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 <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/slab.h>
27 #include <linux/profile.h>
28 #include <linux/interrupt.h>
29 #include <linux/mempolicy.h>
30 #include <linux/migrate.h>
31 #include <linux/task_work.h>
33 #include <trace/events/sched.h>
38 * Targeted preemption latency for CPU-bound tasks:
39 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
41 * NOTE: this latency value is not the same as the concept of
42 * 'timeslice length' - timeslices in CFS are of variable length
43 * and have no persistent notion like in traditional, time-slice
44 * based scheduling concepts.
46 * (to see the precise effective timeslice length of your workload,
47 * run vmstat and monitor the context-switches (cs) field)
49 unsigned int sysctl_sched_latency = 6000000ULL;
50 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
53 * The initial- and re-scaling of tunables is configurable
54 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
57 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
58 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
59 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
61 enum sched_tunable_scaling sysctl_sched_tunable_scaling
62 = SCHED_TUNABLESCALING_LOG;
65 * Minimal preemption granularity for CPU-bound tasks:
66 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 unsigned int sysctl_sched_min_granularity = 750000ULL;
69 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
72 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
74 static unsigned int sched_nr_latency = 8;
77 * After fork, child runs first. If set to 0 (default) then
78 * parent will (try to) run first.
80 unsigned int sysctl_sched_child_runs_first __read_mostly;
83 * SCHED_OTHER wake-up granularity.
84 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
86 * This option delays the preemption effects of decoupled workloads
87 * and reduces their over-scheduling. Synchronous workloads will still
88 * have immediate wakeup/sleep latencies.
90 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
91 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
93 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
96 * The exponential sliding window over which load is averaged for shares
100 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
102 #ifdef CONFIG_CFS_BANDWIDTH
104 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
105 * each time a cfs_rq requests quota.
107 * Note: in the case that the slice exceeds the runtime remaining (either due
108 * to consumption or the quota being specified to be smaller than the slice)
109 * we will always only issue the remaining available time.
111 * default: 5 msec, units: microseconds
113 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
116 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
122 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
128 static inline void update_load_set(struct load_weight *lw, unsigned long w)
135 * Increase the granularity value when there are more CPUs,
136 * because with more CPUs the 'effective latency' as visible
137 * to users decreases. But the relationship is not linear,
138 * so pick a second-best guess by going with the log2 of the
141 * This idea comes from the SD scheduler of Con Kolivas:
143 static int get_update_sysctl_factor(void)
145 unsigned int cpus = min_t(int, num_online_cpus(), 8);
148 switch (sysctl_sched_tunable_scaling) {
149 case SCHED_TUNABLESCALING_NONE:
152 case SCHED_TUNABLESCALING_LINEAR:
155 case SCHED_TUNABLESCALING_LOG:
157 factor = 1 + ilog2(cpus);
164 static void update_sysctl(void)
166 unsigned int factor = get_update_sysctl_factor();
168 #define SET_SYSCTL(name) \
169 (sysctl_##name = (factor) * normalized_sysctl_##name)
170 SET_SYSCTL(sched_min_granularity);
171 SET_SYSCTL(sched_latency);
172 SET_SYSCTL(sched_wakeup_granularity);
176 void sched_init_granularity(void)
181 #define WMULT_CONST (~0U)
182 #define WMULT_SHIFT 32
184 static void __update_inv_weight(struct load_weight *lw)
188 if (likely(lw->inv_weight))
191 w = scale_load_down(lw->weight);
193 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
195 else if (unlikely(!w))
196 lw->inv_weight = WMULT_CONST;
198 lw->inv_weight = WMULT_CONST / w;
202 * delta_exec * weight / lw.weight
204 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
206 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
207 * we're guaranteed shift stays positive because inv_weight is guaranteed to
208 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
210 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
211 * weight/lw.weight <= 1, and therefore our shift will also be positive.
213 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
215 u64 fact = scale_load_down(weight);
216 int shift = WMULT_SHIFT;
218 __update_inv_weight(lw);
220 if (unlikely(fact >> 32)) {
227 /* hint to use a 32x32->64 mul */
228 fact = (u64)(u32)fact * lw->inv_weight;
235 return mul_u64_u32_shr(delta_exec, fact, shift);
239 const struct sched_class fair_sched_class;
241 /**************************************************************
242 * CFS operations on generic schedulable entities:
245 #ifdef CONFIG_FAIR_GROUP_SCHED
247 /* cpu runqueue to which this cfs_rq is attached */
248 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
253 /* An entity is a task if it doesn't "own" a runqueue */
254 #define entity_is_task(se) (!se->my_q)
256 static inline struct task_struct *task_of(struct sched_entity *se)
258 #ifdef CONFIG_SCHED_DEBUG
259 WARN_ON_ONCE(!entity_is_task(se));
261 return container_of(se, struct task_struct, se);
264 /* Walk up scheduling entities hierarchy */
265 #define for_each_sched_entity(se) \
266 for (; se; se = se->parent)
268 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
273 /* runqueue on which this entity is (to be) queued */
274 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
279 /* runqueue "owned" by this group */
280 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
285 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
288 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
290 if (!cfs_rq->on_list) {
292 * Ensure we either appear before our parent (if already
293 * enqueued) or force our parent to appear after us when it is
294 * enqueued. The fact that we always enqueue bottom-up
295 * reduces this to two cases.
297 if (cfs_rq->tg->parent &&
298 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
299 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
300 &rq_of(cfs_rq)->leaf_cfs_rq_list);
302 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
303 &rq_of(cfs_rq)->leaf_cfs_rq_list);
307 /* We should have no load, but we need to update last_decay. */
308 update_cfs_rq_blocked_load(cfs_rq, 0);
312 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
314 if (cfs_rq->on_list) {
315 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
320 /* Iterate thr' all leaf cfs_rq's on a runqueue */
321 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
322 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
324 /* Do the two (enqueued) entities belong to the same group ? */
325 static inline struct cfs_rq *
326 is_same_group(struct sched_entity *se, struct sched_entity *pse)
328 if (se->cfs_rq == pse->cfs_rq)
334 static inline struct sched_entity *parent_entity(struct sched_entity *se)
340 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
342 int se_depth, pse_depth;
345 * preemption test can be made between sibling entities who are in the
346 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
347 * both tasks until we find their ancestors who are siblings of common
351 /* First walk up until both entities are at same depth */
352 se_depth = (*se)->depth;
353 pse_depth = (*pse)->depth;
355 while (se_depth > pse_depth) {
357 *se = parent_entity(*se);
360 while (pse_depth > se_depth) {
362 *pse = parent_entity(*pse);
365 while (!is_same_group(*se, *pse)) {
366 *se = parent_entity(*se);
367 *pse = parent_entity(*pse);
371 #else /* !CONFIG_FAIR_GROUP_SCHED */
373 static inline struct task_struct *task_of(struct sched_entity *se)
375 return container_of(se, struct task_struct, se);
378 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
380 return container_of(cfs_rq, struct rq, cfs);
383 #define entity_is_task(se) 1
385 #define for_each_sched_entity(se) \
386 for (; se; se = NULL)
388 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
390 return &task_rq(p)->cfs;
393 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
395 struct task_struct *p = task_of(se);
396 struct rq *rq = task_rq(p);
401 /* runqueue "owned" by this group */
402 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
407 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
411 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
415 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
416 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
418 static inline struct sched_entity *parent_entity(struct sched_entity *se)
424 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
428 #endif /* CONFIG_FAIR_GROUP_SCHED */
430 static __always_inline
431 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
433 /**************************************************************
434 * Scheduling class tree data structure manipulation methods:
437 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
439 s64 delta = (s64)(vruntime - max_vruntime);
441 max_vruntime = vruntime;
446 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
448 s64 delta = (s64)(vruntime - min_vruntime);
450 min_vruntime = vruntime;
455 static inline int entity_before(struct sched_entity *a,
456 struct sched_entity *b)
458 return (s64)(a->vruntime - b->vruntime) < 0;
461 static void update_min_vruntime(struct cfs_rq *cfs_rq)
463 u64 vruntime = cfs_rq->min_vruntime;
466 vruntime = cfs_rq->curr->vruntime;
468 if (cfs_rq->rb_leftmost) {
469 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
474 vruntime = se->vruntime;
476 vruntime = min_vruntime(vruntime, se->vruntime);
479 /* ensure we never gain time by being placed backwards. */
480 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
483 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
488 * Enqueue an entity into the rb-tree:
490 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
492 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
493 struct rb_node *parent = NULL;
494 struct sched_entity *entry;
498 * Find the right place in the rbtree:
502 entry = rb_entry(parent, struct sched_entity, run_node);
504 * We dont care about collisions. Nodes with
505 * the same key stay together.
507 if (entity_before(se, entry)) {
508 link = &parent->rb_left;
510 link = &parent->rb_right;
516 * Maintain a cache of leftmost tree entries (it is frequently
520 cfs_rq->rb_leftmost = &se->run_node;
522 rb_link_node(&se->run_node, parent, link);
523 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
526 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
528 if (cfs_rq->rb_leftmost == &se->run_node) {
529 struct rb_node *next_node;
531 next_node = rb_next(&se->run_node);
532 cfs_rq->rb_leftmost = next_node;
535 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
538 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
540 struct rb_node *left = cfs_rq->rb_leftmost;
545 return rb_entry(left, struct sched_entity, run_node);
548 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
550 struct rb_node *next = rb_next(&se->run_node);
555 return rb_entry(next, struct sched_entity, run_node);
558 #ifdef CONFIG_SCHED_DEBUG
559 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
561 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
566 return rb_entry(last, struct sched_entity, run_node);
569 /**************************************************************
570 * Scheduling class statistics methods:
573 int sched_proc_update_handler(struct ctl_table *table, int write,
574 void __user *buffer, size_t *lenp,
577 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
578 int factor = get_update_sysctl_factor();
583 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
584 sysctl_sched_min_granularity);
586 #define WRT_SYSCTL(name) \
587 (normalized_sysctl_##name = sysctl_##name / (factor))
588 WRT_SYSCTL(sched_min_granularity);
589 WRT_SYSCTL(sched_latency);
590 WRT_SYSCTL(sched_wakeup_granularity);
600 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
602 if (unlikely(se->load.weight != NICE_0_LOAD))
603 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
609 * The idea is to set a period in which each task runs once.
611 * When there are too many tasks (sched_nr_latency) we have to stretch
612 * this period because otherwise the slices get too small.
614 * p = (nr <= nl) ? l : l*nr/nl
616 static u64 __sched_period(unsigned long nr_running)
618 u64 period = sysctl_sched_latency;
619 unsigned long nr_latency = sched_nr_latency;
621 if (unlikely(nr_running > nr_latency)) {
622 period = sysctl_sched_min_granularity;
623 period *= nr_running;
630 * We calculate the wall-time slice from the period by taking a part
631 * proportional to the weight.
635 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
637 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
639 for_each_sched_entity(se) {
640 struct load_weight *load;
641 struct load_weight lw;
643 cfs_rq = cfs_rq_of(se);
644 load = &cfs_rq->load;
646 if (unlikely(!se->on_rq)) {
649 update_load_add(&lw, se->load.weight);
652 slice = __calc_delta(slice, se->load.weight, load);
658 * We calculate the vruntime slice of a to-be-inserted task.
662 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
664 return calc_delta_fair(sched_slice(cfs_rq, se), se);
668 static unsigned long task_h_load(struct task_struct *p);
670 static inline void __update_task_entity_contrib(struct sched_entity *se);
672 /* Give new task start runnable values to heavy its load in infant time */
673 void init_task_runnable_average(struct task_struct *p)
677 p->se.avg.decay_count = 0;
678 slice = sched_slice(task_cfs_rq(p), &p->se) >> 10;
679 p->se.avg.runnable_avg_sum = slice;
680 p->se.avg.runnable_avg_period = slice;
681 __update_task_entity_contrib(&p->se);
684 void init_task_runnable_average(struct task_struct *p)
690 * Update the current task's runtime statistics.
