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 delta = p->se.avg.load_sum / p->se.load.weight;
1727 *period = LOAD_AVG_MAX;
1730 p->last_sum_exec_runtime = runtime;
1731 p->last_task_numa_placement = now;
1737 * Determine the preferred nid for a task in a numa_group. This needs to
1738 * be done in a way that produces consistent results with group_weight,
1739 * otherwise workloads might not converge.
1741 static int preferred_group_nid(struct task_struct *p, int nid)
1746 /* Direct connections between all NUMA nodes. */
1747 if (sched_numa_topology_type == NUMA_DIRECT)
1751 * On a system with glueless mesh NUMA topology, group_weight
1752 * scores nodes according to the number of NUMA hinting faults on
1753 * both the node itself, and on nearby nodes.
1755 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1756 unsigned long score, max_score = 0;
1757 int node, max_node = nid;
1759 dist = sched_max_numa_distance;
1761 for_each_online_node(node) {
1762 score = group_weight(p, node, dist);
1763 if (score > max_score) {
1772 * Finding the preferred nid in a system with NUMA backplane
1773 * interconnect topology is more involved. The goal is to locate
1774 * tasks from numa_groups near each other in the system, and
1775 * untangle workloads from different sides of the system. This requires
1776 * searching down the hierarchy of node groups, recursively searching
1777 * inside the highest scoring group of nodes. The nodemask tricks
1778 * keep the complexity of the search down.
1780 nodes = node_online_map;
1781 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1782 unsigned long max_faults = 0;
1783 nodemask_t max_group = NODE_MASK_NONE;
1786 /* Are there nodes at this distance from each other? */
1787 if (!find_numa_distance(dist))
1790 for_each_node_mask(a, nodes) {
1791 unsigned long faults = 0;
1792 nodemask_t this_group;
1793 nodes_clear(this_group);
1795 /* Sum group's NUMA faults; includes a==b case. */
1796 for_each_node_mask(b, nodes) {
1797 if (node_distance(a, b) < dist) {
1798 faults += group_faults(p, b);
1799 node_set(b, this_group);
1800 node_clear(b, nodes);
1804 /* Remember the top group. */
1805 if (faults > max_faults) {
1806 max_faults = faults;
1807 max_group = this_group;
1809 * subtle: at the smallest distance there is
1810 * just one node left in each "group", the
1811 * winner is the preferred nid.
1816 /* Next round, evaluate the nodes within max_group. */
1824 static void task_numa_placement(struct task_struct *p)
1826 int seq, nid, max_nid = -1, max_group_nid = -1;
1827 unsigned long max_faults = 0, max_group_faults = 0;
1828 unsigned long fault_types[2] = { 0, 0 };
1829 unsigned long total_faults;
1830 u64 runtime, period;
1831 spinlock_t *group_lock = NULL;
1834 * The p->mm->numa_scan_seq field gets updated without
1835 * exclusive access. Use READ_ONCE() here to ensure
1836 * that the field is read in a single access:
1838 seq = READ_ONCE(p->mm->numa_scan_seq);
1839 if (p->numa_scan_seq == seq)
1841 p->numa_scan_seq = seq;
1842 p->numa_scan_period_max = task_scan_max(p);
1844 total_faults = p->numa_faults_locality[0] +
1845 p->numa_faults_locality[1];
1846 runtime = numa_get_avg_runtime(p, &period);
1848 /* If the task is part of a group prevent parallel updates to group stats */
1849 if (p->numa_group) {
1850 group_lock = &p->numa_group->lock;
1851 spin_lock_irq(group_lock);
1854 /* Find the node with the highest number of faults */
1855 for_each_online_node(nid) {
1856 /* Keep track of the offsets in numa_faults array */
1857 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1858 unsigned long faults = 0, group_faults = 0;
1861 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1862 long diff, f_diff, f_weight;
1864 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1865 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1866 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1867 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1869 /* Decay existing window, copy faults since last scan */
1870 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1871 fault_types[priv] += p->numa_faults[membuf_idx];
1872 p->numa_faults[membuf_idx] = 0;
1875 * Normalize the faults_from, so all tasks in a group
1876 * count according to CPU use, instead of by the raw
1877 * number of faults. Tasks with little runtime have
1878 * little over-all impact on throughput, and thus their
1879 * faults are less important.
1881 f_weight = div64_u64(runtime << 16, period + 1);
1882 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1884 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1885 p->numa_faults[cpubuf_idx] = 0;
1887 p->numa_faults[mem_idx] += diff;
1888 p->numa_faults[cpu_idx] += f_diff;
1889 faults += p->numa_faults[mem_idx];
1890 p->total_numa_faults += diff;
1891 if (p->numa_group) {
1893 * safe because we can only change our own group
1895 * mem_idx represents the offset for a given
1896 * nid and priv in a specific region because it
1897 * is at the beginning of the numa_faults array.
1899 p->numa_group->faults[mem_idx] += diff;
1900 p->numa_group->faults_cpu[mem_idx] += f_diff;
1901 p->numa_group->total_faults += diff;
1902 group_faults += p->numa_group->faults[mem_idx];
1906 if (faults > max_faults) {
1907 max_faults = faults;
1911 if (group_faults > max_group_faults) {
1912 max_group_faults = group_faults;
1913 max_group_nid = nid;
1917 update_task_scan_period(p, fault_types[0], fault_types[1]);
1919 if (p->numa_group) {
1920 update_numa_active_node_mask(p->numa_group);
1921 spin_unlock_irq(group_lock);
1922 max_nid = preferred_group_nid(p, max_group_nid);
1926 /* Set the new preferred node */
1927 if (max_nid != p->numa_preferred_nid)
1928 sched_setnuma(p, max_nid);
1930 if (task_node(p) != p->numa_preferred_nid)
1931 numa_migrate_preferred(p);
1935 static inline int get_numa_group(struct numa_group *grp)
1937 return atomic_inc_not_zero(&grp->refcount);
1940 static inline void put_numa_group(struct numa_group *grp)
1942 if (atomic_dec_and_test(&grp->refcount))
1943 kfree_rcu(grp, rcu);
1946 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1949 struct numa_group *grp, *my_grp;
1950 struct task_struct *tsk;
1952 int cpu = cpupid_to_cpu(cpupid);
1955 if (unlikely(!p->numa_group)) {
1956 unsigned int size = sizeof(struct numa_group) +
1957 4*nr_node_ids*sizeof(unsigned long);
1959 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1963 atomic_set(&grp->refcount, 1);
1964 spin_lock_init(&grp->lock);
1966 /* Second half of the array tracks nids where faults happen */
1967 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1970 node_set(task_node(current), grp->active_nodes);
1972 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1973 grp->faults[i] = p->numa_faults[i];
1975 grp->total_faults = p->total_numa_faults;
1978 rcu_assign_pointer(p->numa_group, grp);
1982 tsk = READ_ONCE(cpu_rq(cpu)->curr);
1984 if (!cpupid_match_pid(tsk, cpupid))
1987 grp = rcu_dereference(tsk->numa_group);
1991 my_grp = p->numa_group;
1996 * Only join the other group if its bigger; if we're the bigger group,
1997 * the other task will join us.
1999 if (my_grp->nr_tasks > grp->nr_tasks)
2003 * Tie-break on the grp address.
2005 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2008 /* Always join threads in the same process. */
2009 if (tsk->mm == current->mm)
2012 /* Simple filter to avoid false positives due to PID collisions */
2013 if (flags & TNF_SHARED)
2016 /* Update priv based on whether false sharing was detected */
2019 if (join && !get_numa_group(grp))
2027 BUG_ON(irqs_disabled());
2028 double_lock_irq(&my_grp->lock, &grp->lock);
2030 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2031 my_grp->faults[i] -= p->numa_faults[i];
2032 grp->faults[i] += p->numa_faults[i];
2034 my_grp->total_faults -= p->total_numa_faults;
2035 grp->total_faults += p->total_numa_faults;
2040 spin_unlock(&my_grp->lock);
2041 spin_unlock_irq(&grp->lock);
2043 rcu_assign_pointer(p->numa_group, grp);
2045 put_numa_group(my_grp);
2053 void task_numa_free(struct task_struct *p)
2055 struct numa_group *grp = p->numa_group;
2056 void *numa_faults = p->numa_faults;
2057 unsigned long flags;
2061 spin_lock_irqsave(&grp->lock, flags);
2062 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2063 grp->faults[i] -= p->numa_faults[i];
2064 grp->total_faults -= p->total_numa_faults;
2067 spin_unlock_irqrestore(&grp->lock, flags);
2068 RCU_INIT_POINTER(p->numa_group, NULL);
2069 put_numa_group(grp);
2072 p->numa_faults = NULL;
2077 * Got a PROT_NONE fault for a page on @node.
2079 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2081 struct task_struct *p = current;
2082 bool migrated = flags & TNF_MIGRATED;
2083 int cpu_node = task_node(current);
2084 int local = !!(flags & TNF_FAULT_LOCAL);
2087 if (!static_branch_likely(&sched_numa_balancing))
2090 /* for example, ksmd faulting in a user's mm */
2094 /* Allocate buffer to track faults on a per-node basis */
2095 if (unlikely(!p->numa_faults)) {
2096 int size = sizeof(*p->numa_faults) *
2097 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2099 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2100 if (!p->numa_faults)
2103 p->total_numa_faults = 0;
2104 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2108 * First accesses are treated as private, otherwise consider accesses
2109 * to be private if the accessing pid has not changed
2111 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2114 priv = cpupid_match_pid(p, last_cpupid);
2115 if (!priv && !(flags & TNF_NO_GROUP))
2116 task_numa_group(p, last_cpupid, flags, &priv);
2120 * If a workload spans multiple NUMA nodes, a shared fault that
2121 * occurs wholly within the set of nodes that the workload is
2122 * actively using should be counted as local. This allows the
2123 * scan rate to slow down when a workload has settled down.
2125 if (!priv && !local && p->numa_group &&
2126 node_isset(cpu_node, p->numa_group->active_nodes) &&
2127 node_isset(mem_node, p->numa_group->active_nodes))
2130 task_numa_placement(p);
2133 * Retry task to preferred node migration periodically, in case it
2134 * case it previously failed, or the scheduler moved us.
2136 if (time_after(jiffies, p->numa_migrate_retry))
2137 numa_migrate_preferred(p);
2140 p->numa_pages_migrated += pages;
2141 if (flags & TNF_MIGRATE_FAIL)
2142 p->numa_faults_locality[2] += pages;
2144 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2145 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2146 p->numa_faults_locality[local] += pages;
2149 static void reset_ptenuma_scan(struct task_struct *p)
2152 * We only did a read acquisition of the mmap sem, so
2153 * p->mm->numa_scan_seq is written to without exclusive access
2154 * and the update is not guaranteed to be atomic. That's not
2155 * much of an issue though, since this is just used for
2156 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2157 * expensive, to avoid any form of compiler optimizations:
2159 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2160 p->mm->numa_scan_offset = 0;
2164 * The expensive part of numa migration is done from task_work context.
2165 * Triggered from task_tick_numa().
2167 void task_numa_work(struct callback_head *work)
2169 unsigned long migrate, next_scan, now = jiffies;
2170 struct task_struct *p = current;
2171 struct mm_struct *mm = p->mm;
2172 struct vm_area_struct *vma;
2173 unsigned long start, end;
2174 unsigned long nr_pte_updates = 0;
2175 long pages, virtpages;
2177 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2179 work->next = work; /* protect against double add */
2181 * Who cares about NUMA placement when they're dying.
2183 * NOTE: make sure not to dereference p->mm before this check,
2184 * exit_task_work() happens _after_ exit_mm() so we could be called
2185 * without p->mm even though we still had it when we enqueued this
2188 if (p->flags & PF_EXITING)
2191 if (!mm->numa_next_scan) {
2192 mm->numa_next_scan = now +
2193 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2197 * Enforce maximal scan/migration frequency..
2199 migrate = mm->numa_next_scan;
2200 if (time_before(now, migrate))
2203 if (p->numa_scan_period == 0) {
2204 p->numa_scan_period_max = task_scan_max(p);
2205 p->numa_scan_period = task_scan_min(p);
2208 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2209 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2213 * Delay this task enough that another task of this mm will likely win
2214 * the next time around.
2216 p->node_stamp += 2 * TICK_NSEC;
2218 start = mm->numa_scan_offset;
2219 pages = sysctl_numa_balancing_scan_size;
2220 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2221 virtpages = pages * 8; /* Scan up to this much virtual space */
2226 if (!down_read_trylock(&mm->mmap_sem))
2228 vma = find_vma(mm, start);
2230 reset_ptenuma_scan(p);
2234 for (; vma; vma = vma->vm_next) {
2235 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2236 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2241 * Shared library pages mapped by multiple processes are not
2242 * migrated as it is expected they are cache replicated. Avoid
2243 * hinting faults in read-only file-backed mappings or the vdso
2244 * as migrating the pages will be of marginal benefit.
2247 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2251 * Skip inaccessible VMAs to avoid any confusion between
2252 * PROT_NONE and NUMA hinting ptes
2254 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2258 start = max(start, vma->vm_start);
2259 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2260 end = min(end, vma->vm_end);
2261 nr_pte_updates = change_prot_numa(vma, start, end);
2264 * Try to scan sysctl_numa_balancing_size worth of
2265 * hpages that have at least one present PTE that
2266 * is not already pte-numa. If the VMA contains
2267 * areas that are unused or already full of prot_numa
2268 * PTEs, scan up to virtpages, to skip through those
2272 pages -= (end - start) >> PAGE_SHIFT;
2273 virtpages -= (end - start) >> PAGE_SHIFT;
2276 if (pages <= 0 || virtpages <= 0)
2280 } while (end != vma->vm_end);
2285 * It is possible to reach the end of the VMA list but the last few
2286 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2287 * would find the !migratable VMA on the next scan but not reset the
2288 * scanner to the start so check it now.
2291 mm->numa_scan_offset = start;
2293 reset_ptenuma_scan(p);
2294 up_read(&mm->mmap_sem);
2298 * Drive the periodic memory faults..
2300 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2302 struct callback_head *work = &curr->numa_work;
2306 * We don't care about NUMA placement if we don't have memory.
2308 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2312 * Using runtime rather than walltime has the dual advantage that
2313 * we (mostly) drive the selection from busy threads and that the
2314 * task needs to have done some actual work before we bother with
2317 now = curr->se.sum_exec_runtime;
2318 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2320 if (now > curr->node_stamp + period) {
2321 if (!curr->node_stamp)
2322 curr->numa_scan_period = task_scan_min(curr);
2323 curr->node_stamp += period;
2325 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2326 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2327 task_work_add(curr, work, true);
2332 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2336 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2340 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2343 #endif /* CONFIG_NUMA_BALANCING */
2346 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2348 update_load_add(&cfs_rq->load, se->load.weight);
2349 if (!parent_entity(se))
2350 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2352 if (entity_is_task(se)) {
2353 struct rq *rq = rq_of(cfs_rq);
2355 account_numa_enqueue(rq, task_of(se));
2356 list_add(&se->group_node, &rq->cfs_tasks);
2359 cfs_rq->nr_running++;
2363 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2365 update_load_sub(&cfs_rq->load, se->load.weight);
2366 if (!parent_entity(se))
2367 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2368 if (entity_is_task(se)) {
2369 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2370 list_del_init(&se->group_node);
2372 cfs_rq->nr_running--;
2375 #ifdef CONFIG_FAIR_GROUP_SCHED
2377 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2382 * Use this CPU's real-time load instead of the last load contribution
2383 * as the updating of the contribution is delayed, and we will use the
2384 * the real-time load to calc the share. See update_tg_load_avg().
2386 tg_weight = atomic_long_read(&tg->load_avg);
2387 tg_weight -= cfs_rq->tg_load_avg_contrib;
2388 tg_weight += cfs_rq->load.weight;
2393 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2395 long tg_weight, load, shares;
2397 tg_weight = calc_tg_weight(tg, cfs_rq);
2398 load = cfs_rq->load.weight;
2400 shares = (tg->shares * load);
2402 shares /= tg_weight;
2404 if (shares < MIN_SHARES)
2405 shares = MIN_SHARES;
2406 if (shares > tg->shares)
2407 shares = tg->shares;
2411 # else /* CONFIG_SMP */
2412 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2416 # endif /* CONFIG_SMP */
2417 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2418 unsigned long weight)
2421 /* commit outstanding execution time */
2422 if (cfs_rq->curr == se)
2423 update_curr(cfs_rq);
2424 account_entity_dequeue(cfs_rq, se);
2427 update_load_set(&se->load, weight);
2430 account_entity_enqueue(cfs_rq, se);
2433 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2435 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2437 struct task_group *tg;
2438 struct sched_entity *se;
2442 se = tg->se[cpu_of(rq_of(cfs_rq))];
2443 if (!se || throttled_hierarchy(cfs_rq))
2446 if (likely(se->load.weight == tg->shares))
2449 shares = calc_cfs_shares(cfs_rq, tg);
2451 reweight_entity(cfs_rq_of(se), se, shares);
2453 #else /* CONFIG_FAIR_GROUP_SCHED */
2454 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2457 #endif /* CONFIG_FAIR_GROUP_SCHED */
2460 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2461 static const u32 runnable_avg_yN_inv[] = {
2462 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2463 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2464 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2465 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2466 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2467 0x85aac367, 0x82cd8698,
2471 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2472 * over-estimates when re-combining.
2474 static const u32 runnable_avg_yN_sum[] = {
2475 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2476 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2477 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2482 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2484 static __always_inline u64 decay_load(u64 val, u64 n)
2486 unsigned int local_n;
2490 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2493 /* after bounds checking we can collapse to 32-bit */
2497 * As y^PERIOD = 1/2, we can combine
2498 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2499 * With a look-up table which covers y^n (n<PERIOD)
2501 * To achieve constant time decay_load.
2503 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2504 val >>= local_n / LOAD_AVG_PERIOD;
2505 local_n %= LOAD_AVG_PERIOD;
2508 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2513 * For updates fully spanning n periods, the contribution to runnable
2514 * average will be: \Sum 1024*y^n
2516 * We can compute this reasonably efficiently by combining:
2517 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2519 static u32 __compute_runnable_contrib(u64 n)
2523 if (likely(n <= LOAD_AVG_PERIOD))
2524 return runnable_avg_yN_sum[n];
2525 else if (unlikely(n >= LOAD_AVG_MAX_N))
2526 return LOAD_AVG_MAX;
2528 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2530 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2531 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2533 n -= LOAD_AVG_PERIOD;
2534 } while (n > LOAD_AVG_PERIOD);
2536 contrib = decay_load(contrib, n);
2537 return contrib + runnable_avg_yN_sum[n];
2540 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2541 #error "load tracking assumes 2^10 as unit"
2544 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2547 * We can represent the historical contribution to runnable average as the
2548 * coefficients of a geometric series. To do this we sub-divide our runnable
2549 * history into segments of approximately 1ms (1024us); label the segment that
2550 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2552 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2554 * (now) (~1ms ago) (~2ms ago)
2556 * Let u_i denote the fraction of p_i that the entity was runnable.
