *se;
3614 long task_delta, dequeue = 1;
3617 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3619 /* freeze hierarchy runnable averages while throttled */
3621 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3624 task_delta = cfs_rq->h_nr_running;
3625 for_each_sched_entity(se) {
3626 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3627 /* throttled entity or throttle-on-deactivate */
3632 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3633 qcfs_rq->h_nr_running -= task_delta;
3635 if (qcfs_rq->load.weight)
3640 sub_nr_running(rq, task_delta);
3642 cfs_rq->throttled = 1;
3643 cfs_rq->throttled_clock = rq_clock(rq);
3644 raw_spin_lock(&cfs_b->lock);
3645 empty = list_empty(&cfs_b->throttled_cfs_rq);
3648 * Add to the _head_ of the list, so that an already-started
3649 * distribute_cfs_runtime will not see us. If disribute_cfs_runtime is
3650 * not running add to the tail so that later runqueues don't get starved.
3652 if (cfs_b->distribute_running)
3653 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3655 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3658 * If we're the first throttled task, make sure the bandwidth
3662 start_cfs_bandwidth(cfs_b);
3664 raw_spin_unlock(&cfs_b->lock);
3667 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3669 struct rq *rq = rq_of(cfs_rq);
3670 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3671 struct sched_entity *se;
3675 se = cfs_rq->tg->se[cpu_of(rq)];
3677 cfs_rq->throttled = 0;
3679 update_rq_clock(rq);
3681 raw_spin_lock(&cfs_b->lock);
3682 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3683 list_del_rcu(&cfs_rq->throttled_list);
3684 raw_spin_unlock(&cfs_b->lock);
3686 /* update hierarchical throttle state */
3687 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3689 if (!cfs_rq->load.weight)
3692 task_delta = cfs_rq->h_nr_running;
3693 for_each_sched_entity(se) {
3697 cfs_rq = cfs_rq_of(se);
3699 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3700 cfs_rq->h_nr_running += task_delta;
3702 if (cfs_rq_throttled(cfs_rq))
3707 add_nr_running(rq, task_delta);
3709 /* determine whether we need to wake up potentially idle cpu */
3710 if (rq->curr == rq->idle && rq->cfs.nr_running)
3714 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3715 u64 remaining, u64 expires)
3717 struct cfs_rq *cfs_rq;
3719 u64 starting_runtime = remaining;
3722 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3724 struct rq *rq = rq_of(cfs_rq);
3726 raw_spin_lock(&rq->lock);
3727 if (!cfs_rq_throttled(cfs_rq))
3730 runtime = -cfs_rq->runtime_remaining + 1;
3731 if (runtime > remaining)
3732 runtime = remaining;
3733 remaining -= runtime;
3735 cfs_rq->runtime_remaining += runtime;
3736 cfs_rq->runtime_expires = expires;
3738 /* we check whether we're throttled above */
3739 if (cfs_rq->runtime_remaining > 0)
3740 unthrottle_cfs_rq(cfs_rq);
3743 raw_spin_unlock(&rq->lock);
3750 return starting_runtime - remaining;
3754 * Responsible for refilling a task_group's bandwidth and unthrottling its
3755 * cfs_rqs as appropriate. If there has been no activity within the last
3756 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3757 * used to track this state.
3759 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3761 u64 runtime, runtime_expires;
3764 /* no need to continue the timer with no bandwidth constraint */
3765 if (cfs_b->quota == RUNTIME_INF)
3766 goto out_deactivate;
3768 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3769 cfs_b->nr_periods += overrun;
3772 * idle depends on !throttled (for the case of a large deficit), and if
3773 * we're going inactive then everything else can be deferred
3775 if (cfs_b->idle && !throttled)
3776 goto out_deactivate;
3778 __refill_cfs_bandwidth_runtime(cfs_b);
3781 /* mark as potentially idle for the upcoming period */
3786 /* account preceding periods in which throttling occurred */
3787 cfs_b->nr_throttled += overrun;
3789 runtime_expires = cfs_b->runtime_expires;
3792 * This check is repeated as we are holding onto the new bandwidth while
3793 * we unthrottle. This can potentially race with an unthrottled group
3794 * trying to acquire new bandwidth from the global pool. This can result
3795 * in us over-using our runtime if it is all used during this loop, but
3796 * only by limited amounts in that extreme case.
3798 while (throttled && cfs_b->runtime > 0 && !cfs_b->distribute_running) {
3799 runtime = cfs_b->runtime;
3800 cfs_b->distribute_running = 1;
3801 raw_spin_unlock(&cfs_b->lock);
3802 /* we can't nest cfs_b->lock while distributing bandwidth */
3803 runtime = distribute_cfs_runtime(cfs_b, runtime,
3805 raw_spin_lock(&cfs_b->lock);
3807 cfs_b->distribute_running = 0;
3808 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3810 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3814 * While we are ensured activity in the period following an
3815 * unthrottle, this also covers the case in which the new bandwidth is
3816 * insufficient to cover the existing bandwidth deficit. (Forcing the
3817 * timer to remain active while there are any throttled entities.)
3827 /* a cfs_rq won't donate quota below this amount */
3828 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3829 /* minimum remaining period time to redistribute slack quota */
3830 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3831 /* how long we wait to gather additional slack before distributing */
3832 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3835 * Are we near the end of the current quota period?
3837 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3838 * hrtimer base being cleared by hrtimer_start. In the case of
3839 * migrate_hrtimers, base is never cleared, so we are fine.
3841 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3843 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3846 /* if the call-back is running a quota refresh is already occurring */
3847 if (hrtimer_callback_running(refresh_timer))
3850 /* is a quota refresh about to occur? */
3851 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3852 if (remaining < min_expire)
3858 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3860 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3862 /* if there's a quota refresh soon don't bother with slack */
3863 if (runtime_refresh_within(cfs_b, min_left))
3866 hrtimer_start(&cfs_b->slack_timer,
3867 ns_to_ktime(cfs_bandwidth_slack_period),
3871 /* we know any runtime found here is valid as update_curr() precedes return */
3872 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3874 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3875 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3877 if (slack_runtime <= 0)
3880 raw_spin_lock(&cfs_b->lock);
3881 if (cfs_b->quota != RUNTIME_INF &&
3882 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3883 cfs_b->runtime += slack_runtime;
3885 /* we are under rq->lock, defer unthrottling using a timer */
3886 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3887 !list_empty(&cfs_b->throttled_cfs_rq))
3888 start_cfs_slack_bandwidth(cfs_b);
3890 raw_spin_unlock(&cfs_b->lock);
3892 /* even if it's not valid for return we don't want to try again */
3893 cfs_rq->runtime_remaining -= slack_runtime;
3896 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3898 if (!cfs_bandwidth_used())
3901 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3904 __return_cfs_rq_runtime(cfs_rq);
3908 * This is done with a timer (instead of inline with bandwidth return) since
3909 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3911 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3913 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3916 /* confirm we're still not at a refresh boundary */
3917 raw_spin_lock(&cfs_b->lock);
3918 if (cfs_b->distribute_running) {
3919 raw_spin_unlock(&cfs_b->lock);
3923 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3924 raw_spin_unlock(&cfs_b->lock);
3928 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3929 runtime = cfs_b->runtime;
3931 expires = cfs_b->runtime_expires;
3933 cfs_b->distribute_running = 1;
3935 raw_spin_unlock(&cfs_b->lock);
3940 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3942 raw_spin_lock(&cfs_b->lock);
3943 if (expires == cfs_b->runtime_expires)
3944 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3945 cfs_b->distribute_running = 0;
3946 raw_spin_unlock(&cfs_b->lock);
3950 * When a group wakes up we want to make sure that its quota is not already
3951 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3952 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3954 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3956 if (!cfs_bandwidth_used())
3959 /* Synchronize hierarchical throttle counter: */
3960 if (unlikely(!cfs_rq->throttle_uptodate)) {
3961 struct rq *rq = rq_of(cfs_rq);
3962 struct cfs_rq *pcfs_rq;
3963 struct task_group *tg;
3965 cfs_rq->throttle_uptodate = 1;
3967 /* Get closest up-to-date node, because leaves go first: */
3968 for (tg = cfs_rq->tg->parent; tg; tg = tg->parent) {
3969 pcfs_rq = tg->cfs_rq[cpu_of(rq)];
3970 if (pcfs_rq->throttle_uptodate)
3974 cfs_rq->throttle_count = pcfs_rq->throttle_count;
3975 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3979 /* an active group must be handled by the update_curr()->put() path */
3980 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3983 /* ensure the group is not already throttled */
3984 if (cfs_rq_throttled(cfs_rq))
3987 /* update runtime allocation */
3988 account_cfs_rq_runtime(cfs_rq, 0);
3989 if (cfs_rq->runtime_remaining <= 0)
3990 throttle_cfs_rq(cfs_rq);
3993 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3994 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3996 if (!cfs_bandwidth_used())
3999 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4003 * it's possible for a throttled entity to be forced into a running
4004 * state (e.g. set_curr_task), in this case we're finished.
4006 if (cfs_rq_throttled(cfs_rq))
4009 throttle_cfs_rq(cfs_rq);
4013 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4015 struct cfs_bandwidth *cfs_b =
4016 container_of(timer, struct cfs_bandwidth, slack_timer);
4018 do_sched_cfs_slack_timer(cfs_b);
4020 return HRTIMER_NORESTART;
4023 extern const u64 max_cfs_quota_period;
4025 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4027 struct cfs_bandwidth *cfs_b =
4028 container_of(timer, struct cfs_bandwidth, period_timer);
4033 raw_spin_lock(&cfs_b->lock);
4035 overrun = hrtimer_forward_now(timer, cfs_b->period);
4040 u64 new, old = ktime_to_ns(cfs_b->period);
4042 new = (old * 147) / 128; /* ~115% */
4043 new = min(new, max_cfs_quota_period);
4045 cfs_b->period = ns_to_ktime(new);
4047 /* since max is 1s, this is limited to 1e9^2, which fits in u64 */
4048 cfs_b->quota *= new;
4049 cfs_b->quota = div64_u64(cfs_b->quota, old);
4051 pr_warn_ratelimited(
4052 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us %lld, cfs_quota_us = %lld)\n",
4054 div_u64(new, NSEC_PER_USEC),
4055 div_u64(cfs_b->quota, NSEC_PER_USEC));
4057 /* reset count so we don't come right back in here */
4061 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4064 cfs_b->period_active = 0;
4065 raw_spin_unlock(&cfs_b->lock);
4067 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4070 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4072 raw_spin_lock_init(&cfs_b->lock);
4074 cfs_b->quota = RUNTIME_INF;
4075 cfs_b->period = ns_to_ktime(default_cfs_period());
4077 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4078 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4079 cfs_b->period_timer.function = sched_cfs_period_timer;
4080 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4081 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4082 cfs_b->distribute_running = 0;
4085 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4087 cfs_rq->runtime_enabled = 0;
4088 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4091 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4093 lockdep_assert_held(&cfs_b->lock);
4095 if (!cfs_b->period_active) {
4096 cfs_b->period_active = 1;
4097 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4098 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4102 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4104 /* init_cfs_bandwidth() was not called */
4105 if (!cfs_b->throttled_cfs_rq.next)
4108 hrtimer_cancel(&cfs_b->period_timer);
4109 hrtimer_cancel(&cfs_b->slack_timer);
4112 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4114 struct cfs_rq *cfs_rq;
4116 for_each_leaf_cfs_rq(rq, cfs_rq) {
4117 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4119 raw_spin_lock(&cfs_b->lock);
4120 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4121 raw_spin_unlock(&cfs_b->lock);
4125 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4127 struct cfs_rq *cfs_rq;
4129 for_each_leaf_cfs_rq(rq, cfs_rq) {
4130 if (!cfs_rq->runtime_enabled)
4134 * clock_task is not advancing so we just need to make sure
4135 * there's some valid quota amount
4137 cfs_rq->runtime_remaining = 1;
4139 * Offline rq is schedulable till cpu is completely disabled
4140 * in take_cpu_down(), so we prevent new cfs throttling here.
4142 cfs_rq->runtime_enabled = 0;
4144 if (cfs_rq_throttled(cfs_rq))
4145 unthrottle_cfs_rq(cfs_rq);
4149 #else /* CONFIG_CFS_BANDWIDTH */
4150 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4152 return rq_clock_task(rq_of(cfs_rq));
4155 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4156 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4157 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4158 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4160 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4165 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4170 static inline int throttled_lb_pair(struct task_group *tg,
4171 int src_cpu, int dest_cpu)
4176 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4178 #ifdef CONFIG_FAIR_GROUP_SCHED
4179 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4182 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4186 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4187 static inline void update_runtime_enabled(struct rq *rq) {}
4188 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4190 #endif /* CONFIG_CFS_BANDWIDTH */
4192 /**************************************************
4193 * CFS operations on tasks:
4196 #ifdef CONFIG_SCHED_HRTICK
4197 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4199 struct sched_entity *se = &p->se;
4200 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4202 WARN_ON(task_rq(p) != rq);
4204 if (cfs_rq->nr_running > 1) {
4205 u64 slice = sched_slice(cfs_rq, se);
4206 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4207 s64 delta = slice - ran;
4214 hrtick_start(rq, delta);
4219 * called from enqueue/dequeue and updates the hrtick when the
4220 * current task is from our class and nr_running is low enough
4223 static void hrtick_update(struct rq *rq)
4225 struct task_struct *curr = rq->curr;
4227 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4230 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4231 hrtick_start_fair(rq, curr);
4233 #else /* !CONFIG_SCHED_HRTICK */
4235 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4239 static inline void hrtick_update(struct rq *rq)
4245 * The enqueue_task method is called before nr_running is
4246 * increased. Here we update the fair scheduling stats and
4247 * then put the task into the rbtree:
4250 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4252 struct cfs_rq *cfs_rq;
4253 struct sched_entity *se = &p->se;
4255 for_each_sched_entity(se) {
4258 cfs_rq = cfs_rq_of(se);
4259 enqueue_entity(cfs_rq, se, flags);
4262 * end evaluation on encountering a throttled cfs_rq
4264 * note: in the case of encountering a throttled cfs_rq we will
4265 * post the final h_nr_running increment below.