692 static void update_curr(struct cfs_rq *cfs_rq)
694 struct sched_entity *curr = cfs_rq->curr;
695 u64 now = rq_clock_task(rq_of(cfs_rq));
701 delta_exec = now - curr->exec_start;
702 if (unlikely((s64)delta_exec <= 0))
705 curr->exec_start = now;
707 schedstat_set(curr->statistics.exec_max,
708 max(delta_exec, curr->statistics.exec_max));
710 curr->sum_exec_runtime += delta_exec;
711 schedstat_add(cfs_rq, exec_clock, delta_exec);
713 curr->vruntime += calc_delta_fair(delta_exec, curr);
714 update_min_vruntime(cfs_rq);
716 if (entity_is_task(curr)) {
717 struct task_struct *curtask = task_of(curr);
719 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
720 cpuacct_charge(curtask, delta_exec);
721 account_group_exec_runtime(curtask, delta_exec);
724 account_cfs_rq_runtime(cfs_rq, delta_exec);
728 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
730 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
734 * Task is being enqueued - update stats:
736 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
739 * Are we enqueueing a waiting task? (for current tasks
740 * a dequeue/enqueue event is a NOP)
742 if (se != cfs_rq->curr)
743 update_stats_wait_start(cfs_rq, se);
747 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
749 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
750 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
751 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
752 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
753 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
754 #ifdef CONFIG_SCHEDSTATS
755 if (entity_is_task(se)) {
756 trace_sched_stat_wait(task_of(se),
757 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
760 schedstat_set(se->statistics.wait_start, 0);
764 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
767 * Mark the end of the wait period if dequeueing a
770 if (se != cfs_rq->curr)
771 update_stats_wait_end(cfs_rq, se);
775 * We are picking a new current task - update its stats:
778 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
781 * We are starting a new run period:
783 se->exec_start = rq_clock_task(rq_of(cfs_rq));
786 /**************************************************
787 * Scheduling class queueing methods:
790 #ifdef CONFIG_NUMA_BALANCING
792 * Approximate time to scan a full NUMA task in ms. The task scan period is
793 * calculated based on the tasks virtual memory size and
794 * numa_balancing_scan_size.
796 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
797 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
799 /* Portion of address space to scan in MB */
800 unsigned int sysctl_numa_balancing_scan_size = 256;
802 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
803 unsigned int sysctl_numa_balancing_scan_delay = 1000;
805 static unsigned int task_nr_scan_windows(struct task_struct *p)
807 unsigned long rss = 0;
808 unsigned long nr_scan_pages;
811 * Calculations based on RSS as non-present and empty pages are skipped
812 * by the PTE scanner and NUMA hinting faults should be trapped based
815 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
816 rss = get_mm_rss(p->mm);
820 rss = round_up(rss, nr_scan_pages);
821 return rss / nr_scan_pages;
824 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
825 #define MAX_SCAN_WINDOW 2560
827 static unsigned int task_scan_min(struct task_struct *p)
829 unsigned int scan, floor;
830 unsigned int windows = 1;
832 if (sysctl_numa_balancing_scan_size < MAX_SCAN_WINDOW)
833 windows = MAX_SCAN_WINDOW / sysctl_numa_balancing_scan_size;
834 floor = 1000 / windows;
836 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
837 return max_t(unsigned int, floor, scan);
840 static unsigned int task_scan_max(struct task_struct *p)
842 unsigned int smin = task_scan_min(p);
845 /* Watch for min being lower than max due to floor calculations */
846 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
847 return max(smin, smax);
850 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
852 rq->nr_numa_running += (p->numa_preferred_nid != -1);
853 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
856 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
858 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
859 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
865 spinlock_t lock; /* nr_tasks, tasks */
868 struct list_head task_list;
871 nodemask_t active_nodes;
872 unsigned long total_faults;
874 * Faults_cpu is used to decide whether memory should move
875 * towards the CPU. As a consequence, these stats are weighted
876 * more by CPU use than by memory faults.
878 unsigned long *faults_cpu;
879 unsigned long faults[0];
882 /* Shared or private faults. */
883 #define NR_NUMA_HINT_FAULT_TYPES 2
885 /* Memory and CPU locality */
886 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
888 /* Averaged statistics, and temporary buffers. */
889 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
891 pid_t task_numa_group_id(struct task_struct *p)
893 return p->numa_group ? p->numa_group->gid : 0;
896 static inline int task_faults_idx(int nid, int priv)
898 return NR_NUMA_HINT_FAULT_TYPES * nid + priv;
901 static inline unsigned long task_faults(struct task_struct *p, int nid)
903 if (!p->numa_faults_memory)
906 return p->numa_faults_memory[task_faults_idx(nid, 0)] +
907 p->numa_faults_memory[task_faults_idx(nid, 1)];
910 static inline unsigned long group_faults(struct task_struct *p, int nid)
915 return p->numa_group->faults[task_faults_idx(nid, 0)] +
916 p->numa_group->faults[task_faults_idx(nid, 1)];
919 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
921 return group->faults_cpu[task_faults_idx(nid, 0)] +
922 group->faults_cpu[task_faults_idx(nid, 1)];
926 * These return the fraction of accesses done by a particular task, or
927 * task group, on a particular numa node. The group weight is given a
928 * larger multiplier, in order to group tasks together that are almost
929 * evenly spread out between numa nodes.
931 static inline unsigned long task_weight(struct task_struct *p, int nid)
933 unsigned long total_faults;
935 if (!p->numa_faults_memory)
938 total_faults = p->total_numa_faults;
943 return 1000 * task_faults(p, nid) / total_faults;
946 static inline unsigned long group_weight(struct task_struct *p, int nid)
948 if (!p->numa_group || !p->numa_group->total_faults)
951 return 1000 * group_faults(p, nid) / p->numa_group->total_faults;
954 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
955 int src_nid, int dst_cpu)
957 struct numa_group *ng = p->numa_group;
958 int dst_nid = cpu_to_node(dst_cpu);
959 int last_cpupid, this_cpupid;
961 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
964 * Multi-stage node selection is used in conjunction with a periodic
965 * migration fault to build a temporal task<->page relation. By using
966 * a two-stage filter we remove short/unlikely relations.
968 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
969 * a task's usage of a particular page (n_p) per total usage of this
970 * page (n_t) (in a given time-span) to a probability.
972 * Our periodic faults will sample this probability and getting the
973 * same result twice in a row, given these samples are fully
974 * independent, is then given by P(n)^2, provided our sample period
975 * is sufficiently short compared to the usage pattern.
977 * This quadric squishes small probabilities, making it less likely we
978 * act on an unlikely task<->page relation.
980 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
981 if (!cpupid_pid_unset(last_cpupid) &&
982 cpupid_to_nid(last_cpupid) != dst_nid)
985 /* Always allow migrate on private faults */
986 if (cpupid_match_pid(p, last_cpupid))
989 /* A shared fault, but p->numa_group has not been set up yet. */
994 * Do not migrate if the destination is not a node that
995 * is actively used by this numa group.
997 if (!node_isset(dst_nid, ng->active_nodes))
1001 * Source is a node that is not actively used by this
1002 * numa group, while the destination is. Migrate.
1004 if (!node_isset(src_nid, ng->active_nodes))
1008 * Both source and destination are nodes in active
1009 * use by this numa group. Maximize memory bandwidth
1010 * by migrating from more heavily used groups, to less
1011 * heavily used ones, spreading the load around.
1012 * Use a 1/4 hysteresis to avoid spurious page movement.
1014 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1017 static unsigned long weighted_cpuload(const int cpu);
1018 static unsigned long source_load(int cpu, int type);
1019 static unsigned long target_load(int cpu, int type);
1020 static unsigned long power_of(int cpu);
1021 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1023 /* Cached statistics for all CPUs within a node */
1025 unsigned long nr_running;
1028 /* Total compute capacity of CPUs on a node */
1029 unsigned long power;
1031 /* Approximate capacity in terms of runnable tasks on a node */
1032 unsigned long capacity;
1037 * XXX borrowed from update_sg_lb_stats
1039 static void update_numa_stats(struct numa_stats *ns, int nid)
1043 memset(ns, 0, sizeof(*ns));
1044 for_each_cpu(cpu, cpumask_of_node(nid)) {
1045 struct rq *rq = cpu_rq(cpu);
1047 ns->nr_running += rq->nr_running;
1048 ns->load += weighted_cpuload(cpu);
1049 ns->power += power_of(cpu);
1055 * If we raced with hotplug and there are no CPUs left in our mask
1056 * the @ns structure is NULL'ed and task_numa_compare() will
1057 * not find this node attractive.
1059 * We'll either bail at !has_capacity, or we'll detect a huge imbalance
1065 ns->load = (ns->load * SCHED_POWER_SCALE) / ns->power;
1066 ns->capacity = DIV_ROUND_CLOSEST(ns->power, SCHED_POWER_SCALE);
1067 ns->has_capacity = (ns->nr_running < ns->capacity);
1070 struct task_numa_env {
1071 struct task_struct *p;
1073 int src_cpu, src_nid;
1074 int dst_cpu, dst_nid;
1076 struct numa_stats src_stats, dst_stats;
1080 struct task_struct *best_task;
1085 static void task_numa_assign(struct task_numa_env *env,
1086 struct task_struct *p, long imp)
1089 put_task_struct(env->best_task);
1094 env->best_imp = imp;
1095 env->best_cpu = env->dst_cpu;
1099 * This checks if the overall compute and NUMA accesses of the system would
1100 * be improved if the source tasks was migrated to the target dst_cpu taking
1101 * into account that it might be best if task running on the dst_cpu should
1102 * be exchanged with the source task
1104 static void task_numa_compare(struct task_numa_env *env,
1105 long taskimp, long groupimp)
1107 struct rq *src_rq = cpu_rq(env->src_cpu);
1108 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1109 struct task_struct *cur;
1110 long dst_load, src_load;
1112 long imp = (groupimp > 0) ? groupimp : taskimp;
1115 cur = ACCESS_ONCE(dst_rq->curr);
1116 if (cur->pid == 0) /* idle */
1120 * "imp" is the fault differential for the source task between the
1121 * source and destination node. Calculate the total differential for
1122 * the source task and potential destination task. The more negative
1123 * the value is, the more rmeote accesses that would be expected to
1124 * be incurred if the tasks were swapped.
1127 /* Skip this swap candidate if cannot move to the source cpu */
1128 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1132 * If dst and source tasks are in the same NUMA group, or not
1133 * in any group then look only at task weights.
1135 if (cur->numa_group == env->p->numa_group) {
1136 imp = taskimp + task_weight(cur, env->src_nid) -
1137 task_weight(cur, env->dst_nid);
1139 * Add some hysteresis to prevent swapping the
1140 * tasks within a group over tiny differences.
1142 if (cur->numa_group)
1146 * Compare the group weights. If a task is all by
1147 * itself (not part of a group), use the task weight
1150 if (env->p->numa_group)
1155 if (cur->numa_group)
1156 imp += group_weight(cur, env->src_nid) -
1157 group_weight(cur, env->dst_nid);
1159 imp += task_weight(cur, env->src_nid) -
1160 task_weight(cur, env->dst_nid);
1164 if (imp < env->best_imp)
1168 /* Is there capacity at our destination? */
1169 if (env->src_stats.has_capacity &&
1170 !env->dst_stats.has_capacity)
1176 /* Balance doesn't matter much if we're running a task per cpu */
1177 if (src_rq->nr_running == 1 && dst_rq->nr_running == 1)
1181 * In the overloaded case, try and keep the load balanced.
1184 dst_load = env->dst_stats.load;
1185 src_load = env->src_stats.load;
1187 /* XXX missing power terms */
1188 load = task_h_load(env->p);
1193 load = task_h_load(cur);
1198 /* make src_load the smaller */
1199 if (dst_load < src_load)
1200 swap(dst_load, src_load);
1202 if (src_load * env->imbalance_pct < dst_load * 100)
1206 task_numa_assign(env, cur, imp);
1211 static void task_numa_find_cpu(struct task_numa_env *env,
1212 long taskimp, long groupimp)
1216 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1217 /* Skip this CPU if the source task cannot migrate */
1218 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1222 task_numa_compare(env, taskimp, groupimp);
1226 static int task_numa_migrate(struct task_struct *p)
1228 struct task_numa_env env = {
1231 .src_cpu = task_cpu(p),
1232 .src_nid = task_node(p),
1234 .imbalance_pct = 112,
1240 struct sched_domain *sd;
1241 unsigned long taskweight, groupweight;
1243 long taskimp, groupimp;
1246 * Pick the lowest SD_NUMA domain, as that would have the smallest
1247 * imbalance and would be the first to start moving tasks about.