2558 * We then designate the fractions u_i as our co-efficients, yielding the
2559 * following representation of historical load:
2560 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2562 * We choose y based on the with of a reasonably scheduling period, fixing:
2565 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2566 * approximately half as much as the contribution to load within the last ms
2569 * When a period "rolls over" and we have new u_0`, multiplying the previous
2570 * sum again by y is sufficient to update:
2571 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2572 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2574 static __always_inline int
2575 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2576 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2578 u64 delta, scaled_delta, periods;
2580 unsigned int delta_w, scaled_delta_w, decayed = 0;
2581 unsigned long scale_freq, scale_cpu;
2583 delta = now - sa->last_update_time;
2585 * This should only happen when time goes backwards, which it
2586 * unfortunately does during sched clock init when we swap over to TSC.
2588 if ((s64)delta < 0) {
2589 sa->last_update_time = now;
2594 * Use 1024ns as the unit of measurement since it's a reasonable
2595 * approximation of 1us and fast to compute.
2600 sa->last_update_time = now;
2602 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2603 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2605 /* delta_w is the amount already accumulated against our next period */
2606 delta_w = sa->period_contrib;
2607 if (delta + delta_w >= 1024) {
2610 /* how much left for next period will start over, we don't know yet */
2611 sa->period_contrib = 0;
2614 * Now that we know we're crossing a period boundary, figure
2615 * out how much from delta we need to complete the current
2616 * period and accrue it.
2618 delta_w = 1024 - delta_w;
2619 scaled_delta_w = cap_scale(delta_w, scale_freq);
2621 sa->load_sum += weight * scaled_delta_w;
2623 cfs_rq->runnable_load_sum +=
2624 weight * scaled_delta_w;
2628 sa->util_sum += scaled_delta_w * scale_cpu;
2632 /* Figure out how many additional periods this update spans */
2633 periods = delta / 1024;
2636 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2638 cfs_rq->runnable_load_sum =
2639 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2641 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2643 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2644 contrib = __compute_runnable_contrib(periods);
2645 contrib = cap_scale(contrib, scale_freq);
2647 sa->load_sum += weight * contrib;
2649 cfs_rq->runnable_load_sum += weight * contrib;
2652 sa->util_sum += contrib * scale_cpu;
2655 /* Remainder of delta accrued against u_0` */
2656 scaled_delta = cap_scale(delta, scale_freq);
2658 sa->load_sum += weight * scaled_delta;
2660 cfs_rq->runnable_load_sum += weight * scaled_delta;
2663 sa->util_sum += scaled_delta * scale_cpu;
2665 sa->period_contrib += delta;
2668 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2670 cfs_rq->runnable_load_avg =
2671 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2673 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2679 #ifdef CONFIG_FAIR_GROUP_SCHED
2681 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2682 * and effective_load (which is not done because it is too costly).
2684 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2686 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2688 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2689 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2690 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2694 #else /* CONFIG_FAIR_GROUP_SCHED */
2695 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2696 #endif /* CONFIG_FAIR_GROUP_SCHED */
2698 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2701 * Unsigned subtract and clamp on underflow.
2703 * Explicitly do a load-store to ensure the intermediate value never hits
2704 * memory. This allows lockless observations without ever seeing the negative
2707 #define sub_positive(_ptr, _val) do { \
2708 typeof(_ptr) ptr = (_ptr); \
2709 typeof(*ptr) val = (_val); \
2710 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2714 WRITE_ONCE(*ptr, res); \
2717 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2718 static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
2720 struct sched_avg *sa = &cfs_rq->avg;
2721 int decayed, removed = 0;
2723 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2724 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2725 sub_positive(&sa->load_avg, r);
2726 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
2730 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2731 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2732 sub_positive(&sa->util_avg, r);
2733 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
2736 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2737 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2739 #ifndef CONFIG_64BIT
2741 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2744 return decayed || removed;
2747 /* Update task and its cfs_rq load average */
2748 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2750 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2751 u64 now = cfs_rq_clock_task(cfs_rq);
2752 int cpu = cpu_of(rq_of(cfs_rq));
2755 * Track task load average for carrying it to new CPU after migrated, and
2756 * track group sched_entity load average for task_h_load calc in migration
2758 __update_load_avg(now, cpu, &se->avg,
2759 se->on_rq * scale_load_down(se->load.weight),
2760 cfs_rq->curr == se, NULL);
2762 if (update_cfs_rq_load_avg(now, cfs_rq) && update_tg)
2763 update_tg_load_avg(cfs_rq, 0);
2766 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2768 if (!sched_feat(ATTACH_AGE_LOAD))
2772 * If we got migrated (either between CPUs or between cgroups) we'll
2773 * have aged the average right before clearing @last_update_time.
2775 if (se->avg.last_update_time) {
2776 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2777 &se->avg, 0, 0, NULL);
2780 * XXX: we could have just aged the entire load away if we've been
2781 * absent from the fair class for too long.
2786 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2787 cfs_rq->avg.load_avg += se->avg.load_avg;
2788 cfs_rq->avg.load_sum += se->avg.load_sum;
2789 cfs_rq->avg.util_avg += se->avg.util_avg;
2790 cfs_rq->avg.util_sum += se->avg.util_sum;
2793 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2795 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2796 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2797 cfs_rq->curr == se, NULL);
2799 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2800 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
2801 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
2802 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
2805 /* Add the load generated by se into cfs_rq's load average */
2807 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2809 struct sched_avg *sa = &se->avg;
2810 u64 now = cfs_rq_clock_task(cfs_rq);
2811 int migrated, decayed;
2813 migrated = !sa->last_update_time;
2815 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2816 se->on_rq * scale_load_down(se->load.weight),
2817 cfs_rq->curr == se, NULL);
2820 decayed = update_cfs_rq_load_avg(now, cfs_rq);
2822 cfs_rq->runnable_load_avg += sa->load_avg;
2823 cfs_rq->runnable_load_sum += sa->load_sum;
2826 attach_entity_load_avg(cfs_rq, se);
2828 if (decayed || migrated)
2829 update_tg_load_avg(cfs_rq, 0);
2832 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2834 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2836 update_load_avg(se, 1);
2838 cfs_rq->runnable_load_avg =
2839 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2840 cfs_rq->runnable_load_sum =
2841 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2844 #ifndef CONFIG_64BIT
2845 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2847 u64 last_update_time_copy;
2848 u64 last_update_time;
2851 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2853 last_update_time = cfs_rq->avg.last_update_time;
2854 } while (last_update_time != last_update_time_copy);
2856 return last_update_time;
2859 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2861 return cfs_rq->avg.last_update_time;
2866 * Task first catches up with cfs_rq, and then subtract
2867 * itself from the cfs_rq (task must be off the queue now).
2869 void remove_entity_load_avg(struct sched_entity *se)
2871 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2872 u64 last_update_time;
2875 * Newly created task or never used group entity should not be removed
2876 * from its (source) cfs_rq
2878 if (se->avg.last_update_time == 0)
2881 last_update_time = cfs_rq_last_update_time(cfs_rq);
2883 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
2884 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2885 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2889 * Update the rq's load with the elapsed running time before entering
2890 * idle. if the last scheduled task is not a CFS task, idle_enter will
2891 * be the only way to update the runnable statistic.
2893 void idle_enter_fair(struct rq *this_rq)
2898 * Update the rq's load with the elapsed idle time before a task is
2899 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2900 * be the only way to update the runnable statistic.
2902 void idle_exit_fair(struct rq *this_rq)
2906 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
2908 return cfs_rq->runnable_load_avg;
2911 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
2913 return cfs_rq->avg.load_avg;
2916 static int idle_balance(struct rq *this_rq);
2918 #else /* CONFIG_SMP */
2920 static inline void update_load_avg(struct sched_entity *se, int update_tg) {}
2922 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2924 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2925 static inline void remove_entity_load_avg(struct sched_entity *se) {}
2928 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2930 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2932 static inline int idle_balance(struct rq *rq)
2937 #endif /* CONFIG_SMP */
2939 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2941 #ifdef CONFIG_SCHEDSTATS
2942 struct task_struct *tsk = NULL;
2944 if (entity_is_task(se))
2947 if (se->statistics.sleep_start) {
2948 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2953 if (unlikely(delta > se->statistics.sleep_max))
2954 se->statistics.sleep_max = delta;
2956 se->statistics.sleep_start = 0;
2957 se->statistics.sum_sleep_runtime += delta;
2960 account_scheduler_latency(tsk, delta >> 10, 1);
2961 trace_sched_stat_sleep(tsk, delta);
2964 if (se->statistics.block_start) {
2965 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2970 if (unlikely(delta > se->statistics.block_max))
2971 se->statistics.block_max = delta;
2973 se->statistics.block_start = 0;
2974 se->statistics.sum_sleep_runtime += delta;
2977 if (tsk->in_iowait) {
2978 se->statistics.iowait_sum += delta;
2979 se->statistics.iowait_count++;
2980 trace_sched_stat_iowait(tsk, delta);
2983 trace_sched_stat_blocked(tsk, delta);
2986 * Blocking time is in units of nanosecs, so shift by
2987 * 20 to get a milliseconds-range estimation of the
2988 * amount of time that the task spent sleeping:
2990 if (unlikely(prof_on == SLEEP_PROFILING)) {
2991 profile_hits(SLEEP_PROFILING,
2992 (void *)get_wchan(tsk),
2995 account_scheduler_latency(tsk, delta >> 10, 0);
3001 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3003 #ifdef CONFIG_SCHED_DEBUG
3004 s64 d = se->vruntime - cfs_rq->min_vruntime;
3009 if (d > 3*sysctl_sched_latency)
3010 schedstat_inc(cfs_rq, nr_spread_over);
3015 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3017 u64 vruntime = cfs_rq->min_vruntime;
3020 * The 'current' period is already promised to the current tasks,
3021 * however the extra weight of the new task will slow them down a
3022 * little, place the new task so that it fits in the slot that
3023 * stays open at the end.
3025 if (initial && sched_feat(START_DEBIT))
3026 vruntime += sched_vslice(cfs_rq, se);
3028 /* sleeps up to a single latency don't count. */
3030 unsigned long thresh = sysctl_sched_latency;
3033 * Halve their sleep time's effect, to allow
3034 * for a gentler effect of sleepers:
3036 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3042 /* ensure we never gain time by being placed backwards. */
3043 se->vruntime = max_vruntime(se->vruntime, vruntime);
3046 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3049 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3052 * Update the normalized vruntime before updating min_vruntime
3053 * through calling update_curr().
3055 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3056 se->vruntime += cfs_rq->min_vruntime;
3059 * Update run-time statistics of the 'current'.
3061 update_curr(cfs_rq);
3062 enqueue_entity_load_avg(cfs_rq, se);
3063 account_entity_enqueue(cfs_rq, se);
3064 update_cfs_shares(cfs_rq);
3066 if (flags & ENQUEUE_WAKEUP) {
3067 place_entity(cfs_rq, se, 0);
3068 enqueue_sleeper(cfs_rq, se);
3071 update_stats_enqueue(cfs_rq, se);
3072 check_spread(cfs_rq, se);
3073 if (se != cfs_rq->curr)
3074 __enqueue_entity(cfs_rq, se);
3077 if (cfs_rq->nr_running == 1) {
3078 list_add_leaf_cfs_rq(cfs_rq);
3079 check_enqueue_throttle(cfs_rq);
3083 static void __clear_buddies_last(struct sched_entity *se)
3085 for_each_sched_entity(se) {
3086 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3087 if (cfs_rq->last != se)
3090 cfs_rq->last = NULL;
3094 static void __clear_buddies_next(struct sched_entity *se)
3096 for_each_sched_entity(se) {
3097 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3098 if (cfs_rq->next != se)
3101 cfs_rq->next = NULL;
3105 static void __clear_buddies_skip(struct sched_entity *se)
3107 for_each_sched_entity(se) {
3108 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3109 if (cfs_rq->skip != se)
3112 cfs_rq->skip = NULL;
3116 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3118 if (cfs_rq->last == se)
3119 __clear_buddies_last(se);
3121 if (cfs_rq->next == se)
3122 __clear_buddies_next(se);
3124 if (cfs_rq->skip == se)
3125 __clear_buddies_skip(se);
3128 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3131 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3134 * Update run-time statistics of the 'current'.
3136 update_curr(cfs_rq);
3137 dequeue_entity_load_avg(cfs_rq, se);
3139 update_stats_dequeue(cfs_rq, se);
3140 if (flags & DEQUEUE_SLEEP) {
3141 #ifdef CONFIG_SCHEDSTATS
3142 if (entity_is_task(se)) {
3143 struct task_struct *tsk = task_of(se);
3145 if (tsk->state & TASK_INTERRUPTIBLE)
3146 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3147 if (tsk->state & TASK_UNINTERRUPTIBLE)
3148 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3153 clear_buddies(cfs_rq, se);
3155 if (se != cfs_rq->curr)
3156 __dequeue_entity(cfs_rq, se);
3158 account_entity_dequeue(cfs_rq, se);
3161 * Normalize the entity after updating the min_vruntime because the
3162 * update can refer to the ->curr item and we need to reflect this
3163 * movement in our normalized position.
3165 if (!(flags & DEQUEUE_SLEEP))
3166 se->vruntime -= cfs_rq->min_vruntime;
3168 /* return excess runtime on last dequeue */
3169 return_cfs_rq_runtime(cfs_rq);
3171 update_min_vruntime(cfs_rq);
3172 update_cfs_shares(cfs_rq);
3176 * Preempt the current task with a newly woken task if needed:
3179 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3181 unsigned long ideal_runtime, delta_exec;
3182 struct sched_entity *se;
3185 ideal_runtime = sched_slice(cfs_rq, curr);
3186 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3187 if (delta_exec > ideal_runtime) {
3188 resched_curr(rq_of(cfs_rq));
3190 * The current task ran long enough, ensure it doesn't get
3191 * re-elected due to buddy favours.
3193 clear_buddies(cfs_rq, curr);
3198 * Ensure that a task that missed wakeup preemption by a
3199 * narrow margin doesn't have to wait for a full slice.
3200 * This also mitigates buddy induced latencies under load.
3202 if (delta_exec < sysctl_sched_min_granularity)
3205 se = __pick_first_entity(cfs_rq);
3206 delta = curr->vruntime - se->vruntime;
3211 if (delta > ideal_runtime)
3212 resched_curr(rq_of(cfs_rq));
3216 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3218 /* 'current' is not kept within the tree. */
3221 * Any task has to be enqueued before it get to execute on
3222 * a CPU. So account for the time it spent waiting on the
3225 update_stats_wait_end(cfs_rq, se);
3226 __dequeue_entity(cfs_rq, se);
3227 update_load_avg(se, 1);
3230 update_stats_curr_start(cfs_rq, se);
3232 #ifdef CONFIG_SCHEDSTATS
3234 * Track our maximum slice length, if the CPU's load is at
3235 * least twice that of our own weight (i.e. dont track it
3236 * when there are only lesser-weight tasks around):
3238 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3239 se->statistics.slice_max = max(se->statistics.slice_max,
3240 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3243 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3247 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3250 * Pick the next process, keeping these things in mind, in this order:
3251 * 1) keep things fair between processes/task groups
3252 * 2) pick the "next" process, since someone really wants that to run
3253 * 3) pick the "last" process, for cache locality
3254 * 4) do not run the "skip" process, if something else is available
3256 static struct sched_entity *
3257 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3259 struct sched_entity *left = __pick_first_entity(cfs_rq);
3260 struct sched_entity *se;
3263 * If curr is set we have to see if its left of the leftmost entity
3264 * still in the tree, provided there was anything in the tree at all.
3266 if (!left || (curr && entity_before(curr, left)))
3269 se = left; /* ideally we run the leftmost entity */
3272 * Avoid running the skip buddy, if running something else can
3273 * be done without getting too unfair.
3275 if (cfs_rq->skip == se) {
3276 struct sched_entity *second;
3279 second = __pick_first_entity(cfs_rq);
3281 second = __pick_next_entity(se);
3282 if (!second || (curr && entity_before(curr, second)))
3286 if (second && wakeup_preempt_entity(second, left) < 1)
3291 * Prefer last buddy, try to return the CPU to a preempted task.
3293 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3297 * Someone really wants this to run. If it's not unfair, run it.
3299 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3302 clear_buddies(cfs_rq, se);
3307 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3309 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3312 * If still on the runqueue then deactivate_task()
3313 * was not called and update_curr() has to be done:
3316 update_curr(cfs_rq);
3318 /* throttle cfs_rqs exceeding runtime */
3319 check_cfs_rq_runtime(cfs_rq);
3321 check_spread(cfs_rq, prev);
3323 update_stats_wait_start(cfs_rq, prev);
3324 /* Put 'current' back into the tree. */
3325 __enqueue_entity(cfs_rq, prev);
3326 /* in !on_rq case, update occurred at dequeue */
3327 update_load_avg(prev, 0);
3329 cfs_rq->curr = NULL;
3333 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3336 * Update run-time statistics of the 'current'.
3338 update_curr(cfs_rq);
3341 * Ensure that runnable average is periodically updated.
3343 update_load_avg(curr, 1);
3344 update_cfs_shares(cfs_rq);
3346 #ifdef CONFIG_SCHED_HRTICK
3348 * queued ticks are scheduled to match the slice, so don't bother
3349 * validating it and just reschedule.
3352 resched_curr(rq_of(cfs_rq));
3356 * don't let the period tick interfere with the hrtick preemption
3358 if (!sched_feat(DOUBLE_TICK) &&
3359 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3363 if (cfs_rq->nr_running > 1)
3364 check_preempt_tick(cfs_rq, curr);
3368 /**************************************************
3369 * CFS bandwidth control machinery
3372 #ifdef CONFIG_CFS_BANDWIDTH
3374 #ifdef HAVE_JUMP_LABEL
3375 static struct static_key __cfs_bandwidth_used;
3377 static inline bool cfs_bandwidth_used(void)
3379 return static_key_false(&__cfs_bandwidth_used);
3382 void cfs_bandwidth_usage_inc(void)
3384 static_key_slow_inc(&__cfs_bandwidth_used);
3387 void cfs_bandwidth_usage_dec(void)
3389 static_key_slow_dec(&__cfs_bandwidth_used);
3391 #else /* HAVE_JUMP_LABEL */
3392 static bool cfs_bandwidth_used(void)
3397 void cfs_bandwidth_usage_inc(void) {}
3398 void cfs_bandwidth_usage_dec(void) {}
3399 #endif /* HAVE_JUMP_LABEL */
3402 * default period for cfs group bandwidth.
3403 * default: 0.1s, units: nanoseconds
3405 static inline u64 default_cfs_period(void)
3407 return 100000000ULL;
3410 static inline u64 sched_cfs_bandwidth_slice(void)
3412 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3416 * Replenish runtime according to assigned quota and update expiration time.
3417 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3418 * additional synchronization around rq->lock.