4267 if (cfs_rq_throttled(cfs_rq))
4269 cfs_rq->h_nr_running++;
4271 flags = ENQUEUE_WAKEUP;
4274 for_each_sched_entity(se) {
4275 cfs_rq = cfs_rq_of(se);
4276 cfs_rq->h_nr_running++;
4278 if (cfs_rq_throttled(cfs_rq))
4281 update_load_avg(se, 1);
4282 update_cfs_shares(cfs_rq);
4286 add_nr_running(rq, 1);
4291 static void set_next_buddy(struct sched_entity *se);
4294 * The dequeue_task method is called before nr_running is
4295 * decreased. We remove the task from the rbtree and
4296 * update the fair scheduling stats:
4298 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4300 struct cfs_rq *cfs_rq;
4301 struct sched_entity *se = &p->se;
4302 int task_sleep = flags & DEQUEUE_SLEEP;
4304 for_each_sched_entity(se) {
4305 cfs_rq = cfs_rq_of(se);
4306 dequeue_entity(cfs_rq, se, flags);
4309 * end evaluation on encountering a throttled cfs_rq
4311 * note: in the case of encountering a throttled cfs_rq we will
4312 * post the final h_nr_running decrement below.
4314 if (cfs_rq_throttled(cfs_rq))
4316 cfs_rq->h_nr_running--;
4318 /* Don't dequeue parent if it has other entities besides us */
4319 if (cfs_rq->load.weight) {
4320 /* Avoid re-evaluating load for this entity: */
4321 se = parent_entity(se);
4323 * Bias pick_next to pick a task from this cfs_rq, as
4324 * p is sleeping when it is within its sched_slice.
4326 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4330 flags |= DEQUEUE_SLEEP;
4333 for_each_sched_entity(se) {
4334 cfs_rq = cfs_rq_of(se);
4335 cfs_rq->h_nr_running--;
4337 if (cfs_rq_throttled(cfs_rq))
4340 update_load_avg(se, 1);
4341 update_cfs_shares(cfs_rq);
4345 sub_nr_running(rq, 1);
4353 * per rq 'load' arrray crap; XXX kill this.
4357 * The exact cpuload at various idx values, calculated at every tick would be
4358 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4360 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4361 * on nth tick when cpu may be busy, then we have:
4362 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4363 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4365 * decay_load_missed() below does efficient calculation of
4366 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4367 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4369 * The calculation is approximated on a 128 point scale.
4370 * degrade_zero_ticks is the number of ticks after which load at any
4371 * particular idx is approximated to be zero.
4372 * degrade_factor is a precomputed table, a row for each load idx.
4373 * Each column corresponds to degradation factor for a power of two ticks,
4374 * based on 128 point scale.
4376 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4377 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4379 * With this power of 2 load factors, we can degrade the load n times
4380 * by looking at 1 bits in n and doing as many mult/shift instead of
4381 * n mult/shifts needed by the exact degradation.
4383 #define DEGRADE_SHIFT 7
4384 static const unsigned char
4385 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4386 static const unsigned char
4387 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4388 {0, 0, 0, 0, 0, 0, 0, 0},
4389 {64, 32, 8, 0, 0, 0, 0, 0},
4390 {96, 72, 40, 12, 1, 0, 0},
4391 {112, 98, 75, 43, 15, 1, 0},
4392 {120, 112, 98, 76, 45, 16, 2} };
4395 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4396 * would be when CPU is idle and so we just decay the old load without
4397 * adding any new load.
4399 static unsigned long
4400 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4404 if (!missed_updates)
4407 if (missed_updates >= degrade_zero_ticks[idx])
4411 return load >> missed_updates;
4413 while (missed_updates) {
4414 if (missed_updates % 2)
4415 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4417 missed_updates >>= 1;
4424 * Update rq->cpu_load[] statistics. This function is usually called every
4425 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4426 * every tick. We fix it up based on jiffies.
4428 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4429 unsigned long pending_updates)
4433 this_rq->nr_load_updates++;
4435 /* Update our load: */
4436 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4437 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4438 unsigned long old_load, new_load;
4440 /* scale is effectively 1 << i now, and >> i divides by scale */
4442 old_load = this_rq->cpu_load[i];
4443 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4444 new_load = this_load;
4446 * Round up the averaging division if load is increasing. This
4447 * prevents us from getting stuck on 9 if the load is 10, for
4450 if (new_load > old_load)
4451 new_load += scale - 1;
4453 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4456 sched_avg_update(this_rq);
4459 /* Used instead of source_load when we know the type == 0 */
4460 static unsigned long weighted_cpuload(const int cpu)
4462 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4465 #ifdef CONFIG_NO_HZ_COMMON
4467 * There is no sane way to deal with nohz on smp when using jiffies because the
4468 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4469 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4471 * Therefore we cannot use the delta approach from the regular tick since that
4472 * would seriously skew the load calculation. However we'll make do for those
4473 * updates happening while idle (nohz_idle_balance) or coming out of idle
4474 * (tick_nohz_idle_exit).
4476 * This means we might still be one tick off for nohz periods.
4480 * Called from nohz_idle_balance() to update the load ratings before doing the
4483 static void update_idle_cpu_load(struct rq *this_rq)
4485 unsigned long curr_jiffies = READ_ONCE(jiffies);
4486 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4487 unsigned long pending_updates;
4490 * bail if there's load or we're actually up-to-date.
4492 if (load || curr_jiffies == this_rq->last_load_update_tick)
4495 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4496 this_rq->last_load_update_tick = curr_jiffies;
4498 __update_cpu_load(this_rq, load, pending_updates);
4502 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4504 void update_cpu_load_nohz(void)
4506 struct rq *this_rq = this_rq();
4507 unsigned long curr_jiffies = READ_ONCE(jiffies);
4508 unsigned long pending_updates;
4510 if (curr_jiffies == this_rq->last_load_update_tick)
4513 raw_spin_lock(&this_rq->lock);
4514 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4515 if (pending_updates) {
4516 this_rq->last_load_update_tick = curr_jiffies;
4518 * We were idle, this means load 0, the current load might be
4519 * !0 due to remote wakeups and the sort.
4521 __update_cpu_load(this_rq, 0, pending_updates);
4523 raw_spin_unlock(&this_rq->lock);
4525 #endif /* CONFIG_NO_HZ */
4528 * Called from scheduler_tick()
4530 void update_cpu_load_active(struct rq *this_rq)
4532 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4534 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4536 this_rq->last_load_update_tick = jiffies;
4537 __update_cpu_load(this_rq, load, 1);
4541 * Return a low guess at the load of a migration-source cpu weighted
4542 * according to the scheduling class and "nice" value.
4544 * We want to under-estimate the load of migration sources, to
4545 * balance conservatively.
4547 static unsigned long source_load(int cpu, int type)
4549 struct rq *rq = cpu_rq(cpu);
4550 unsigned long total = weighted_cpuload(cpu);
4552 if (type == 0 || !sched_feat(LB_BIAS))
4555 return min(rq->cpu_load[type-1], total);
4559 * Return a high guess at the load of a migration-target cpu weighted
4560 * according to the scheduling class and "nice" value.
4562 static unsigned long target_load(int cpu, int type)
4564 struct rq *rq = cpu_rq(cpu);
4565 unsigned long total = weighted_cpuload(cpu);
4567 if (type == 0 || !sched_feat(LB_BIAS))
4570 return max(rq->cpu_load[type-1], total);
4573 static unsigned long capacity_of(int cpu)
4575 return cpu_rq(cpu)->cpu_capacity;
4578 static unsigned long capacity_orig_of(int cpu)
4580 return cpu_rq(cpu)->cpu_capacity_orig;
4583 static unsigned long cpu_avg_load_per_task(int cpu)
4585 struct rq *rq = cpu_rq(cpu);
4586 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4587 unsigned long load_avg = weighted_cpuload(cpu);
4590 return load_avg / nr_running;
4595 static void record_wakee(struct task_struct *p)
4598 * Rough decay (wiping) for cost saving, don't worry
4599 * about the boundary, really active task won't care
4602 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4603 current->wakee_flips >>= 1;
4604 current->wakee_flip_decay_ts = jiffies;
4607 if (current->last_wakee != p) {
4608 current->last_wakee = p;
4609 current->wakee_flips++;
4613 static void task_waking_fair(struct task_struct *p)
4615 struct sched_entity *se = &p->se;
4616 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4619 #ifndef CONFIG_64BIT
4620 u64 min_vruntime_copy;
4623 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4625 min_vruntime = cfs_rq->min_vruntime;
4626 } while (min_vruntime != min_vruntime_copy);
4628 min_vruntime = cfs_rq->min_vruntime;
4631 se->vruntime -= min_vruntime;
4635 #ifdef CONFIG_FAIR_GROUP_SCHED
4637 * effective_load() calculates the load change as seen from the root_task_group
4639 * Adding load to a group doesn't make a group heavier, but can cause movement
4640 * of group shares between cpus. Assuming the shares were perfectly aligned one
4641 * can calculate the shift in shares.
4643 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4644 * on this @cpu and results in a total addition (subtraction) of @wg to the
4645 * total group weight.
4647 * Given a runqueue weight distribution (rw_i) we can compute a shares
4648 * distribution (s_i) using:
4650 * s_i = rw_i / \Sum rw_j (1)
4652 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4653 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4654 * shares distribution (s_i):
4656 * rw_i = { 2, 4, 1, 0 }
4657 * s_i = { 2/7, 4/7, 1/7, 0 }
4659 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4660 * task used to run on and the CPU the waker is running on), we need to
4661 * compute the effect of waking a task on either CPU and, in case of a sync
4662 * wakeup, compute the effect of the current task going to sleep.
4664 * So for a change of @wl to the local @cpu with an overall group weight change
4665 * of @wl we can compute the new shares distribution (s'_i) using:
4667 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4669 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4670 * differences in waking a task to CPU 0. The additional task changes the
4671 * weight and shares distributions like:
4673 * rw'_i = { 3, 4, 1, 0 }
4674 * s'_i = { 3/8, 4/8, 1/8, 0 }
4676 * We can then compute the difference in effective weight by using:
4678 * dw_i = S * (s'_i - s_i) (3)
4680 * Where 'S' is the group weight as seen by its parent.
4682 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4683 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4684 * 4/7) times the weight of the group.
4686 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4688 struct sched_entity *se = tg->se[cpu];
4690 if (!tg->parent) /* the trivial, non-cgroup case */
4693 for_each_sched_entity(se) {
4694 struct cfs_rq *cfs_rq = se->my_q;
4695 long W, w = cfs_rq_load_avg(cfs_rq);
4700 * W = @wg + \Sum rw_j
4702 W = wg + atomic_long_read(&tg->load_avg);
4704 /* Ensure \Sum rw_j >= rw_i */
4705 W -= cfs_rq->tg_load_avg_contrib;
4714 * wl = S * s'_i; see (2)
4717 wl = (w * (long)tg->shares) / W;
4722 * Per the above, wl is the new se->load.weight value; since
4723 * those are clipped to [MIN_SHARES, ...) do so now. See
4724 * calc_cfs_shares().
4726 if (wl < MIN_SHARES)
4730 * wl = dw_i = S * (s'_i - s_i); see (3)
4732 wl -= se->avg.load_avg;
4735 * Recursively apply this logic to all parent groups to compute
4736 * the final effective load change on the root group. Since
4737 * only the @tg group gets extra weight, all parent groups can
4738 * only redistribute existing shares. @wl is the shift in shares
4739 * resulting from this level per the above.
4748 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4756 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
4757 * A waker of many should wake a different task than the one last awakened
4758 * at a frequency roughly N times higher than one of its wakees. In order
4759 * to determine whether we should let the load spread vs consolodating to
4760 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
4761 * partner, and a factor of lls_size higher frequency in the other. With
4762 * both conditions met, we can be relatively sure that the relationship is
4763 * non-monogamous, with partner count exceeding socket size. Waker/wakee
4764 * being client/server, worker/dispatcher, interrupt source or whatever is
4765 * irrelevant, spread criteria is apparent partner count exceeds socket size.
4767 static int wake_wide(struct task_struct *p)
4769 unsigned int master = current->wakee_flips;
4770 unsigned int slave = p->wakee_flips;
4771 int factor = this_cpu_read(sd_llc_size);
4774 swap(master, slave);
4775 if (slave < factor || master < slave * factor)
4780 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4782 s64 this_load, load;
4783 s64 this_eff_load, prev_eff_load;
4784 int idx, this_cpu, prev_cpu;
4785 struct task_group *tg;
4786 unsigned long weight;
4790 this_cpu = smp_processor_id();
4791 prev_cpu = task_cpu(p);
4792 load = source_load(prev_cpu, idx);
4793 this_load = target_load(this_cpu, idx);
4796 * If sync wakeup then subtract the (maximum possible)
4797 * effect of the currently running task from the load
4798 * of the current CPU:
4801 tg = task_group(current);
4802 weight = current->se.avg.load_avg;
4804 this_load += effective_load(tg, this_cpu, -weight, -weight);
4805 load += effective_load(tg, prev_cpu, 0, -weight);
4809 weight = p->se.avg.load_avg;
4812 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4813 * due to the sync cause above having dropped this_load to 0, we'll
4814 * always have an imbalance, but there's really nothing you can do
4815 * about that, so that's good too.
4817 * Otherwise check if either cpus are near enough in load to allow this
4818 * task to be woken on this_cpu.