1249 * And we want to avoid any moving of tasks about, as that would create
1250 * random movement of tasks -- counter the numa conditions we're trying
1254 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1256 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1260 * Cpusets can break the scheduler domain tree into smaller
1261 * balance domains, some of which do not cross NUMA boundaries.
1262 * Tasks that are "trapped" in such domains cannot be migrated
1263 * elsewhere, so there is no point in (re)trying.
1265 if (unlikely(!sd)) {
1266 p->numa_preferred_nid = task_node(p);
1270 taskweight = task_weight(p, env.src_nid);
1271 groupweight = group_weight(p, env.src_nid);
1272 update_numa_stats(&env.src_stats, env.src_nid);
1273 env.dst_nid = p->numa_preferred_nid;
1274 taskimp = task_weight(p, env.dst_nid) - taskweight;
1275 groupimp = group_weight(p, env.dst_nid) - groupweight;
1276 update_numa_stats(&env.dst_stats, env.dst_nid);
1278 /* If the preferred nid has capacity, try to use it. */
1279 if (env.dst_stats.has_capacity)
1280 task_numa_find_cpu(&env, taskimp, groupimp);
1282 /* No space available on the preferred nid. Look elsewhere. */
1283 if (env.best_cpu == -1) {
1284 for_each_online_node(nid) {
1285 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1288 /* Only consider nodes where both task and groups benefit */
1289 taskimp = task_weight(p, nid) - taskweight;
1290 groupimp = group_weight(p, nid) - groupweight;
1291 if (taskimp < 0 && groupimp < 0)
1295 update_numa_stats(&env.dst_stats, env.dst_nid);
1296 task_numa_find_cpu(&env, taskimp, groupimp);
1300 /* No better CPU than the current one was found. */
1301 if (env.best_cpu == -1)
1304 sched_setnuma(p, env.dst_nid);
1307 * Reset the scan period if the task is being rescheduled on an
1308 * alternative node to recheck if the tasks is now properly placed.
1310 p->numa_scan_period = task_scan_min(p);
1312 if (env.best_task == NULL) {
1313 ret = migrate_task_to(p, env.best_cpu);
1315 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1319 ret = migrate_swap(p, env.best_task);
1321 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1322 put_task_struct(env.best_task);
1326 /* Attempt to migrate a task to a CPU on the preferred node. */
1327 static void numa_migrate_preferred(struct task_struct *p)
1329 unsigned long interval = HZ;
1331 /* This task has no NUMA fault statistics yet */
1332 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults_memory))
1335 /* Periodically retry migrating the task to the preferred node */
1336 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1337 p->numa_migrate_retry = jiffies + interval;
1339 /* Success if task is already running on preferred CPU */
1340 if (task_node(p) == p->numa_preferred_nid)
1343 /* Otherwise, try migrate to a CPU on the preferred node */
1344 task_numa_migrate(p);
1348 * Find the nodes on which the workload is actively running. We do this by
1349 * tracking the nodes from which NUMA hinting faults are triggered. This can
1350 * be different from the set of nodes where the workload's memory is currently
1353 * The bitmask is used to make smarter decisions on when to do NUMA page
1354 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1355 * are added when they cause over 6/16 of the maximum number of faults, but
1356 * only removed when they drop below 3/16.
1358 static void update_numa_active_node_mask(struct numa_group *numa_group)
1360 unsigned long faults, max_faults = 0;
1363 for_each_online_node(nid) {
1364 faults = group_faults_cpu(numa_group, nid);
1365 if (faults > max_faults)
1366 max_faults = faults;
1369 for_each_online_node(nid) {
1370 faults = group_faults_cpu(numa_group, nid);
1371 if (!node_isset(nid, numa_group->active_nodes)) {
1372 if (faults > max_faults * 6 / 16)
1373 node_set(nid, numa_group->active_nodes);
1374 } else if (faults < max_faults * 3 / 16)
1375 node_clear(nid, numa_group->active_nodes);
1380 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1381 * increments. The more local the fault statistics are, the higher the scan
1382 * period will be for the next scan window. If local/remote ratio is below
1383 * NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) the
1384 * scan period will decrease
1386 #define NUMA_PERIOD_SLOTS 10
1387 #define NUMA_PERIOD_THRESHOLD 3
1390 * Increase the scan period (slow down scanning) if the majority of
1391 * our memory is already on our local node, or if the majority of
1392 * the page accesses are shared with other processes.
1393 * Otherwise, decrease the scan period.
1395 static void update_task_scan_period(struct task_struct *p,
1396 unsigned long shared, unsigned long private)
1398 unsigned int period_slot;
1402 unsigned long remote = p->numa_faults_locality[0];
1403 unsigned long local = p->numa_faults_locality[1];
1406 * If there were no record hinting faults then either the task is
1407 * completely idle or all activity is areas that are not of interest
1408 * to automatic numa balancing. Scan slower
1410 if (local + shared == 0) {
1411 p->numa_scan_period = min(p->numa_scan_period_max,
1412 p->numa_scan_period << 1);
1414 p->mm->numa_next_scan = jiffies +
1415 msecs_to_jiffies(p->numa_scan_period);
1421 * Prepare to scale scan period relative to the current period.
1422 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1423 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1424 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1426 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1427 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1428 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1429 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1432 diff = slot * period_slot;
1434 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1437 * Scale scan rate increases based on sharing. There is an
1438 * inverse relationship between the degree of sharing and
1439 * the adjustment made to the scanning period. Broadly
1440 * speaking the intent is that there is little point
1441 * scanning faster if shared accesses dominate as it may
1442 * simply bounce migrations uselessly
1444 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared));
1445 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1448 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1449 task_scan_min(p), task_scan_max(p));
1450 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1454 * Get the fraction of time the task has been running since the last
1455 * NUMA placement cycle. The scheduler keeps similar statistics, but
1456 * decays those on a 32ms period, which is orders of magnitude off
1457 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1458 * stats only if the task is so new there are no NUMA statistics yet.
1460 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1462 u64 runtime, delta, now;
1463 /* Use the start of this time slice to avoid calculations. */
1464 now = p->se.exec_start;
1465 runtime = p->se.sum_exec_runtime;
1467 if (p->last_task_numa_placement) {
1468 delta = runtime - p->last_sum_exec_runtime;
1469 *period = now - p->last_task_numa_placement;
1471 delta = p->se.avg.runnable_avg_sum;
1472 *period = p->se.avg.runnable_avg_period;
1475 p->last_sum_exec_runtime = runtime;
1476 p->last_task_numa_placement = now;
1481 static void task_numa_placement(struct task_struct *p)
1483 int seq, nid, max_nid = -1, max_group_nid = -1;
1484 unsigned long max_faults = 0, max_group_faults = 0;
1485 unsigned long fault_types[2] = { 0, 0 };
1486 unsigned long total_faults;
1487 u64 runtime, period;
1488 spinlock_t *group_lock = NULL;
1490 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1491 if (p->numa_scan_seq == seq)
1493 p->numa_scan_seq = seq;
1494 p->numa_scan_period_max = task_scan_max(p);
1496 total_faults = p->numa_faults_locality[0] +
1497 p->numa_faults_locality[1];
1498 runtime = numa_get_avg_runtime(p, &period);
1500 /* If the task is part of a group prevent parallel updates to group stats */
1501 if (p->numa_group) {
1502 group_lock = &p->numa_group->lock;
1503 spin_lock_irq(group_lock);
1506 /* Find the node with the highest number of faults */
1507 for_each_online_node(nid) {
1508 unsigned long faults = 0, group_faults = 0;
1511 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1512 long diff, f_diff, f_weight;
1514 i = task_faults_idx(nid, priv);
1516 /* Decay existing window, copy faults since last scan */
1517 diff = p->numa_faults_buffer_memory[i] - p->numa_faults_memory[i] / 2;
1518 fault_types[priv] += p->numa_faults_buffer_memory[i];
1519 p->numa_faults_buffer_memory[i] = 0;
1522 * Normalize the faults_from, so all tasks in a group
1523 * count according to CPU use, instead of by the raw
1524 * number of faults. Tasks with little runtime have
1525 * little over-all impact on throughput, and thus their
1526 * faults are less important.
1528 f_weight = div64_u64(runtime << 16, period + 1);
1529 f_weight = (f_weight * p->numa_faults_buffer_cpu[i]) /
1531 f_diff = f_weight - p->numa_faults_cpu[i] / 2;
1532 p->numa_faults_buffer_cpu[i] = 0;
1534 p->numa_faults_memory[i] += diff;
1535 p->numa_faults_cpu[i] += f_diff;
1536 faults += p->numa_faults_memory[i];
1537 p->total_numa_faults += diff;
1538 if (p->numa_group) {
1539 /* safe because we can only change our own group */
1540 p->numa_group->faults[i] += diff;
1541 p->numa_group->faults_cpu[i] += f_diff;
1542 p->numa_group->total_faults += diff;
1543 group_faults += p->numa_group->faults[i];
1547 if (faults > max_faults) {
1548 max_faults = faults;
1552 if (group_faults > max_group_faults) {
1553 max_group_faults = group_faults;
1554 max_group_nid = nid;
1558 update_task_scan_period(p, fault_types[0], fault_types[1]);
1560 if (p->numa_group) {
1561 update_numa_active_node_mask(p->numa_group);
1563 * If the preferred task and group nids are different,
1564 * iterate over the nodes again to find the best place.
1566 if (max_nid != max_group_nid) {
1567 unsigned long weight, max_weight = 0;
1569 for_each_online_node(nid) {
1570 weight = task_weight(p, nid) + group_weight(p, nid);
1571 if (weight > max_weight) {
1572 max_weight = weight;
1578 spin_unlock_irq(group_lock);
1581 /* Preferred node as the node with the most faults */
1582 if (max_faults && max_nid != p->numa_preferred_nid) {
1583 /* Update the preferred nid and migrate task if possible */
1584 sched_setnuma(p, max_nid);
1585 numa_migrate_preferred(p);
1589 static inline int get_numa_group(struct numa_group *grp)
1591 return atomic_inc_not_zero(&grp->refcount);
1594 static inline void put_numa_group(struct numa_group *grp)
1596 if (atomic_dec_and_test(&grp->refcount))
1597 kfree_rcu(grp, rcu);
1600 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1603 struct numa_group *grp, *my_grp;
1604 struct task_struct *tsk;
1606 int cpu = cpupid_to_cpu(cpupid);
1609 if (unlikely(!p->numa_group)) {
1610 unsigned int size = sizeof(struct numa_group) +
1611 4*nr_node_ids*sizeof(unsigned long);
1613 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1617 atomic_set(&grp->refcount, 1);
1618 spin_lock_init(&grp->lock);
1619 INIT_LIST_HEAD(&grp->task_list);
1621 /* Second half of the array tracks nids where faults happen */
1622 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1625 node_set(task_node(current), grp->active_nodes);
1627 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1628 grp->faults[i] = p->numa_faults_memory[i];
1630 grp->total_faults = p->total_numa_faults;
1632 list_add(&p->numa_entry, &grp->task_list);
1634 rcu_assign_pointer(p->numa_group, grp);
1638 tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1640 if (!cpupid_match_pid(tsk, cpupid))
1643 grp = rcu_dereference(tsk->numa_group);
1647 my_grp = p->numa_group;
1652 * Only join the other group if its bigger; if we're the bigger group,
1653 * the other task will join us.
1655 if (my_grp->nr_tasks > grp->nr_tasks)
1659 * Tie-break on the grp address.