3420 * requires cfs_b->lock
3422 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3426 if (cfs_b->quota == RUNTIME_INF)
3429 now = sched_clock_cpu(smp_processor_id());
3430 cfs_b->runtime = cfs_b->quota;
3431 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3434 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3436 return &tg->cfs_bandwidth;
3439 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3440 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3442 if (unlikely(cfs_rq->throttle_count))
3443 return cfs_rq->throttled_clock_task;
3445 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3448 /* returns 0 on failure to allocate runtime */
3449 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3451 struct task_group *tg = cfs_rq->tg;
3452 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3453 u64 amount = 0, min_amount, expires;
3455 /* note: this is a positive sum as runtime_remaining <= 0 */
3456 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3458 raw_spin_lock(&cfs_b->lock);
3459 if (cfs_b->quota == RUNTIME_INF)
3460 amount = min_amount;
3462 start_cfs_bandwidth(cfs_b);
3464 if (cfs_b->runtime > 0) {
3465 amount = min(cfs_b->runtime, min_amount);
3466 cfs_b->runtime -= amount;
3470 expires = cfs_b->runtime_expires;
3471 raw_spin_unlock(&cfs_b->lock);
3473 cfs_rq->runtime_remaining += amount;
3475 * we may have advanced our local expiration to account for allowed
3476 * spread between our sched_clock and the one on which runtime was
3479 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3480 cfs_rq->runtime_expires = expires;
3482 return cfs_rq->runtime_remaining > 0;
3486 * Note: This depends on the synchronization provided by sched_clock and the
3487 * fact that rq->clock snapshots this value.
3489 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3491 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3493 /* if the deadline is ahead of our clock, nothing to do */
3494 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3497 if (cfs_rq->runtime_remaining < 0)
3501 * If the local deadline has passed we have to consider the
3502 * possibility that our sched_clock is 'fast' and the global deadline
3503 * has not truly expired.
3505 * Fortunately we can check determine whether this the case by checking
3506 * whether the global deadline has advanced. It is valid to compare
3507 * cfs_b->runtime_expires without any locks since we only care about
3508 * exact equality, so a partial write will still work.
3511 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3512 /* extend local deadline, drift is bounded above by 2 ticks */
3513 cfs_rq->runtime_expires += TICK_NSEC;
3515 /* global deadline is ahead, expiration has passed */
3516 cfs_rq->runtime_remaining = 0;
3520 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3522 /* dock delta_exec before expiring quota (as it could span periods) */
3523 cfs_rq->runtime_remaining -= delta_exec;
3524 expire_cfs_rq_runtime(cfs_rq);
3526 if (likely(cfs_rq->runtime_remaining > 0))
3530 * if we're unable to extend our runtime we resched so that the active
3531 * hierarchy can be throttled
3533 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3534 resched_curr(rq_of(cfs_rq));
3537 static __always_inline
3538 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3540 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3543 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3546 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3548 return cfs_bandwidth_used() && cfs_rq->throttled;
3551 /* check whether cfs_rq, or any parent, is throttled */
3552 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3554 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3558 * Ensure that neither of the group entities corresponding to src_cpu or
3559 * dest_cpu are members of a throttled hierarchy when performing group
3560 * load-balance operations.
3562 static inline int throttled_lb_pair(struct task_group *tg,
3563 int src_cpu, int dest_cpu)
3565 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3567 src_cfs_rq = tg->cfs_rq[src_cpu];
3568 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3570 return throttled_hierarchy(src_cfs_rq) ||
3571 throttled_hierarchy(dest_cfs_rq);
3574 /* updated child weight may affect parent so we have to do this bottom up */
3575 static int tg_unthrottle_up(struct task_group *tg, void *data)
3577 struct rq *rq = data;
3578 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3580 cfs_rq->throttle_count--;
3582 if (!cfs_rq->throttle_count) {
3583 /* adjust cfs_rq_clock_task() */
3584 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3585 cfs_rq->throttled_clock_task;
3592 static int tg_throttle_down(struct task_group *tg, void *data)
3594 struct rq *rq = data;
3595 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3597 /* group is entering throttled state, stop time */
3598 if (!cfs_rq->throttle_count)
3599 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3600 cfs_rq->throttle_count++;
3605 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3607 struct rq *rq = rq_of(cfs_rq);
3608 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3609 struct sched_entity *se;
3610 long task_delta, dequeue = 1;
3613 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3615 /* freeze hierarchy runnable averages while throttled */
3617 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3620 task_delta = cfs_rq->h_nr_running;
3621 for_each_sched_entity(se) {
3622 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3623 /* throttled entity or throttle-on-deactivate */
3628 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3629 qcfs_rq->h_nr_running -= task_delta;
3631 if (qcfs_rq->load.weight)
3636 sub_nr_running(rq, task_delta);
3638 cfs_rq->throttled = 1;
3639 cfs_rq->throttled_clock = rq_clock(rq);
3640 raw_spin_lock(&cfs_b->lock);
3641 empty = list_empty(&cfs_b->throttled_cfs_rq);
3644 * Add to the _head_ of the list, so that an already-started
3645 * distribute_cfs_runtime will not see us. If disribute_cfs_runtime is
3646 * not running add to the tail so that later runqueues don't get starved.
3648 if (cfs_b->distribute_running)
3649 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3651 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3654 * If we're the first throttled task, make sure the bandwidth
3658 start_cfs_bandwidth(cfs_b);
3660 raw_spin_unlock(&cfs_b->lock);
3663 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3665 struct rq *rq = rq_of(cfs_rq);
3666 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3667 struct sched_entity *se;
3671 se = cfs_rq->tg->se[cpu_of(rq)];
3673 cfs_rq->throttled = 0;
3675 update_rq_clock(rq);
3677 raw_spin_lock(&cfs_b->lock);
3678 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3679 list_del_rcu(&cfs_rq->throttled_list);
3680 raw_spin_unlock(&cfs_b->lock);
3682 /* update hierarchical throttle state */
3683 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3685 if (!cfs_rq->load.weight)
3688 task_delta = cfs_rq->h_nr_running;
3689 for_each_sched_entity(se) {
3693 cfs_rq = cfs_rq_of(se);
3695 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3696 cfs_rq->h_nr_running += task_delta;
3698 if (cfs_rq_throttled(cfs_rq))
3703 add_nr_running(rq, task_delta);
3705 /* determine whether we need to wake up potentially idle cpu */
3706 if (rq->curr == rq->idle && rq->cfs.nr_running)
3710 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3711 u64 remaining, u64 expires)
3713 struct cfs_rq *cfs_rq;
3715 u64 starting_runtime = remaining;
3718 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3720 struct rq *rq = rq_of(cfs_rq);
3722 raw_spin_lock(&rq->lock);
3723 if (!cfs_rq_throttled(cfs_rq))
3726 runtime = -cfs_rq->runtime_remaining + 1;
3727 if (runtime > remaining)
3728 runtime = remaining;
3729 remaining -= runtime;
3731 cfs_rq->runtime_remaining += runtime;
3732 cfs_rq->runtime_expires = expires;
3734 /* we check whether we're throttled above */
3735 if (cfs_rq->runtime_remaining > 0)
3736 unthrottle_cfs_rq(cfs_rq);
3739 raw_spin_unlock(&rq->lock);
3746 return starting_runtime - remaining;
3750 * Responsible for refilling a task_group's bandwidth and unthrottling its
3751 * cfs_rqs as appropriate. If there has been no activity within the last
3752 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3753 * used to track this state.
3755 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3757 u64 runtime, runtime_expires;
3760 /* no need to continue the timer with no bandwidth constraint */
3761 if (cfs_b->quota == RUNTIME_INF)
3762 goto out_deactivate;
3764 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3765 cfs_b->nr_periods += overrun;
3768 * idle depends on !throttled (for the case of a large deficit), and if
3769 * we're going inactive then everything else can be deferred
3771 if (cfs_b->idle && !throttled)
3772 goto out_deactivate;
3774 __refill_cfs_bandwidth_runtime(cfs_b);
3777 /* mark as potentially idle for the upcoming period */
3782 /* account preceding periods in which throttling occurred */
3783 cfs_b->nr_throttled += overrun;
3785 runtime_expires = cfs_b->runtime_expires;
3788 * This check is repeated as we are holding onto the new bandwidth while
3789 * we unthrottle. This can potentially race with an unthrottled group
3790 * trying to acquire new bandwidth from the global pool. This can result
3791 * in us over-using our runtime if it is all used during this loop, but
3792 * only by limited amounts in that extreme case.
3794 while (throttled && cfs_b->runtime > 0 && !cfs_b->distribute_running) {
3795 runtime = cfs_b->runtime;
3796 cfs_b->distribute_running = 1;
3797 raw_spin_unlock(&cfs_b->lock);
3798 /* we can't nest cfs_b->lock while distributing bandwidth */
3799 runtime = distribute_cfs_runtime(cfs_b, runtime,
3801 raw_spin_lock(&cfs_b->lock);
3803 cfs_b->distribute_running = 0;
3804 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3806 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3810 * While we are ensured activity in the period following an
3811 * unthrottle, this also covers the case in which the new bandwidth is
3812 * insufficient to cover the existing bandwidth deficit. (Forcing the
3813 * timer to remain active while there are any throttled entities.)
3823 /* a cfs_rq won't donate quota below this amount */
3824 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3825 /* minimum remaining period time to redistribute slack quota */
3826 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3827 /* how long we wait to gather additional slack before distributing */
3828 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3831 * Are we near the end of the current quota period?
3833 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3834 * hrtimer base being cleared by hrtimer_start. In the case of
3835 * migrate_hrtimers, base is never cleared, so we are fine.
3837 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3839 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3842 /* if the call-back is running a quota refresh is already occurring */
3843 if (hrtimer_callback_running(refresh_timer))
3846 /* is a quota refresh about to occur? */
3847 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3848 if (remaining < min_expire)
3854 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3856 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3858 /* if there's a quota refresh soon don't bother with slack */
3859 if (runtime_refresh_within(cfs_b, min_left))
3862 hrtimer_start(&cfs_b->slack_timer,
3863 ns_to_ktime(cfs_bandwidth_slack_period),
3867 /* we know any runtime found here is valid as update_curr() precedes return */
3868 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3870 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3871 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3873 if (slack_runtime <= 0)
3876 raw_spin_lock(&cfs_b->lock);
3877 if (cfs_b->quota != RUNTIME_INF &&
3878 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3879 cfs_b->runtime += slack_runtime;
3881 /* we are under rq->lock, defer unthrottling using a timer */
3882 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3883 !list_empty(&cfs_b->throttled_cfs_rq))
3884 start_cfs_slack_bandwidth(cfs_b);
3886 raw_spin_unlock(&cfs_b->lock);
3888 /* even if it's not valid for return we don't want to try again */
3889 cfs_rq->runtime_remaining -= slack_runtime;
3892 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3894 if (!cfs_bandwidth_used())
3897 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3900 __return_cfs_rq_runtime(cfs_rq);
3904 * This is done with a timer (instead of inline with bandwidth return) since
3905 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3907 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3909 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3912 /* confirm we're still not at a refresh boundary */
3913 raw_spin_lock(&cfs_b->lock);
3914 if (cfs_b->distribute_running) {
3915 raw_spin_unlock(&cfs_b->lock);
3919 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3920 raw_spin_unlock(&cfs_b->lock);
3924 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3925 runtime = cfs_b->runtime;
3927 expires = cfs_b->runtime_expires;
3929 cfs_b->distribute_running = 1;
3931 raw_spin_unlock(&cfs_b->lock);
3936 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3938 raw_spin_lock(&cfs_b->lock);
3939 if (expires == cfs_b->runtime_expires)
3940 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3941 cfs_b->distribute_running = 0;
3942 raw_spin_unlock(&cfs_b->lock);
3946 * When a group wakes up we want to make sure that its quota is not already
3947 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3948 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3950 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3952 if (!cfs_bandwidth_used())
3955 /* Synchronize hierarchical throttle counter: */
3956 if (unlikely(!cfs_rq->throttle_uptodate)) {
3957 struct rq *rq = rq_of(cfs_rq);
3958 struct cfs_rq *pcfs_rq;
3959 struct task_group *tg;
3961 cfs_rq->throttle_uptodate = 1;
3963 /* Get closest up-to-date node, because leaves go first: */
3964 for (tg = cfs_rq->tg->parent; tg; tg = tg->parent) {
3965 pcfs_rq = tg->cfs_rq[cpu_of(rq)];
3966 if (pcfs_rq->throttle_uptodate)
3970 cfs_rq->throttle_count = pcfs_rq->throttle_count;
3971 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3975 /* an active group must be handled by the update_curr()->put() path */
3976 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3979 /* ensure the group is not already throttled */
3980 if (cfs_rq_throttled(cfs_rq))
3983 /* update runtime allocation */
3984 account_cfs_rq_runtime(cfs_rq, 0);
3985 if (cfs_rq->runtime_remaining <= 0)
3986 throttle_cfs_rq(cfs_rq);
3989 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3990 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3992 if (!cfs_bandwidth_used())
3995 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3999 * it's possible for a throttled entity to be forced into a running
4000 * state (e.g. set_curr_task), in this case we're finished.
4002 if (cfs_rq_throttled(cfs_rq))
4005 throttle_cfs_rq(cfs_rq);
4009 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4011 struct cfs_bandwidth *cfs_b =
4012 container_of(timer, struct cfs_bandwidth, slack_timer);
4014 do_sched_cfs_slack_timer(cfs_b);
4016 return HRTIMER_NORESTART;
4019 extern const u64 max_cfs_quota_period;
4021 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4023 struct cfs_bandwidth *cfs_b =
4024 container_of(timer, struct cfs_bandwidth, period_timer);
4029 raw_spin_lock(&cfs_b->lock);
4031 overrun = hrtimer_forward_now(timer, cfs_b->period);
4036 u64 new, old = ktime_to_ns(cfs_b->period);
4038 new = (old * 147) / 128; /* ~115% */
4039 new = min(new, max_cfs_quota_period);
4041 cfs_b->period = ns_to_ktime(new);
4043 /* since max is 1s, this is limited to 1e9^2, which fits in u64 */
4044 cfs_b->quota *= new;
4045 cfs_b->quota = div64_u64(cfs_b->quota, old);
4047 pr_warn_ratelimited(
4048 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us %lld, cfs_quota_us = %lld)\n",
4050 div_u64(new, NSEC_PER_USEC),
4051 div_u64(cfs_b->quota, NSEC_PER_USEC));
4053 /* reset count so we don't come right back in here */
4057 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4060 cfs_b->period_active = 0;
4061 raw_spin_unlock(&cfs_b->lock);
4063 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4066 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4068 raw_spin_lock_init(&cfs_b->lock);
4070 cfs_b->quota = RUNTIME_INF;
4071 cfs_b->period = ns_to_ktime(default_cfs_period());
4073 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4074 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4075 cfs_b->period_timer.function = sched_cfs_period_timer;
4076 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4077 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4078 cfs_b->distribute_running = 0;
4081 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4083 cfs_rq->runtime_enabled = 0;
4084 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4087 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4089 lockdep_assert_held(&cfs_b->lock);
4091 if (!cfs_b->period_active) {
4092 cfs_b->period_active = 1;
4093 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4094 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4098 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4100 /* init_cfs_bandwidth() was not called */
4101 if (!cfs_b->throttled_cfs_rq.next)
4104 hrtimer_cancel(&cfs_b->period_timer);
4105 hrtimer_cancel(&cfs_b->slack_timer);
4108 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4110 struct cfs_rq *cfs_rq;
4112 for_each_leaf_cfs_rq(rq, cfs_rq) {
4113 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4115 raw_spin_lock(&cfs_b->lock);
4116 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4117 raw_spin_unlock(&cfs_b->lock);
4121 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4123 struct cfs_rq *cfs_rq;
4125 for_each_leaf_cfs_rq(rq, cfs_rq) {
4126 if (!cfs_rq->runtime_enabled)
4130 * clock_task is not advancing so we just need to make sure
4131 * there's some valid quota amount
4133 cfs_rq->runtime_remaining = 1;
4135 * Offline rq is schedulable till cpu is completely disabled
4136 * in take_cpu_down(), so we prevent new cfs throttling here.
4138 cfs_rq->runtime_enabled = 0;
4140 if (cfs_rq_throttled(cfs_rq))
4141 unthrottle_cfs_rq(cfs_rq);
4145 #else /* CONFIG_CFS_BANDWIDTH */
4146 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4148 return rq_clock_task(rq_of(cfs_rq));
4151 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4152 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4153 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4154 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4156 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4161 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4166 static inline int throttled_lb_pair(struct task_group *tg,
4167 int src_cpu, int dest_cpu)
4172 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4174 #ifdef CONFIG_FAIR_GROUP_SCHED
4175 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4178 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4182 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4183 static inline void update_runtime_enabled(struct rq *rq) {}
4184 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4186 #endif /* CONFIG_CFS_BANDWIDTH */
4188 /**************************************************
4189 * CFS operations on tasks:
4192 #ifdef CONFIG_SCHED_HRTICK
4193 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4195 struct sched_entity *se = &p->se;
4196 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4198 WARN_ON(task_rq(p) != rq);
4200 if (cfs_rq->nr_running > 1) {
4201 u64 slice = sched_slice(cfs_rq, se);
4202 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4203 s64 delta = slice - ran;
4210 hrtick_start(rq, delta);
4215 * called from enqueue/dequeue and updates the hrtick when the
4216 * current task is from our class and nr_running is low enough
4219 static void hrtick_update(struct rq *rq)
4221 struct task_struct *curr = rq->curr;
4223 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4226 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4227 hrtick_start_fair(rq, curr);
4229 #else /* !CONFIG_SCHED_HRTICK */
4231 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4235 static inline void hrtick_update(struct rq *rq)
4241 * The enqueue_task method is called before nr_running is
4242 * increased. Here we update the fair scheduling stats and
4243 * then put the task into the rbtree:
4246 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4248 struct cfs_rq *cfs_rq;
4249 struct sched_entity *se = &p->se;
4251 for_each_sched_entity(se) {
4254 cfs_rq = cfs_rq_of(se);
4255 enqueue_entity(cfs_rq, se, flags);
4258 * end evaluation on encountering a throttled cfs_rq
4260 * note: in the case of encountering a throttled cfs_rq we will
4261 * post the final h_nr_running increment below.
4263 if (cfs_rq_throttled(cfs_rq))
4265 cfs_rq->h_nr_running++;
4267 flags = ENQUEUE_WAKEUP;
4270 for_each_sched_entity(se) {
4271 cfs_rq = cfs_rq_of(se);
4272 cfs_rq->h_nr_running++;
4274 if (cfs_rq_throttled(cfs_rq))
4277 update_load_avg(se, 1);
4278 update_cfs_shares(cfs_rq);
4282 add_nr_running(rq, 1);
4287 static void set_next_buddy(struct sched_entity *se);
4290 * The dequeue_task method is called before nr_running is
4291 * decreased. We remove the task from the rbtree and
4292 * update the fair scheduling stats:
4294 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4296 struct cfs_rq *cfs_rq;
4297 struct sched_entity *se = &p->se;
4298 int task_sleep = flags & DEQUEUE_SLEEP;
4300 for_each_sched_entity(se) {
4301 cfs_rq = cfs_rq_of(se);
4302 dequeue_entity(cfs_rq, se, flags);
4305 * end evaluation on encountering a throttled cfs_rq
4307 * note: in the case of encountering a throttled cfs_rq we will
4308 * post the final h_nr_running decrement below.
4310 if (cfs_rq_throttled(cfs_rq))
4312 cfs_rq->h_nr_running--;
4314 /* Don't dequeue parent if it has other entities besides us */
4315 if (cfs_rq->load.weight) {
4316 /* Avoid re-evaluating load for this entity: */
4317 se = parent_entity(se);
4319 * Bias pick_next to pick a task from this cfs_rq, as
4320 * p is sleeping when it is within its sched_slice.