4820 this_eff_load = 100;
4821 this_eff_load *= capacity_of(prev_cpu);
4823 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4824 prev_eff_load *= capacity_of(this_cpu);
4826 if (this_load > 0) {
4827 this_eff_load *= this_load +
4828 effective_load(tg, this_cpu, weight, weight);
4830 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4833 balanced = this_eff_load <= prev_eff_load;
4835 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4840 schedstat_inc(sd, ttwu_move_affine);
4841 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4847 * find_idlest_group finds and returns the least busy CPU group within the
4850 static struct sched_group *
4851 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4852 int this_cpu, int sd_flag)
4854 struct sched_group *idlest = NULL, *group = sd->groups;
4855 unsigned long min_load = ULONG_MAX, this_load = 0;
4856 int load_idx = sd->forkexec_idx;
4857 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4859 if (sd_flag & SD_BALANCE_WAKE)
4860 load_idx = sd->wake_idx;
4863 unsigned long load, avg_load;
4867 /* Skip over this group if it has no CPUs allowed */
4868 if (!cpumask_intersects(sched_group_cpus(group),
4869 tsk_cpus_allowed(p)))
4872 local_group = cpumask_test_cpu(this_cpu,
4873 sched_group_cpus(group));
4875 /* Tally up the load of all CPUs in the group */
4878 for_each_cpu(i, sched_group_cpus(group)) {
4879 /* Bias balancing toward cpus of our domain */
4881 load = source_load(i, load_idx);
4883 load = target_load(i, load_idx);
4888 /* Adjust by relative CPU capacity of the group */
4889 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
4892 this_load = avg_load;
4893 } else if (avg_load < min_load) {
4894 min_load = avg_load;
4897 } while (group = group->next, group != sd->groups);
4899 if (!idlest || 100*this_load < imbalance*min_load)
4905 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4908 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4910 unsigned long load, min_load = ULONG_MAX;
4911 unsigned int min_exit_latency = UINT_MAX;
4912 u64 latest_idle_timestamp = 0;
4913 int least_loaded_cpu = this_cpu;
4914 int shallowest_idle_cpu = -1;
4917 /* Traverse only the allowed CPUs */
4918 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4920 struct rq *rq = cpu_rq(i);
4921 struct cpuidle_state *idle = idle_get_state(rq);
4922 if (idle && idle->exit_latency < min_exit_latency) {
4924 * We give priority to a CPU whose idle state
4925 * has the smallest exit latency irrespective
4926 * of any idle timestamp.
4928 min_exit_latency = idle->exit_latency;
4929 latest_idle_timestamp = rq->idle_stamp;
4930 shallowest_idle_cpu = i;
4931 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
4932 rq->idle_stamp > latest_idle_timestamp) {
4934 * If equal or no active idle state, then
4935 * the most recently idled CPU might have
4938 latest_idle_timestamp = rq->idle_stamp;
4939 shallowest_idle_cpu = i;
4941 } else if (shallowest_idle_cpu == -1) {
4942 load = weighted_cpuload(i);
4943 if (load < min_load || (load == min_load && i == this_cpu)) {
4945 least_loaded_cpu = i;
4950 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
4954 * Try and locate an idle CPU in the sched_domain.
4956 static int select_idle_sibling(struct task_struct *p, int target)
4958 struct sched_domain *sd;
4959 struct sched_group *sg;
4960 int i = task_cpu(p);
4962 if (idle_cpu(target))
4966 * If the prevous cpu is cache affine and idle, don't be stupid.
4968 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4972 * Otherwise, iterate the domains and find an elegible idle cpu.
4974 sd = rcu_dereference(per_cpu(sd_llc, target));
4975 for_each_lower_domain(sd) {
4978 if (!cpumask_intersects(sched_group_cpus(sg),
4979 tsk_cpus_allowed(p)))
4982 for_each_cpu(i, sched_group_cpus(sg)) {
4983 if (i == target || !idle_cpu(i))
4987 target = cpumask_first_and(sched_group_cpus(sg),
4988 tsk_cpus_allowed(p));
4992 } while (sg != sd->groups);
4999 * cpu_util returns the amount of capacity of a CPU that is used by CFS
5000 * tasks. The unit of the return value must be the one of capacity so we can
5001 * compare the utilization with the capacity of the CPU that is available for
5002 * CFS task (ie cpu_capacity).
5004 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5005 * recent utilization of currently non-runnable tasks on a CPU. It represents
5006 * the amount of utilization of a CPU in the range [0..capacity_orig] where
5007 * capacity_orig is the cpu_capacity available at the highest frequency
5008 * (arch_scale_freq_capacity()).
5009 * The utilization of a CPU converges towards a sum equal to or less than the
5010 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5011 * the running time on this CPU scaled by capacity_curr.
5013 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5014 * higher than capacity_orig because of unfortunate rounding in
5015 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5016 * the average stabilizes with the new running time. We need to check that the
5017 * utilization stays within the range of [0..capacity_orig] and cap it if
5018 * necessary. Without utilization capping, a group could be seen as overloaded
5019 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5020 * available capacity. We allow utilization to overshoot capacity_curr (but not
5021 * capacity_orig) as it useful for predicting the capacity required after task
5022 * migrations (scheduler-driven DVFS).
5024 static int cpu_util(int cpu)
5026 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
5027 unsigned long capacity = capacity_orig_of(cpu);
5029 return (util >= capacity) ? capacity : util;
5033 * select_task_rq_fair: Select target runqueue for the waking task in domains
5034 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5035 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5037 * Balances load by selecting the idlest cpu in the idlest group, or under
5038 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5040 * Returns the target cpu number.
5042 * preempt must be disabled.
5045 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5047 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5048 int cpu = smp_processor_id();
5049 int new_cpu = prev_cpu;
5050 int want_affine = 0;
5051 int sync = wake_flags & WF_SYNC;
5053 if (sd_flag & SD_BALANCE_WAKE)
5054 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
5057 for_each_domain(cpu, tmp) {
5058 if (!(tmp->flags & SD_LOAD_BALANCE))
5062 * If both cpu and prev_cpu are part of this domain,
5063 * cpu is a valid SD_WAKE_AFFINE target.
5065 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5066 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5071 if (tmp->flags & sd_flag)
5073 else if (!want_affine)
5078 sd = NULL; /* Prefer wake_affine over balance flags */
5079 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5084 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5085 new_cpu = select_idle_sibling(p, new_cpu);
5088 struct sched_group *group;
5091 if (!(sd->flags & sd_flag)) {
5096 group = find_idlest_group(sd, p, cpu, sd_flag);
5102 new_cpu = find_idlest_cpu(group, p, cpu);
5103 if (new_cpu == -1 || new_cpu == cpu) {
5104 /* Now try balancing at a lower domain level of cpu */
5109 /* Now try balancing at a lower domain level of new_cpu */
5111 weight = sd->span_weight;
5113 for_each_domain(cpu, tmp) {
5114 if (weight <= tmp->span_weight)
5116 if (tmp->flags & sd_flag)
5119 /* while loop will break here if sd == NULL */
5127 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5128 * cfs_rq_of(p) references at time of call are still valid and identify the
5129 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
5130 * other assumptions, including the state of rq->lock, should be made.
5132 static void migrate_task_rq_fair(struct task_struct *p)
5135 * We are supposed to update the task to "current" time, then its up to date
5136 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5137 * what current time is, so simply throw away the out-of-date time. This
5138 * will result in the wakee task is less decayed, but giving the wakee more
5139 * load sounds not bad.
5141 remove_entity_load_avg(&p->se);
5143 /* Tell new CPU we are migrated */
5144 p->se.avg.last_update_time = 0;
5146 /* We have migrated, no longer consider this task hot */
5147 p->se.exec_start = 0;
5150 static void task_dead_fair(struct task_struct *p)
5152 remove_entity_load_avg(&p->se);
5154 #endif /* CONFIG_SMP */
5156 static unsigned long
5157 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5159 unsigned long gran = sysctl_sched_wakeup_granularity;
5162 * Since its curr running now, convert the gran from real-time
5163 * to virtual-time in his units.
5165 * By using 'se' instead of 'curr' we penalize light tasks, so
5166 * they get preempted easier. That is, if 'se' < 'curr' then
5167 * the resulting gran will be larger, therefore penalizing the
5168 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5169 * be smaller, again penalizing the lighter task.
5171 * This is especially important for buddies when the leftmost
5172 * task is higher priority than the buddy.
5174 return calc_delta_fair(gran, se);
5178 * Should 'se' preempt 'curr'.
5192 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5194 s64 gran, vdiff = curr->vruntime - se->vruntime;
5199 gran = wakeup_gran(curr, se);
5206 static void set_last_buddy(struct sched_entity *se)
5208 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5211 for_each_sched_entity(se)
5212 cfs_rq_of(se)->last = se;
5215 static void set_next_buddy(struct sched_entity *se)
5217 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5220 for_each_sched_entity(se)
5221 cfs_rq_of(se)->next = se;
5224 static void set_skip_buddy(struct sched_entity *se)
5226 for_each_sched_entity(se)
5227 cfs_rq_of(se)->skip = se;
5231 * Preempt the current task with a newly woken task if needed:
5233 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5235 struct task_struct *curr = rq->curr;
5236 struct sched_entity *se = &curr->se, *pse = &p->se;
5237 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5238 int scale = cfs_rq->nr_running >= sched_nr_latency;
5239 int next_buddy_marked = 0;
5241 if (unlikely(se == pse))
5245 * This is possible from callers such as attach_tasks(), in which we
5246 * unconditionally check_prempt_curr() after an enqueue (which may have
5247 * lead to a throttle). This both saves work and prevents false
5248 * next-buddy nomination below.
5250 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5253 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5254 set_next_buddy(pse);
5255 next_buddy_marked = 1;
5259 * We can come here with TIF_NEED_RESCHED already set from new task
5262 * Note: this also catches the edge-case of curr being in a throttled
5263 * group (e.g. via set_curr_task), since update_curr() (in the
5264 * enqueue of curr) will have resulted in resched being set. This
5265 * prevents us from potentially nominating it as a false LAST_BUDDY
5268 if (test_tsk_need_resched(curr))
5271 /* Idle tasks are by definition preempted by non-idle tasks. */
5272 if (unlikely(curr->policy == SCHED_IDLE) &&
5273 likely(p->policy != SCHED_IDLE))
5277 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5278 * is driven by the tick):
5280 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5283 find_matching_se(&se, &pse);
5284 update_curr(cfs_rq_of(se));
5286 if (wakeup_preempt_entity(se, pse) == 1) {
5288 * Bias pick_next to pick the sched entity that is
5289 * triggering this preemption.
5291 if (!next_buddy_marked)
5292 set_next_buddy(pse);
5301 * Only set the backward buddy when the current task is still
5302 * on the rq. This can happen when a wakeup gets interleaved
5303 * with schedule on the ->pre_schedule() or idle_balance()
5304 * point, either of which can * drop the rq lock.
5306 * Also, during early boot the idle thread is in the fair class,
5307 * for obvious reasons its a bad idea to schedule back to it.
5309 if (unlikely(!se->on_rq || curr == rq->idle))
5312 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5316 static struct task_struct *
5317 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
5319 struct cfs_rq *cfs_rq = &rq->cfs;
5320 struct sched_entity *se;
5321 struct task_struct *p;
5325 #ifdef CONFIG_FAIR_GROUP_SCHED
5326 if (!cfs_rq->nr_running)
5329 if (prev->sched_class != &fair_sched_class)
5333 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5334 * likely that a next task is from the same cgroup as the current.
5336 * Therefore attempt to avoid putting and setting the entire cgroup
5337 * hierarchy, only change the part that actually changes.
5341 struct sched_entity *curr = cfs_rq->curr;
5344 * Since we got here without doing put_prev_entity() we also
5345 * have to consider cfs_rq->curr. If it is still a runnable
5346 * entity, update_curr() will update its vruntime, otherwise
5347 * forget we've ever seen it.
5351 update_curr(cfs_rq);
5356 * This call to check_cfs_rq_runtime() will do the
5357 * throttle and dequeue its entity in the parent(s).
5358 * Therefore the 'simple' nr_running test will indeed
5361 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5365 se = pick_next_entity(cfs_rq, curr);
5366 cfs_rq = group_cfs_rq(se);
5372 * Since we haven't yet done put_prev_entity and if the selected task
5373 * is a different task than we started out with, try and touch the
5374 * least amount of cfs_rqs.
5377 struct sched_entity *pse = &prev->se;
5379 while (!(cfs_rq = is_same_group(se, pse))) {
5380 int se_depth = se->depth;
5381 int pse_depth = pse->depth;
5383 if (se_depth <= pse_depth) {
5384 put_prev_entity(cfs_rq_of(pse), pse);
5385 pse = parent_entity(pse);
5387 if (se_depth >= pse_depth) {
5388 set_next_entity(cfs_rq_of(se), se);
5389 se = parent_entity(se);
5393 put_prev_entity(cfs_rq, pse);
5394 set_next_entity(cfs_rq, se);
5397 if (hrtick_enabled(rq))
5398 hrtick_start_fair(rq, p);
5405 if (!cfs_rq->nr_running)
5408 put_prev_task(rq, prev);
5411 se = pick_next_entity(cfs_rq, NULL);
5412 set_next_entity(cfs_rq, se);
5413 cfs_rq = group_cfs_rq(se);
5418 if (hrtick_enabled(rq))
5419 hrtick_start_fair(rq, p);
5425 * This is OK, because current is on_cpu, which avoids it being picked
5426 * for load-balance and preemption/IRQs are still disabled avoiding
5427 * further scheduler activity on it and we're being very careful to
5428 * re-start the picking loop.
5430 lockdep_unpin_lock(&rq->lock);
5431 new_tasks = idle_balance(rq);
5432 lockdep_pin_lock(&rq->lock);
5434 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5435 * possible for any higher priority task to appear. In that case we
5436 * must re-start the pick_next_entity() loop.
5448 * Account for a descheduled task:
5450 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5452 struct sched_entity *se = &prev->se;
5453 struct cfs_rq *cfs_rq;
5455 for_each_sched_entity(se) {
5456 cfs_rq = cfs_rq_of(se);
5457 put_prev_entity(cfs_rq, se);
5462 * sched_yield() is very simple
5464 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5466 static void yield_task_fair(struct rq *rq)
5468 struct task_struct *curr = rq->curr;
5469 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5470 struct sched_entity *se = &curr->se;
5473 * Are we the only task in the tree?
5475 if (unlikely(rq->nr_running == 1))
5478 clear_buddies(cfs_rq, se);
5480 if (curr->policy != SCHED_BATCH) {
5481 update_rq_clock(rq);
5483 * Update run-time statistics of the 'current'.