1661 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1664 /* Always join threads in the same process. */
1665 if (tsk->mm == current->mm)
1668 /* Simple filter to avoid false positives due to PID collisions */
1669 if (flags & TNF_SHARED)
1672 /* Update priv based on whether false sharing was detected */
1675 if (join && !get_numa_group(grp))
1683 BUG_ON(irqs_disabled());
1684 double_lock_irq(&my_grp->lock, &grp->lock);
1686 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
1687 my_grp->faults[i] -= p->numa_faults_memory[i];
1688 grp->faults[i] += p->numa_faults_memory[i];
1690 my_grp->total_faults -= p->total_numa_faults;
1691 grp->total_faults += p->total_numa_faults;
1693 list_move(&p->numa_entry, &grp->task_list);
1697 spin_unlock(&my_grp->lock);
1698 spin_unlock_irq(&grp->lock);
1700 rcu_assign_pointer(p->numa_group, grp);
1702 put_numa_group(my_grp);
1710 void task_numa_free(struct task_struct *p)
1712 struct numa_group *grp = p->numa_group;
1714 void *numa_faults = p->numa_faults_memory;
1717 spin_lock_irq(&grp->lock);
1718 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1719 grp->faults[i] -= p->numa_faults_memory[i];
1720 grp->total_faults -= p->total_numa_faults;
1722 list_del(&p->numa_entry);
1724 spin_unlock_irq(&grp->lock);
1725 rcu_assign_pointer(p->numa_group, NULL);
1726 put_numa_group(grp);
1729 p->numa_faults_memory = NULL;
1730 p->numa_faults_buffer_memory = NULL;
1731 p->numa_faults_cpu= NULL;
1732 p->numa_faults_buffer_cpu = NULL;
1737 * Got a PROT_NONE fault for a page on @node.
1739 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
1741 struct task_struct *p = current;
1742 bool migrated = flags & TNF_MIGRATED;
1743 int cpu_node = task_node(current);
1744 int local = !!(flags & TNF_FAULT_LOCAL);
1747 if (!numabalancing_enabled)
1750 /* for example, ksmd faulting in a user's mm */
1754 /* Do not worry about placement if exiting */
1755 if (p->state == TASK_DEAD)
1758 /* Allocate buffer to track faults on a per-node basis */
1759 if (unlikely(!p->numa_faults_memory)) {
1760 int size = sizeof(*p->numa_faults_memory) *
1761 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
1763 p->numa_faults_memory = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
1764 if (!p->numa_faults_memory)
1767 BUG_ON(p->numa_faults_buffer_memory);
1769 * The averaged statistics, shared & private, memory & cpu,
1770 * occupy the first half of the array. The second half of the
1771 * array is for current counters, which are averaged into the
1772 * first set by task_numa_placement.
1774 p->numa_faults_cpu = p->numa_faults_memory + (2 * nr_node_ids);
1775 p->numa_faults_buffer_memory = p->numa_faults_memory + (4 * nr_node_ids);
1776 p->numa_faults_buffer_cpu = p->numa_faults_memory + (6 * nr_node_ids);
1777 p->total_numa_faults = 0;
1778 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1782 * First accesses are treated as private, otherwise consider accesses
1783 * to be private if the accessing pid has not changed
1785 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
1788 priv = cpupid_match_pid(p, last_cpupid);
1789 if (!priv && !(flags & TNF_NO_GROUP))
1790 task_numa_group(p, last_cpupid, flags, &priv);
1794 * If a workload spans multiple NUMA nodes, a shared fault that
1795 * occurs wholly within the set of nodes that the workload is
1796 * actively using should be counted as local. This allows the
1797 * scan rate to slow down when a workload has settled down.
1799 if (!priv && !local && p->numa_group &&
1800 node_isset(cpu_node, p->numa_group->active_nodes) &&
1801 node_isset(mem_node, p->numa_group->active_nodes))
1804 task_numa_placement(p);
1807 * Retry task to preferred node migration periodically, in case it
1808 * case it previously failed, or the scheduler moved us.
1810 if (time_after(jiffies, p->numa_migrate_retry))
1811 numa_migrate_preferred(p);
1814 p->numa_pages_migrated += pages;
1816 p->numa_faults_buffer_memory[task_faults_idx(mem_node, priv)] += pages;
1817 p->numa_faults_buffer_cpu[task_faults_idx(cpu_node, priv)] += pages;
1818 p->numa_faults_locality[local] += pages;
1821 static void reset_ptenuma_scan(struct task_struct *p)
1823 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1824 p->mm->numa_scan_offset = 0;
1828 * The expensive part of numa migration is done from task_work context.
1829 * Triggered from task_tick_numa().
1831 void task_numa_work(struct callback_head *work)
1833 unsigned long migrate, next_scan, now = jiffies;
1834 struct task_struct *p = current;
1835 struct mm_struct *mm = p->mm;
1836 struct vm_area_struct *vma;
1837 unsigned long start, end;
1838 unsigned long nr_pte_updates = 0;
1841 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1843 work->next = work; /* protect against double add */
1845 * Who cares about NUMA placement when they're dying.
1847 * NOTE: make sure not to dereference p->mm before this check,
1848 * exit_task_work() happens _after_ exit_mm() so we could be called
1849 * without p->mm even though we still had it when we enqueued this
1852 if (p->flags & PF_EXITING)
1855 if (!mm->numa_next_scan) {
1856 mm->numa_next_scan = now +
1857 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1861 * Enforce maximal scan/migration frequency..
1863 migrate = mm->numa_next_scan;
1864 if (time_before(now, migrate))
1867 if (p->numa_scan_period == 0) {
1868 p->numa_scan_period_max = task_scan_max(p);
1869 p->numa_scan_period = task_scan_min(p);
1872 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1873 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1877 * Delay this task enough that another task of this mm will likely win
1878 * the next time around.
1880 p->node_stamp += 2 * TICK_NSEC;
1882 start = mm->numa_scan_offset;
1883 pages = sysctl_numa_balancing_scan_size;
1884 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1888 down_read(&mm->mmap_sem);
1889 vma = find_vma(mm, start);
1891 reset_ptenuma_scan(p);
1895 for (; vma; vma = vma->vm_next) {
1896 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1900 * Shared library pages mapped by multiple processes are not
1901 * migrated as it is expected they are cache replicated. Avoid
1902 * hinting faults in read-only file-backed mappings or the vdso
1903 * as migrating the pages will be of marginal benefit.
1906 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1910 * Skip inaccessible VMAs to avoid any confusion between
1911 * PROT_NONE and NUMA hinting ptes
1913 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
1917 start = max(start, vma->vm_start);
1918 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1919 end = min(end, vma->vm_end);
1920 nr_pte_updates += change_prot_numa(vma, start, end);
1923 * Scan sysctl_numa_balancing_scan_size but ensure that
1924 * at least one PTE is updated so that unused virtual
1925 * address space is quickly skipped.
1928 pages -= (end - start) >> PAGE_SHIFT;
1935 } while (end != vma->vm_end);
1940 * It is possible to reach the end of the VMA list but the last few
1941 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1942 * would find the !migratable VMA on the next scan but not reset the
1943 * scanner to the start so check it now.
1946 mm->numa_scan_offset = start;
1948 reset_ptenuma_scan(p);
1949 up_read(&mm->mmap_sem);
1953 * Drive the periodic memory faults..
1955 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1957 struct callback_head *work = &curr->numa_work;
1961 * We don't care about NUMA placement if we don't have memory.
1963 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1967 * Using runtime rather than walltime has the dual advantage that
1968 * we (mostly) drive the selection from busy threads and that the
1969 * task needs to have done some actual work before we bother with
1972 now = curr->se.sum_exec_runtime;
1973 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1975 if (now - curr->node_stamp > period) {
1976 if (!curr->node_stamp)
1977 curr->numa_scan_period = task_scan_min(curr);
1978 curr->node_stamp += period;
1980 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1981 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1982 task_work_add(curr, work, true);
1987 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1991 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1995 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1998 #endif /* CONFIG_NUMA_BALANCING */
2001 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2003 update_load_add(&cfs_rq->load, se->load.weight);
2004 if (!parent_entity(se))
2005 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2007 if (entity_is_task(se)) {
2008 struct rq *rq = rq_of(cfs_rq);
2010 account_numa_enqueue(rq, task_of(se));
2011 list_add(&se->group_node, &rq->cfs_tasks);
2014 cfs_rq->nr_running++;
2018 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2020 update_load_sub(&cfs_rq->load, se->load.weight);
2021 if (!parent_entity(se))
2022 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2023 if (entity_is_task(se)) {
2024 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2025 list_del_init(&se->group_node);
2027 cfs_rq->nr_running--;
2030 #ifdef CONFIG_FAIR_GROUP_SCHED
2032 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2037 * Use this CPU's actual weight instead of the last load_contribution
2038 * to gain a more accurate current total weight. See
2039 * update_cfs_rq_load_contribution().
2041 tg_weight = atomic_long_read(&tg->load_avg);
2042 tg_weight -= cfs_rq->tg_load_contrib;
2043 tg_weight += cfs_rq->load.weight;
2048 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2050 long tg_weight, load, shares;
2052 tg_weight = calc_tg_weight(tg, cfs_rq);
2053 load = cfs_rq->load.weight;
2055 shares = (tg->shares * load);
2057 shares /= tg_weight;
2059 if (shares < MIN_SHARES)
2060 shares = MIN_SHARES;
2061 if (shares > tg->shares)
2062 shares = tg->shares;
2066 # else /* CONFIG_SMP */
2067 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2071 # endif /* CONFIG_SMP */
2072 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2073 unsigned long weight)
2076 /* commit outstanding execution time */
2077 if (cfs_rq->curr == se)
2078 update_curr(cfs_rq);
2079 account_entity_dequeue(cfs_rq, se);
2082 update_load_set(&se->load, weight);
2085 account_entity_enqueue(cfs_rq, se);
2088 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2090 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2092 struct task_group *tg;
2093 struct sched_entity *se;
2097 se = tg->se[cpu_of(rq_of(cfs_rq))];
2098 if (!se || throttled_hierarchy(cfs_rq))
2101 if (likely(se->load.weight == tg->shares))
2104 shares = calc_cfs_shares(cfs_rq, tg);
2106 reweight_entity(cfs_rq_of(se), se, shares);
2108 #else /* CONFIG_FAIR_GROUP_SCHED */
2109 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2112 #endif /* CONFIG_FAIR_GROUP_SCHED */
2116 * We choose a half-life close to 1 scheduling period.
2117 * Note: The tables below are dependent on this value.
2119 #define LOAD_AVG_PERIOD 32
2120 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2121 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2123 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2124 static const u32 runnable_avg_yN_inv[] = {
2125 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2126 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2127 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2128 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2129 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2130 0x85aac367, 0x82cd8698,
2134 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2135 * over-estimates when re-combining.
2137 static const u32 runnable_avg_yN_sum[] = {
2138 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2139 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2140 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2145 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2147 static __always_inline u64 decay_load(u64 val, u64 n)
2149 unsigned int local_n;
2153 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2156 /* after bounds checking we can collapse to 32-bit */
2160 * As y^PERIOD = 1/2, we can combine
2161 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
2162 * With a look-up table which covers k^n (n<PERIOD)
2164 * To achieve constant time decay_load.
2166 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2167 val >>= local_n / LOAD_AVG_PERIOD;
2168 local_n %= LOAD_AVG_PERIOD;
2171 val *= runnable_avg_yN_inv[local_n];
2172 /* We don't use SRR here since we always want to round down. */
2177 * For updates fully spanning n periods, the contribution to runnable
2178 * average will be: \Sum 1024*y^n
2180 * We can compute this reasonably efficiently by combining:
2181 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2183 static u32 __compute_runnable_contrib(u64 n)
2187 if (likely(n <= LOAD_AVG_PERIOD))
2188 return runnable_avg_yN_sum[n];
2189 else if (unlikely(n >= LOAD_AVG_MAX_N))
2190 return LOAD_AVG_MAX;
2192 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2194 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2195 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2197 n -= LOAD_AVG_PERIOD;
2198 } while (n > LOAD_AVG_PERIOD);
2200 contrib = decay_load(contrib, n);
2201 return contrib + runnable_avg_yN_sum[n];
2205 * We can represent the historical contribution to runnable average as the
2206 * coefficients of a geometric series. To do this we sub-divide our runnable
2207 * history into segments of approximately 1ms (1024us); label the segment that
2208 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2210 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2212 * (now) (~1ms ago) (~2ms ago)
2214 * Let u_i denote the fraction of p_i that the entity was runnable.