4322 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4326 flags |= DEQUEUE_SLEEP;
4329 for_each_sched_entity(se) {
4330 cfs_rq = cfs_rq_of(se);
4331 cfs_rq->h_nr_running--;
4333 if (cfs_rq_throttled(cfs_rq))
4336 update_load_avg(se, 1);
4337 update_cfs_shares(cfs_rq);
4341 sub_nr_running(rq, 1);
4349 * per rq 'load' arrray crap; XXX kill this.
4353 * The exact cpuload at various idx values, calculated at every tick would be
4354 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4356 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4357 * on nth tick when cpu may be busy, then we have:
4358 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4359 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4361 * decay_load_missed() below does efficient calculation of
4362 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4363 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4365 * The calculation is approximated on a 128 point scale.
4366 * degrade_zero_ticks is the number of ticks after which load at any
4367 * particular idx is approximated to be zero.
4368 * degrade_factor is a precomputed table, a row for each load idx.
4369 * Each column corresponds to degradation factor for a power of two ticks,
4370 * based on 128 point scale.
4372 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4373 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4375 * With this power of 2 load factors, we can degrade the load n times
4376 * by looking at 1 bits in n and doing as many mult/shift instead of
4377 * n mult/shifts needed by the exact degradation.
4379 #define DEGRADE_SHIFT 7
4380 static const unsigned char
4381 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4382 static const unsigned char
4383 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4384 {0, 0, 0, 0, 0, 0, 0, 0},
4385 {64, 32, 8, 0, 0, 0, 0, 0},
4386 {96, 72, 40, 12, 1, 0, 0},
4387 {112, 98, 75, 43, 15, 1, 0},
4388 {120, 112, 98, 76, 45, 16, 2} };
4391 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4392 * would be when CPU is idle and so we just decay the old load without
4393 * adding any new load.
4395 static unsigned long
4396 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4400 if (!missed_updates)
4403 if (missed_updates >= degrade_zero_ticks[idx])
4407 return load >> missed_updates;
4409 while (missed_updates) {
4410 if (missed_updates % 2)
4411 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4413 missed_updates >>= 1;
4420 * Update rq->cpu_load[] statistics. This function is usually called every
4421 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4422 * every tick. We fix it up based on jiffies.
4424 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4425 unsigned long pending_updates)
4429 this_rq->nr_load_updates++;
4431 /* Update our load: */
4432 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4433 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4434 unsigned long old_load, new_load;
4436 /* scale is effectively 1 << i now, and >> i divides by scale */
4438 old_load = this_rq->cpu_load[i];
4439 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4440 new_load = this_load;
4442 * Round up the averaging division if load is increasing. This
4443 * prevents us from getting stuck on 9 if the load is 10, for
4446 if (new_load > old_load)
4447 new_load += scale - 1;
4449 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4452 sched_avg_update(this_rq);
4455 /* Used instead of source_load when we know the type == 0 */
4456 static unsigned long weighted_cpuload(const int cpu)
4458 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4461 #ifdef CONFIG_NO_HZ_COMMON
4463 * There is no sane way to deal with nohz on smp when using jiffies because the
4464 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4465 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4467 * Therefore we cannot use the delta approach from the regular tick since that
4468 * would seriously skew the load calculation. However we'll make do for those
4469 * updates happening while idle (nohz_idle_balance) or coming out of idle
4470 * (tick_nohz_idle_exit).
4472 * This means we might still be one tick off for nohz periods.
4476 * Called from nohz_idle_balance() to update the load ratings before doing the
4479 static void update_idle_cpu_load(struct rq *this_rq)
4481 unsigned long curr_jiffies = READ_ONCE(jiffies);
4482 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4483 unsigned long pending_updates;
4486 * bail if there's load or we're actually up-to-date.
4488 if (load || curr_jiffies == this_rq->last_load_update_tick)
4491 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4492 this_rq->last_load_update_tick = curr_jiffies;
4494 __update_cpu_load(this_rq, load, pending_updates);
4498 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4500 void update_cpu_load_nohz(void)
4502 struct rq *this_rq = this_rq();
4503 unsigned long curr_jiffies = READ_ONCE(jiffies);
4504 unsigned long pending_updates;
4506 if (curr_jiffies == this_rq->last_load_update_tick)
4509 raw_spin_lock(&this_rq->lock);
4510 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4511 if (pending_updates) {
4512 this_rq->last_load_update_tick = curr_jiffies;
4514 * We were idle, this means load 0, the current load might be
4515 * !0 due to remote wakeups and the sort.
4517 __update_cpu_load(this_rq, 0, pending_updates);
4519 raw_spin_unlock(&this_rq->lock);
4521 #endif /* CONFIG_NO_HZ */
4524 * Called from scheduler_tick()
4526 void update_cpu_load_active(struct rq *this_rq)
4528 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4530 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4532 this_rq->last_load_update_tick = jiffies;
4533 __update_cpu_load(this_rq, load, 1);
4537 * Return a low guess at the load of a migration-source cpu weighted
4538 * according to the scheduling class and "nice" value.
4540 * We want to under-estimate the load of migration sources, to
4541 * balance conservatively.
4543 static unsigned long source_load(int cpu, int type)
4545 struct rq *rq = cpu_rq(cpu);
4546 unsigned long total = weighted_cpuload(cpu);
4548 if (type == 0 || !sched_feat(LB_BIAS))
4551 return min(rq->cpu_load[type-1], total);
4555 * Return a high guess at the load of a migration-target cpu weighted
4556 * according to the scheduling class and "nice" value.
4558 static unsigned long target_load(int cpu, int type)
4560 struct rq *rq = cpu_rq(cpu);
4561 unsigned long total = weighted_cpuload(cpu);
4563 if (type == 0 || !sched_feat(LB_BIAS))
4566 return max(rq->cpu_load[type-1], total);
4569 static unsigned long capacity_of(int cpu)
4571 return cpu_rq(cpu)->cpu_capacity;
4574 static unsigned long capacity_orig_of(int cpu)
4576 return cpu_rq(cpu)->cpu_capacity_orig;
4579 static unsigned long cpu_avg_load_per_task(int cpu)
4581 struct rq *rq = cpu_rq(cpu);
4582 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4583 unsigned long load_avg = weighted_cpuload(cpu);
4586 return load_avg / nr_running;
4591 static void record_wakee(struct task_struct *p)
4594 * Rough decay (wiping) for cost saving, don't worry
4595 * about the boundary, really active task won't care
4598 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4599 current->wakee_flips >>= 1;
4600 current->wakee_flip_decay_ts = jiffies;
4603 if (current->last_wakee != p) {
4604 current->last_wakee = p;
4605 current->wakee_flips++;
4609 static void task_waking_fair(struct task_struct *p)
4611 struct sched_entity *se = &p->se;
4612 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4615 #ifndef CONFIG_64BIT
4616 u64 min_vruntime_copy;
4619 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4621 min_vruntime = cfs_rq->min_vruntime;
4622 } while (min_vruntime != min_vruntime_copy);
4624 min_vruntime = cfs_rq->min_vruntime;
4627 se->vruntime -= min_vruntime;
4631 #ifdef CONFIG_FAIR_GROUP_SCHED
4633 * effective_load() calculates the load change as seen from the root_task_group
4635 * Adding load to a group doesn't make a group heavier, but can cause movement
4636 * of group shares between cpus. Assuming the shares were perfectly aligned one
4637 * can calculate the shift in shares.
4639 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4640 * on this @cpu and results in a total addition (subtraction) of @wg to the
4641 * total group weight.
4643 * Given a runqueue weight distribution (rw_i) we can compute a shares
4644 * distribution (s_i) using:
4646 * s_i = rw_i / \Sum rw_j (1)
4648 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4649 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4650 * shares distribution (s_i):
4652 * rw_i = { 2, 4, 1, 0 }
4653 * s_i = { 2/7, 4/7, 1/7, 0 }
4655 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4656 * task used to run on and the CPU the waker is running on), we need to
4657 * compute the effect of waking a task on either CPU and, in case of a sync
4658 * wakeup, compute the effect of the current task going to sleep.
4660 * So for a change of @wl to the local @cpu with an overall group weight change
4661 * of @wl we can compute the new shares distribution (s'_i) using:
4663 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4665 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4666 * differences in waking a task to CPU 0. The additional task changes the
4667 * weight and shares distributions like:
4669 * rw'_i = { 3, 4, 1, 0 }
4670 * s'_i = { 3/8, 4/8, 1/8, 0 }
4672 * We can then compute the difference in effective weight by using:
4674 * dw_i = S * (s'_i - s_i) (3)
4676 * Where 'S' is the group weight as seen by its parent.
4678 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4679 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4680 * 4/7) times the weight of the group.
4682 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4684 struct sched_entity *se = tg->se[cpu];
4686 if (!tg->parent) /* the trivial, non-cgroup case */
4689 for_each_sched_entity(se) {
4690 struct cfs_rq *cfs_rq = se->my_q;
4691 long W, w = cfs_rq_load_avg(cfs_rq);
4696 * W = @wg + \Sum rw_j
4698 W = wg + atomic_long_read(&tg->load_avg);
4700 /* Ensure \Sum rw_j >= rw_i */
4701 W -= cfs_rq->tg_load_avg_contrib;
4710 * wl = S * s'_i; see (2)
4713 wl = (w * (long)tg->shares) / W;
4718 * Per the above, wl is the new se->load.weight value; since
4719 * those are clipped to [MIN_SHARES, ...) do so now. See
4720 * calc_cfs_shares().
4722 if (wl < MIN_SHARES)
4726 * wl = dw_i = S * (s'_i - s_i); see (3)
4728 wl -= se->avg.load_avg;
4731 * Recursively apply this logic to all parent groups to compute
4732 * the final effective load change on the root group. Since
4733 * only the @tg group gets extra weight, all parent groups can
4734 * only redistribute existing shares. @wl is the shift in shares
4735 * resulting from this level per the above.
4744 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4752 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
4753 * A waker of many should wake a different task than the one last awakened
4754 * at a frequency roughly N times higher than one of its wakees. In order
4755 * to determine whether we should let the load spread vs consolodating to
4756 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
4757 * partner, and a factor of lls_size higher frequency in the other. With
4758 * both conditions met, we can be relatively sure that the relationship is
4759 * non-monogamous, with partner count exceeding socket size. Waker/wakee
4760 * being client/server, worker/dispatcher, interrupt source or whatever is
4761 * irrelevant, spread criteria is apparent partner count exceeds socket size.
4763 static int wake_wide(struct task_struct *p)
4765 unsigned int master = current->wakee_flips;
4766 unsigned int slave = p->wakee_flips;
4767 int factor = this_cpu_read(sd_llc_size);
4770 swap(master, slave);
4771 if (slave < factor || master < slave * factor)
4776 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4778 s64 this_load, load;
4779 s64 this_eff_load, prev_eff_load;
4780 int idx, this_cpu, prev_cpu;
4781 struct task_group *tg;
4782 unsigned long weight;
4786 this_cpu = smp_processor_id();
4787 prev_cpu = task_cpu(p);
4788 load = source_load(prev_cpu, idx);
4789 this_load = target_load(this_cpu, idx);
4792 * If sync wakeup then subtract the (maximum possible)
4793 * effect of the currently running task from the load
4794 * of the current CPU:
4797 tg = task_group(current);
4798 weight = current->se.avg.load_avg;
4800 this_load += effective_load(tg, this_cpu, -weight, -weight);
4801 load += effective_load(tg, prev_cpu, 0, -weight);
4805 weight = p->se.avg.load_avg;
4808 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4809 * due to the sync cause above having dropped this_load to 0, we'll
4810 * always have an imbalance, but there's really nothing you can do
4811 * about that, so that's good too.
4813 * Otherwise check if either cpus are near enough in load to allow this
4814 * task to be woken on this_cpu.
4816 this_eff_load = 100;
4817 this_eff_load *= capacity_of(prev_cpu);
4819 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4820 prev_eff_load *= capacity_of(this_cpu);
4822 if (this_load > 0) {
4823 this_eff_load *= this_load +
4824 effective_load(tg, this_cpu, weight, weight);
4826 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4829 balanced = this_eff_load <= prev_eff_load;
4831 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4836 schedstat_inc(sd, ttwu_move_affine);
4837 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4843 * find_idlest_group finds and returns the least busy CPU group within the
4846 static struct sched_group *
4847 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4848 int this_cpu, int sd_flag)
4850 struct sched_group *idlest = NULL, *group = sd->groups;
4851 unsigned long min_load = ULONG_MAX, this_load = 0;
4852 int load_idx = sd->forkexec_idx;
4853 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4855 if (sd_flag & SD_BALANCE_WAKE)
4856 load_idx = sd->wake_idx;
4859 unsigned long load, avg_load;
4863 /* Skip over this group if it has no CPUs allowed */
4864 if (!cpumask_intersects(sched_group_cpus(group),
4865 tsk_cpus_allowed(p)))
4868 local_group = cpumask_test_cpu(this_cpu,
4869 sched_group_cpus(group));
4871 /* Tally up the load of all CPUs in the group */
4874 for_each_cpu(i, sched_group_cpus(group)) {
4875 /* Bias balancing toward cpus of our domain */
4877 load = source_load(i, load_idx);
4879 load = target_load(i, load_idx);
4884 /* Adjust by relative CPU capacity of the group */
4885 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
4888 this_load = avg_load;
4889 } else if (avg_load < min_load) {
4890 min_load = avg_load;
4893 } while (group = group->next, group != sd->groups);
4895 if (!idlest || 100*this_load < imbalance*min_load)
4901 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4904 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4906 unsigned long load, min_load = ULONG_MAX;
4907 unsigned int min_exit_latency = UINT_MAX;
4908 u64 latest_idle_timestamp = 0;
4909 int least_loaded_cpu = this_cpu;
4910 int shallowest_idle_cpu = -1;
4913 /* Traverse only the allowed CPUs */
4914 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4916 struct rq *rq = cpu_rq(i);
4917 struct cpuidle_state *idle = idle_get_state(rq);
4918 if (idle && idle->exit_latency < min_exit_latency) {
4920 * We give priority to a CPU whose idle state
4921 * has the smallest exit latency irrespective
4922 * of any idle timestamp.
4924 min_exit_latency = idle->exit_latency;
4925 latest_idle_timestamp = rq->idle_stamp;
4926 shallowest_idle_cpu = i;
4927 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
4928 rq->idle_stamp > latest_idle_timestamp) {
4930 * If equal or no active idle state, then
4931 * the most recently idled CPU might have
4934 latest_idle_timestamp = rq->idle_stamp;
4935 shallowest_idle_cpu = i;
4937 } else if (shallowest_idle_cpu == -1) {
4938 load = weighted_cpuload(i);
4939 if (load < min_load || (load == min_load && i == this_cpu)) {
4941 least_loaded_cpu = i;
4946 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
4950 * Try and locate an idle CPU in the sched_domain.
4952 static int select_idle_sibling(struct task_struct *p, int target)
4954 struct sched_domain *sd;
4955 struct sched_group *sg;
4956 int i = task_cpu(p);
4958 if (idle_cpu(target))
4962 * If the prevous cpu is cache affine and idle, don't be stupid.
4964 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4968 * Otherwise, iterate the domains and find an elegible idle cpu.
4970 sd = rcu_dereference(per_cpu(sd_llc, target));
4971 for_each_lower_domain(sd) {
4974 if (!cpumask_intersects(sched_group_cpus(sg),
4975 tsk_cpus_allowed(p)))
4978 for_each_cpu(i, sched_group_cpus(sg)) {
4979 if (i == target || !idle_cpu(i))
4983 target = cpumask_first_and(sched_group_cpus(sg),
4984 tsk_cpus_allowed(p));
4988 } while (sg != sd->groups);
4995 * cpu_util returns the amount of capacity of a CPU that is used by CFS
4996 * tasks. The unit of the return value must be the one of capacity so we can
4997 * compare the utilization with the capacity of the CPU that is available for
4998 * CFS task (ie cpu_capacity).
5000 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5001 * recent utilization of currently non-runnable tasks on a CPU. It represents
5002 * the amount of utilization of a CPU in the range [0..capacity_orig] where
5003 * capacity_orig is the cpu_capacity available at the highest frequency
5004 * (arch_scale_freq_capacity()).
5005 * The utilization of a CPU converges towards a sum equal to or less than the
5006 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5007 * the running time on this CPU scaled by capacity_curr.
5009 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5010 * higher than capacity_orig because of unfortunate rounding in
5011 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5012 * the average stabilizes with the new running time. We need to check that the
5013 * utilization stays within the range of [0..capacity_orig] and cap it if
5014 * necessary. Without utilization capping, a group could be seen as overloaded
5015 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5016 * available capacity. We allow utilization to overshoot capacity_curr (but not
5017 * capacity_orig) as it useful for predicting the capacity required after task
5018 * migrations (scheduler-driven DVFS).
5020 static int cpu_util(int cpu)
5022 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
5023 unsigned long capacity = capacity_orig_of(cpu);
5025 return (util >= capacity) ? capacity : util;
5029 * select_task_rq_fair: Select target runqueue for the waking task in domains
5030 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5031 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5033 * Balances load by selecting the idlest cpu in the idlest group, or under
5034 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5036 * Returns the target cpu number.
5038 * preempt must be disabled.
5041 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5043 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5044 int cpu = smp_processor_id();
5045 int new_cpu = prev_cpu;
5046 int want_affine = 0;
5047 int sync = wake_flags & WF_SYNC;
5049 if (sd_flag & SD_BALANCE_WAKE)
5050 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
5053 for_each_domain(cpu, tmp) {
5054 if (!(tmp->flags & SD_LOAD_BALANCE))
5058 * If both cpu and prev_cpu are part of this domain,
5059 * cpu is a valid SD_WAKE_AFFINE target.
5061 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5062 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5067 if (tmp->flags & sd_flag)
5069 else if (!want_affine)
5074 sd = NULL; /* Prefer wake_affine over balance flags */
5075 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5080 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5081 new_cpu = select_idle_sibling(p, new_cpu);
5084 struct sched_group *group;
5087 if (!(sd->flags & sd_flag)) {
5092 group = find_idlest_group(sd, p, cpu, sd_flag);
5098 new_cpu = find_idlest_cpu(group, p, cpu);
5099 if (new_cpu == -1 || new_cpu == cpu) {
5100 /* Now try balancing at a lower domain level of cpu */
5105 /* Now try balancing at a lower domain level of new_cpu */
5107 weight = sd->span_weight;
5109 for_each_domain(cpu, tmp) {
5110 if (weight <= tmp->span_weight)
5112 if (tmp->flags & sd_flag)
5115 /* while loop will break here if sd == NULL */
5123 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5124 * cfs_rq_of(p) references at time of call are still valid and identify the
5125 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
5126 * other assumptions, including the state of rq->lock, should be made.
5128 static void migrate_task_rq_fair(struct task_struct *p)
5131 * We are supposed to update the task to "current" time, then its up to date
5132 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5133 * what current time is, so simply throw away the out-of-date time. This
5134 * will result in the wakee task is less decayed, but giving the wakee more
5135 * load sounds not bad.