5485 update_curr(cfs_rq);
5487 * Tell update_rq_clock() that we've just updated,
5488 * so we don't do microscopic update in schedule()
5489 * and double the fastpath cost.
5491 rq_clock_skip_update(rq, true);
5497 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5499 struct sched_entity *se = &p->se;
5501 /* throttled hierarchies are not runnable */
5502 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5505 /* Tell the scheduler that we'd really like pse to run next. */
5508 yield_task_fair(rq);
5514 /**************************************************
5515 * Fair scheduling class load-balancing methods.
5519 * The purpose of load-balancing is to achieve the same basic fairness the
5520 * per-cpu scheduler provides, namely provide a proportional amount of compute
5521 * time to each task. This is expressed in the following equation:
5523 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5525 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5526 * W_i,0 is defined as:
5528 * W_i,0 = \Sum_j w_i,j (2)
5530 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5531 * is derived from the nice value as per prio_to_weight[].
5533 * The weight average is an exponential decay average of the instantaneous
5536 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5538 * C_i is the compute capacity of cpu i, typically it is the
5539 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5540 * can also include other factors [XXX].
5542 * To achieve this balance we define a measure of imbalance which follows
5543 * directly from (1):
5545 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5547 * We them move tasks around to minimize the imbalance. In the continuous
5548 * function space it is obvious this converges, in the discrete case we get
5549 * a few fun cases generally called infeasible weight scenarios.
5552 * - infeasible weights;
5553 * - local vs global optima in the discrete case. ]
5558 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5559 * for all i,j solution, we create a tree of cpus that follows the hardware
5560 * topology where each level pairs two lower groups (or better). This results
5561 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5562 * tree to only the first of the previous level and we decrease the frequency
5563 * of load-balance at each level inv. proportional to the number of cpus in
5569 * \Sum { --- * --- * 2^i } = O(n) (5)
5571 * `- size of each group
5572 * | | `- number of cpus doing load-balance
5574 * `- sum over all levels
5576 * Coupled with a limit on how many tasks we can migrate every balance pass,
5577 * this makes (5) the runtime complexity of the balancer.
5579 * An important property here is that each CPU is still (indirectly) connected
5580 * to every other cpu in at most O(log n) steps:
5582 * The adjacency matrix of the resulting graph is given by:
5585 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5588 * And you'll find that:
5590 * A^(log_2 n)_i,j != 0 for all i,j (7)
5592 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5593 * The task movement gives a factor of O(m), giving a convergence complexity
5596 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5601 * In order to avoid CPUs going idle while there's still work to do, new idle
5602 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5603 * tree itself instead of relying on other CPUs to bring it work.
5605 * This adds some complexity to both (5) and (8) but it reduces the total idle
5613 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5616 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5621 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5623 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5625 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5628 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5629 * rewrite all of this once again.]
5632 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5634 enum fbq_type { regular, remote, all };
5636 #define LBF_ALL_PINNED 0x01
5637 #define LBF_NEED_BREAK 0x02
5638 #define LBF_DST_PINNED 0x04
5639 #define LBF_SOME_PINNED 0x08
5642 struct sched_domain *sd;
5650 struct cpumask *dst_grpmask;
5652 enum cpu_idle_type idle;
5654 /* The set of CPUs under consideration for load-balancing */
5655 struct cpumask *cpus;
5660 unsigned int loop_break;
5661 unsigned int loop_max;
5663 enum fbq_type fbq_type;
5664 struct list_head tasks;
5668 * Is this task likely cache-hot:
5670 static int task_hot(struct task_struct *p, struct lb_env *env)
5674 lockdep_assert_held(&env->src_rq->lock);
5676 if (p->sched_class != &fair_sched_class)
5679 if (unlikely(p->policy == SCHED_IDLE))
5683 * Buddy candidates are cache hot:
5685 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
5686 (&p->se == cfs_rq_of(&p->se)->next ||
5687 &p->se == cfs_rq_of(&p->se)->last))
5690 if (sysctl_sched_migration_cost == -1)
5692 if (sysctl_sched_migration_cost == 0)
5695 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
5697 return delta < (s64)sysctl_sched_migration_cost;
5700 #ifdef CONFIG_NUMA_BALANCING
5702 * Returns 1, if task migration degrades locality
5703 * Returns 0, if task migration improves locality i.e migration preferred.
5704 * Returns -1, if task migration is not affected by locality.
5706 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5708 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5709 unsigned long src_faults, dst_faults;
5710 int src_nid, dst_nid;
5712 if (!static_branch_likely(&sched_numa_balancing))
5715 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
5718 src_nid = cpu_to_node(env->src_cpu);
5719 dst_nid = cpu_to_node(env->dst_cpu);
5721 if (src_nid == dst_nid)
5724 /* Migrating away from the preferred node is always bad. */
5725 if (src_nid == p->numa_preferred_nid) {
5726 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
5732 /* Encourage migration to the preferred node. */
5733 if (dst_nid == p->numa_preferred_nid)
5737 src_faults = group_faults(p, src_nid);
5738 dst_faults = group_faults(p, dst_nid);
5740 src_faults = task_faults(p, src_nid);
5741 dst_faults = task_faults(p, dst_nid);
5744 return dst_faults < src_faults;
5748 static inline int migrate_degrades_locality(struct task_struct *p,
5756 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5759 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5763 lockdep_assert_held(&env->src_rq->lock);
5766 * We do not migrate tasks that are:
5767 * 1) throttled_lb_pair, or
5768 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5769 * 3) running (obviously), or
5770 * 4) are cache-hot on their current CPU.
5772 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5775 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5778 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5780 env->flags |= LBF_SOME_PINNED;
5783 * Remember if this task can be migrated to any other cpu in
5784 * our sched_group. We may want to revisit it if we couldn't
5785 * meet load balance goals by pulling other tasks on src_cpu.
5787 * Also avoid computing new_dst_cpu if we have already computed
5788 * one in current iteration.
5790 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5793 /* Prevent to re-select dst_cpu via env's cpus */
5794 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5795 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5796 env->flags |= LBF_DST_PINNED;
5797 env->new_dst_cpu = cpu;
5805 /* Record that we found atleast one task that could run on dst_cpu */
5806 env->flags &= ~LBF_ALL_PINNED;
5808 if (task_running(env->src_rq, p)) {
5809 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5814 * Aggressive migration if:
5815 * 1) destination numa is preferred
5816 * 2) task is cache cold, or
5817 * 3) too many balance attempts have failed.
5819 tsk_cache_hot = migrate_degrades_locality(p, env);
5820 if (tsk_cache_hot == -1)
5821 tsk_cache_hot = task_hot(p, env);
5823 if (tsk_cache_hot <= 0 ||
5824 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5825 if (tsk_cache_hot == 1) {
5826 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5827 schedstat_inc(p, se.statistics.nr_forced_migrations);
5832 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5837 * detach_task() -- detach the task for the migration specified in env
5839 static void detach_task(struct task_struct *p, struct lb_env *env)
5841 lockdep_assert_held(&env->src_rq->lock);
5843 deactivate_task(env->src_rq, p, 0);
5844 p->on_rq = TASK_ON_RQ_MIGRATING;
5845 set_task_cpu(p, env->dst_cpu);
5849 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5850 * part of active balancing operations within "domain".
5852 * Returns a task if successful and NULL otherwise.
5854 static struct task_struct *detach_one_task(struct lb_env *env)
5856 struct task_struct *p, *n;
5858 lockdep_assert_held(&env->src_rq->lock);
5860 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5861 if (!can_migrate_task(p, env))
5864 detach_task(p, env);
5867 * Right now, this is only the second place where
5868 * lb_gained[env->idle] is updated (other is detach_tasks)
5869 * so we can safely collect stats here rather than
5870 * inside detach_tasks().
5872 schedstat_inc(env->sd, lb_gained[env->idle]);
5878 static const unsigned int sched_nr_migrate_break = 32;
5881 * detach_tasks() -- tries to detach up to imbalance weighted load from
5882 * busiest_rq, as part of a balancing operation within domain "sd".
5884 * Returns number of detached tasks if successful and 0 otherwise.
5886 static int detach_tasks(struct lb_env *env)
5888 struct list_head *tasks = &env->src_rq->cfs_tasks;
5889 struct task_struct *p;
5893 lockdep_assert_held(&env->src_rq->lock);
5895 if (env->imbalance <= 0)
5898 while (!list_empty(tasks)) {
5900 * We don't want to steal all, otherwise we may be treated likewise,
5901 * which could at worst lead to a livelock crash.
5903 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
5906 p = list_first_entry(tasks, struct task_struct, se.group_node);
5909 /* We've more or less seen every task there is, call it quits */
5910 if (env->loop > env->loop_max)
5913 /* take a breather every nr_migrate tasks */
5914 if (env->loop > env->loop_break) {
5915 env->loop_break += sched_nr_migrate_break;
5916 env->flags |= LBF_NEED_BREAK;
5920 if (!can_migrate_task(p, env))
5923 load = task_h_load(p);
5925 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5928 if ((load / 2) > env->imbalance)
5931 detach_task(p, env);
5932 list_add(&p->se.group_node, &env->tasks);
5935 env->imbalance -= load;
5937 #ifdef CONFIG_PREEMPT
5939 * NEWIDLE balancing is a source of latency, so preemptible
5940 * kernels will stop after the first task is detached to minimize
5941 * the critical section.
5943 if (env->idle == CPU_NEWLY_IDLE)
5948 * We only want to steal up to the prescribed amount of
5951 if (env->imbalance <= 0)
5956 list_move_tail(&p->se.group_node, tasks);
5960 * Right now, this is one of only two places we collect this stat
5961 * so we can safely collect detach_one_task() stats here rather
5962 * than inside detach_one_task().
5964 schedstat_add(env->sd, lb_gained[env->idle], detached);
5970 * attach_task() -- attach the task detached by detach_task() to its new rq.
5972 static void attach_task(struct rq *rq, struct task_struct *p)
5974 lockdep_assert_held(&rq->lock);
5976 BUG_ON(task_rq(p) != rq);
5977 p->on_rq = TASK_ON_RQ_QUEUED;
5978 activate_task(rq, p, 0);
5979 check_preempt_curr(rq, p, 0);
5983 * attach_one_task() -- attaches the task returned from detach_one_task() to
5986 static void attach_one_task(struct rq *rq, struct task_struct *p)
5988 raw_spin_lock(&rq->lock);
5990 raw_spin_unlock(&rq->lock);
5994 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
5997 static void attach_tasks(struct lb_env *env)
5999 struct list_head *tasks = &env->tasks;
6000 struct task_struct *p;
6002 raw_spin_lock(&env->dst_rq->lock);
6004 while (!list_empty(tasks)) {
6005 p = list_first_entry(tasks, struct task_struct, se.group_node);
6006 list_del_init(&p->se.group_node);
6008 attach_task(env->dst_rq, p);
6011 raw_spin_unlock(&env->dst_rq->lock);
6014 #ifdef CONFIG_FAIR_GROUP_SCHED
6015 static void update_blocked_averages(int cpu)
6017 struct rq *rq = cpu_rq(cpu);
6018 struct cfs_rq *cfs_rq;
6019 unsigned long flags;
6021 raw_spin_lock_irqsave(&rq->lock, flags);
6022 update_rq_clock(rq);
6025 * Iterates the task_group tree in a bottom up fashion, see
6026 * list_add_leaf_cfs_rq() for details.
6028 for_each_leaf_cfs_rq(rq, cfs_rq) {
6029 /* throttled entities do not contribute to load */
6030 if (throttled_hierarchy(cfs_rq))
6033 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
6034 update_tg_load_avg(cfs_rq, 0);
6036 raw_spin_unlock_irqrestore(&rq->lock, flags);
6040 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6041 * This needs to be done in a top-down fashion because the load of a child
6042 * group is a fraction of its parents load.
6044 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6046 struct rq *rq = rq_of(cfs_rq);
6047 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6048 unsigned long now = jiffies;
6051 if (cfs_rq->last_h_load_update == now)
6054 WRITE_ONCE(cfs_rq->h_load_next, NULL);
6055 for_each_sched_entity(se) {
6056 cfs_rq = cfs_rq_of(se);
6057 WRITE_ONCE(cfs_rq->h_load_next, se);
6058 if (cfs_rq->last_h_load_update == now)
6063 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6064 cfs_rq->last_h_load_update = now;
6067 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
6068 load = cfs_rq->h_load;
6069 load = div64_ul(load * se->avg.load_avg,
6070 cfs_rq_load_avg(cfs_rq) + 1);
6071 cfs_rq = group_cfs_rq(se);
6072 cfs_rq->h_load = load;
6073 cfs_rq->last_h_load_update = now;
6077 static unsigned long task_h_load(struct task_struct *p)
6079 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6081 update_cfs_rq_h_load(cfs_rq);
6082 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6083 cfs_rq_load_avg(cfs_rq) + 1);
6086 static inline void update_blocked_averages(int cpu)
6088 struct rq *rq = cpu_rq(cpu);
6089 struct cfs_rq *cfs_rq = &rq->cfs;
6090 unsigned long flags;
6092 raw_spin_lock_irqsave(&rq->lock, flags);
6093 update_rq_clock(rq);
6094 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
6095 raw_spin_unlock_irqrestore(&rq->lock, flags);
6098 static unsigned long task_h_load(struct task_struct *p)
6100 return p->se.avg.load_avg;
6104 /********** Helpers for find_busiest_group ************************/
6113 * sg_lb_stats - stats of a sched_group required for load_balancing
6115 struct sg_lb_stats {
6116 unsigned long avg_load; /*Avg load across the CPUs of the group */
6117 unsigned long group_load; /* Total load over the CPUs of the group */
6118 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6119 unsigned long load_per_task;
6120 unsigned long group_capacity;
6121 unsigned long group_util; /* Total utilization of the group */
6122 unsigned int sum_nr_running; /* Nr tasks running in the group */
6123 unsigned int idle_cpus;
6124 unsigned int group_weight;
6125 enum group_type group_type;
6126 int group_no_capacity;
6127 #ifdef CONFIG_NUMA_BALANCING
6128 unsigned int nr_numa_running;
6129 unsigned int nr_preferred_running;
6134 * sd_lb_stats - Structure to store the statistics of a sched_domain
6135 * during load balancing.