2216 * We then designate the fractions u_i as our co-efficients, yielding the
2217 * following representation of historical load:
2218 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2220 * We choose y based on the with of a reasonably scheduling period, fixing:
2223 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2224 * approximately half as much as the contribution to load within the last ms
2227 * When a period "rolls over" and we have new u_0`, multiplying the previous
2228 * sum again by y is sufficient to update:
2229 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2230 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2232 static __always_inline int __update_entity_runnable_avg(u64 now,
2233 struct sched_avg *sa,
2237 u32 runnable_contrib;
2238 int delta_w, decayed = 0;
2240 delta = now - sa->last_runnable_update;
2242 * This should only happen when time goes backwards, which it
2243 * unfortunately does during sched clock init when we swap over to TSC.
2245 if ((s64)delta < 0) {
2246 sa->last_runnable_update = now;
2251 * Use 1024ns as the unit of measurement since it's a reasonable
2252 * approximation of 1us and fast to compute.
2257 sa->last_runnable_update = now;
2259 /* delta_w is the amount already accumulated against our next period */
2260 delta_w = sa->runnable_avg_period % 1024;
2261 if (delta + delta_w >= 1024) {
2262 /* period roll-over */
2266 * Now that we know we're crossing a period boundary, figure
2267 * out how much from delta we need to complete the current
2268 * period and accrue it.
2270 delta_w = 1024 - delta_w;
2272 sa->runnable_avg_sum += delta_w;
2273 sa->runnable_avg_period += delta_w;
2277 /* Figure out how many additional periods this update spans */
2278 periods = delta / 1024;
2281 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2283 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
2286 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2287 runnable_contrib = __compute_runnable_contrib(periods);
2289 sa->runnable_avg_sum += runnable_contrib;
2290 sa->runnable_avg_period += runnable_contrib;
2293 /* Remainder of delta accrued against u_0` */
2295 sa->runnable_avg_sum += delta;
2296 sa->runnable_avg_period += delta;
2301 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2302 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2304 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2305 u64 decays = atomic64_read(&cfs_rq->decay_counter);
2307 decays -= se->avg.decay_count;
2311 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2312 se->avg.decay_count = 0;
2317 #ifdef CONFIG_FAIR_GROUP_SCHED
2318 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2321 struct task_group *tg = cfs_rq->tg;
2324 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2325 tg_contrib -= cfs_rq->tg_load_contrib;
2327 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2328 atomic_long_add(tg_contrib, &tg->load_avg);
2329 cfs_rq->tg_load_contrib += tg_contrib;
2334 * Aggregate cfs_rq runnable averages into an equivalent task_group
2335 * representation for computing load contributions.
2337 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2338 struct cfs_rq *cfs_rq)
2340 struct task_group *tg = cfs_rq->tg;
2343 /* The fraction of a cpu used by this cfs_rq */
2344 contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT,
2345 sa->runnable_avg_period + 1);
2346 contrib -= cfs_rq->tg_runnable_contrib;
2348 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2349 atomic_add(contrib, &tg->runnable_avg);
2350 cfs_rq->tg_runnable_contrib += contrib;
2354 static inline void __update_group_entity_contrib(struct sched_entity *se)
2356 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2357 struct task_group *tg = cfs_rq->tg;
2362 contrib = cfs_rq->tg_load_contrib * tg->shares;
2363 se->avg.load_avg_contrib = div_u64(contrib,
2364 atomic_long_read(&tg->load_avg) + 1);
2367 * For group entities we need to compute a correction term in the case
2368 * that they are consuming <1 cpu so that we would contribute the same
2369 * load as a task of equal weight.
2371 * Explicitly co-ordinating this measurement would be expensive, but
2372 * fortunately the sum of each cpus contribution forms a usable
2373 * lower-bound on the true value.
2375 * Consider the aggregate of 2 contributions. Either they are disjoint
2376 * (and the sum represents true value) or they are disjoint and we are
2377 * understating by the aggregate of their overlap.
2379 * Extending this to N cpus, for a given overlap, the maximum amount we
2380 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2381 * cpus that overlap for this interval and w_i is the interval width.
2383 * On a small machine; the first term is well-bounded which bounds the
2384 * total error since w_i is a subset of the period. Whereas on a
2385 * larger machine, while this first term can be larger, if w_i is the
2386 * of consequential size guaranteed to see n_i*w_i quickly converge to
2387 * our upper bound of 1-cpu.
2389 runnable_avg = atomic_read(&tg->runnable_avg);
2390 if (runnable_avg < NICE_0_LOAD) {
2391 se->avg.load_avg_contrib *= runnable_avg;
2392 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2396 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2398 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
2399 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2401 #else /* CONFIG_FAIR_GROUP_SCHED */
2402 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2403 int force_update) {}
2404 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2405 struct cfs_rq *cfs_rq) {}
2406 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2407 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2408 #endif /* CONFIG_FAIR_GROUP_SCHED */
2410 static inline void __update_task_entity_contrib(struct sched_entity *se)
2414 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2415 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2416 contrib /= (se->avg.runnable_avg_period + 1);
2417 se->avg.load_avg_contrib = scale_load(contrib);
2420 /* Compute the current contribution to load_avg by se, return any delta */
2421 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2423 long old_contrib = se->avg.load_avg_contrib;
2425 if (entity_is_task(se)) {
2426 __update_task_entity_contrib(se);
2428 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2429 __update_group_entity_contrib(se);
2432 return se->avg.load_avg_contrib - old_contrib;
2435 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2438 if (likely(load_contrib < cfs_rq->blocked_load_avg))
2439 cfs_rq->blocked_load_avg -= load_contrib;
2441 cfs_rq->blocked_load_avg = 0;
2444 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2446 /* Update a sched_entity's runnable average */
2447 static inline void update_entity_load_avg(struct sched_entity *se,
2450 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2455 * For a group entity we need to use their owned cfs_rq_clock_task() in
2456 * case they are the parent of a throttled hierarchy.
2458 if (entity_is_task(se))
2459 now = cfs_rq_clock_task(cfs_rq);
2461 now = cfs_rq_clock_task(group_cfs_rq(se));
2463 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
2466 contrib_delta = __update_entity_load_avg_contrib(se);
2472 cfs_rq->runnable_load_avg += contrib_delta;
2474 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2478 * Decay the load contributed by all blocked children and account this so that
2479 * their contribution may appropriately discounted when they wake up.
2481 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2483 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2486 decays = now - cfs_rq->last_decay;
2487 if (!decays && !force_update)
2490 if (atomic_long_read(&cfs_rq->removed_load)) {
2491 unsigned long removed_load;
2492 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2493 subtract_blocked_load_contrib(cfs_rq, removed_load);
2497 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2499 atomic64_add(decays, &cfs_rq->decay_counter);
2500 cfs_rq->last_decay = now;
2503 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2506 /* Add the load generated by se into cfs_rq's child load-average */
2507 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2508 struct sched_entity *se,
2512 * We track migrations using entity decay_count <= 0, on a wake-up
2513 * migration we use a negative decay count to track the remote decays
2514 * accumulated while sleeping.
2516 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2517 * are seen by enqueue_entity_load_avg() as a migration with an already
2518 * constructed load_avg_contrib.
2520 if (unlikely(se->avg.decay_count <= 0)) {
2521 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2522 if (se->avg.decay_count) {
2524 * In a wake-up migration we have to approximate the
2525 * time sleeping. This is because we can't synchronize
2526 * clock_task between the two cpus, and it is not
2527 * guaranteed to be read-safe. Instead, we can
2528 * approximate this using our carried decays, which are
2529 * explicitly atomically readable.
2531 se->avg.last_runnable_update -= (-se->avg.decay_count)
2533 update_entity_load_avg(se, 0);
2534 /* Indicate that we're now synchronized and on-rq */
2535 se->avg.decay_count = 0;
2539 __synchronize_entity_decay(se);
2542 /* migrated tasks did not contribute to our blocked load */
2544 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2545 update_entity_load_avg(se, 0);
2548 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2549 /* we force update consideration on load-balancer moves */
2550 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2554 * Remove se's load from this cfs_rq child load-average, if the entity is
2555 * transitioning to a blocked state we track its projected decay using
2558 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2559 struct sched_entity *se,
2562 update_entity_load_avg(se, 1);
2563 /* we force update consideration on load-balancer moves */
2564 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2566 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2568 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2569 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2570 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2574 * Update the rq's load with the elapsed running time before entering
2575 * idle. if the last scheduled task is not a CFS task, idle_enter will
2576 * be the only way to update the runnable statistic.
2578 void idle_enter_fair(struct rq *this_rq)
2580 update_rq_runnable_avg(this_rq, 1);
2584 * Update the rq's load with the elapsed idle time before a task is
2585 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2586 * be the only way to update the runnable statistic.
2588 void idle_exit_fair(struct rq *this_rq)
2590 update_rq_runnable_avg(this_rq, 0);
2593 static int idle_balance(struct rq *this_rq);
2595 #else /* CONFIG_SMP */
2597 static inline void update_entity_load_avg(struct sched_entity *se,
2598 int update_cfs_rq) {}
2599 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2600 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2601 struct sched_entity *se,
2603 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2604 struct sched_entity *se,
2606 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2607 int force_update) {}
2609 static inline int idle_balance(struct rq *rq)
2614 #endif /* CONFIG_SMP */
2616 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2618 #ifdef CONFIG_SCHEDSTATS
2619 struct task_struct *tsk = NULL;
2621 if (entity_is_task(se))
2624 if (se->statistics.sleep_start) {
2625 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2630 if (unlikely(delta > se->statistics.sleep_max))
2631 se->statistics.sleep_max = delta;
2633 se->statistics.sleep_start = 0;
2634 se->statistics.sum_sleep_runtime += delta;
2637 account_scheduler_latency(tsk, delta >> 10, 1);
2638 trace_sched_stat_sleep(tsk, delta);
2641 if (se->statistics.block_start) {
2642 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2647 if (unlikely(delta > se->statistics.block_max))
2648 se->statistics.block_max = delta;
2650 se->statistics.block_start = 0;
2651 se->statistics.sum_sleep_runtime += delta;
2654 if (tsk->in_iowait) {
2655 se->statistics.iowait_sum += delta;
2656 se->statistics.iowait_count++;
2657 trace_sched_stat_iowait(tsk, delta);
2660 trace_sched_stat_blocked(tsk, delta);
2663 * Blocking time is in units of nanosecs, so shift by
2664 * 20 to get a milliseconds-range estimation of the
2665 * amount of time that the task spent sleeping:
2667 if (unlikely(prof_on == SLEEP_PROFILING)) {
2668 profile_hits(SLEEP_PROFILING,
2669 (void *)get_wchan(tsk),
2672 account_scheduler_latency(tsk, delta >> 10, 0);
2678 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2680 #ifdef CONFIG_SCHED_DEBUG
2681 s64 d = se->vruntime - cfs_rq->min_vruntime;
2686 if (d > 3*sysctl_sched_latency)
2687 schedstat_inc(cfs_rq, nr_spread_over);
2692 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2694 u64 vruntime = cfs_rq->min_vruntime;
2697 * The 'current' period is already promised to the current tasks,
2698 * however the extra weight of the new task will slow them down a
2699 * little, place the new task so that it fits in the slot that
2700 * stays open at the end.
2702 if (initial && sched_feat(START_DEBIT))
2703 vruntime += sched_vslice(cfs_rq, se);
2705 /* sleeps up to a single latency don't count. */
2707 unsigned long thresh = sysctl_sched_latency;
2710 * Halve their sleep time's effect, to allow
2711 * for a gentler effect of sleepers:
2713 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2719 /* ensure we never gain time by being placed backwards. */
2720 se->vruntime = max_vruntime(se->vruntime, vruntime);
2723 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2726 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2729 * Update the normalized vruntime before updating min_vruntime
2730 * through calling update_curr().
2732 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2733 se->vruntime += cfs_rq->min_vruntime;
2736 * Update run-time statistics of the 'current'.