5137 remove_entity_load_avg(&p->se);
5139 /* Tell new CPU we are migrated */
5140 p->se.avg.last_update_time = 0;
5142 /* We have migrated, no longer consider this task hot */
5143 p->se.exec_start = 0;
5146 static void task_dead_fair(struct task_struct *p)
5148 remove_entity_load_avg(&p->se);
5150 #endif /* CONFIG_SMP */
5152 static unsigned long
5153 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5155 unsigned long gran = sysctl_sched_wakeup_granularity;
5158 * Since its curr running now, convert the gran from real-time
5159 * to virtual-time in his units.
5161 * By using 'se' instead of 'curr' we penalize light tasks, so
5162 * they get preempted easier. That is, if 'se' < 'curr' then
5163 * the resulting gran will be larger, therefore penalizing the
5164 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5165 * be smaller, again penalizing the lighter task.
5167 * This is especially important for buddies when the leftmost
5168 * task is higher priority than the buddy.
5170 return calc_delta_fair(gran, se);
5174 * Should 'se' preempt 'curr'.
5188 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5190 s64 gran, vdiff = curr->vruntime - se->vruntime;
5195 gran = wakeup_gran(curr, se);
5202 static void set_last_buddy(struct sched_entity *se)
5204 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5207 for_each_sched_entity(se)
5208 cfs_rq_of(se)->last = se;
5211 static void set_next_buddy(struct sched_entity *se)
5213 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5216 for_each_sched_entity(se)
5217 cfs_rq_of(se)->next = se;
5220 static void set_skip_buddy(struct sched_entity *se)
5222 for_each_sched_entity(se)
5223 cfs_rq_of(se)->skip = se;
5227 * Preempt the current task with a newly woken task if needed:
5229 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5231 struct task_struct *curr = rq->curr;
5232 struct sched_entity *se = &curr->se, *pse = &p->se;
5233 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5234 int scale = cfs_rq->nr_running >= sched_nr_latency;
5235 int next_buddy_marked = 0;
5237 if (unlikely(se == pse))
5241 * This is possible from callers such as attach_tasks(), in which we
5242 * unconditionally check_prempt_curr() after an enqueue (which may have
5243 * lead to a throttle). This both saves work and prevents false
5244 * next-buddy nomination below.
5246 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5249 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5250 set_next_buddy(pse);
5251 next_buddy_marked = 1;
5255 * We can come here with TIF_NEED_RESCHED already set from new task
5258 * Note: this also catches the edge-case of curr being in a throttled
5259 * group (e.g. via set_curr_task), since update_curr() (in the
5260 * enqueue of curr) will have resulted in resched being set. This
5261 * prevents us from potentially nominating it as a false LAST_BUDDY
5264 if (test_tsk_need_resched(curr))
5267 /* Idle tasks are by definition preempted by non-idle tasks. */
5268 if (unlikely(curr->policy == SCHED_IDLE) &&
5269 likely(p->policy != SCHED_IDLE))
5273 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5274 * is driven by the tick):
5276 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5279 find_matching_se(&se, &pse);
5280 update_curr(cfs_rq_of(se));
5282 if (wakeup_preempt_entity(se, pse) == 1) {
5284 * Bias pick_next to pick the sched entity that is
5285 * triggering this preemption.
5287 if (!next_buddy_marked)
5288 set_next_buddy(pse);
5297 * Only set the backward buddy when the current task is still
5298 * on the rq. This can happen when a wakeup gets interleaved
5299 * with schedule on the ->pre_schedule() or idle_balance()
5300 * point, either of which can * drop the rq lock.
5302 * Also, during early boot the idle thread is in the fair class,
5303 * for obvious reasons its a bad idea to schedule back to it.
5305 if (unlikely(!se->on_rq || curr == rq->idle))
5308 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5312 static struct task_struct *
5313 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
5315 struct cfs_rq *cfs_rq = &rq->cfs;
5316 struct sched_entity *se;
5317 struct task_struct *p;
5321 #ifdef CONFIG_FAIR_GROUP_SCHED
5322 if (!cfs_rq->nr_running)
5325 if (prev->sched_class != &fair_sched_class)
5329 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5330 * likely that a next task is from the same cgroup as the current.
5332 * Therefore attempt to avoid putting and setting the entire cgroup
5333 * hierarchy, only change the part that actually changes.
5337 struct sched_entity *curr = cfs_rq->curr;
5340 * Since we got here without doing put_prev_entity() we also
5341 * have to consider cfs_rq->curr. If it is still a runnable
5342 * entity, update_curr() will update its vruntime, otherwise
5343 * forget we've ever seen it.
5347 update_curr(cfs_rq);
5352 * This call to check_cfs_rq_runtime() will do the
5353 * throttle and dequeue its entity in the parent(s).
5354 * Therefore the 'simple' nr_running test will indeed
5357 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5361 se = pick_next_entity(cfs_rq, curr);
5362 cfs_rq = group_cfs_rq(se);
5368 * Since we haven't yet done put_prev_entity and if the selected task
5369 * is a different task than we started out with, try and touch the
5370 * least amount of cfs_rqs.
5373 struct sched_entity *pse = &prev->se;
5375 while (!(cfs_rq = is_same_group(se, pse))) {
5376 int se_depth = se->depth;
5377 int pse_depth = pse->depth;
5379 if (se_depth <= pse_depth) {
5380 put_prev_entity(cfs_rq_of(pse), pse);
5381 pse = parent_entity(pse);
5383 if (se_depth >= pse_depth) {
5384 set_next_entity(cfs_rq_of(se), se);
5385 se = parent_entity(se);
5389 put_prev_entity(cfs_rq, pse);
5390 set_next_entity(cfs_rq, se);
5393 if (hrtick_enabled(rq))
5394 hrtick_start_fair(rq, p);
5401 if (!cfs_rq->nr_running)
5404 put_prev_task(rq, prev);
5407 se = pick_next_entity(cfs_rq, NULL);
5408 set_next_entity(cfs_rq, se);
5409 cfs_rq = group_cfs_rq(se);
5414 if (hrtick_enabled(rq))
5415 hrtick_start_fair(rq, p);
5421 * This is OK, because current is on_cpu, which avoids it being picked
5422 * for load-balance and preemption/IRQs are still disabled avoiding
5423 * further scheduler activity on it and we're being very careful to
5424 * re-start the picking loop.
5426 lockdep_unpin_lock(&rq->lock);
5427 new_tasks = idle_balance(rq);
5428 lockdep_pin_lock(&rq->lock);
5430 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5431 * possible for any higher priority task to appear. In that case we
5432 * must re-start the pick_next_entity() loop.
5444 * Account for a descheduled task:
5446 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5448 struct sched_entity *se = &prev->se;
5449 struct cfs_rq *cfs_rq;
5451 for_each_sched_entity(se) {
5452 cfs_rq = cfs_rq_of(se);
5453 put_prev_entity(cfs_rq, se);
5458 * sched_yield() is very simple
5460 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5462 static void yield_task_fair(struct rq *rq)
5464 struct task_struct *curr = rq->curr;
5465 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5466 struct sched_entity *se = &curr->se;
5469 * Are we the only task in the tree?
5471 if (unlikely(rq->nr_running == 1))
5474 clear_buddies(cfs_rq, se);
5476 if (curr->policy != SCHED_BATCH) {
5477 update_rq_clock(rq);
5479 * Update run-time statistics of the 'current'.
5481 update_curr(cfs_rq);
5483 * Tell update_rq_clock() that we've just updated,
5484 * so we don't do microscopic update in schedule()
5485 * and double the fastpath cost.
5487 rq_clock_skip_update(rq, true);
5493 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5495 struct sched_entity *se = &p->se;
5497 /* throttled hierarchies are not runnable */
5498 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5501 /* Tell the scheduler that we'd really like pse to run next. */
5504 yield_task_fair(rq);
5510 /**************************************************
5511 * Fair scheduling class load-balancing methods.
5515 * The purpose of load-balancing is to achieve the same basic fairness the
5516 * per-cpu scheduler provides, namely provide a proportional amount of compute
5517 * time to each task. This is expressed in the following equation:
5519 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5521 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5522 * W_i,0 is defined as:
5524 * W_i,0 = \Sum_j w_i,j (2)
5526 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5527 * is derived from the nice value as per prio_to_weight[].
5529 * The weight average is an exponential decay average of the instantaneous
5532 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5534 * C_i is the compute capacity of cpu i, typically it is the
5535 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5536 * can also include other factors [XXX].
5538 * To achieve this balance we define a measure of imbalance which follows
5539 * directly from (1):
5541 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5543 * We them move tasks around to minimize the imbalance. In the continuous
5544 * function space it is obvious this converges, in the discrete case we get
5545 * a few fun cases generally called infeasible weight scenarios.
5548 * - infeasible weights;
5549 * - local vs global optima in the discrete case. ]
5554 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5555 * for all i,j solution, we create a tree of cpus that follows the hardware
5556 * topology where each level pairs two lower groups (or better). This results
5557 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5558 * tree to only the first of the previous level and we decrease the frequency
5559 * of load-balance at each level inv. proportional to the number of cpus in
5565 * \Sum { --- * --- * 2^i } = O(n) (5)
5567 * `- size of each group
5568 * | | `- number of cpus doing load-balance
5570 * `- sum over all levels
5572 * Coupled with a limit on how many tasks we can migrate every balance pass,
5573 * this makes (5) the runtime complexity of the balancer.
5575 * An important property here is that each CPU is still (indirectly) connected
5576 * to every other cpu in at most O(log n) steps:
5578 * The adjacency matrix of the resulting graph is given by:
5581 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5584 * And you'll find that:
5586 * A^(log_2 n)_i,j != 0 for all i,j (7)
5588 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5589 * The task movement gives a factor of O(m), giving a convergence complexity
5592 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5597 * In order to avoid CPUs going idle while there's still work to do, new idle
5598 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5599 * tree itself instead of relying on other CPUs to bring it work.
5601 * This adds some complexity to both (5) and (8) but it reduces the total idle
5609 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5612 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5617 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5619 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5621 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5624 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5625 * rewrite all of this once again.]
5628 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5630 enum fbq_type { regular, remote, all };
5632 #define LBF_ALL_PINNED 0x01
5633 #define LBF_NEED_BREAK 0x02
5634 #define LBF_DST_PINNED 0x04
5635 #define LBF_SOME_PINNED 0x08
5638 struct sched_domain *sd;
5646 struct cpumask *dst_grpmask;
5648 enum cpu_idle_type idle;
5650 /* The set of CPUs under consideration for load-balancing */
5651 struct cpumask *cpus;
5656 unsigned int loop_break;
5657 unsigned int loop_max;
5659 enum fbq_type fbq_type;
5660 struct list_head tasks;
5664 * Is this task likely cache-hot:
5666 static int task_hot(struct task_struct *p, struct lb_env *env)
5670 lockdep_assert_held(&env->src_rq->lock);
5672 if (p->sched_class != &fair_sched_class)
5675 if (unlikely(p->policy == SCHED_IDLE))
5679 * Buddy candidates are cache hot:
5681 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
5682 (&p->se == cfs_rq_of(&p->se)->next ||
5683 &p->se == cfs_rq_of(&p->se)->last))
5686 if (sysctl_sched_migration_cost == -1)
5688 if (sysctl_sched_migration_cost == 0)
5691 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
5693 return delta < (s64)sysctl_sched_migration_cost;
5696 #ifdef CONFIG_NUMA_BALANCING
5698 * Returns 1, if task migration degrades locality
5699 * Returns 0, if task migration improves locality i.e migration preferred.
5700 * Returns -1, if task migration is not affected by locality.
5702 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5704 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5705 unsigned long src_faults, dst_faults;
5706 int src_nid, dst_nid;
5708 if (!static_branch_likely(&sched_numa_balancing))
5711 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
5714 src_nid = cpu_to_node(env->src_cpu);
5715 dst_nid = cpu_to_node(env->dst_cpu);
5717 if (src_nid == dst_nid)
5720 /* Migrating away from the preferred node is always bad. */
5721 if (src_nid == p->numa_preferred_nid) {
5722 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
5728 /* Encourage migration to the preferred node. */
5729 if (dst_nid == p->numa_preferred_nid)
5733 src_faults = group_faults(p, src_nid);
5734 dst_faults = group_faults(p, dst_nid);
5736 src_faults = task_faults(p, src_nid);
5737 dst_faults = task_faults(p, dst_nid);
5740 return dst_faults < src_faults;
5744 static inline int migrate_degrades_locality(struct task_struct *p,
5752 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5755 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5759 lockdep_assert_held(&env->src_rq->lock);
5762 * We do not migrate tasks that are:
5763 * 1) throttled_lb_pair, or
5764 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5765 * 3) running (obviously), or
5766 * 4) are cache-hot on their current CPU.
5768 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5771 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5774 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5776 env->flags |= LBF_SOME_PINNED;
5779 * Remember if this task can be migrated to any other cpu in
5780 * our sched_group. We may want to revisit it if we couldn't
5781 * meet load balance goals by pulling other tasks on src_cpu.
5783 * Also avoid computing new_dst_cpu if we have already computed
5784 * one in current iteration.
5786 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5789 /* Prevent to re-select dst_cpu via env's cpus */
5790 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5791 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5792 env->flags |= LBF_DST_PINNED;
5793 env->new_dst_cpu = cpu;
5801 /* Record that we found atleast one task that could run on dst_cpu */
5802 env->flags &= ~LBF_ALL_PINNED;
5804 if (task_running(env->src_rq, p)) {
5805 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5810 * Aggressive migration if:
5811 * 1) destination numa is preferred
5812 * 2) task is cache cold, or
5813 * 3) too many balance attempts have failed.
5815 tsk_cache_hot = migrate_degrades_locality(p, env);
5816 if (tsk_cache_hot == -1)
5817 tsk_cache_hot = task_hot(p, env);
5819 if (tsk_cache_hot <= 0 ||
5820 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5821 if (tsk_cache_hot == 1) {
5822 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5823 schedstat_inc(p, se.statistics.nr_forced_migrations);
5828 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5833 * detach_task() -- detach the task for the migration specified in env
5835 static void detach_task(struct task_struct *p, struct lb_env *env)
5837 lockdep_assert_held(&env->src_rq->lock);
5839 deactivate_task(env->src_rq, p, 0);
5840 p->on_rq = TASK_ON_RQ_MIGRATING;
5841 set_task_cpu(p, env->dst_cpu);
5845 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5846 * part of active balancing operations within "domain".
5848 * Returns a task if successful and NULL otherwise.
5850 static struct task_struct *detach_one_task(struct lb_env *env)
5852 struct task_struct *p, *n;
5854 lockdep_assert_held(&env->src_rq->lock);
5856 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5857 if (!can_migrate_task(p, env))
5860 detach_task(p, env);
5863 * Right now, this is only the second place where
5864 * lb_gained[env->idle] is updated (other is detach_tasks)
5865 * so we can safely collect stats here rather than
5866 * inside detach_tasks().
5868 schedstat_inc(env->sd, lb_gained[env->idle]);
5874 static const unsigned int sched_nr_migrate_break = 32;
5877 * detach_tasks() -- tries to detach up to imbalance weighted load from
5878 * busiest_rq, as part of a balancing operation within domain "sd".
5880 * Returns number of detached tasks if successful and 0 otherwise.
5882 static int detach_tasks(struct lb_env *env)
5884 struct list_head *tasks = &env->src_rq->cfs_tasks;
5885 struct task_struct *p;
5889 lockdep_assert_held(&env->src_rq->lock);
5891 if (env->imbalance <= 0)
5894 while (!list_empty(tasks)) {
5896 * We don't want to steal all, otherwise we may be treated likewise,
5897 * which could at worst lead to a livelock crash.
5899 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
5902 p = list_first_entry(tasks, struct task_struct, se.group_node);
5905 /* We've more or less seen every task there is, call it quits */
5906 if (env->loop > env->loop_max)
5909 /* take a breather every nr_migrate tasks */
5910 if (env->loop > env->loop_break) {
5911 env->loop_break += sched_nr_migrate_break;
5912 env->flags |= LBF_NEED_BREAK;
5916 if (!can_migrate_task(p, env))
5919 load = task_h_load(p);
5921 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5924 if ((load / 2) > env->imbalance)
5927 detach_task(p, env);
5928 list_add(&p->se.group_node, &env->tasks);
5931 env->imbalance -= load;
5933 #ifdef CONFIG_PREEMPT
5935 * NEWIDLE balancing is a source of latency, so preemptible
5936 * kernels will stop after the first task is detached to minimize
5937 * the critical section.
5939 if (env->idle == CPU_NEWLY_IDLE)
5944 * We only want to steal up to the prescribed amount of
5947 if (env->imbalance <= 0)
5952 list_move_tail(&p->se.group_node, tasks);
5956 * Right now, this is one of only two places we collect this stat
5957 * so we can safely collect detach_one_task() stats here rather
5958 * than inside detach_one_task().
5960 schedstat_add(env->sd, lb_gained[env->idle], detached);
5966 * attach_task() -- attach the task detached by detach_task() to its new rq.
5968 static void attach_task(struct rq *rq, struct task_struct *p)
5970 lockdep_assert_held(&rq->lock);
5972 BUG_ON(task_rq(p) != rq);
5973 p->on_rq = TASK_ON_RQ_QUEUED;
5974 activate_task(rq, p, 0);
5975 check_preempt_curr(rq, p, 0);
5979 * attach_one_task() -- attaches the task returned from detach_one_task() to
5982 static void attach_one_task(struct rq *rq, struct task_struct *p)
5984 raw_spin_lock(&rq->lock);
5986 raw_spin_unlock(&rq->lock);
5990 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
5993 static void attach_tasks(struct lb_env *env)
5995 struct list_head *tasks = &env->tasks;
5996 struct task_struct *p;
5998 raw_spin_lock(&env->dst_rq->lock);
6000 while (!list_empty(tasks)) {
6001 p = list_first_entry(tasks, struct task_struct, se.group_node);
6002 list_del_init(&p->se.group_node);
6004 attach_task(env->dst_rq, p);
6007 raw_spin_unlock(&env->dst_rq->lock);
6010 #ifdef CONFIG_FAIR_GROUP_SCHED
6011 static void update_blocked_averages(int cpu)
6013 struct rq *rq = cpu_rq(cpu);
6014 struct cfs_rq *cfs_rq;
6015 unsigned long flags;
6017 raw_spin_lock_irqsave(&rq->lock, flags);
6018 update_rq_clock(rq);
6021 * Iterates the task_group tree in a bottom up fashion, see
6022 * list_add_leaf_cfs_rq() for details.
6024 for_each_leaf_cfs_rq(rq, cfs_rq) {
6025 /* throttled entities do not contribute to load */
6026 if (throttled_hierarchy(cfs_rq))
6029 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
6030 update_tg_load_avg(cfs_rq, 0);
6032 raw_spin_unlock_irqrestore(&rq->lock, flags);
6036 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6037 * This needs to be done in a top-down fashion because the load of a child
6038 * group is a fraction of its parents load.