6137 struct sd_lb_stats {
6138 struct sched_group *busiest; /* Busiest group in this sd */
6139 struct sched_group *local; /* Local group in this sd */
6140 unsigned long total_load; /* Total load of all groups in sd */
6141 unsigned long total_capacity; /* Total capacity of all groups in sd */
6142 unsigned long avg_load; /* Average load across all groups in sd */
6144 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6145 struct sg_lb_stats local_stat; /* Statistics of the local group */
6148 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6151 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6152 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6153 * We must however clear busiest_stat::avg_load because
6154 * update_sd_pick_busiest() reads this before assignment.
6156 *sds = (struct sd_lb_stats){
6160 .total_capacity = 0UL,
6163 .sum_nr_running = 0,
6164 .group_type = group_other,
6170 * get_sd_load_idx - Obtain the load index for a given sched domain.
6171 * @sd: The sched_domain whose load_idx is to be obtained.
6172 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6174 * Return: The load index.
6176 static inline int get_sd_load_idx(struct sched_domain *sd,
6177 enum cpu_idle_type idle)
6183 load_idx = sd->busy_idx;
6186 case CPU_NEWLY_IDLE:
6187 load_idx = sd->newidle_idx;
6190 load_idx = sd->idle_idx;
6197 static unsigned long scale_rt_capacity(int cpu)
6199 struct rq *rq = cpu_rq(cpu);
6200 u64 total, used, age_stamp, avg;
6204 * Since we're reading these variables without serialization make sure
6205 * we read them once before doing sanity checks on them.
6207 age_stamp = READ_ONCE(rq->age_stamp);
6208 avg = READ_ONCE(rq->rt_avg);
6209 delta = __rq_clock_broken(rq) - age_stamp;
6211 if (unlikely(delta < 0))
6214 total = sched_avg_period() + delta;
6216 used = div_u64(avg, total);
6218 if (likely(used < SCHED_CAPACITY_SCALE))
6219 return SCHED_CAPACITY_SCALE - used;
6224 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6226 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6227 struct sched_group *sdg = sd->groups;
6229 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6231 capacity *= scale_rt_capacity(cpu);
6232 capacity >>= SCHED_CAPACITY_SHIFT;
6237 cpu_rq(cpu)->cpu_capacity = capacity;
6238 sdg->sgc->capacity = capacity;
6241 void update_group_capacity(struct sched_domain *sd, int cpu)
6243 struct sched_domain *child = sd->child;
6244 struct sched_group *group, *sdg = sd->groups;
6245 unsigned long capacity;
6246 unsigned long interval;
6248 interval = msecs_to_jiffies(sd->balance_interval);
6249 interval = clamp(interval, 1UL, max_load_balance_interval);
6250 sdg->sgc->next_update = jiffies + interval;
6253 update_cpu_capacity(sd, cpu);
6259 if (child->flags & SD_OVERLAP) {
6261 * SD_OVERLAP domains cannot assume that child groups
6262 * span the current group.
6265 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6266 struct sched_group_capacity *sgc;
6267 struct rq *rq = cpu_rq(cpu);
6270 * build_sched_domains() -> init_sched_groups_capacity()
6271 * gets here before we've attached the domains to the
6274 * Use capacity_of(), which is set irrespective of domains
6275 * in update_cpu_capacity().
6277 * This avoids capacity from being 0 and
6278 * causing divide-by-zero issues on boot.
6280 if (unlikely(!rq->sd)) {
6281 capacity += capacity_of(cpu);
6285 sgc = rq->sd->groups->sgc;
6286 capacity += sgc->capacity;
6290 * !SD_OVERLAP domains can assume that child groups
6291 * span the current group.
6294 group = child->groups;
6296 capacity += group->sgc->capacity;
6297 group = group->next;
6298 } while (group != child->groups);
6301 sdg->sgc->capacity = capacity;
6305 * Check whether the capacity of the rq has been noticeably reduced by side
6306 * activity. The imbalance_pct is used for the threshold.
6307 * Return true is the capacity is reduced
6310 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6312 return ((rq->cpu_capacity * sd->imbalance_pct) <
6313 (rq->cpu_capacity_orig * 100));
6317 * Group imbalance indicates (and tries to solve) the problem where balancing
6318 * groups is inadequate due to tsk_cpus_allowed() constraints.
6320 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6321 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6324 * { 0 1 2 3 } { 4 5 6 7 }
6327 * If we were to balance group-wise we'd place two tasks in the first group and
6328 * two tasks in the second group. Clearly this is undesired as it will overload
6329 * cpu 3 and leave one of the cpus in the second group unused.
6331 * The current solution to this issue is detecting the skew in the first group
6332 * by noticing the lower domain failed to reach balance and had difficulty
6333 * moving tasks due to affinity constraints.
6335 * When this is so detected; this group becomes a candidate for busiest; see
6336 * update_sd_pick_busiest(). And calculate_imbalance() and
6337 * find_busiest_group() avoid some of the usual balance conditions to allow it
6338 * to create an effective group imbalance.
6340 * This is a somewhat tricky proposition since the next run might not find the
6341 * group imbalance and decide the groups need to be balanced again. A most
6342 * subtle and fragile situation.
6345 static inline int sg_imbalanced(struct sched_group *group)
6347 return group->sgc->imbalance;
6351 * group_has_capacity returns true if the group has spare capacity that could
6352 * be used by some tasks.
6353 * We consider that a group has spare capacity if the * number of task is
6354 * smaller than the number of CPUs or if the utilization is lower than the
6355 * available capacity for CFS tasks.
6356 * For the latter, we use a threshold to stabilize the state, to take into
6357 * account the variance of the tasks' load and to return true if the available
6358 * capacity in meaningful for the load balancer.
6359 * As an example, an available capacity of 1% can appear but it doesn't make
6360 * any benefit for the load balance.
6363 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6365 if (sgs->sum_nr_running < sgs->group_weight)
6368 if ((sgs->group_capacity * 100) >
6369 (sgs->group_util * env->sd->imbalance_pct))
6376 * group_is_overloaded returns true if the group has more tasks than it can
6378 * group_is_overloaded is not equals to !group_has_capacity because a group
6379 * with the exact right number of tasks, has no more spare capacity but is not
6380 * overloaded so both group_has_capacity and group_is_overloaded return
6384 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6386 if (sgs->sum_nr_running <= sgs->group_weight)
6389 if ((sgs->group_capacity * 100) <
6390 (sgs->group_util * env->sd->imbalance_pct))
6397 group_type group_classify(struct sched_group *group,
6398 struct sg_lb_stats *sgs)
6400 if (sgs->group_no_capacity)
6401 return group_overloaded;
6403 if (sg_imbalanced(group))
6404 return group_imbalanced;
6410 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6411 * @env: The load balancing environment.
6412 * @group: sched_group whose statistics are to be updated.
6413 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6414 * @local_group: Does group contain this_cpu.
6415 * @sgs: variable to hold the statistics for this group.
6416 * @overload: Indicate more than one runnable task for any CPU.
6418 static inline void update_sg_lb_stats(struct lb_env *env,
6419 struct sched_group *group, int load_idx,
6420 int local_group, struct sg_lb_stats *sgs,
6426 memset(sgs, 0, sizeof(*sgs));
6428 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6429 struct rq *rq = cpu_rq(i);
6431 /* Bias balancing toward cpus of our domain */
6433 load = target_load(i, load_idx);
6435 load = source_load(i, load_idx);
6437 sgs->group_load += load;
6438 sgs->group_util += cpu_util(i);
6439 sgs->sum_nr_running += rq->cfs.h_nr_running;
6441 if (rq->nr_running > 1)
6444 #ifdef CONFIG_NUMA_BALANCING
6445 sgs->nr_numa_running += rq->nr_numa_running;
6446 sgs->nr_preferred_running += rq->nr_preferred_running;
6448 sgs->sum_weighted_load += weighted_cpuload(i);
6453 /* Adjust by relative CPU capacity of the group */
6454 sgs->group_capacity = group->sgc->capacity;
6455 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6457 if (sgs->sum_nr_running)
6458 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6460 sgs->group_weight = group->group_weight;
6462 sgs->group_no_capacity = group_is_overloaded(env, sgs);
6463 sgs->group_type = group_classify(group, sgs);
6467 * update_sd_pick_busiest - return 1 on busiest group
6468 * @env: The load balancing environment.
6469 * @sds: sched_domain statistics
6470 * @sg: sched_group candidate to be checked for being the busiest
6471 * @sgs: sched_group statistics
6473 * Determine if @sg is a busier group than the previously selected
6476 * Return: %true if @sg is a busier group than the previously selected
6477 * busiest group. %false otherwise.
6479 static bool update_sd_pick_busiest(struct lb_env *env,
6480 struct sd_lb_stats *sds,
6481 struct sched_group *sg,
6482 struct sg_lb_stats *sgs)
6484 struct sg_lb_stats *busiest = &sds->busiest_stat;
6486 if (sgs->group_type > busiest->group_type)
6489 if (sgs->group_type < busiest->group_type)
6492 if (sgs->avg_load <= busiest->avg_load)
6495 /* This is the busiest node in its class. */
6496 if (!(env->sd->flags & SD_ASYM_PACKING))
6500 * ASYM_PACKING needs to move all the work to the lowest
6501 * numbered CPUs in the group, therefore mark all groups
6502 * higher than ourself as busy.
6504 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6508 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
6515 #ifdef CONFIG_NUMA_BALANCING
6516 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6518 if (sgs->sum_nr_running > sgs->nr_numa_running)
6520 if (sgs->sum_nr_running > sgs->nr_preferred_running)
6525 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6527 if (rq->nr_running > rq->nr_numa_running)
6529 if (rq->nr_running > rq->nr_preferred_running)
6534 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6539 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6543 #endif /* CONFIG_NUMA_BALANCING */
6546 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6547 * @env: The load balancing environment.
6548 * @sds: variable to hold the statistics for this sched_domain.
6550 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6552 struct sched_domain *child = env->sd->child;
6553 struct sched_group *sg = env->sd->groups;
6554 struct sg_lb_stats tmp_sgs;
6555 int load_idx, prefer_sibling = 0;
6556 bool overload = false;
6558 if (child && child->flags & SD_PREFER_SIBLING)
6561 load_idx = get_sd_load_idx(env->sd, env->idle);
6564 struct sg_lb_stats *sgs = &tmp_sgs;
6567 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6570 sgs = &sds->local_stat;
6572 if (env->idle != CPU_NEWLY_IDLE ||
6573 time_after_eq(jiffies, sg->sgc->next_update))
6574 update_group_capacity(env->sd, env->dst_cpu);
6577 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6584 * In case the child domain prefers tasks go to siblings
6585 * first, lower the sg capacity so that we'll try
6586 * and move all the excess tasks away. We lower the capacity
6587 * of a group only if the local group has the capacity to fit
6588 * these excess tasks. The extra check prevents the case where
6589 * you always pull from the heaviest group when it is already
6590 * under-utilized (possible with a large weight task outweighs
6591 * the tasks on the system).
6593 if (prefer_sibling && sds->local &&
6594 group_has_capacity(env, &sds->local_stat) &&
6595 (sgs->sum_nr_running > 1)) {
6596 sgs->group_no_capacity = 1;
6597 sgs->group_type = group_classify(sg, sgs);
6600 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6602 sds->busiest_stat = *sgs;
6606 /* Now, start updating sd_lb_stats */
6607 sds->total_load += sgs->group_load;
6608 sds->total_capacity += sgs->group_capacity;
6611 } while (sg != env->sd->groups);
6613 if (env->sd->flags & SD_NUMA)
6614 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6616 if (!env->sd->parent) {
6617 /* update overload indicator if we are at root domain */
6618 if (env->dst_rq->rd->overload != overload)
6619 env->dst_rq->rd->overload = overload;
6625 * check_asym_packing - Check to see if the group is packed into the
6628 * This is primarily intended to used at the sibling level. Some
6629 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6630 * case of POWER7, it can move to lower SMT modes only when higher
6631 * threads are idle. When in lower SMT modes, the threads will
6632 * perform better since they share less core resources. Hence when we
6633 * have idle threads, we want them to be the higher ones.
6635 * This packing function is run on idle threads. It checks to see if
6636 * the busiest CPU in this domain (core in the P7 case) has a higher
6637 * CPU number than the packing function is being run on. Here we are
6638 * assuming lower CPU number will be equivalent to lower a SMT thread
6641 * Return: 1 when packing is required and a task should be moved to
6642 * this CPU. The amount of the imbalance is returned in *imbalance.
6644 * @env: The load balancing environment.
6645 * @sds: Statistics of the sched_domain which is to be packed
6647 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6651 if (!(env->sd->flags & SD_ASYM_PACKING))
6657 busiest_cpu = group_first_cpu(sds->busiest);
6658 if (env->dst_cpu > busiest_cpu)
6661 env->imbalance = DIV_ROUND_CLOSEST(
6662 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6663 SCHED_CAPACITY_SCALE);
6669 * fix_small_imbalance - Calculate the minor imbalance that exists
6670 * amongst the groups of a sched_domain, during
6672 * @env: The load balancing environment.