2738 update_curr(cfs_rq);
2739 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2740 account_entity_enqueue(cfs_rq, se);
2741 update_cfs_shares(cfs_rq);
2743 if (flags & ENQUEUE_WAKEUP) {
2744 place_entity(cfs_rq, se, 0);
2745 enqueue_sleeper(cfs_rq, se);
2748 update_stats_enqueue(cfs_rq, se);
2749 check_spread(cfs_rq, se);
2750 if (se != cfs_rq->curr)
2751 __enqueue_entity(cfs_rq, se);
2754 if (cfs_rq->nr_running == 1) {
2755 list_add_leaf_cfs_rq(cfs_rq);
2756 check_enqueue_throttle(cfs_rq);
2760 static void __clear_buddies_last(struct sched_entity *se)
2762 for_each_sched_entity(se) {
2763 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2764 if (cfs_rq->last != se)
2767 cfs_rq->last = NULL;
2771 static void __clear_buddies_next(struct sched_entity *se)
2773 for_each_sched_entity(se) {
2774 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2775 if (cfs_rq->next != se)
2778 cfs_rq->next = NULL;
2782 static void __clear_buddies_skip(struct sched_entity *se)
2784 for_each_sched_entity(se) {
2785 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2786 if (cfs_rq->skip != se)
2789 cfs_rq->skip = NULL;
2793 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2795 if (cfs_rq->last == se)
2796 __clear_buddies_last(se);
2798 if (cfs_rq->next == se)
2799 __clear_buddies_next(se);
2801 if (cfs_rq->skip == se)
2802 __clear_buddies_skip(se);
2805 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2808 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2811 * Update run-time statistics of the 'current'.
2813 update_curr(cfs_rq);
2814 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2816 update_stats_dequeue(cfs_rq, se);
2817 if (flags & DEQUEUE_SLEEP) {
2818 #ifdef CONFIG_SCHEDSTATS
2819 if (entity_is_task(se)) {
2820 struct task_struct *tsk = task_of(se);
2822 if (tsk->state & TASK_INTERRUPTIBLE)
2823 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2824 if (tsk->state & TASK_UNINTERRUPTIBLE)
2825 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2830 clear_buddies(cfs_rq, se);
2832 if (se != cfs_rq->curr)
2833 __dequeue_entity(cfs_rq, se);
2835 account_entity_dequeue(cfs_rq, se);
2838 * Normalize the entity after updating the min_vruntime because the
2839 * update can refer to the ->curr item and we need to reflect this
2840 * movement in our normalized position.
2842 if (!(flags & DEQUEUE_SLEEP))
2843 se->vruntime -= cfs_rq->min_vruntime;
2845 /* return excess runtime on last dequeue */
2846 return_cfs_rq_runtime(cfs_rq);
2848 update_min_vruntime(cfs_rq);
2849 update_cfs_shares(cfs_rq);
2853 * Preempt the current task with a newly woken task if needed:
2856 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2858 unsigned long ideal_runtime, delta_exec;
2859 struct sched_entity *se;
2862 ideal_runtime = sched_slice(cfs_rq, curr);
2863 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2864 if (delta_exec > ideal_runtime) {
2865 resched_task(rq_of(cfs_rq)->curr);
2867 * The current task ran long enough, ensure it doesn't get
2868 * re-elected due to buddy favours.
2870 clear_buddies(cfs_rq, curr);
2875 * Ensure that a task that missed wakeup preemption by a
2876 * narrow margin doesn't have to wait for a full slice.
2877 * This also mitigates buddy induced latencies under load.
2879 if (delta_exec < sysctl_sched_min_granularity)
2882 se = __pick_first_entity(cfs_rq);
2883 delta = curr->vruntime - se->vruntime;
2888 if (delta > ideal_runtime)
2889 resched_task(rq_of(cfs_rq)->curr);
2893 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2895 /* 'current' is not kept within the tree. */
2898 * Any task has to be enqueued before it get to execute on
2899 * a CPU. So account for the time it spent waiting on the
2902 update_stats_wait_end(cfs_rq, se);
2903 __dequeue_entity(cfs_rq, se);
2906 update_stats_curr_start(cfs_rq, se);
2908 #ifdef CONFIG_SCHEDSTATS
2910 * Track our maximum slice length, if the CPU's load is at
2911 * least twice that of our own weight (i.e. dont track it
2912 * when there are only lesser-weight tasks around):
2914 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2915 se->statistics.slice_max = max(se->statistics.slice_max,
2916 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2919 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2923 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2926 * Pick the next process, keeping these things in mind, in this order:
2927 * 1) keep things fair between processes/task groups
2928 * 2) pick the "next" process, since someone really wants that to run
2929 * 3) pick the "last" process, for cache locality
2930 * 4) do not run the "skip" process, if something else is available
2932 static struct sched_entity *
2933 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2935 struct sched_entity *left = __pick_first_entity(cfs_rq);
2936 struct sched_entity *se;
2939 * If curr is set we have to see if its left of the leftmost entity
2940 * still in the tree, provided there was anything in the tree at all.
2942 if (!left || (curr && entity_before(curr, left)))
2945 se = left; /* ideally we run the leftmost entity */
2948 * Avoid running the skip buddy, if running something else can
2949 * be done without getting too unfair.
2951 if (cfs_rq->skip == se) {
2952 struct sched_entity *second;
2955 second = __pick_first_entity(cfs_rq);
2957 second = __pick_next_entity(se);
2958 if (!second || (curr && entity_before(curr, second)))
2962 if (second && wakeup_preempt_entity(second, left) < 1)
2967 * Prefer last buddy, try to return the CPU to a preempted task.
2969 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2973 * Someone really wants this to run. If it's not unfair, run it.
2975 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2978 clear_buddies(cfs_rq, se);
2983 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2985 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2988 * If still on the runqueue then deactivate_task()
2989 * was not called and update_curr() has to be done:
2992 update_curr(cfs_rq);
2994 /* throttle cfs_rqs exceeding runtime */
2995 check_cfs_rq_runtime(cfs_rq);
2997 check_spread(cfs_rq, prev);
2999 update_stats_wait_start(cfs_rq, prev);
3000 /* Put 'current' back into the tree. */
3001 __enqueue_entity(cfs_rq, prev);
3002 /* in !on_rq case, update occurred at dequeue */
3003 update_entity_load_avg(prev, 1);
3005 cfs_rq->curr = NULL;
3009 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3012 * Update run-time statistics of the 'current'.
3014 update_curr(cfs_rq);
3017 * Ensure that runnable average is periodically updated.
3019 update_entity_load_avg(curr, 1);
3020 update_cfs_rq_blocked_load(cfs_rq, 1);
3021 update_cfs_shares(cfs_rq);
3023 #ifdef CONFIG_SCHED_HRTICK
3025 * queued ticks are scheduled to match the slice, so don't bother
3026 * validating it and just reschedule.
3029 resched_task(rq_of(cfs_rq)->curr);
3033 * don't let the period tick interfere with the hrtick preemption
3035 if (!sched_feat(DOUBLE_TICK) &&
3036 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3040 if (cfs_rq->nr_running > 1)
3041 check_preempt_tick(cfs_rq, curr);
3045 /**************************************************
3046 * CFS bandwidth control machinery
3049 #ifdef CONFIG_CFS_BANDWIDTH
3051 #ifdef HAVE_JUMP_LABEL
3052 static struct static_key __cfs_bandwidth_used;
3054 static inline bool cfs_bandwidth_used(void)
3056 return static_key_false(&__cfs_bandwidth_used);
3059 void cfs_bandwidth_usage_inc(void)
3061 static_key_slow_inc(&__cfs_bandwidth_used);
3064 void cfs_bandwidth_usage_dec(void)
3066 static_key_slow_dec(&__cfs_bandwidth_used);
3068 #else /* HAVE_JUMP_LABEL */
3069 static bool cfs_bandwidth_used(void)
3074 void cfs_bandwidth_usage_inc(void) {}
3075 void cfs_bandwidth_usage_dec(void) {}
3076 #endif /* HAVE_JUMP_LABEL */
3079 * default period for cfs group bandwidth.
3080 * default: 0.1s, units: nanoseconds
3082 static inline u64 default_cfs_period(void)
3084 return 100000000ULL;
3087 static inline u64 sched_cfs_bandwidth_slice(void)
3089 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3093 * Replenish runtime according to assigned quota and update expiration time.
3094 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3095 * additional synchronization around rq->lock.
3097 * requires cfs_b->lock
3099 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3103 if (cfs_b->quota == RUNTIME_INF)
3106 now = sched_clock_cpu(smp_processor_id());
3107 cfs_b->runtime = cfs_b->quota;
3108 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3111 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3113 return &tg->cfs_bandwidth;
3116 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3117 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3119 if (unlikely(cfs_rq->throttle_count))
3120 return cfs_rq->throttled_clock_task;
3122 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3125 /* returns 0 on failure to allocate runtime */
3126 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3128 struct task_group *tg = cfs_rq->tg;
3129 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3130 u64 amount = 0, min_amount, expires;
3132 /* note: this is a positive sum as runtime_remaining <= 0 */
3133 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3135 raw_spin_lock(&cfs_b->lock);
3136 if (cfs_b->quota == RUNTIME_INF)
3137 amount = min_amount;
3140 * If the bandwidth pool has become inactive, then at least one
3141 * period must have elapsed since the last consumption.
3142 * Refresh the global state and ensure bandwidth timer becomes
3145 if (!cfs_b->timer_active) {
3146 __refill_cfs_bandwidth_runtime(cfs_b);
3147 __start_cfs_bandwidth(cfs_b);
3150 if (cfs_b->runtime > 0) {
3151 amount = min(cfs_b->runtime, min_amount);
3152 cfs_b->runtime -= amount;
3156 expires = cfs_b->runtime_expires;
3157 raw_spin_unlock(&cfs_b->lock);
3159 cfs_rq->runtime_remaining += amount;
3161 * we may have advanced our local expiration to account for allowed
3162 * spread between our sched_clock and the one on which runtime was
3165 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3166 cfs_rq->runtime_expires = expires;
3168 return cfs_rq->runtime_remaining > 0;
3172 * Note: This depends on the synchronization provided by sched_clock and the
3173 * fact that rq->clock snapshots this value.
3175 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3177 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3179 /* if the deadline is ahead of our clock, nothing to do */
3180 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3183 if (cfs_rq->runtime_remaining < 0)
3187 * If the local deadline has passed we have to consider the
3188 * possibility that our sched_clock is 'fast' and the global deadline
3189 * has not truly expired.
3191 * Fortunately we can check determine whether this the case by checking
3192 * whether the global deadline has advanced.
3195 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
3196 /* extend local deadline, drift is bounded above by 2 ticks */
3197 cfs_rq->runtime_expires += TICK_NSEC;
3199 /* global deadline is ahead, expiration has passed */
3200 cfs_rq->runtime_remaining = 0;
3204 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3206 /* dock delta_exec before expiring quota (as it could span periods) */
3207 cfs_rq->runtime_remaining -= delta_exec;
3208 expire_cfs_rq_runtime(cfs_rq);
3210 if (likely(cfs_rq->runtime_remaining > 0))
3214 * if we're unable to extend our runtime we resched so that the active
3215 * hierarchy can be throttled
3217 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3218 resched_task(rq_of(cfs_rq)->curr);
3221 static __always_inline
3222 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3224 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3227 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3230 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3232 return cfs_bandwidth_used() && cfs_rq->throttled;
3235 /* check whether cfs_rq, or any parent, is throttled */
3236 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3238 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3242 * Ensure that neither of the group entities corresponding to src_cpu or
3243 * dest_cpu are members of a throttled hierarchy when performing group
3244 * load-balance operations.