6040 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6042 struct rq *rq = rq_of(cfs_rq);
6043 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6044 unsigned long now = jiffies;
6047 if (cfs_rq->last_h_load_update == now)
6050 WRITE_ONCE(cfs_rq->h_load_next, NULL);
6051 for_each_sched_entity(se) {
6052 cfs_rq = cfs_rq_of(se);
6053 WRITE_ONCE(cfs_rq->h_load_next, se);
6054 if (cfs_rq->last_h_load_update == now)
6059 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6060 cfs_rq->last_h_load_update = now;
6063 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
6064 load = cfs_rq->h_load;
6065 load = div64_ul(load * se->avg.load_avg,
6066 cfs_rq_load_avg(cfs_rq) + 1);
6067 cfs_rq = group_cfs_rq(se);
6068 cfs_rq->h_load = load;
6069 cfs_rq->last_h_load_update = now;
6073 static unsigned long task_h_load(struct task_struct *p)
6075 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6077 update_cfs_rq_h_load(cfs_rq);
6078 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6079 cfs_rq_load_avg(cfs_rq) + 1);
6082 static inline void update_blocked_averages(int cpu)
6084 struct rq *rq = cpu_rq(cpu);
6085 struct cfs_rq *cfs_rq = &rq->cfs;
6086 unsigned long flags;
6088 raw_spin_lock_irqsave(&rq->lock, flags);
6089 update_rq_clock(rq);
6090 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
6091 raw_spin_unlock_irqrestore(&rq->lock, flags);
6094 static unsigned long task_h_load(struct task_struct *p)
6096 return p->se.avg.load_avg;
6100 /********** Helpers for find_busiest_group ************************/
6109 * sg_lb_stats - stats of a sched_group required for load_balancing
6111 struct sg_lb_stats {
6112 unsigned long avg_load; /*Avg load across the CPUs of the group */
6113 unsigned long group_load; /* Total load over the CPUs of the group */
6114 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6115 unsigned long load_per_task;
6116 unsigned long group_capacity;
6117 unsigned long group_util; /* Total utilization of the group */
6118 unsigned int sum_nr_running; /* Nr tasks running in the group */
6119 unsigned int idle_cpus;
6120 unsigned int group_weight;
6121 enum group_type group_type;
6122 int group_no_capacity;
6123 #ifdef CONFIG_NUMA_BALANCING
6124 unsigned int nr_numa_running;
6125 unsigned int nr_preferred_running;
6130 * sd_lb_stats - Structure to store the statistics of a sched_domain
6131 * during load balancing.
6133 struct sd_lb_stats {
6134 struct sched_group *busiest; /* Busiest group in this sd */
6135 struct sched_group *local; /* Local group in this sd */
6136 unsigned long total_load; /* Total load of all groups in sd */
6137 unsigned long total_capacity; /* Total capacity of all groups in sd */
6138 unsigned long avg_load; /* Average load across all groups in sd */
6140 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6141 struct sg_lb_stats local_stat; /* Statistics of the local group */
6144 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6147 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6148 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6149 * We must however clear busiest_stat::avg_load because
6150 * update_sd_pick_busiest() reads this before assignment.
6152 *sds = (struct sd_lb_stats){
6156 .total_capacity = 0UL,
6159 .sum_nr_running = 0,
6160 .group_type = group_other,
6166 * get_sd_load_idx - Obtain the load index for a given sched domain.
6167 * @sd: The sched_domain whose load_idx is to be obtained.
6168 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6170 * Return: The load index.
6172 static inline int get_sd_load_idx(struct sched_domain *sd,
6173 enum cpu_idle_type idle)
6179 load_idx = sd->busy_idx;
6182 case CPU_NEWLY_IDLE:
6183 load_idx = sd->newidle_idx;
6186 load_idx = sd->idle_idx;
6193 static unsigned long scale_rt_capacity(int cpu)
6195 struct rq *rq = cpu_rq(cpu);
6196 u64 total, used, age_stamp, avg;
6200 * Since we're reading these variables without serialization make sure
6201 * we read them once before doing sanity checks on them.
6203 age_stamp = READ_ONCE(rq->age_stamp);
6204 avg = READ_ONCE(rq->rt_avg);
6205 delta = __rq_clock_broken(rq) - age_stamp;
6207 if (unlikely(delta < 0))
6210 total = sched_avg_period() + delta;
6212 used = div_u64(avg, total);
6214 if (likely(used < SCHED_CAPACITY_SCALE))
6215 return SCHED_CAPACITY_SCALE - used;
6220 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6222 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6223 struct sched_group *sdg = sd->groups;
6225 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6227 capacity *= scale_rt_capacity(cpu);
6228 capacity >>= SCHED_CAPACITY_SHIFT;
6233 cpu_rq(cpu)->cpu_capacity = capacity;
6234 sdg->sgc->capacity = capacity;
6237 void update_group_capacity(struct sched_domain *sd, int cpu)
6239 struct sched_domain *child = sd->child;
6240 struct sched_group *group, *sdg = sd->groups;
6241 unsigned long capacity;
6242 unsigned long interval;
6244 interval = msecs_to_jiffies(sd->balance_interval);
6245 interval = clamp(interval, 1UL, max_load_balance_interval);
6246 sdg->sgc->next_update = jiffies + interval;
6249 update_cpu_capacity(sd, cpu);
6255 if (child->flags & SD_OVERLAP) {
6257 * SD_OVERLAP domains cannot assume that child groups
6258 * span the current group.
6261 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6262 struct sched_group_capacity *sgc;
6263 struct rq *rq = cpu_rq(cpu);
6266 * build_sched_domains() -> init_sched_groups_capacity()
6267 * gets here before we've attached the domains to the
6270 * Use capacity_of(), which is set irrespective of domains
6271 * in update_cpu_capacity().
6273 * This avoids capacity from being 0 and
6274 * causing divide-by-zero issues on boot.
6276 if (unlikely(!rq->sd)) {
6277 capacity += capacity_of(cpu);
6281 sgc = rq->sd->groups->sgc;
6282 capacity += sgc->capacity;
6286 * !SD_OVERLAP domains can assume that child groups
6287 * span the current group.
6290 group = child->groups;
6292 capacity += group->sgc->capacity;
6293 group = group->next;
6294 } while (group != child->groups);
6297 sdg->sgc->capacity = capacity;
6301 * Check whether the capacity of the rq has been noticeably reduced by side
6302 * activity. The imbalance_pct is used for the threshold.
6303 * Return true is the capacity is reduced
6306 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6308 return ((rq->cpu_capacity * sd->imbalance_pct) <
6309 (rq->cpu_capacity_orig * 100));
6313 * Group imbalance indicates (and tries to solve) the problem where balancing
6314 * groups is inadequate due to tsk_cpus_allowed() constraints.
6316 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6317 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6320 * { 0 1 2 3 } { 4 5 6 7 }
6323 * If we were to balance group-wise we'd place two tasks in the first group and
6324 * two tasks in the second group. Clearly this is undesired as it will overload
6325 * cpu 3 and leave one of the cpus in the second group unused.
6327 * The current solution to this issue is detecting the skew in the first group
6328 * by noticing the lower domain failed to reach balance and had difficulty
6329 * moving tasks due to affinity constraints.
6331 * When this is so detected; this group becomes a candidate for busiest; see
6332 * update_sd_pick_busiest(). And calculate_imbalance() and
6333 * find_busiest_group() avoid some of the usual balance conditions to allow it
6334 * to create an effective group imbalance.
6336 * This is a somewhat tricky proposition since the next run might not find the
6337 * group imbalance and decide the groups need to be balanced again. A most
6338 * subtle and fragile situation.
6341 static inline int sg_imbalanced(struct sched_group *group)
6343 return group->sgc->imbalance;
6347 * group_has_capacity returns true if the group has spare capacity that could
6348 * be used by some tasks.
6349 * We consider that a group has spare capacity if the * number of task is
6350 * smaller than the number of CPUs or if the utilization is lower than the
6351 * available capacity for CFS tasks.
6352 * For the latter, we use a threshold to stabilize the state, to take into
6353 * account the variance of the tasks' load and to return true if the available
6354 * capacity in meaningful for the load balancer.
6355 * As an example, an available capacity of 1% can appear but it doesn't make
6356 * any benefit for the load balance.
6359 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6361 if (sgs->sum_nr_running < sgs->group_weight)
6364 if ((sgs->group_capacity * 100) >
6365 (sgs->group_util * env->sd->imbalance_pct))
6372 * group_is_overloaded returns true if the group has more tasks than it can
6374 * group_is_overloaded is not equals to !group_has_capacity because a group
6375 * with the exact right number of tasks, has no more spare capacity but is not
6376 * overloaded so both group_has_capacity and group_is_overloaded return
6380 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6382 if (sgs->sum_nr_running <= sgs->group_weight)
6385 if ((sgs->group_capacity * 100) <
6386 (sgs->group_util * env->sd->imbalance_pct))
6393 group_type group_classify(struct sched_group *group,
6394 struct sg_lb_stats *sgs)
6396 if (sgs->group_no_capacity)
6397 return group_overloaded;
6399 if (sg_imbalanced(group))
6400 return group_imbalanced;
6406 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6407 * @env: The load balancing environment.
6408 * @group: sched_group whose statistics are to be updated.
6409 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6410 * @local_group: Does group contain this_cpu.
6411 * @sgs: variable to hold the statistics for this group.
6412 * @overload: Indicate more than one runnable task for any CPU.
6414 static inline void update_sg_lb_stats(struct lb_env *env,
6415 struct sched_group *group, int load_idx,
6416 int local_group, struct sg_lb_stats *sgs,
6422 memset(sgs, 0, sizeof(*sgs));
6424 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6425 struct rq *rq = cpu_rq(i);
6427 /* Bias balancing toward cpus of our domain */
6429 load = target_load(i, load_idx);
6431 load = source_load(i, load_idx);
6433 sgs->group_load += load;
6434 sgs->group_util += cpu_util(i);
6435 sgs->sum_nr_running += rq->cfs.h_nr_running;
6437 if (rq->nr_running > 1)
6440 #ifdef CONFIG_NUMA_BALANCING
6441 sgs->nr_numa_running += rq->nr_numa_running;
6442 sgs->nr_preferred_running += rq->nr_preferred_running;
6444 sgs->sum_weighted_load += weighted_cpuload(i);
6449 /* Adjust by relative CPU capacity of the group */
6450 sgs->group_capacity = group->sgc->capacity;
6451 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6453 if (sgs->sum_nr_running)
6454 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6456 sgs->group_weight = group->group_weight;
6458 sgs->group_no_capacity = group_is_overloaded(env, sgs);
6459 sgs->group_type = group_classify(group, sgs);
6463 * update_sd_pick_busiest - return 1 on busiest group
6464 * @env: The load balancing environment.
6465 * @sds: sched_domain statistics
6466 * @sg: sched_group candidate to be checked for being the busiest
6467 * @sgs: sched_group statistics
6469 * Determine if @sg is a busier group than the previously selected
6472 * Return: %true if @sg is a busier group than the previously selected
6473 * busiest group. %false otherwise.
6475 static bool update_sd_pick_busiest(struct lb_env *env,
6476 struct sd_lb_stats *sds,
6477 struct sched_group *sg,
6478 struct sg_lb_stats *sgs)
6480 struct sg_lb_stats *busiest = &sds->busiest_stat;
6482 if (sgs->group_type > busiest->group_type)
6485 if (sgs->group_type < busiest->group_type)
6488 if (sgs->avg_load <= busiest->avg_load)
6491 /* This is the busiest node in its class. */
6492 if (!(env->sd->flags & SD_ASYM_PACKING))
6496 * ASYM_PACKING needs to move all the work to the lowest
6497 * numbered CPUs in the group, therefore mark all groups
6498 * higher than ourself as busy.
6500 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6504 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
6511 #ifdef CONFIG_NUMA_BALANCING
6512 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6514 if (sgs->sum_nr_running > sgs->nr_numa_running)
6516 if (sgs->sum_nr_running > sgs->nr_preferred_running)
6521 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6523 if (rq->nr_running > rq->nr_numa_running)
6525 if (rq->nr_running > rq->nr_preferred_running)
6530 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6535 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6539 #endif /* CONFIG_NUMA_BALANCING */
6542 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6543 * @env: The load balancing environment.
6544 * @sds: variable to hold the statistics for this sched_domain.
6546 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6548 struct sched_domain *child = env->sd->child;
6549 struct sched_group *sg = env->sd->groups;
6550 struct sg_lb_stats tmp_sgs;
6551 int load_idx, prefer_sibling = 0;
6552 bool overload = false;
6554 if (child && child->flags & SD_PREFER_SIBLING)
6557 load_idx = get_sd_load_idx(env->sd, env->idle);
6560 struct sg_lb_stats *sgs = &tmp_sgs;
6563 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6566 sgs = &sds->local_stat;
6568 if (env->idle != CPU_NEWLY_IDLE ||
6569 time_after_eq(jiffies, sg->sgc->next_update))
6570 update_group_capacity(env->sd, env->dst_cpu);
6573 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6580 * In case the child domain prefers tasks go to siblings
6581 * first, lower the sg capacity so that we'll try
6582 * and move all the excess tasks away. We lower the capacity
6583 * of a group only if the local group has the capacity to fit
6584 * these excess tasks. The extra check prevents the case where
6585 * you always pull from the heaviest group when it is already
6586 * under-utilized (possible with a large weight task outweighs
6587 * the tasks on the system).
6589 if (prefer_sibling && sds->local &&
6590 group_has_capacity(env, &sds->local_stat) &&
6591 (sgs->sum_nr_running > 1)) {
6592 sgs->group_no_capacity = 1;
6593 sgs->group_type = group_classify(sg, sgs);
6596 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6598 sds->busiest_stat = *sgs;
6602 /* Now, start updating sd_lb_stats */
6603 sds->total_load += sgs->group_load;
6604 sds->total_capacity += sgs->group_capacity;
6607 } while (sg != env->sd->groups);
6609 if (env->sd->flags & SD_NUMA)
6610 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6612 if (!env->sd->parent) {
6613 /* update overload indicator if we are at root domain */
6614 if (env->dst_rq->rd->overload != overload)
6615 env->dst_rq->rd->overload = overload;
6621 * check_asym_packing - Check to see if the group is packed into the
6624 * This is primarily intended to used at the sibling level. Some
6625 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6626 * case of POWER7, it can move to lower SMT modes only when higher
6627 * threads are idle. When in lower SMT modes, the threads will
6628 * perform better since they share less core resources. Hence when we
6629 * have idle threads, we want them to be the higher ones.
6631 * This packing function is run on idle threads. It checks to see if
6632 * the busiest CPU in this domain (core in the P7 case) has a higher
6633 * CPU number than the packing function is being run on. Here we are
6634 * assuming lower CPU number will be equivalent to lower a SMT thread
6637 * Return: 1 when packing is required and a task should be moved to
6638 * this CPU. The amount of the imbalance is returned in *imbalance.
6640 * @env: The load balancing environment.
6641 * @sds: Statistics of the sched_domain which is to be packed
6643 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6647 if (!(env->sd->flags & SD_ASYM_PACKING))
6653 busiest_cpu = group_first_cpu(sds->busiest);
6654 if (env->dst_cpu > busiest_cpu)
6657 env->imbalance = DIV_ROUND_CLOSEST(
6658 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6659 SCHED_CAPACITY_SCALE);
6665 * fix_small_imbalance - Calculate the minor imbalance that exists
6666 * amongst the groups of a sched_domain, during
6668 * @env: The load balancing environment.
6669 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6672 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6674 unsigned long tmp, capa_now = 0, capa_move = 0;
6675 unsigned int imbn = 2;
6676 unsigned long scaled_busy_load_per_task;
6677 struct sg_lb_stats *local, *busiest;
6679 local = &sds->local_stat;
6680 busiest = &sds->busiest_stat;
6682 if (!local->sum_nr_running)
6683 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6684 else if (busiest->load_per_task > local->load_per_task)
6687 scaled_busy_load_per_task =
6688 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6689 busiest->group_capacity;
6691 if (busiest->avg_load + scaled_busy_load_per_task >=
6692 local->avg_load + (scaled_busy_load_per_task * imbn)) {
6693 env->imbalance = busiest->load_per_task;
6698 * OK, we don't have enough imbalance to justify moving tasks,
6699 * however we may be able to increase total CPU capacity used by
6703 capa_now += busiest->group_capacity *
6704 min(busiest->load_per_task, busiest->avg_load);
6705 capa_now += local->group_capacity *
6706 min(local->load_per_task, local->avg_load);
6707 capa_now /= SCHED_CAPACITY_SCALE;
6709 /* Amount of load we'd subtract */
6710 if (busiest->avg_load > scaled_busy_load_per_task) {
6711 capa_move += busiest->group_capacity *
6712 min(busiest->load_per_task,
6713 busiest->avg_load - scaled_busy_load_per_task);
6716 /* Amount of load we'd add */
6717 if (busiest->avg_load * busiest->group_capacity <
6718 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
6719 tmp = (busiest->avg_load * busiest->group_capacity) /
6720 local->group_capacity;
6722 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6723 local->group_capacity;
6725 capa_move += local->group_capacity *
6726 min(local->load_per_task, local->avg_load + tmp);
6727 capa_move /= SCHED_CAPACITY_SCALE;
6729 /* Move if we gain throughput */
6730 if (capa_move > capa_now)
6731 env->imbalance = busiest->load_per_task;
6735 * calculate_imbalance - Calculate the amount of imbalance present within the
6736 * groups of a given sched_domain during load balance.
6737 * @env: load balance environment
6738 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6740 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6742 unsigned long max_pull, load_above_capacity = ~0UL;
6743 struct sg_lb_stats *local, *busiest;
6745 local = &sds->local_stat;
6746 busiest = &sds->busiest_stat;
6748 if (busiest->group_type == group_imbalanced) {
6750 * In the group_imb case we cannot rely on group-wide averages
6751 * to ensure cpu-load equilibrium, look at wider averages. XXX
6753 busiest->load_per_task =
6754 min(busiest->load_per_task, sds->avg_load);
6758 * In the presence of smp nice balancing, certain scenarios can have
6759 * max load less than avg load(as we skip the groups at or below
6760 * its cpu_capacity, while calculating max_load..)
6762 if (busiest->avg_load <= sds->avg_load ||
6763 local->avg_load >= sds->avg_load) {
6765 return fix_small_imbalance(env, sds);
6769 * If there aren't any idle cpus, avoid creating some.
6771 if (busiest->group_type == group_overloaded &&
6772 local->group_type == group_overloaded) {
6773 load_above_capacity = busiest->sum_nr_running *
6775 if (load_above_capacity > busiest->group_capacity)
6776 load_above_capacity -= busiest->group_capacity;
6778 load_above_capacity = ~0UL;
6782 * We're trying to get all the cpus to the average_load, so we don't
6783 * want to push ourselves above the average load, nor do we wish to
6784 * reduce the max loaded cpu below the average load. At the same time,
6785 * we also don't want to reduce the group load below the group capacity
6786 * (so that we can implement power-savings policies etc). Thus we look
6787 * for the minimum possible imbalance.
6789 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6791 /* How much load to actually move to equalise the imbalance */
6792 env->imbalance = min(
6793 max_pull * busiest->group_capacity,
6794 (sds->avg_load - local->avg_load) * local->group_capacity
6795 ) / SCHED_CAPACITY_SCALE;
6798 * if *imbalance is less than the average load per runnable task
6799 * there is no guarantee that any tasks will be moved so we'll have
6800 * a think about bumping its value to force at least one task to be
6803 if (env->imbalance < busiest->load_per_task)
6804 return fix_small_imbalance(env, sds);
6807 /******* find_busiest_group() helpers end here *********************/
6810 * find_busiest_group - Returns the busiest group within the sched_domain
6811 * if there is an imbalance. If there isn't an imbalance, and
6812 * the user has opted for power-savings, it returns a group whose
6813 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6814 * such a group exists.