6673 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6676 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6678 unsigned long tmp, capa_now = 0, capa_move = 0;
6679 unsigned int imbn = 2;
6680 unsigned long scaled_busy_load_per_task;
6681 struct sg_lb_stats *local, *busiest;
6683 local = &sds->local_stat;
6684 busiest = &sds->busiest_stat;
6686 if (!local->sum_nr_running)
6687 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6688 else if (busiest->load_per_task > local->load_per_task)
6691 scaled_busy_load_per_task =
6692 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6693 busiest->group_capacity;
6695 if (busiest->avg_load + scaled_busy_load_per_task >=
6696 local->avg_load + (scaled_busy_load_per_task * imbn)) {
6697 env->imbalance = busiest->load_per_task;
6702 * OK, we don't have enough imbalance to justify moving tasks,
6703 * however we may be able to increase total CPU capacity used by
6707 capa_now += busiest->group_capacity *
6708 min(busiest->load_per_task, busiest->avg_load);
6709 capa_now += local->group_capacity *
6710 min(local->load_per_task, local->avg_load);
6711 capa_now /= SCHED_CAPACITY_SCALE;
6713 /* Amount of load we'd subtract */
6714 if (busiest->avg_load > scaled_busy_load_per_task) {
6715 capa_move += busiest->group_capacity *
6716 min(busiest->load_per_task,
6717 busiest->avg_load - scaled_busy_load_per_task);
6720 /* Amount of load we'd add */
6721 if (busiest->avg_load * busiest->group_capacity <
6722 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
6723 tmp = (busiest->avg_load * busiest->group_capacity) /
6724 local->group_capacity;
6726 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6727 local->group_capacity;
6729 capa_move += local->group_capacity *
6730 min(local->load_per_task, local->avg_load + tmp);
6731 capa_move /= SCHED_CAPACITY_SCALE;
6733 /* Move if we gain throughput */
6734 if (capa_move > capa_now)
6735 env->imbalance = busiest->load_per_task;
6739 * calculate_imbalance - Calculate the amount of imbalance present within the
6740 * groups of a given sched_domain during load balance.
6741 * @env: load balance environment
6742 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6744 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6746 unsigned long max_pull, load_above_capacity = ~0UL;
6747 struct sg_lb_stats *local, *busiest;
6749 local = &sds->local_stat;
6750 busiest = &sds->busiest_stat;
6752 if (busiest->group_type == group_imbalanced) {
6754 * In the group_imb case we cannot rely on group-wide averages
6755 * to ensure cpu-load equilibrium, look at wider averages. XXX
6757 busiest->load_per_task =
6758 min(busiest->load_per_task, sds->avg_load);
6762 * In the presence of smp nice balancing, certain scenarios can have
6763 * max load less than avg load(as we skip the groups at or below
6764 * its cpu_capacity, while calculating max_load..)
6766 if (busiest->avg_load <= sds->avg_load ||
6767 local->avg_load >= sds->avg_load) {
6769 return fix_small_imbalance(env, sds);
6773 * If there aren't any idle cpus, avoid creating some.
6775 if (busiest->group_type == group_overloaded &&
6776 local->group_type == group_overloaded) {
6777 load_above_capacity = busiest->sum_nr_running *
6779 if (load_above_capacity > busiest->group_capacity)
6780 load_above_capacity -= busiest->group_capacity;
6782 load_above_capacity = ~0UL;
6786 * We're trying to get all the cpus to the average_load, so we don't
6787 * want to push ourselves above the average load, nor do we wish to
6788 * reduce the max loaded cpu below the average load. At the same time,
6789 * we also don't want to reduce the group load below the group capacity
6790 * (so that we can implement power-savings policies etc). Thus we look
6791 * for the minimum possible imbalance.
6793 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6795 /* How much load to actually move to equalise the imbalance */
6796 env->imbalance = min(
6797 max_pull * busiest->group_capacity,
6798 (sds->avg_load - local->avg_load) * local->group_capacity
6799 ) / SCHED_CAPACITY_SCALE;
6802 * if *imbalance is less than the average load per runnable task
6803 * there is no guarantee that any tasks will be moved so we'll have
6804 * a think about bumping its value to force at least one task to be
6807 if (env->imbalance < busiest->load_per_task)
6808 return fix_small_imbalance(env, sds);
6811 /******* find_busiest_group() helpers end here *********************/
6814 * find_busiest_group - Returns the busiest group within the sched_domain
6815 * if there is an imbalance. If there isn't an imbalance, and
6816 * the user has opted for power-savings, it returns a group whose
6817 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6818 * such a group exists.
6820 * Also calculates the amount of weighted load which should be moved
6821 * to restore balance.
6823 * @env: The load balancing environment.
6825 * Return: - The busiest group if imbalance exists.
6826 * - If no imbalance and user has opted for power-savings balance,
6827 * return the least loaded group whose CPUs can be
6828 * put to idle by rebalancing its tasks onto our group.
6830 static struct sched_group *find_busiest_group(struct lb_env *env)
6832 struct sg_lb_stats *local, *busiest;
6833 struct sd_lb_stats sds;
6835 init_sd_lb_stats(&sds);
6838 * Compute the various statistics relavent for load balancing at
6841 update_sd_lb_stats(env, &sds);
6842 local = &sds.local_stat;
6843 busiest = &sds.busiest_stat;
6845 /* ASYM feature bypasses nice load balance check */
6846 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6847 check_asym_packing(env, &sds))
6850 /* There is no busy sibling group to pull tasks from */
6851 if (!sds.busiest || busiest->sum_nr_running == 0)
6854 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
6855 / sds.total_capacity;
6858 * If the busiest group is imbalanced the below checks don't
6859 * work because they assume all things are equal, which typically
6860 * isn't true due to cpus_allowed constraints and the like.
6862 if (busiest->group_type == group_imbalanced)
6865 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6866 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
6867 busiest->group_no_capacity)
6871 * If the local group is busier than the selected busiest group
6872 * don't try and pull any tasks.
6874 if (local->avg_load >= busiest->avg_load)
6878 * Don't pull any tasks if this group is already above the domain
6881 if (local->avg_load >= sds.avg_load)
6884 if (env->idle == CPU_IDLE) {
6886 * This cpu is idle. If the busiest group is not overloaded
6887 * and there is no imbalance between this and busiest group
6888 * wrt idle cpus, it is balanced. The imbalance becomes
6889 * significant if the diff is greater than 1 otherwise we
6890 * might end up to just move the imbalance on another group
6892 if ((busiest->group_type != group_overloaded) &&
6893 (local->idle_cpus <= (busiest->idle_cpus + 1)))
6897 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6898 * imbalance_pct to be conservative.
6900 if (100 * busiest->avg_load <=
6901 env->sd->imbalance_pct * local->avg_load)
6906 /* Looks like there is an imbalance. Compute it */
6907 calculate_imbalance(env, &sds);
6916 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6918 static struct rq *find_busiest_queue(struct lb_env *env,
6919 struct sched_group *group)
6921 struct rq *busiest = NULL, *rq;
6922 unsigned long busiest_load = 0, busiest_capacity = 1;
6925 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6926 unsigned long capacity, wl;
6930 rt = fbq_classify_rq(rq);
6933 * We classify groups/runqueues into three groups:
6934 * - regular: there are !numa tasks
6935 * - remote: there are numa tasks that run on the 'wrong' node
6936 * - all: there is no distinction
6938 * In order to avoid migrating ideally placed numa tasks,
6939 * ignore those when there's better options.
6941 * If we ignore the actual busiest queue to migrate another
6942 * task, the next balance pass can still reduce the busiest
6943 * queue by moving tasks around inside the node.
6945 * If we cannot move enough load due to this classification
6946 * the next pass will adjust the group classification and
6947 * allow migration of more tasks.
6949 * Both cases only affect the total convergence complexity.
6951 if (rt > env->fbq_type)
6954 capacity = capacity_of(i);
6956 wl = weighted_cpuload(i);
6959 * When comparing with imbalance, use weighted_cpuload()
6960 * which is not scaled with the cpu capacity.
6963 if (rq->nr_running == 1 && wl > env->imbalance &&
6964 !check_cpu_capacity(rq, env->sd))
6968 * For the load comparisons with the other cpu's, consider
6969 * the weighted_cpuload() scaled with the cpu capacity, so
6970 * that the load can be moved away from the cpu that is
6971 * potentially running at a lower capacity.
6973 * Thus we're looking for max(wl_i / capacity_i), crosswise
6974 * multiplication to rid ourselves of the division works out
6975 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
6976 * our previous maximum.
6978 if (wl * busiest_capacity > busiest_load * capacity) {
6980 busiest_capacity = capacity;
6989 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6990 * so long as it is large enough.
6992 #define MAX_PINNED_INTERVAL 512
6994 /* Working cpumask for load_balance and load_balance_newidle. */
6995 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6997 static int need_active_balance(struct lb_env *env)
6999 struct sched_domain *sd = env->sd;
7001 if (env->idle == CPU_NEWLY_IDLE) {
7004 * ASYM_PACKING needs to force migrate tasks from busy but
7005 * higher numbered CPUs in order to pack all tasks in the
7006 * lowest numbered CPUs.
7008 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7013 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7014 * It's worth migrating the task if the src_cpu's capacity is reduced
7015 * because of other sched_class or IRQs if more capacity stays
7016 * available on dst_cpu.
7018 if ((env->idle != CPU_NOT_IDLE) &&
7019 (env->src_rq->cfs.h_nr_running == 1)) {
7020 if ((check_cpu_capacity(env->src_rq, sd)) &&
7021 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7025 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7028 static int active_load_balance_cpu_stop(void *data);
7030 static int should_we_balance(struct lb_env *env)
7032 struct sched_group *sg = env->sd->groups;
7033 struct cpumask *sg_cpus, *sg_mask;
7034 int cpu, balance_cpu = -1;
7037 * In the newly idle case, we will allow all the cpu's
7038 * to do the newly idle load balance.
7040 if (env->idle == CPU_NEWLY_IDLE)
7043 sg_cpus = sched_group_cpus(sg);
7044 sg_mask = sched_group_mask(sg);
7045 /* Try to find first idle cpu */
7046 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7047 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7054 if (balance_cpu == -1)
7055 balance_cpu = group_balance_cpu(sg);
7058 * First idle cpu or the first cpu(busiest) in this sched group
7059 * is eligible for doing load balancing at this and above domains.
7061 return balance_cpu == env->dst_cpu;
7065 * Check this_cpu to ensure it is balanced within domain. Attempt to move
7066 * tasks if there is an imbalance.
7068 static int load_balance(int this_cpu, struct rq *this_rq,
7069 struct sched_domain *sd, enum cpu_idle_type idle,
7070 int *continue_balancing)
7072 int ld_moved, cur_ld_moved, active_balance = 0;
7073 struct sched_domain *sd_parent = sd->parent;
7074 struct sched_group *group;
7076 unsigned long flags;
7077 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
7079 struct lb_env env = {
7081 .dst_cpu = this_cpu,
7083 .dst_grpmask = sched_group_cpus(sd->groups),
7085 .loop_break = sched_nr_migrate_break,
7088 .tasks = LIST_HEAD_INIT(env.tasks),
7092 * For NEWLY_IDLE load_balancing, we don't need to consider
7093 * other cpus in our group
7095 if (idle == CPU_NEWLY_IDLE)
7096 env.dst_grpmask = NULL;
7098 cpumask_copy(cpus, cpu_active_mask);
7100 schedstat_inc(sd, lb_count[idle]);
7103 if (!should_we_balance(&env)) {
7104 *continue_balancing = 0;
7108 group = find_busiest_group(&env);
7110 schedstat_inc(sd, lb_nobusyg[idle]);
7114 busiest = find_busiest_queue(&env, group);
7116 schedstat_inc(sd, lb_nobusyq[idle]);
7120 BUG_ON(busiest == env.dst_rq);
7122 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
7124 env.src_cpu = busiest->cpu;
7125 env.src_rq = busiest;
7128 if (busiest->nr_running > 1) {
7130 * Attempt to move tasks. If find_busiest_group has found
7131 * an imbalance but busiest->nr_running <= 1, the group is
7132 * still unbalanced. ld_moved simply stays zero, so it is
7133 * correctly treated as an imbalance.
7135 env.flags |= LBF_ALL_PINNED;
7136 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7139 raw_spin_lock_irqsave(&busiest->lock, flags);
7142 * cur_ld_moved - load moved in current iteration
7143 * ld_moved - cumulative load moved across iterations
7145 cur_ld_moved = detach_tasks(&env);
7148 * We've detached some tasks from busiest_rq. Every
7149 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7150 * unlock busiest->lock, and we are able to be sure
7151 * that nobody can manipulate the tasks in parallel.
7152 * See task_rq_lock() family for the details.
7155 raw_spin_unlock(&busiest->lock);
7159 ld_moved += cur_ld_moved;
7162 local_irq_restore(flags);
7164 if (env.flags & LBF_NEED_BREAK) {
7165 env.flags &= ~LBF_NEED_BREAK;
7170 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7171 * us and move them to an alternate dst_cpu in our sched_group
7172 * where they can run. The upper limit on how many times we
7173 * iterate on same src_cpu is dependent on number of cpus in our
7176 * This changes load balance semantics a bit on who can move
7177 * load to a given_cpu. In addition to the given_cpu itself
7178 * (or a ilb_cpu acting on its behalf where given_cpu is
7179 * nohz-idle), we now have balance_cpu in a position to move
7180 * load to given_cpu. In rare situations, this may cause
7181 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7182 * _independently_ and at _same_ time to move some load to
7183 * given_cpu) causing exceess load to be moved to given_cpu.
7184 * This however should not happen so much in practice and
7185 * moreover subsequent load balance cycles should correct the
7186 * excess load moved.
7188 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7190 /* Prevent to re-select dst_cpu via env's cpus */
7191 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7193 env.dst_rq = cpu_rq(env.new_dst_cpu);
7194 env.dst_cpu = env.new_dst_cpu;
7195 env.flags &= ~LBF_DST_PINNED;
7197 env.loop_break = sched_nr_migrate_break;
7200 * Go back to "more_balance" rather than "redo" since we
7201 * need to continue with same src_cpu.
7207 * We failed to reach balance because of affinity.
7210 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7212 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7213 *group_imbalance = 1;
7216 /* All tasks on this runqueue were pinned by CPU affinity */
7217 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7218 cpumask_clear_cpu(cpu_of(busiest), cpus);
7219 if (!cpumask_empty(cpus)) {
7221 env.loop_break = sched_nr_migrate_break;
7224 goto out_all_pinned;
7229 schedstat_inc(sd, lb_failed[idle]);
7231 * Increment the failure counter only on periodic balance.
7232 * We do not want newidle balance, which can be very
7233 * frequent, pollute the failure counter causing
7234 * excessive cache_hot migrations and active balances.