3246 static inline int throttled_lb_pair(struct task_group *tg,
3247 int src_cpu, int dest_cpu)
3249 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3251 src_cfs_rq = tg->cfs_rq[src_cpu];
3252 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3254 return throttled_hierarchy(src_cfs_rq) ||
3255 throttled_hierarchy(dest_cfs_rq);
3258 /* updated child weight may affect parent so we have to do this bottom up */
3259 static int tg_unthrottle_up(struct task_group *tg, void *data)
3261 struct rq *rq = data;
3262 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3264 cfs_rq->throttle_count--;
3266 if (!cfs_rq->throttle_count) {
3267 /* adjust cfs_rq_clock_task() */
3268 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3269 cfs_rq->throttled_clock_task;
3276 static int tg_throttle_down(struct task_group *tg, void *data)
3278 struct rq *rq = data;
3279 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3281 /* group is entering throttled state, stop time */
3282 if (!cfs_rq->throttle_count)
3283 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3284 cfs_rq->throttle_count++;
3289 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3291 struct rq *rq = rq_of(cfs_rq);
3292 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3293 struct sched_entity *se;
3294 long task_delta, dequeue = 1;
3296 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3298 /* freeze hierarchy runnable averages while throttled */
3300 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3303 task_delta = cfs_rq->h_nr_running;
3304 for_each_sched_entity(se) {
3305 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3306 /* throttled entity or throttle-on-deactivate */
3311 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3312 qcfs_rq->h_nr_running -= task_delta;
3314 if (qcfs_rq->load.weight)
3319 rq->nr_running -= task_delta;
3321 cfs_rq->throttled = 1;
3322 cfs_rq->throttled_clock = rq_clock(rq);
3323 raw_spin_lock(&cfs_b->lock);
3324 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3325 if (!cfs_b->timer_active)
3326 __start_cfs_bandwidth(cfs_b);
3327 raw_spin_unlock(&cfs_b->lock);
3330 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3332 struct rq *rq = rq_of(cfs_rq);
3333 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3334 struct sched_entity *se;
3338 se = cfs_rq->tg->se[cpu_of(rq)];
3340 cfs_rq->throttled = 0;
3342 update_rq_clock(rq);
3344 raw_spin_lock(&cfs_b->lock);
3345 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3346 list_del_rcu(&cfs_rq->throttled_list);
3347 raw_spin_unlock(&cfs_b->lock);
3349 /* update hierarchical throttle state */
3350 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3352 if (!cfs_rq->load.weight)
3355 task_delta = cfs_rq->h_nr_running;
3356 for_each_sched_entity(se) {
3360 cfs_rq = cfs_rq_of(se);
3362 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3363 cfs_rq->h_nr_running += task_delta;
3365 if (cfs_rq_throttled(cfs_rq))
3370 rq->nr_running += task_delta;
3372 /* determine whether we need to wake up potentially idle cpu */
3373 if (rq->curr == rq->idle && rq->cfs.nr_running)
3374 resched_task(rq->curr);
3377 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3378 u64 remaining, u64 expires)
3380 struct cfs_rq *cfs_rq;
3381 u64 runtime = remaining;
3384 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3386 struct rq *rq = rq_of(cfs_rq);
3388 raw_spin_lock(&rq->lock);
3389 if (!cfs_rq_throttled(cfs_rq))
3392 runtime = -cfs_rq->runtime_remaining + 1;
3393 if (runtime > remaining)
3394 runtime = remaining;
3395 remaining -= runtime;
3397 cfs_rq->runtime_remaining += runtime;
3398 cfs_rq->runtime_expires = expires;
3400 /* we check whether we're throttled above */
3401 if (cfs_rq->runtime_remaining > 0)
3402 unthrottle_cfs_rq(cfs_rq);
3405 raw_spin_unlock(&rq->lock);
3416 * Responsible for refilling a task_group's bandwidth and unthrottling its
3417 * cfs_rqs as appropriate. If there has been no activity within the last
3418 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3419 * used to track this state.
3421 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3423 u64 runtime, runtime_expires;
3424 int idle = 1, throttled;
3426 raw_spin_lock(&cfs_b->lock);
3427 /* no need to continue the timer with no bandwidth constraint */
3428 if (cfs_b->quota == RUNTIME_INF)
3431 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3432 /* idle depends on !throttled (for the case of a large deficit) */
3433 idle = cfs_b->idle && !throttled;
3434 cfs_b->nr_periods += overrun;
3436 /* if we're going inactive then everything else can be deferred */
3441 * if we have relooped after returning idle once, we need to update our
3442 * status as actually running, so that other cpus doing
3443 * __start_cfs_bandwidth will stop trying to cancel us.
3445 cfs_b->timer_active = 1;
3447 __refill_cfs_bandwidth_runtime(cfs_b);
3450 /* mark as potentially idle for the upcoming period */
3455 /* account preceding periods in which throttling occurred */
3456 cfs_b->nr_throttled += overrun;
3459 * There are throttled entities so we must first use the new bandwidth
3460 * to unthrottle them before making it generally available. This
3461 * ensures that all existing debts will be paid before a new cfs_rq is
3464 runtime = cfs_b->runtime;
3465 runtime_expires = cfs_b->runtime_expires;
3469 * This check is repeated as we are holding onto the new bandwidth
3470 * while we unthrottle. This can potentially race with an unthrottled
3471 * group trying to acquire new bandwidth from the global pool.
3473 while (throttled && runtime > 0) {
3474 raw_spin_unlock(&cfs_b->lock);
3475 /* we can't nest cfs_b->lock while distributing bandwidth */
3476 runtime = distribute_cfs_runtime(cfs_b, runtime,
3478 raw_spin_lock(&cfs_b->lock);
3480 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3483 /* return (any) remaining runtime */
3484 cfs_b->runtime = runtime;
3486 * While we are ensured activity in the period following an
3487 * unthrottle, this also covers the case in which the new bandwidth is
3488 * insufficient to cover the existing bandwidth deficit. (Forcing the
3489 * timer to remain active while there are any throttled entities.)
3494 cfs_b->timer_active = 0;
3495 raw_spin_unlock(&cfs_b->lock);
3500 /* a cfs_rq won't donate quota below this amount */
3501 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3502 /* minimum remaining period time to redistribute slack quota */
3503 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3504 /* how long we wait to gather additional slack before distributing */
3505 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3508 * Are we near the end of the current quota period?
3510 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3511 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3512 * migrate_hrtimers, base is never cleared, so we are fine.
3514 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3516 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3519 /* if the call-back is running a quota refresh is already occurring */
3520 if (hrtimer_callback_running(refresh_timer))
3523 /* is a quota refresh about to occur? */
3524 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3525 if (remaining < min_expire)
3531 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3533 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3535 /* if there's a quota refresh soon don't bother with slack */
3536 if (runtime_refresh_within(cfs_b, min_left))
3539 start_bandwidth_timer(&cfs_b->slack_timer,
3540 ns_to_ktime(cfs_bandwidth_slack_period));
3543 /* we know any runtime found here is valid as update_curr() precedes return */
3544 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3546 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3547 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3549 if (slack_runtime <= 0)
3552 raw_spin_lock(&cfs_b->lock);
3553 if (cfs_b->quota != RUNTIME_INF &&
3554 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3555 cfs_b->runtime += slack_runtime;
3557 /* we are under rq->lock, defer unthrottling using a timer */
3558 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3559 !list_empty(&cfs_b->throttled_cfs_rq))
3560 start_cfs_slack_bandwidth(cfs_b);
3562 raw_spin_unlock(&cfs_b->lock);
3564 /* even if it's not valid for return we don't want to try again */
3565 cfs_rq->runtime_remaining -= slack_runtime;
3568 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3570 if (!cfs_bandwidth_used())
3573 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3576 __return_cfs_rq_runtime(cfs_rq);
3580 * This is done with a timer (instead of inline with bandwidth return) since
3581 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3583 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3585 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3588 /* confirm we're still not at a refresh boundary */
3589 raw_spin_lock(&cfs_b->lock);
3590 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3591 raw_spin_unlock(&cfs_b->lock);
3595 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
3596 runtime = cfs_b->runtime;
3599 expires = cfs_b->runtime_expires;
3600 raw_spin_unlock(&cfs_b->lock);
3605 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3607 raw_spin_lock(&cfs_b->lock);
3608 if (expires == cfs_b->runtime_expires)
3609 cfs_b->runtime = runtime;
3610 raw_spin_unlock(&cfs_b->lock);
3614 * When a group wakes up we want to make sure that its quota is not already
3615 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3616 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3618 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3620 if (!cfs_bandwidth_used())
3623 /* an active group must be handled by the update_curr()->put() path */
3624 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3627 /* ensure the group is not already throttled */
3628 if (cfs_rq_throttled(cfs_rq))
3631 /* update runtime allocation */
3632 account_cfs_rq_runtime(cfs_rq, 0);
3633 if (cfs_rq->runtime_remaining <= 0)
3634 throttle_cfs_rq(cfs_rq);
3637 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3638 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3640 if (!cfs_bandwidth_used())
3643 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3647 * it's possible for a throttled entity to be forced into a running
3648 * state (e.g. set_curr_task), in this case we're finished.
3650 if (cfs_rq_throttled(cfs_rq))
3653 throttle_cfs_rq(cfs_rq);
3657 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3659 struct cfs_bandwidth *cfs_b =
3660 container_of(timer, struct cfs_bandwidth, slack_timer);
3661 do_sched_cfs_slack_timer(cfs_b);
3663 return HRTIMER_NORESTART;
3666 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3668 struct cfs_bandwidth *cfs_b =
3669 container_of(timer, struct cfs_bandwidth, period_timer);
3675 now = hrtimer_cb_get_time(timer);
3676 overrun = hrtimer_forward(timer, now, cfs_b->period);
3681 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3684 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3687 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3689 raw_spin_lock_init(&cfs_b->lock);
3691 cfs_b->quota = RUNTIME_INF;
3692 cfs_b->period = ns_to_ktime(default_cfs_period());
3694 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3695 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3696 cfs_b->period_timer.function = sched_cfs_period_timer;
3697 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3698 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3701 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3703 cfs_rq->runtime_enabled = 0;
3704 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3707 /* requires cfs_b->lock, may release to reprogram timer */
3708 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3711 * The timer may be active because we're trying to set a new bandwidth
3712 * period or because we're racing with the tear-down path
3713 * (timer_active==0 becomes visible before the hrtimer call-back
3714 * terminates). In either case we ensure that it's re-programmed
3716 while (unlikely(hrtimer_active(&cfs_b->period_timer)) &&
3717 hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) {
3718 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3719 raw_spin_unlock(&cfs_b->lock);
3721 raw_spin_lock(&cfs_b->lock);
3722 /* if someone else restarted the timer then we're done */
3723 if (cfs_b->timer_active)
3727 cfs_b->timer_active = 1;
3728 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3731 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3733 hrtimer_cancel(&cfs_b->period_timer);
3734 hrtimer_cancel(&cfs_b->slack_timer);
3737 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3739 struct cfs_rq *cfs_rq;
3741 for_each_leaf_cfs_rq(rq, cfs_rq) {
3742 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3744 if (!cfs_rq->runtime_enabled)
3748 * clock_task is not advancing so we just need to make sure
3749 * there's some valid quota amount
3751 cfs_rq->runtime_remaining = cfs_b->quota;
3752 if (cfs_rq_throttled(cfs_rq))
3753 unthrottle_cfs_rq(cfs_rq);
3757 #else /* CONFIG_CFS_BANDWIDTH */
3758 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3760 return rq_clock_task(rq_of(cfs_rq));
3763 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
3764 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
3765 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3766 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3768 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3773 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3778 static inline int throttled_lb_pair(struct task_group *tg,
3779 int src_cpu, int dest_cpu)
3784 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3786 #ifdef CONFIG_FAIR_GROUP_SCHED
3787 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3790 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3794 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3795 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3797 #endif /* CONFIG_CFS_BANDWIDTH */
3799 /**************************************************
3800 * CFS operations on tasks:
3803 #ifdef CONFIG_SCHED_HRTICK
3804 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3806 struct sched_entity *se = &p->se;
3807 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3809 WARN_ON(task_rq(p) != rq);
3811 if (cfs_rq->nr_running > 1) {
3812 u64 slice = sched_slice(cfs_rq, se);
3813 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3814 s64 delta = slice - ran;
3823 * Don't schedule slices shorter than 10000ns, that just
3824 * doesn't make sense. Rely on vruntime for fairness.