6816 * Also calculates the amount of weighted load which should be moved
6817 * to restore balance.
6819 * @env: The load balancing environment.
6821 * Return: - The busiest group if imbalance exists.
6822 * - If no imbalance and user has opted for power-savings balance,
6823 * return the least loaded group whose CPUs can be
6824 * put to idle by rebalancing its tasks onto our group.
6826 static struct sched_group *find_busiest_group(struct lb_env *env)
6828 struct sg_lb_stats *local, *busiest;
6829 struct sd_lb_stats sds;
6831 init_sd_lb_stats(&sds);
6834 * Compute the various statistics relavent for load balancing at
6837 update_sd_lb_stats(env, &sds);
6838 local = &sds.local_stat;
6839 busiest = &sds.busiest_stat;
6841 /* ASYM feature bypasses nice load balance check */
6842 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6843 check_asym_packing(env, &sds))
6846 /* There is no busy sibling group to pull tasks from */
6847 if (!sds.busiest || busiest->sum_nr_running == 0)
6850 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
6851 / sds.total_capacity;
6854 * If the busiest group is imbalanced the below checks don't
6855 * work because they assume all things are equal, which typically
6856 * isn't true due to cpus_allowed constraints and the like.
6858 if (busiest->group_type == group_imbalanced)
6861 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6862 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
6863 busiest->group_no_capacity)
6867 * If the local group is busier than the selected busiest group
6868 * don't try and pull any tasks.
6870 if (local->avg_load >= busiest->avg_load)
6874 * Don't pull any tasks if this group is already above the domain
6877 if (local->avg_load >= sds.avg_load)
6880 if (env->idle == CPU_IDLE) {
6882 * This cpu is idle. If the busiest group is not overloaded
6883 * and there is no imbalance between this and busiest group
6884 * wrt idle cpus, it is balanced. The imbalance becomes
6885 * significant if the diff is greater than 1 otherwise we
6886 * might end up to just move the imbalance on another group
6888 if ((busiest->group_type != group_overloaded) &&
6889 (local->idle_cpus <= (busiest->idle_cpus + 1)))
6893 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6894 * imbalance_pct to be conservative.
6896 if (100 * busiest->avg_load <=
6897 env->sd->imbalance_pct * local->avg_load)
6902 /* Looks like there is an imbalance. Compute it */
6903 calculate_imbalance(env, &sds);
6912 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6914 static struct rq *find_busiest_queue(struct lb_env *env,
6915 struct sched_group *group)
6917 struct rq *busiest = NULL, *rq;
6918 unsigned long busiest_load = 0, busiest_capacity = 1;
6921 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6922 unsigned long capacity, wl;
6926 rt = fbq_classify_rq(rq);
6929 * We classify groups/runqueues into three groups:
6930 * - regular: there are !numa tasks
6931 * - remote: there are numa tasks that run on the 'wrong' node
6932 * - all: there is no distinction
6934 * In order to avoid migrating ideally placed numa tasks,
6935 * ignore those when there's better options.
6937 * If we ignore the actual busiest queue to migrate another
6938 * task, the next balance pass can still reduce the busiest
6939 * queue by moving tasks around inside the node.
6941 * If we cannot move enough load due to this classification
6942 * the next pass will adjust the group classification and
6943 * allow migration of more tasks.
6945 * Both cases only affect the total convergence complexity.
6947 if (rt > env->fbq_type)
6950 capacity = capacity_of(i);
6952 wl = weighted_cpuload(i);
6955 * When comparing with imbalance, use weighted_cpuload()
6956 * which is not scaled with the cpu capacity.
6959 if (rq->nr_running == 1 && wl > env->imbalance &&
6960 !check_cpu_capacity(rq, env->sd))
6964 * For the load comparisons with the other cpu's, consider
6965 * the weighted_cpuload() scaled with the cpu capacity, so
6966 * that the load can be moved away from the cpu that is
6967 * potentially running at a lower capacity.
6969 * Thus we're looking for max(wl_i / capacity_i), crosswise
6970 * multiplication to rid ourselves of the division works out
6971 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
6972 * our previous maximum.
6974 if (wl * busiest_capacity > busiest_load * capacity) {
6976 busiest_capacity = capacity;
6985 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6986 * so long as it is large enough.
6988 #define MAX_PINNED_INTERVAL 512
6990 /* Working cpumask for load_balance and load_balance_newidle. */
6991 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6993 static int need_active_balance(struct lb_env *env)
6995 struct sched_domain *sd = env->sd;
6997 if (env->idle == CPU_NEWLY_IDLE) {
7000 * ASYM_PACKING needs to force migrate tasks from busy but
7001 * higher numbered CPUs in order to pack all tasks in the
7002 * lowest numbered CPUs.
7004 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7009 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7010 * It's worth migrating the task if the src_cpu's capacity is reduced
7011 * because of other sched_class or IRQs if more capacity stays
7012 * available on dst_cpu.
7014 if ((env->idle != CPU_NOT_IDLE) &&
7015 (env->src_rq->cfs.h_nr_running == 1)) {
7016 if ((check_cpu_capacity(env->src_rq, sd)) &&
7017 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7021 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7024 static int active_load_balance_cpu_stop(void *data);
7026 static int should_we_balance(struct lb_env *env)
7028 struct sched_group *sg = env->sd->groups;
7029 struct cpumask *sg_cpus, *sg_mask;
7030 int cpu, balance_cpu = -1;
7033 * In the newly idle case, we will allow all the cpu's
7034 * to do the newly idle load balance.
7036 if (env->idle == CPU_NEWLY_IDLE)
7039 sg_cpus = sched_group_cpus(sg);
7040 sg_mask = sched_group_mask(sg);
7041 /* Try to find first idle cpu */
7042 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7043 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7050 if (balance_cpu == -1)
7051 balance_cpu = group_balance_cpu(sg);
7054 * First idle cpu or the first cpu(busiest) in this sched group
7055 * is eligible for doing load balancing at this and above domains.
7057 return balance_cpu == env->dst_cpu;
7061 * Check this_cpu to ensure it is balanced within domain. Attempt to move
7062 * tasks if there is an imbalance.
7064 static int load_balance(int this_cpu, struct rq *this_rq,
7065 struct sched_domain *sd, enum cpu_idle_type idle,
7066 int *continue_balancing)
7068 int ld_moved, cur_ld_moved, active_balance = 0;
7069 struct sched_domain *sd_parent = sd->parent;
7070 struct sched_group *group;
7072 unsigned long flags;
7073 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
7075 struct lb_env env = {
7077 .dst_cpu = this_cpu,
7079 .dst_grpmask = sched_group_cpus(sd->groups),
7081 .loop_break = sched_nr_migrate_break,
7084 .tasks = LIST_HEAD_INIT(env.tasks),
7088 * For NEWLY_IDLE load_balancing, we don't need to consider
7089 * other cpus in our group
7091 if (idle == CPU_NEWLY_IDLE)
7092 env.dst_grpmask = NULL;
7094 cpumask_copy(cpus, cpu_active_mask);
7096 schedstat_inc(sd, lb_count[idle]);
7099 if (!should_we_balance(&env)) {
7100 *continue_balancing = 0;
7104 group = find_busiest_group(&env);
7106 schedstat_inc(sd, lb_nobusyg[idle]);
7110 busiest = find_busiest_queue(&env, group);
7112 schedstat_inc(sd, lb_nobusyq[idle]);
7116 BUG_ON(busiest == env.dst_rq);
7118 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
7120 env.src_cpu = busiest->cpu;
7121 env.src_rq = busiest;
7124 if (busiest->nr_running > 1) {
7126 * Attempt to move tasks. If find_busiest_group has found
7127 * an imbalance but busiest->nr_running <= 1, the group is
7128 * still unbalanced. ld_moved simply stays zero, so it is
7129 * correctly treated as an imbalance.
7131 env.flags |= LBF_ALL_PINNED;
7132 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7135 raw_spin_lock_irqsave(&busiest->lock, flags);
7138 * cur_ld_moved - load moved in current iteration
7139 * ld_moved - cumulative load moved across iterations
7141 cur_ld_moved = detach_tasks(&env);
7144 * We've detached some tasks from busiest_rq. Every
7145 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7146 * unlock busiest->lock, and we are able to be sure
7147 * that nobody can manipulate the tasks in parallel.
7148 * See task_rq_lock() family for the details.
7151 raw_spin_unlock(&busiest->lock);
7155 ld_moved += cur_ld_moved;
7158 local_irq_restore(flags);
7160 if (env.flags & LBF_NEED_BREAK) {
7161 env.flags &= ~LBF_NEED_BREAK;
7166 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7167 * us and move them to an alternate dst_cpu in our sched_group
7168 * where they can run. The upper limit on how many times we
7169 * iterate on same src_cpu is dependent on number of cpus in our
7172 * This changes load balance semantics a bit on who can move
7173 * load to a given_cpu. In addition to the given_cpu itself
7174 * (or a ilb_cpu acting on its behalf where given_cpu is
7175 * nohz-idle), we now have balance_cpu in a position to move
7176 * load to given_cpu. In rare situations, this may cause
7177 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7178 * _independently_ and at _same_ time to move some load to
7179 * given_cpu) causing exceess load to be moved to given_cpu.
7180 * This however should not happen so much in practice and
7181 * moreover subsequent load balance cycles should correct the
7182 * excess load moved.
7184 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7186 /* Prevent to re-select dst_cpu via env's cpus */
7187 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7189 env.dst_rq = cpu_rq(env.new_dst_cpu);
7190 env.dst_cpu = env.new_dst_cpu;
7191 env.flags &= ~LBF_DST_PINNED;
7193 env.loop_break = sched_nr_migrate_break;
7196 * Go back to "more_balance" rather than "redo" since we
7197 * need to continue with same src_cpu.
7203 * We failed to reach balance because of affinity.
7206 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7208 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7209 *group_imbalance = 1;
7212 /* All tasks on this runqueue were pinned by CPU affinity */
7213 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7214 cpumask_clear_cpu(cpu_of(busiest), cpus);
7215 if (!cpumask_empty(cpus)) {
7217 env.loop_break = sched_nr_migrate_break;
7220 goto out_all_pinned;
7225 schedstat_inc(sd, lb_failed[idle]);
7227 * Increment the failure counter only on periodic balance.
7228 * We do not want newidle balance, which can be very
7229 * frequent, pollute the failure counter causing
7230 * excessive cache_hot migrations and active balances.
7232 if (idle != CPU_NEWLY_IDLE)
7233 sd->nr_balance_failed++;
7235 if (need_active_balance(&env)) {
7236 raw_spin_lock_irqsave(&busiest->lock, flags);
7238 /* don't kick the active_load_balance_cpu_stop,
7239 * if the curr task on busiest cpu can't be
7242 if (!cpumask_test_cpu(this_cpu,
7243 tsk_cpus_allowed(busiest->curr))) {
7244 raw_spin_unlock_irqrestore(&busiest->lock,
7246 env.flags |= LBF_ALL_PINNED;
7247 goto out_one_pinned;
7251 * ->active_balance synchronizes accesses to
7252 * ->active_balance_work. Once set, it's cleared
7253 * only after active load balance is finished.
7255 if (!busiest->active_balance) {
7256 busiest->active_balance = 1;
7257 busiest->push_cpu = this_cpu;
7260 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7262 if (active_balance) {
7263 stop_one_cpu_nowait(cpu_of(busiest),
7264 active_load_balance_cpu_stop, busiest,
7265 &busiest->active_balance_work);
7269 * We've kicked active balancing, reset the failure
7272 sd->nr_balance_failed = sd->cache_nice_tries+1;
7275 sd->nr_balance_failed = 0;
7277 if (likely(!active_balance)) {
7278 /* We were unbalanced, so reset the balancing interval */
7279 sd->balance_interval = sd->min_interval;
7282 * If we've begun active balancing, start to back off. This
7283 * case may not be covered by the all_pinned logic if there
7284 * is only 1 task on the busy runqueue (because we don't call
7287 if (sd->balance_interval < sd->max_interval)
7288 sd->balance_interval *= 2;
7295 * We reach balance although we may have faced some affinity
7296 * constraints. Clear the imbalance flag if it was set.
7299 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7301 if (*group_imbalance)
7302 *group_imbalance = 0;
7307 * We reach balance because all tasks are pinned at this level so
7308 * we can't migrate them. Let the imbalance flag set so parent level
7309 * can try to migrate them.
7311 schedstat_inc(sd, lb_balanced[idle]);
7313 sd->nr_balance_failed = 0;
7316 /* tune up the balancing interval */
7317 if (((env.flags & LBF_ALL_PINNED) &&
7318 sd->balance_interval < MAX_PINNED_INTERVAL) ||
7319 (sd->balance_interval < sd->max_interval))
7320 sd->balance_interval *= 2;
7327 static inline unsigned long
7328 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7330 unsigned long interval = sd->balance_interval;
7333 interval *= sd->busy_factor;
7335 /* scale ms to jiffies */
7336 interval = msecs_to_jiffies(interval);
7337 interval = clamp(interval, 1UL, max_load_balance_interval);
7343 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
7345 unsigned long interval, next;
7347 interval = get_sd_balance_interval(sd, cpu_busy);
7348 next = sd->last_balance + interval;
7350 if (time_after(*next_balance, next))
7351 *next_balance = next;
7355 * idle_balance is called by schedule() if this_cpu is about to become
7356 * idle. Attempts to pull tasks from other CPUs.
7358 static int idle_balance(struct rq *this_rq)
7360 unsigned long next_balance = jiffies + HZ;
7361 int this_cpu = this_rq->cpu;
7362 struct sched_domain *sd;
7363 int pulled_task = 0;
7366 idle_enter_fair(this_rq);
7369 * We must set idle_stamp _before_ calling idle_balance(), such that we
7370 * measure the duration of idle_balance() as idle time.
7372 this_rq->idle_stamp = rq_clock(this_rq);
7374 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
7375 !this_rq->rd->overload) {
7377 sd = rcu_dereference_check_sched_domain(this_rq->sd);
7379 update_next_balance(sd, 0, &next_balance);
7385 raw_spin_unlock(&this_rq->lock);
7387 update_blocked_averages(this_cpu);
7389 for_each_domain(this_cpu, sd) {
7390 int continue_balancing = 1;
7391 u64 t0, domain_cost;
7393 if (!(sd->flags & SD_LOAD_BALANCE))
7396 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7397 update_next_balance(sd, 0, &next_balance);
7401 if (sd->flags & SD_BALANCE_NEWIDLE) {
7402 t0 = sched_clock_cpu(this_cpu);
7404 pulled_task = load_balance(this_cpu, this_rq,
7406 &continue_balancing);
7408 domain_cost = sched_clock_cpu(this_cpu) - t0;
7409 if (domain_cost > sd->max_newidle_lb_cost)
7410 sd->max_newidle_lb_cost = domain_cost;
7412 curr_cost += domain_cost;
7415 update_next_balance(sd, 0, &next_balance);
7418 * Stop searching for tasks to pull if there are
7419 * now runnable tasks on this rq.
7421 if (pulled_task || this_rq->nr_running > 0)
7426 raw_spin_lock(&this_rq->lock);
7428 if (curr_cost > this_rq->max_idle_balance_cost)
7429 this_rq->max_idle_balance_cost = curr_cost;
7432 * While browsing the domains, we released the rq lock, a task could
7433 * have been enqueued in the meantime. Since we're not going idle,
7434 * pretend we pulled a task.
7436 if (this_rq->cfs.h_nr_running && !pulled_task)
7440 /* Move the next balance forward */
7441 if (time_after(this_rq->next_balance, next_balance))
7442 this_rq->next_balance = next_balance;
7444 /* Is there a task of a high priority class? */
7445 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7449 idle_exit_fair(this_rq);
7450 this_rq->idle_stamp = 0;
7457 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7458 * running tasks off the busiest CPU onto idle CPUs. It requires at
7459 * least 1 task to be running on each physical CPU where possible, and
7460 * avoids physical / logical imbalances.
7462 static int active_load_balance_cpu_stop(void *data)
7464 struct rq *busiest_rq = data;
7465 int busiest_cpu = cpu_of(busiest_rq);
7466 int target_cpu = busiest_rq->push_cpu;
7467 struct rq *target_rq = cpu_rq(target_cpu);
7468 struct sched_domain *sd;
7469 struct task_struct *p = NULL;
7471 raw_spin_lock_irq(&busiest_rq->lock);
7473 /* make sure the requested cpu hasn't gone down in the meantime */
7474 if (unlikely(busiest_cpu != smp_processor_id() ||
7475 !busiest_rq->active_balance))
7478 /* Is there any task to move? */
7479 if (busiest_rq->nr_running <= 1)
7483 * This condition is "impossible", if it occurs
7484 * we need to fix it. Originally reported by
7485 * Bjorn Helgaas on a 128-cpu setup.
7487 BUG_ON(busiest_rq == target_rq);
7489 /* Search for an sd spanning us and the target CPU. */
7491 for_each_domain(target_cpu, sd) {
7492 if ((sd->flags & SD_LOAD_BALANCE) &&
7493 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7498 struct lb_env env = {
7500 .dst_cpu = target_cpu,
7501 .dst_rq = target_rq,
7502 .src_cpu = busiest_rq->cpu,
7503 .src_rq = busiest_rq,
7507 schedstat_inc(sd, alb_count);
7509 p = detach_one_task(&env);
7511 schedstat_inc(sd, alb_pushed);
7513 schedstat_inc(sd, alb_failed);
7517 busiest_rq->active_balance = 0;
7518 raw_spin_unlock(&busiest_rq->lock);
7521 attach_one_task(target_rq, p);
7528 static inline int on_null_domain(struct rq *rq)
7530 return unlikely(!rcu_dereference_sched(rq->sd));
7533 #ifdef CONFIG_NO_HZ_COMMON
7535 * idle load balancing details
7536 * - When one of the busy CPUs notice that there may be an idle rebalancing
7537 * needed, they will kick the idle load balancer, which then does idle
7538 * load balancing for all the idle CPUs.
7541 cpumask_var_t idle_cpus_mask;
7543 unsigned long next_balance; /* in jiffy units */
7544 } nohz ____cacheline_aligned;
7546 static inline int find_new_ilb(void)
7548 int ilb = cpumask_first(nohz.idle_cpus_mask);
7550 if (ilb < nr_cpu_ids && idle_cpu(ilb))
7557 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7558 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7559 * CPU (if there is one).
7561 static void nohz_balancer_kick(void)
7565 nohz.next_balance++;
7567 ilb_cpu = find_new_ilb();
7569 if (ilb_cpu >= nr_cpu_ids)
7572 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7575 * Use smp_send_reschedule() instead of resched_cpu().
7576 * This way we generate a sched IPI on the target cpu which
7577 * is idle. And the softirq performing nohz idle load balance
7578 * will be run before returning from the IPI.
7580 smp_send_reschedule(ilb_cpu);
7584 static inline void nohz_balance_exit_idle(int cpu)
7586 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7588 * Completely isolated CPUs don't ever set, so we must test.