7236 if (idle != CPU_NEWLY_IDLE)
7237 sd->nr_balance_failed++;
7239 if (need_active_balance(&env)) {
7240 raw_spin_lock_irqsave(&busiest->lock, flags);
7242 /* don't kick the active_load_balance_cpu_stop,
7243 * if the curr task on busiest cpu can't be
7246 if (!cpumask_test_cpu(this_cpu,
7247 tsk_cpus_allowed(busiest->curr))) {
7248 raw_spin_unlock_irqrestore(&busiest->lock,
7250 env.flags |= LBF_ALL_PINNED;
7251 goto out_one_pinned;
7255 * ->active_balance synchronizes accesses to
7256 * ->active_balance_work. Once set, it's cleared
7257 * only after active load balance is finished.
7259 if (!busiest->active_balance) {
7260 busiest->active_balance = 1;
7261 busiest->push_cpu = this_cpu;
7264 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7266 if (active_balance) {
7267 stop_one_cpu_nowait(cpu_of(busiest),
7268 active_load_balance_cpu_stop, busiest,
7269 &busiest->active_balance_work);
7273 * We've kicked active balancing, reset the failure
7276 sd->nr_balance_failed = sd->cache_nice_tries+1;
7279 sd->nr_balance_failed = 0;
7281 if (likely(!active_balance)) {
7282 /* We were unbalanced, so reset the balancing interval */
7283 sd->balance_interval = sd->min_interval;
7286 * If we've begun active balancing, start to back off. This
7287 * case may not be covered by the all_pinned logic if there
7288 * is only 1 task on the busy runqueue (because we don't call
7291 if (sd->balance_interval < sd->max_interval)
7292 sd->balance_interval *= 2;
7299 * We reach balance although we may have faced some affinity
7300 * constraints. Clear the imbalance flag if it was set.
7303 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7305 if (*group_imbalance)
7306 *group_imbalance = 0;
7311 * We reach balance because all tasks are pinned at this level so
7312 * we can't migrate them. Let the imbalance flag set so parent level
7313 * can try to migrate them.
7315 schedstat_inc(sd, lb_balanced[idle]);
7317 sd->nr_balance_failed = 0;
7320 /* tune up the balancing interval */
7321 if (((env.flags & LBF_ALL_PINNED) &&
7322 sd->balance_interval < MAX_PINNED_INTERVAL) ||
7323 (sd->balance_interval < sd->max_interval))
7324 sd->balance_interval *= 2;
7331 static inline unsigned long
7332 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7334 unsigned long interval = sd->balance_interval;
7337 interval *= sd->busy_factor;
7339 /* scale ms to jiffies */
7340 interval = msecs_to_jiffies(interval);
7341 interval = clamp(interval, 1UL, max_load_balance_interval);
7347 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
7349 unsigned long interval, next;
7351 interval = get_sd_balance_interval(sd, cpu_busy);
7352 next = sd->last_balance + interval;
7354 if (time_after(*next_balance, next))
7355 *next_balance = next;
7359 * idle_balance is called by schedule() if this_cpu is about to become
7360 * idle. Attempts to pull tasks from other CPUs.
7362 static int idle_balance(struct rq *this_rq)
7364 unsigned long next_balance = jiffies + HZ;
7365 int this_cpu = this_rq->cpu;
7366 struct sched_domain *sd;
7367 int pulled_task = 0;
7370 idle_enter_fair(this_rq);
7373 * We must set idle_stamp _before_ calling idle_balance(), such that we
7374 * measure the duration of idle_balance() as idle time.
7376 this_rq->idle_stamp = rq_clock(this_rq);
7378 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
7379 !this_rq->rd->overload) {
7381 sd = rcu_dereference_check_sched_domain(this_rq->sd);
7383 update_next_balance(sd, 0, &next_balance);
7389 raw_spin_unlock(&this_rq->lock);
7391 update_blocked_averages(this_cpu);
7393 for_each_domain(this_cpu, sd) {
7394 int continue_balancing = 1;
7395 u64 t0, domain_cost;
7397 if (!(sd->flags & SD_LOAD_BALANCE))
7400 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7401 update_next_balance(sd, 0, &next_balance);
7405 if (sd->flags & SD_BALANCE_NEWIDLE) {
7406 t0 = sched_clock_cpu(this_cpu);
7408 pulled_task = load_balance(this_cpu, this_rq,
7410 &continue_balancing);
7412 domain_cost = sched_clock_cpu(this_cpu) - t0;
7413 if (domain_cost > sd->max_newidle_lb_cost)
7414 sd->max_newidle_lb_cost = domain_cost;
7416 curr_cost += domain_cost;
7419 update_next_balance(sd, 0, &next_balance);
7422 * Stop searching for tasks to pull if there are
7423 * now runnable tasks on this rq.
7425 if (pulled_task || this_rq->nr_running > 0)
7430 raw_spin_lock(&this_rq->lock);
7432 if (curr_cost > this_rq->max_idle_balance_cost)
7433 this_rq->max_idle_balance_cost = curr_cost;
7436 * While browsing the domains, we released the rq lock, a task could
7437 * have been enqueued in the meantime. Since we're not going idle,
7438 * pretend we pulled a task.
7440 if (this_rq->cfs.h_nr_running && !pulled_task)
7444 /* Move the next balance forward */
7445 if (time_after(this_rq->next_balance, next_balance))
7446 this_rq->next_balance = next_balance;
7448 /* Is there a task of a high priority class? */
7449 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7453 idle_exit_fair(this_rq);
7454 this_rq->idle_stamp = 0;
7461 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7462 * running tasks off the busiest CPU onto idle CPUs. It requires at
7463 * least 1 task to be running on each physical CPU where possible, and
7464 * avoids physical / logical imbalances.
7466 static int active_load_balance_cpu_stop(void *data)
7468 struct rq *busiest_rq = data;
7469 int busiest_cpu = cpu_of(busiest_rq);
7470 int target_cpu = busiest_rq->push_cpu;
7471 struct rq *target_rq = cpu_rq(target_cpu);
7472 struct sched_domain *sd;
7473 struct task_struct *p = NULL;
7475 raw_spin_lock_irq(&busiest_rq->lock);
7477 /* make sure the requested cpu hasn't gone down in the meantime */
7478 if (unlikely(busiest_cpu != smp_processor_id() ||
7479 !busiest_rq->active_balance))
7482 /* Is there any task to move? */
7483 if (busiest_rq->nr_running <= 1)
7487 * This condition is "impossible", if it occurs
7488 * we need to fix it. Originally reported by
7489 * Bjorn Helgaas on a 128-cpu setup.
7491 BUG_ON(busiest_rq == target_rq);
7493 /* Search for an sd spanning us and the target CPU. */
7495 for_each_domain(target_cpu, sd) {
7496 if ((sd->flags & SD_LOAD_BALANCE) &&
7497 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7502 struct lb_env env = {
7504 .dst_cpu = target_cpu,
7505 .dst_rq = target_rq,
7506 .src_cpu = busiest_rq->cpu,
7507 .src_rq = busiest_rq,
7511 schedstat_inc(sd, alb_count);
7513 p = detach_one_task(&env);
7515 schedstat_inc(sd, alb_pushed);
7517 schedstat_inc(sd, alb_failed);
7521 busiest_rq->active_balance = 0;
7522 raw_spin_unlock(&busiest_rq->lock);
7525 attach_one_task(target_rq, p);
7532 static inline int on_null_domain(struct rq *rq)
7534 return unlikely(!rcu_dereference_sched(rq->sd));
7537 #ifdef CONFIG_NO_HZ_COMMON
7539 * idle load balancing details
7540 * - When one of the busy CPUs notice that there may be an idle rebalancing
7541 * needed, they will kick the idle load balancer, which then does idle
7542 * load balancing for all the idle CPUs.
7545 cpumask_var_t idle_cpus_mask;
7547 unsigned long next_balance; /* in jiffy units */
7548 } nohz ____cacheline_aligned;
7550 static inline int find_new_ilb(void)
7552 int ilb = cpumask_first(nohz.idle_cpus_mask);
7554 if (ilb < nr_cpu_ids && idle_cpu(ilb))
7561 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7562 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7563 * CPU (if there is one).
7565 static void nohz_balancer_kick(void)
7569 nohz.next_balance++;
7571 ilb_cpu = find_new_ilb();
7573 if (ilb_cpu >= nr_cpu_ids)
7576 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7579 * Use smp_send_reschedule() instead of resched_cpu().
7580 * This way we generate a sched IPI on the target cpu which
7581 * is idle. And the softirq performing nohz idle load balance
7582 * will be run before returning from the IPI.
7584 smp_send_reschedule(ilb_cpu);
7588 static inline void nohz_balance_exit_idle(int cpu)
7590 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7592 * Completely isolated CPUs don't ever set, so we must test.
7594 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7595 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7596 atomic_dec(&nohz.nr_cpus);
7598 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7602 static inline void set_cpu_sd_state_busy(void)
7604 struct sched_domain *sd;
7605 int cpu = smp_processor_id();
7608 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7610 if (!sd || !sd->nohz_idle)
7614 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7619 void set_cpu_sd_state_idle(void)
7621 struct sched_domain *sd;
7622 int cpu = smp_processor_id();
7625 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7627 if (!sd || sd->nohz_idle)
7631 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7637 * This routine will record that the cpu is going idle with tick stopped.
7638 * This info will be used in performing idle load balancing in the future.
7640 void nohz_balance_enter_idle(int cpu)
7643 * If this cpu is going down, then nothing needs to be done.
7645 if (!cpu_active(cpu))
7648 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7652 * If we're a completely isolated CPU, we don't play.
7654 if (on_null_domain(cpu_rq(cpu)))
7657 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7658 atomic_inc(&nohz.nr_cpus);
7659 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7662 static int sched_ilb_notifier(struct notifier_block *nfb,
7663 unsigned long action, void *hcpu)
7665 switch (action & ~CPU_TASKS_FROZEN) {
7667 nohz_balance_exit_idle(smp_processor_id());
7675 static DEFINE_SPINLOCK(balancing);
7678 * Scale the max load_balance interval with the number of CPUs in the system.
7679 * This trades load-balance latency on larger machines for less cross talk.
7681 void update_max_interval(void)
7683 max_load_balance_interval = HZ*num_online_cpus()/10;
7687 * It checks each scheduling domain to see if it is due to be balanced,
7688 * and initiates a balancing operation if so.
7690 * Balancing parameters are set up in init_sched_domains.
7692 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7694 int continue_balancing = 1;
7696 unsigned long interval;
7697 struct sched_domain *sd;
7698 /* Earliest time when we have to do rebalance again */
7699 unsigned long next_balance = jiffies + 60*HZ;
7700 int update_next_balance = 0;
7701 int need_serialize, need_decay = 0;
7704 update_blocked_averages(cpu);
7707 for_each_domain(cpu, sd) {
7709 * Decay the newidle max times here because this is a regular
7710 * visit to all the domains. Decay ~1% per second.
7712 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7713 sd->max_newidle_lb_cost =
7714 (sd->max_newidle_lb_cost * 253) / 256;
7715 sd->next_decay_max_lb_cost = jiffies + HZ;
7718 max_cost += sd->max_newidle_lb_cost;
7720 if (!(sd->flags & SD_LOAD_BALANCE))
7724 * Stop the load balance at this level. There is another
7725 * CPU in our sched group which is doing load balancing more
7728 if (!continue_balancing) {
7734 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7736 need_serialize = sd->flags & SD_SERIALIZE;
7737 if (need_serialize) {
7738 if (!spin_trylock(&balancing))
7742 if (time_after_eq(jiffies, sd->last_balance + interval)) {
7743 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7745 * The LBF_DST_PINNED logic could have changed
7746 * env->dst_cpu, so we can't know our idle
7747 * state even if we migrated tasks. Update it.
7749 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7751 sd->last_balance = jiffies;
7752 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7755 spin_unlock(&balancing);
7757 if (time_after(next_balance, sd->last_balance + interval)) {
7758 next_balance = sd->last_balance + interval;
7759 update_next_balance = 1;
7764 * Ensure the rq-wide value also decays but keep it at a
7765 * reasonable floor to avoid funnies with rq->avg_idle.
7767 rq->max_idle_balance_cost =
7768 max((u64)sysctl_sched_migration_cost, max_cost);
7773 * next_balance will be updated only when there is a need.
7774 * When the cpu is attached to null domain for ex, it will not be
7777 if (likely(update_next_balance)) {
7778 rq->next_balance = next_balance;
7780 #ifdef CONFIG_NO_HZ_COMMON
7782 * If this CPU has been elected to perform the nohz idle
7783 * balance. Other idle CPUs have already rebalanced with
7784 * nohz_idle_balance() and nohz.next_balance has been
7785 * updated accordingly. This CPU is now running the idle load
7786 * balance for itself and we need to update the
7787 * nohz.next_balance accordingly.
7789 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
7790 nohz.next_balance = rq->next_balance;
7795 #ifdef CONFIG_NO_HZ_COMMON
7797 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7798 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7800 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7802 int this_cpu = this_rq->cpu;
7805 /* Earliest time when we have to do rebalance again */
7806 unsigned long next_balance = jiffies + 60*HZ;
7807 int update_next_balance = 0;
7809 if (idle != CPU_IDLE ||
7810 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7813 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7814 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7818 * If this cpu gets work to do, stop the load balancing
7819 * work being done for other cpus. Next load
7820 * balancing owner will pick it up.
7825 rq = cpu_rq(balance_cpu);
7828 * If time for next balance is due,
7831 if (time_after_eq(jiffies, rq->next_balance)) {
7832 raw_spin_lock_irq(&rq->lock);
7833 update_rq_clock(rq);
7834 update_idle_cpu_load(rq);
7835 raw_spin_unlock_irq(&rq->lock);
7836 rebalance_domains(rq, CPU_IDLE);
7839 if (time_after(next_balance, rq->next_balance)) {
7840 next_balance = rq->next_balance;
7841 update_next_balance = 1;
7846 * next_balance will be updated only when there is a need.