7590 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7591 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7592 atomic_dec(&nohz.nr_cpus);
7594 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7598 static inline void set_cpu_sd_state_busy(void)
7600 struct sched_domain *sd;
7601 int cpu = smp_processor_id();
7604 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7606 if (!sd || !sd->nohz_idle)
7610 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7615 void set_cpu_sd_state_idle(void)
7617 struct sched_domain *sd;
7618 int cpu = smp_processor_id();
7621 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7623 if (!sd || sd->nohz_idle)
7627 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7633 * This routine will record that the cpu is going idle with tick stopped.
7634 * This info will be used in performing idle load balancing in the future.
7636 void nohz_balance_enter_idle(int cpu)
7639 * If this cpu is going down, then nothing needs to be done.
7641 if (!cpu_active(cpu))
7644 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7648 * If we're a completely isolated CPU, we don't play.
7650 if (on_null_domain(cpu_rq(cpu)))
7653 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7654 atomic_inc(&nohz.nr_cpus);
7655 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7658 static int sched_ilb_notifier(struct notifier_block *nfb,
7659 unsigned long action, void *hcpu)
7661 switch (action & ~CPU_TASKS_FROZEN) {
7663 nohz_balance_exit_idle(smp_processor_id());
7671 static DEFINE_SPINLOCK(balancing);
7674 * Scale the max load_balance interval with the number of CPUs in the system.
7675 * This trades load-balance latency on larger machines for less cross talk.
7677 void update_max_interval(void)
7679 max_load_balance_interval = HZ*num_online_cpus()/10;
7683 * It checks each scheduling domain to see if it is due to be balanced,
7684 * and initiates a balancing operation if so.
7686 * Balancing parameters are set up in init_sched_domains.
7688 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7690 int continue_balancing = 1;
7692 unsigned long interval;
7693 struct sched_domain *sd;
7694 /* Earliest time when we have to do rebalance again */
7695 unsigned long next_balance = jiffies + 60*HZ;
7696 int update_next_balance = 0;
7697 int need_serialize, need_decay = 0;
7700 update_blocked_averages(cpu);
7703 for_each_domain(cpu, sd) {
7705 * Decay the newidle max times here because this is a regular
7706 * visit to all the domains. Decay ~1% per second.
7708 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7709 sd->max_newidle_lb_cost =
7710 (sd->max_newidle_lb_cost * 253) / 256;
7711 sd->next_decay_max_lb_cost = jiffies + HZ;
7714 max_cost += sd->max_newidle_lb_cost;
7716 if (!(sd->flags & SD_LOAD_BALANCE))
7720 * Stop the load balance at this level. There is another
7721 * CPU in our sched group which is doing load balancing more
7724 if (!continue_balancing) {
7730 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7732 need_serialize = sd->flags & SD_SERIALIZE;
7733 if (need_serialize) {
7734 if (!spin_trylock(&balancing))
7738 if (time_after_eq(jiffies, sd->last_balance + interval)) {
7739 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7741 * The LBF_DST_PINNED logic could have changed
7742 * env->dst_cpu, so we can't know our idle
7743 * state even if we migrated tasks. Update it.
7745 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7747 sd->last_balance = jiffies;
7748 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7751 spin_unlock(&balancing);
7753 if (time_after(next_balance, sd->last_balance + interval)) {
7754 next_balance = sd->last_balance + interval;
7755 update_next_balance = 1;
7760 * Ensure the rq-wide value also decays but keep it at a
7761 * reasonable floor to avoid funnies with rq->avg_idle.
7763 rq->max_idle_balance_cost =
7764 max((u64)sysctl_sched_migration_cost, max_cost);
7769 * next_balance will be updated only when there is a need.
7770 * When the cpu is attached to null domain for ex, it will not be
7773 if (likely(update_next_balance)) {
7774 rq->next_balance = next_balance;
7776 #ifdef CONFIG_NO_HZ_COMMON
7778 * If this CPU has been elected to perform the nohz idle
7779 * balance. Other idle CPUs have already rebalanced with
7780 * nohz_idle_balance() and nohz.next_balance has been
7781 * updated accordingly. This CPU is now running the idle load
7782 * balance for itself and we need to update the
7783 * nohz.next_balance accordingly.
7785 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
7786 nohz.next_balance = rq->next_balance;
7791 #ifdef CONFIG_NO_HZ_COMMON
7793 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7794 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7796 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7798 int this_cpu = this_rq->cpu;
7801 /* Earliest time when we have to do rebalance again */
7802 unsigned long next_balance = jiffies + 60*HZ;
7803 int update_next_balance = 0;
7805 if (idle != CPU_IDLE ||
7806 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7809 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7810 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7814 * If this cpu gets work to do, stop the load balancing
7815 * work being done for other cpus. Next load
7816 * balancing owner will pick it up.
7821 rq = cpu_rq(balance_cpu);
7824 * If time for next balance is due,
7827 if (time_after_eq(jiffies, rq->next_balance)) {
7828 raw_spin_lock_irq(&rq->lock);
7829 update_rq_clock(rq);
7830 update_idle_cpu_load(rq);
7831 raw_spin_unlock_irq(&rq->lock);
7832 rebalance_domains(rq, CPU_IDLE);
7835 if (time_after(next_balance, rq->next_balance)) {
7836 next_balance = rq->next_balance;
7837 update_next_balance = 1;
7842 * next_balance will be updated only when there is a need.
7843 * When the CPU is attached to null domain for ex, it will not be
7846 if (likely(update_next_balance))
7847 nohz.next_balance = next_balance;
7849 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7853 * Current heuristic for kicking the idle load balancer in the presence
7854 * of an idle cpu in the system.
7855 * - This rq has more than one task.
7856 * - This rq has at least one CFS task and the capacity of the CPU is
7857 * significantly reduced because of RT tasks or IRQs.
7858 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
7859 * multiple busy cpu.
7860 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7861 * domain span are idle.
7863 static inline bool nohz_kick_needed(struct rq *rq)
7865 unsigned long now = jiffies;
7866 struct sched_domain *sd;
7867 struct sched_group_capacity *sgc;
7868 int nr_busy, cpu = rq->cpu;
7871 if (unlikely(rq->idle_balance))
7875 * We may be recently in ticked or tickless idle mode. At the first
7876 * busy tick after returning from idle, we will update the busy stats.
7878 set_cpu_sd_state_busy();
7879 nohz_balance_exit_idle(cpu);
7882 * None are in tickless mode and hence no need for NOHZ idle load
7885 if (likely(!atomic_read(&nohz.nr_cpus)))
7888 if (time_before(now, nohz.next_balance))
7891 if (rq->nr_running >= 2)
7895 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7897 sgc = sd->groups->sgc;
7898 nr_busy = atomic_read(&sgc->nr_busy_cpus);
7907 sd = rcu_dereference(rq->sd);
7909 if ((rq->cfs.h_nr_running >= 1) &&
7910 check_cpu_capacity(rq, sd)) {
7916 sd = rcu_dereference(per_cpu(sd_asym, cpu));
7917 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7918 sched_domain_span(sd)) < cpu)) {
7928 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7932 * run_rebalance_domains is triggered when needed from the scheduler tick.
7933 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7935 static void run_rebalance_domains(struct softirq_action *h)
7937 struct rq *this_rq = this_rq();
7938 enum cpu_idle_type idle = this_rq->idle_balance ?
7939 CPU_IDLE : CPU_NOT_IDLE;
7942 * If this cpu has a pending nohz_balance_kick, then do the
7943 * balancing on behalf of the other idle cpus whose ticks are
7944 * stopped. Do nohz_idle_balance *before* rebalance_domains to
7945 * give the idle cpus a chance to load balance. Else we may
7946 * load balance only within the local sched_domain hierarchy
7947 * and abort nohz_idle_balance altogether if we pull some load.
7949 nohz_idle_balance(this_rq, idle);
7950 rebalance_domains(this_rq, idle);
7954 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7956 void trigger_load_balance(struct rq *rq)
7958 /* Don't need to rebalance while attached to NULL domain */
7959 if (unlikely(on_null_domain(rq)))
7962 if (time_after_eq(jiffies, rq->next_balance))
7963 raise_softirq(SCHED_SOFTIRQ);
7964 #ifdef CONFIG_NO_HZ_COMMON
7965 if (nohz_kick_needed(rq))
7966 nohz_balancer_kick();
7970 static void rq_online_fair(struct rq *rq)
7974 update_runtime_enabled(rq);
7977 static void rq_offline_fair(struct rq *rq)
7981 /* Ensure any throttled groups are reachable by pick_next_task */
7982 unthrottle_offline_cfs_rqs(rq);
7985 #endif /* CONFIG_SMP */
7988 * scheduler tick hitting a task of our scheduling class:
7990 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7992 struct cfs_rq *cfs_rq;
7993 struct sched_entity *se = &curr->se;
7995 for_each_sched_entity(se) {
7996 cfs_rq = cfs_rq_of(se);
7997 entity_tick(cfs_rq, se, queued);
8000 if (static_branch_unlikely(&sched_numa_balancing))
8001 task_tick_numa(rq, curr);
8005 * called on fork with the child task as argument from the parent's context
8006 * - child not yet on the tasklist
8007 * - preemption disabled
8009 static void task_fork_fair(struct task_struct *p)
8011 struct cfs_rq *cfs_rq;
8012 struct sched_entity *se = &p->se, *curr;
8013 int this_cpu = smp_processor_id();
8014 struct rq *rq = this_rq();
8015 unsigned long flags;
8017 raw_spin_lock_irqsave(&rq->lock, flags);
8019 update_rq_clock(rq);
8021 cfs_rq = task_cfs_rq(current);
8022 curr = cfs_rq->curr;
8025 * Not only the cpu but also the task_group of the parent might have
8026 * been changed after parent->se.parent,cfs_rq were copied to
8027 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
8028 * of child point to valid ones.
8031 __set_task_cpu(p, this_cpu);
8034 update_curr(cfs_rq);
8037 se->vruntime = curr->vruntime;
8038 place_entity(cfs_rq, se, 1);
8040 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8042 * Upon rescheduling, sched_class::put_prev_task() will place
8043 * 'current' within the tree based on its new key value.
8045 swap(curr->vruntime, se->vruntime);
8049 se->vruntime -= cfs_rq->min_vruntime;
8051 raw_spin_unlock_irqrestore(&rq->lock, flags);
8055 * Priority of the task has changed. Check to see if we preempt
8059 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
8061 if (!task_on_rq_queued(p))
8065 * Reschedule if we are currently running on this runqueue and
8066 * our priority decreased, or if we are not currently running on
8067 * this runqueue and our priority is higher than the current's
8069 if (rq->curr == p) {
8070 if (p->prio > oldprio)
8073 check_preempt_curr(rq, p, 0);
8076 static inline bool vruntime_normalized(struct task_struct *p)
8078 struct sched_entity *se = &p->se;
8081 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
8082 * the dequeue_entity(.flags=0) will already have normalized the
8089 * When !on_rq, vruntime of the task has usually NOT been normalized.
8090 * But there are some cases where it has already been normalized:
8092 * - A forked child which is waiting for being woken up by
8093 * wake_up_new_task().
8094 * - A task which has been woken up by try_to_wake_up() and
8095 * waiting for actually being woken up by sched_ttwu_pending().
8097 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
8103 static void detach_task_cfs_rq(struct task_struct *p)
8105 struct sched_entity *se = &p->se;
8106 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8108 if (!vruntime_normalized(p)) {
8110 * Fix up our vruntime so that the current sleep doesn't
8111 * cause 'unlimited' sleep bonus.
8113 place_entity(cfs_rq, se, 0);
8114 se->vruntime -= cfs_rq->min_vruntime;
8117 /* Catch up with the cfs_rq and remove our load when we leave */
8118 detach_entity_load_avg(cfs_rq, se);
8121 static void attach_task_cfs_rq(struct task_struct *p)
8123 struct sched_entity *se = &p->se;
8124 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8126 #ifdef CONFIG_FAIR_GROUP_SCHED
8128 * Since the real-depth could have been changed (only FAIR
8129 * class maintain depth value), reset depth properly.
8131 se->depth = se->parent ? se->parent->depth + 1 : 0;
8134 /* Synchronize task with its cfs_rq */
8135 attach_entity_load_avg(cfs_rq, se);
8137 if (!vruntime_normalized(p))
8138 se->vruntime += cfs_rq->min_vruntime;
8141 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8143 detach_task_cfs_rq(p);
8146 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8148 attach_task_cfs_rq(p);
8150 if (task_on_rq_queued(p)) {
8152 * We were most likely switched from sched_rt, so
8153 * kick off the schedule if running, otherwise just see
8154 * if we can still preempt the current task.
8159 check_preempt_curr(rq, p, 0);
8163 /* Account for a task changing its policy or group.
8165 * This routine is mostly called to set cfs_rq->curr field when a task
8166 * migrates between groups/classes.
8168 static void set_curr_task_fair(struct rq *rq)
8170 struct sched_entity *se = &rq->curr->se;
8172 for_each_sched_entity(se) {
8173 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8175 set_next_entity(cfs_rq, se);
8176 /* ensure bandwidth has been allocated on our new cfs_rq */
8177 account_cfs_rq_runtime(cfs_rq, 0);
8181 void init_cfs_rq(struct cfs_rq *cfs_rq)
8183 cfs_rq->tasks_timeline = RB_ROOT;
8184 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8185 #ifndef CONFIG_64BIT
8186 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8189 atomic_long_set(&cfs_rq->removed_load_avg, 0);
8190 atomic_long_set(&cfs_rq->removed_util_avg, 0);
8194 #ifdef CONFIG_FAIR_GROUP_SCHED
8195 static void task_move_group_fair(struct task_struct *p)
8197 detach_task_cfs_rq(p);
8198 set_task_rq(p, task_cpu(p));
8201 /* Tell se's cfs_rq has been changed -- migrated */
8202 p->se.avg.last_update_time = 0;
8204 attach_task_cfs_rq(p);
8207 void free_fair_sched_group(struct task_group *tg)
8211 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8213 for_each_possible_cpu(i) {
8215 kfree(tg->cfs_rq[i]);
8224 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8226 struct cfs_rq *cfs_rq;
8227 struct sched_entity *se;
8230 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8233 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8237 tg->shares = NICE_0_LOAD;
8239 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8241 for_each_possible_cpu(i) {
8242 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8243 GFP_KERNEL, cpu_to_node(i));
8247 se = kzalloc_node(sizeof(struct sched_entity),
8248 GFP_KERNEL, cpu_to_node(i));
8252 init_cfs_rq(cfs_rq);
8253 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8254 init_entity_runnable_average(se);
8265 void unregister_fair_sched_group(struct task_group *tg)
8267 unsigned long flags;
8271 for_each_possible_cpu(cpu) {
8273 remove_entity_load_avg(tg->se[cpu]);
8276 * Only empty task groups can be destroyed; so we can speculatively
8277 * check on_list without danger of it being re-added.
8279 if (!tg->cfs_rq[cpu]->on_list)
8284 raw_spin_lock_irqsave(&rq->lock, flags);
8285 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8286 raw_spin_unlock_irqrestore(&rq->lock, flags);
8290 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8291 struct sched_entity *se, int cpu,
8292 struct sched_entity *parent)
8294 struct rq *rq = cpu_rq(cpu);
8298 init_cfs_rq_runtime(cfs_rq);
8300 tg->cfs_rq[cpu] = cfs_rq;
8303 /* se could be NULL for root_task_group */
8308 se->cfs_rq = &rq->cfs;
8311 se->cfs_rq = parent->my_q;
8312 se->depth = parent->depth + 1;
8316 /* guarantee group entities always have weight */
8317 update_load_set(&se->load, NICE_0_LOAD);
8318 se->parent = parent;
8321 static DEFINE_MUTEX(shares_mutex);
8323 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8326 unsigned long flags;
8329 * We can't change the weight of the root cgroup.
8334 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8336 mutex_lock(&shares_mutex);
8337 if (tg->shares == shares)
8340 tg->shares = shares;
8341 for_each_possible_cpu(i) {
8342 struct rq *rq = cpu_rq(i);
8343 struct sched_entity *se;
8346 /* Propagate contribution to hierarchy */
8347 raw_spin_lock_irqsave(&rq->lock, flags);
8349 /* Possible calls to update_curr() need rq clock */
8350 update_rq_clock(rq);
8351 for_each_sched_entity(se)
8352 update_cfs_shares(group_cfs_rq(se));
8353 raw_spin_unlock_irqrestore(&rq->lock, flags);
8357 mutex_unlock(&shares_mutex);
8360 #else /* CONFIG_FAIR_GROUP_SCHED */
8362 void free_fair_sched_group(struct task_group *tg) { }
8364 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8369 void unregister_fair_sched_group(struct task_group *tg) { }
8371 #endif /* CONFIG_FAIR_GROUP_SCHED */
8374 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
8376 struct sched_entity *se = &task->se;
8377 unsigned int rr_interval = 0;
8380 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8383 if (rq->cfs.load.weight)
8384 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
8390 * All the scheduling class methods:
8392 const struct sched_class fair_sched_class = {
8393 .next = &idle_sched_class,
8394 .enqueue_task = enqueue_task_fair,
8395 .dequeue_task = dequeue_task_fair,
8396 .yield_task = yield_task_fair,
8397 .yield_to_task = yield_to_task_fair,
8399 .check_preempt_curr = check_preempt_wakeup,
8401 .pick_next_task = pick_next_task_fair,
8402 .put_prev_task = put_prev_task_fair,
8405 .select_task_rq = select_task_rq_fair,
8406 .migrate_task_rq = migrate_task_rq_fair,
8408 .rq_online = rq_online_fair,
8409 .rq_offline = rq_offline_fair,
8411 .task_waking = task_waking_fair,
8412 .task_dead = task_dead_fair,
8413 .set_cpus_allowed = set_cpus_allowed_common,
8416 .set_curr_task = set_curr_task_fair,
8417 .task_tick = task_tick_fair,
8418 .task_fork = task_fork_fair,
8420 .prio_changed = prio_changed_fair,
8421 .switched_from = switched_from_fair,
8422 .switched_to = switched_to_fair,
8424 .get_rr_interval = get_rr_interval_fair,
8426 .update_curr = update_curr_fair,
8428 #ifdef CONFIG_FAIR_GROUP_SCHED
8429 .task_move_group = task_move_group_fair,
8433 #ifdef CONFIG_SCHED_DEBUG
8434 void print_cfs_stats(struct seq_file *m, int cpu)
8436 struct cfs_rq *cfs_rq;
8439 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
8440 print_cfs_rq(m, cpu, cfs_rq);
8444 #ifdef CONFIG_NUMA_BALANCING
8445 void show_numa_stats(struct task_struct *p, struct seq_file *m)
8448 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
8450 for_each_online_node(node) {
8451 if (p->numa_faults) {
8452 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
8453 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
8455 if (p->numa_group) {
8456 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
8457 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
8459 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
8462 #endif /* CONFIG_NUMA_BALANCING */
8463 #endif /* CONFIG_SCHED_DEBUG */
8465 __init void init_sched_fair_class(void)
8468 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8470 #ifdef CONFIG_NO_HZ_COMMON
8471 nohz.next_balance = jiffies;
8472 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8473 cpu_notifier(sched_ilb_notifier, 0);