7847 * When the CPU is attached to null domain for ex, it will not be
7850 if (likely(update_next_balance))
7851 nohz.next_balance = next_balance;
7853 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7857 * Current heuristic for kicking the idle load balancer in the presence
7858 * of an idle cpu in the system.
7859 * - This rq has more than one task.
7860 * - This rq has at least one CFS task and the capacity of the CPU is
7861 * significantly reduced because of RT tasks or IRQs.
7862 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
7863 * multiple busy cpu.
7864 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7865 * domain span are idle.
7867 static inline bool nohz_kick_needed(struct rq *rq)
7869 unsigned long now = jiffies;
7870 struct sched_domain *sd;
7871 struct sched_group_capacity *sgc;
7872 int nr_busy, cpu = rq->cpu;
7875 if (unlikely(rq->idle_balance))
7879 * We may be recently in ticked or tickless idle mode. At the first
7880 * busy tick after returning from idle, we will update the busy stats.
7882 set_cpu_sd_state_busy();
7883 nohz_balance_exit_idle(cpu);
7886 * None are in tickless mode and hence no need for NOHZ idle load
7889 if (likely(!atomic_read(&nohz.nr_cpus)))
7892 if (time_before(now, nohz.next_balance))
7895 if (rq->nr_running >= 2)
7899 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7901 sgc = sd->groups->sgc;
7902 nr_busy = atomic_read(&sgc->nr_busy_cpus);
7911 sd = rcu_dereference(rq->sd);
7913 if ((rq->cfs.h_nr_running >= 1) &&
7914 check_cpu_capacity(rq, sd)) {
7920 sd = rcu_dereference(per_cpu(sd_asym, cpu));
7921 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7922 sched_domain_span(sd)) < cpu)) {
7932 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7936 * run_rebalance_domains is triggered when needed from the scheduler tick.
7937 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7939 static void run_rebalance_domains(struct softirq_action *h)
7941 struct rq *this_rq = this_rq();
7942 enum cpu_idle_type idle = this_rq->idle_balance ?
7943 CPU_IDLE : CPU_NOT_IDLE;
7946 * If this cpu has a pending nohz_balance_kick, then do the
7947 * balancing on behalf of the other idle cpus whose ticks are
7948 * stopped. Do nohz_idle_balance *before* rebalance_domains to
7949 * give the idle cpus a chance to load balance. Else we may
7950 * load balance only within the local sched_domain hierarchy
7951 * and abort nohz_idle_balance altogether if we pull some load.
7953 nohz_idle_balance(this_rq, idle);
7954 rebalance_domains(this_rq, idle);
7958 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7960 void trigger_load_balance(struct rq *rq)
7962 /* Don't need to rebalance while attached to NULL domain */
7963 if (unlikely(on_null_domain(rq)))
7966 if (time_after_eq(jiffies, rq->next_balance))
7967 raise_softirq(SCHED_SOFTIRQ);
7968 #ifdef CONFIG_NO_HZ_COMMON
7969 if (nohz_kick_needed(rq))
7970 nohz_balancer_kick();
7974 static void rq_online_fair(struct rq *rq)
7978 update_runtime_enabled(rq);
7981 static void rq_offline_fair(struct rq *rq)
7985 /* Ensure any throttled groups are reachable by pick_next_task */
7986 unthrottle_offline_cfs_rqs(rq);
7989 #endif /* CONFIG_SMP */
7992 * scheduler tick hitting a task of our scheduling class:
7994 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7996 struct cfs_rq *cfs_rq;
7997 struct sched_entity *se = &curr->se;
7999 for_each_sched_entity(se) {
8000 cfs_rq = cfs_rq_of(se);
8001 entity_tick(cfs_rq, se, queued);
8004 if (static_branch_unlikely(&sched_numa_balancing))
8005 task_tick_numa(rq, curr);
8009 * called on fork with the child task as argument from the parent's context
8010 * - child not yet on the tasklist
8011 * - preemption disabled
8013 static void task_fork_fair(struct task_struct *p)
8015 struct cfs_rq *cfs_rq;
8016 struct sched_entity *se = &p->se, *curr;
8017 int this_cpu = smp_processor_id();
8018 struct rq *rq = this_rq();
8019 unsigned long flags;
8021 raw_spin_lock_irqsave(&rq->lock, flags);
8023 update_rq_clock(rq);
8025 cfs_rq = task_cfs_rq(current);
8026 curr = cfs_rq->curr;
8029 * Not only the cpu but also the task_group of the parent might have
8030 * been changed after parent->se.parent,cfs_rq were copied to
8031 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
8032 * of child point to valid ones.
8035 __set_task_cpu(p, this_cpu);
8038 update_curr(cfs_rq);
8041 se->vruntime = curr->vruntime;
8042 place_entity(cfs_rq, se, 1);
8044 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8046 * Upon rescheduling, sched_class::put_prev_task() will place
8047 * 'current' within the tree based on its new key value.
8049 swap(curr->vruntime, se->vruntime);
8053 se->vruntime -= cfs_rq->min_vruntime;
8055 raw_spin_unlock_irqrestore(&rq->lock, flags);
8059 * Priority of the task has changed. Check to see if we preempt
8063 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
8065 if (!task_on_rq_queued(p))
8069 * Reschedule if we are currently running on this runqueue and
8070 * our priority decreased, or if we are not currently running on
8071 * this runqueue and our priority is higher than the current's
8073 if (rq->curr == p) {
8074 if (p->prio > oldprio)
8077 check_preempt_curr(rq, p, 0);
8080 static inline bool vruntime_normalized(struct task_struct *p)
8082 struct sched_entity *se = &p->se;
8085 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
8086 * the dequeue_entity(.flags=0) will already have normalized the
8093 * When !on_rq, vruntime of the task has usually NOT been normalized.
8094 * But there are some cases where it has already been normalized:
8096 * - A forked child which is waiting for being woken up by
8097 * wake_up_new_task().
8098 * - A task which has been woken up by try_to_wake_up() and
8099 * waiting for actually being woken up by sched_ttwu_pending().
8101 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
8107 static void detach_task_cfs_rq(struct task_struct *p)
8109 struct sched_entity *se = &p->se;
8110 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8112 if (!vruntime_normalized(p)) {
8114 * Fix up our vruntime so that the current sleep doesn't
8115 * cause 'unlimited' sleep bonus.
8117 place_entity(cfs_rq, se, 0);
8118 se->vruntime -= cfs_rq->min_vruntime;
8121 /* Catch up with the cfs_rq and remove our load when we leave */
8122 detach_entity_load_avg(cfs_rq, se);
8125 static void attach_task_cfs_rq(struct task_struct *p)
8127 struct sched_entity *se = &p->se;
8128 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8130 #ifdef CONFIG_FAIR_GROUP_SCHED
8132 * Since the real-depth could have been changed (only FAIR
8133 * class maintain depth value), reset depth properly.
8135 se->depth = se->parent ? se->parent->depth + 1 : 0;
8138 /* Synchronize task with its cfs_rq */
8139 attach_entity_load_avg(cfs_rq, se);
8141 if (!vruntime_normalized(p))
8142 se->vruntime += cfs_rq->min_vruntime;
8145 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8147 detach_task_cfs_rq(p);
8150 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8152 attach_task_cfs_rq(p);
8154 if (task_on_rq_queued(p)) {
8156 * We were most likely switched from sched_rt, so
8157 * kick off the schedule if running, otherwise just see
8158 * if we can still preempt the current task.
8163 check_preempt_curr(rq, p, 0);
8167 /* Account for a task changing its policy or group.
8169 * This routine is mostly called to set cfs_rq->curr field when a task
8170 * migrates between groups/classes.
8172 static void set_curr_task_fair(struct rq *rq)
8174 struct sched_entity *se = &rq->curr->se;
8176 for_each_sched_entity(se) {
8177 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8179 set_next_entity(cfs_rq, se);
8180 /* ensure bandwidth has been allocated on our new cfs_rq */
8181 account_cfs_rq_runtime(cfs_rq, 0);
8185 void init_cfs_rq(struct cfs_rq *cfs_rq)
8187 cfs_rq->tasks_timeline = RB_ROOT;
8188 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8189 #ifndef CONFIG_64BIT
8190 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8193 atomic_long_set(&cfs_rq->removed_load_avg, 0);
8194 atomic_long_set(&cfs_rq->removed_util_avg, 0);
8198 #ifdef CONFIG_FAIR_GROUP_SCHED
8199 static void task_move_group_fair(struct task_struct *p)
8201 detach_task_cfs_rq(p);
8202 set_task_rq(p, task_cpu(p));
8205 /* Tell se's cfs_rq has been changed -- migrated */
8206 p->se.avg.last_update_time = 0;
8208 attach_task_cfs_rq(p);
8211 void free_fair_sched_group(struct task_group *tg)
8215 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8217 for_each_possible_cpu(i) {
8219 kfree(tg->cfs_rq[i]);
8228 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8230 struct cfs_rq *cfs_rq;
8231 struct sched_entity *se;
8234 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8237 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8241 tg->shares = NICE_0_LOAD;
8243 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8245 for_each_possible_cpu(i) {
8246 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8247 GFP_KERNEL, cpu_to_node(i));
8251 se = kzalloc_node(sizeof(struct sched_entity),
8252 GFP_KERNEL, cpu_to_node(i));
8256 init_cfs_rq(cfs_rq);
8257 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8258 init_entity_runnable_average(se);
8269 void unregister_fair_sched_group(struct task_group *tg)
8271 unsigned long flags;
8275 for_each_possible_cpu(cpu) {
8277 remove_entity_load_avg(tg->se[cpu]);
8280 * Only empty task groups can be destroyed; so we can speculatively
8281 * check on_list without danger of it being re-added.
8283 if (!tg->cfs_rq[cpu]->on_list)
8288 raw_spin_lock_irqsave(&rq->lock, flags);
8289 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8290 raw_spin_unlock_irqrestore(&rq->lock, flags);
8294 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8295 struct sched_entity *se, int cpu,
8296 struct sched_entity *parent)
8298 struct rq *rq = cpu_rq(cpu);
8302 init_cfs_rq_runtime(cfs_rq);
8304 tg->cfs_rq[cpu] = cfs_rq;
8307 /* se could be NULL for root_task_group */
8312 se->cfs_rq = &rq->cfs;
8315 se->cfs_rq = parent->my_q;
8316 se->depth = parent->depth + 1;
8320 /* guarantee group entities always have weight */
8321 update_load_set(&se->load, NICE_0_LOAD);
8322 se->parent = parent;
8325 static DEFINE_MUTEX(shares_mutex);
8327 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8330 unsigned long flags;
8333 * We can't change the weight of the root cgroup.
8338 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8340 mutex_lock(&shares_mutex);
8341 if (tg->shares == shares)
8344 tg->shares = shares;
8345 for_each_possible_cpu(i) {
8346 struct rq *rq = cpu_rq(i);
8347 struct sched_entity *se;
8350 /* Propagate contribution to hierarchy */
8351 raw_spin_lock_irqsave(&rq->lock, flags);
8353 /* Possible calls to update_curr() need rq clock */
8354 update_rq_clock(rq);
8355 for_each_sched_entity(se)
8356 update_cfs_shares(group_cfs_rq(se));
8357 raw_spin_unlock_irqrestore(&rq->lock, flags);
8361 mutex_unlock(&shares_mutex);
8364 #else /* CONFIG_FAIR_GROUP_SCHED */
8366 void free_fair_sched_group(struct task_group *tg) { }
8368 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8373 void unregister_fair_sched_group(struct task_group *tg) { }
8375 #endif /* CONFIG_FAIR_GROUP_SCHED */
8378 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
8380 struct sched_entity *se = &task->se;
8381 unsigned int rr_interval = 0;
8384 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8387 if (rq->cfs.load.weight)
8388 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
8394 * All the scheduling class methods:
8396 const struct sched_class fair_sched_class = {
8397 .next = &idle_sched_class,
8398 .enqueue_task = enqueue_task_fair,
8399 .dequeue_task = dequeue_task_fair,
8400 .yield_task = yield_task_fair,
8401 .yield_to_task = yield_to_task_fair,
8403 .check_preempt_curr = check_preempt_wakeup,
8405 .pick_next_task = pick_next_task_fair,
8406 .put_prev_task = put_prev_task_fair,
8409 .select_task_rq = select_task_rq_fair,
8410 .migrate_task_rq = migrate_task_rq_fair,
8412 .rq_online = rq_online_fair,
8413 .rq_offline = rq_offline_fair,
8415 .task_waking = task_waking_fair,
8416 .task_dead = task_dead_fair,
8417 .set_cpus_allowed = set_cpus_allowed_common,
8420 .set_curr_task = set_curr_task_fair,
8421 .task_tick = task_tick_fair,
8422 .task_fork = task_fork_fair,
8424 .prio_changed = prio_changed_fair,
8425 .switched_from = switched_from_fair,
8426 .switched_to = switched_to_fair,
8428 .get_rr_interval = get_rr_interval_fair,
8430 .update_curr = update_curr_fair,
8432 #ifdef CONFIG_FAIR_GROUP_SCHED
8433 .task_move_group = task_move_group_fair,
8437 #ifdef CONFIG_SCHED_DEBUG
8438 void print_cfs_stats(struct seq_file *m, int cpu)
8440 struct cfs_rq *cfs_rq;
8443 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
8444 print_cfs_rq(m, cpu, cfs_rq);
8448 #ifdef CONFIG_NUMA_BALANCING
8449 void show_numa_stats(struct task_struct *p, struct seq_file *m)
8452 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
8454 for_each_online_node(node) {
8455 if (p->numa_faults) {
8456 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
8457 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
8459 if (p->numa_group) {
8460 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
8461 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
8463 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
8466 #endif /* CONFIG_NUMA_BALANCING */
8467 #endif /* CONFIG_SCHED_DEBUG */
8469 __init void init_sched_fair_class(void)
8472 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8474 #ifdef CONFIG_NO_HZ_COMMON
8475 nohz.next_balance = jiffies;
8476 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8477 cpu_notifier(sched_ilb_notifier, 0);