3827 delta = max_t(s64, 10000LL, delta);
3829 hrtick_start(rq, delta);
3834 * called from enqueue/dequeue and updates the hrtick when the
3835 * current task is from our class and nr_running is low enough
3838 static void hrtick_update(struct rq *rq)
3840 struct task_struct *curr = rq->curr;
3842 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3845 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3846 hrtick_start_fair(rq, curr);
3848 #else /* !CONFIG_SCHED_HRTICK */
3850 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3854 static inline void hrtick_update(struct rq *rq)
3860 * The enqueue_task method is called before nr_running is
3861 * increased. Here we update the fair scheduling stats and
3862 * then put the task into the rbtree:
3865 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3867 struct cfs_rq *cfs_rq;
3868 struct sched_entity *se = &p->se;
3870 for_each_sched_entity(se) {
3873 cfs_rq = cfs_rq_of(se);
3874 enqueue_entity(cfs_rq, se, flags);
3877 * end evaluation on encountering a throttled cfs_rq
3879 * note: in the case of encountering a throttled cfs_rq we will
3880 * post the final h_nr_running increment below.
3882 if (cfs_rq_throttled(cfs_rq))
3884 cfs_rq->h_nr_running++;
3886 flags = ENQUEUE_WAKEUP;
3889 for_each_sched_entity(se) {
3890 cfs_rq = cfs_rq_of(se);
3891 cfs_rq->h_nr_running++;
3893 if (cfs_rq_throttled(cfs_rq))
3896 update_cfs_shares(cfs_rq);
3897 update_entity_load_avg(se, 1);
3901 update_rq_runnable_avg(rq, rq->nr_running);
3907 static void set_next_buddy(struct sched_entity *se);
3910 * The dequeue_task method is called before nr_running is
3911 * decreased. We remove the task from the rbtree and
3912 * update the fair scheduling stats:
3914 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3916 struct cfs_rq *cfs_rq;
3917 struct sched_entity *se = &p->se;
3918 int task_sleep = flags & DEQUEUE_SLEEP;
3920 for_each_sched_entity(se) {
3921 cfs_rq = cfs_rq_of(se);
3922 dequeue_entity(cfs_rq, se, flags);
3925 * end evaluation on encountering a throttled cfs_rq
3927 * note: in the case of encountering a throttled cfs_rq we will
3928 * post the final h_nr_running decrement below.
3930 if (cfs_rq_throttled(cfs_rq))
3932 cfs_rq->h_nr_running--;
3934 /* Don't dequeue parent if it has other entities besides us */
3935 if (cfs_rq->load.weight) {
3937 * Bias pick_next to pick a task from this cfs_rq, as
3938 * p is sleeping when it is within its sched_slice.
3940 if (task_sleep && parent_entity(se))
3941 set_next_buddy(parent_entity(se));
3943 /* avoid re-evaluating load for this entity */
3944 se = parent_entity(se);
3947 flags |= DEQUEUE_SLEEP;
3950 for_each_sched_entity(se) {
3951 cfs_rq = cfs_rq_of(se);
3952 cfs_rq->h_nr_running--;
3954 if (cfs_rq_throttled(cfs_rq))
3957 update_cfs_shares(cfs_rq);
3958 update_entity_load_avg(se, 1);
3963 update_rq_runnable_avg(rq, 1);
3969 /* Used instead of source_load when we know the type == 0 */
3970 static unsigned long weighted_cpuload(const int cpu)
3972 return cpu_rq(cpu)->cfs.runnable_load_avg;
3976 * Return a low guess at the load of a migration-source cpu weighted
3977 * according to the scheduling class and "nice" value.
3979 * We want to under-estimate the load of migration sources, to
3980 * balance conservatively.
3982 static unsigned long source_load(int cpu, int type)
3984 struct rq *rq = cpu_rq(cpu);
3985 unsigned long total = weighted_cpuload(cpu);
3987 if (type == 0 || !sched_feat(LB_BIAS))
3990 return min(rq->cpu_load[type-1], total);
3994 * Return a high guess at the load of a migration-target cpu weighted
3995 * according to the scheduling class and "nice" value.
3997 static unsigned long target_load(int cpu, int type)
3999 struct rq *rq = cpu_rq(cpu);
4000 unsigned long total = weighted_cpuload(cpu);
4002 if (type == 0 || !sched_feat(LB_BIAS))
4005 return max(rq->cpu_load[type-1], total);
4008 static unsigned long power_of(int cpu)
4010 return cpu_rq(cpu)->cpu_power;
4013 static unsigned long cpu_avg_load_per_task(int cpu)
4015 struct rq *rq = cpu_rq(cpu);
4016 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
4017 unsigned long load_avg = rq->cfs.runnable_load_avg;
4020 return load_avg / nr_running;
4025 static void record_wakee(struct task_struct *p)
4028 * Rough decay (wiping) for cost saving, don't worry
4029 * about the boundary, really active task won't care
4032 if (jiffies > current->wakee_flip_decay_ts + HZ) {
4033 current->wakee_flips = 0;
4034 current->wakee_flip_decay_ts = jiffies;
4037 if (current->last_wakee != p) {
4038 current->last_wakee = p;
4039 current->wakee_flips++;
4043 static void task_waking_fair(struct task_struct *p)
4045 struct sched_entity *se = &p->se;
4046 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4049 #ifndef CONFIG_64BIT
4050 u64 min_vruntime_copy;
4053 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4055 min_vruntime = cfs_rq->min_vruntime;
4056 } while (min_vruntime != min_vruntime_copy);
4058 min_vruntime = cfs_rq->min_vruntime;
4061 se->vruntime -= min_vruntime;
4065 #ifdef CONFIG_FAIR_GROUP_SCHED
4067 * effective_load() calculates the load change as seen from the root_task_group
4069 * Adding load to a group doesn't make a group heavier, but can cause movement
4070 * of group shares between cpus. Assuming the shares were perfectly aligned one
4071 * can calculate the shift in shares.
4073 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4074 * on this @cpu and results in a total addition (subtraction) of @wg to the
4075 * total group weight.
4077 * Given a runqueue weight distribution (rw_i) we can compute a shares
4078 * distribution (s_i) using:
4080 * s_i = rw_i / \Sum rw_j (1)
4082 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4083 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4084 * shares distribution (s_i):
4086 * rw_i = { 2, 4, 1, 0 }
4087 * s_i = { 2/7, 4/7, 1/7, 0 }
4089 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4090 * task used to run on and the CPU the waker is running on), we need to
4091 * compute the effect of waking a task on either CPU and, in case of a sync
4092 * wakeup, compute the effect of the current task going to sleep.
4094 * So for a change of @wl to the local @cpu with an overall group weight change
4095 * of @wl we can compute the new shares distribution (s'_i) using:
4097 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4099 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4100 * differences in waking a task to CPU 0. The additional task changes the
4101 * weight and shares distributions like:
4103 * rw'_i = { 3, 4, 1, 0 }
4104 * s'_i = { 3/8, 4/8, 1/8, 0 }
4106 * We can then compute the difference in effective weight by using:
4108 * dw_i = S * (s'_i - s_i) (3)
4110 * Where 'S' is the group weight as seen by its parent.
4112 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4113 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4114 * 4/7) times the weight of the group.
4116 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4118 struct sched_entity *se = tg->se[cpu];
4120 if (!tg->parent) /* the trivial, non-cgroup case */
4123 for_each_sched_entity(se) {
4129 * W = @wg + \Sum rw_j
4131 W = wg + calc_tg_weight(tg, se->my_q);
4136 w = se->my_q->load.weight + wl;
4139 * wl = S * s'_i; see (2)
4142 wl = (w * tg->shares) / W;
4147 * Per the above, wl is the new se->load.weight value; since
4148 * those are clipped to [MIN_SHARES, ...) do so now. See
4149 * calc_cfs_shares().
4151 if (wl < MIN_SHARES)
4155 * wl = dw_i = S * (s'_i - s_i); see (3)
4157 wl -= se->load.weight;
4160 * Recursively apply this logic to all parent groups to compute
4161 * the final effective load change on the root group. Since
4162 * only the @tg group gets extra weight, all parent groups can
4163 * only redistribute existing shares. @wl is the shift in shares
4164 * resulting from this level per the above.
4173 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4180 static int wake_wide(struct task_struct *p)
4182 int factor = this_cpu_read(sd_llc_size);
4185 * Yeah, it's the switching-frequency, could means many wakee or
4186 * rapidly switch, use factor here will just help to automatically
4187 * adjust the loose-degree, so bigger node will lead to more pull.
4189 if (p->wakee_flips > factor) {
4191 * wakee is somewhat hot, it needs certain amount of cpu
4192 * resource, so if waker is far more hot, prefer to leave
4195 if (current->wakee_flips > (factor * p->wakee_flips))
4202 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4204 s64 this_load, load;
4205 int idx, this_cpu, prev_cpu;
4206 unsigned long tl_per_task;
4207 struct task_group *tg;
4208 unsigned long weight;
4212 * If we wake multiple tasks be careful to not bounce
4213 * ourselves around too much.
4219 this_cpu = smp_processor_id();
4220 prev_cpu = task_cpu(p);
4221 load = source_load(prev_cpu, idx);
4222 this_load = target_load(this_cpu, idx);
4225 * If sync wakeup then subtract the (maximum possible)
4226 * effect of the currently running task from the load
4227 * of the current CPU:
4230 tg = task_group(current);
4231 weight = current->se.load.weight;
4233 this_load += effective_load(tg, this_cpu, -weight, -weight);
4234 load += effective_load(tg, prev_cpu, 0, -weight);
4238 weight = p->se.load.weight;
4241 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4242 * due to the sync cause above having dropped this_load to 0, we'll
4243 * always have an imbalance, but there's really nothing you can do
4244 * about that, so that's good too.
4246 * Otherwise check if either cpus are near enough in load to allow this
4247 * task to be woken on this_cpu.
4249 if (this_load > 0) {
4250 s64 this_eff_load, prev_eff_load;
4252 this_eff_load = 100;
4253 this_eff_load *= power_of(prev_cpu);
4254 this_eff_load *= this_load +
4255 effective_load(tg, this_cpu, weight, weight);
4257 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4258 prev_eff_load *= power_of(this_cpu);
4259 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4261 balanced = this_eff_load <= prev_eff_load;
4266 * If the currently running task will sleep within
4267 * a reasonable amount of time then attract this newly
4270 if (sync && balanced)
4273 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4274 tl_per_task = cpu_avg_load_per_task(this_cpu);
4277 (this_load <= load &&
4278 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
4280 * This domain has SD_WAKE_AFFINE and
4281 * p is cache cold in this domain, and
4282 * there is no bad imbalance.
4284 schedstat_inc(sd, ttwu_move_affine);
4285 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4293 * find_idlest_group finds and returns the least busy CPU group within the
4296 static struct sched_group *
4297 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4298 int this_cpu, int sd_flag)
4300 struct sched_group *idlest = NULL, *group = sd->groups;
4301 unsigned long min_load = ULONG_MAX, this_load = 0;
4302 int load_idx = sd->forkexec_idx;
4303 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4305 if (sd_flag & SD_BALANCE_WAKE)
4306 load_idx = sd->wake_idx;
4309 unsigned long load, avg_load;
4313 /* Skip over this group if it has no CPUs allowed */
4314 if (!cpumask_intersects(sched_group_cpus(group),
4315 tsk_cpus_allowed(p)))
4318 local_group = cpumask_test_cpu(this_cpu,
4319 sched_group_cpus(group));
4321 /* Tally up the load of all CPUs in the group */
4324 for_each_cpu(i, sched_group_cpus(group)) {
4325 /* Bias balancing toward cpus of our domain */
4327 load = source_load(i, load_idx);
4329 load = target_load(i, load_idx);
4334 /* Adjust by relative CPU power of the group */
4335 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
4338 this_load = avg_load;
4339 } else if (avg_load < min_load) {
4340 min_load = avg_load;
4343 } while (group = group->next, group != sd->groups);
4345 if (!idlest || 100*this_load < imbalance*min_load)
4351 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4354 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4356 unsigned long load, min_load = ULONG_MAX;
4360 /* Traverse only the allowed CPUs */
4361 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4362 load = weighted_cpuload(i);
4364 if (load < min_load || (load == min_load && i == this_cpu)) {
4374 * Try and locate an idle CPU in the sched_domain.
4376 static int select_idle_sibling(struct task_struct *p, int target)
4378 struct sched_domain *sd;
4379 struct sched_group *sg;
4380 int i = task_cpu(p);
4382 if (idle_cpu(target))
4386 * If the prevous cpu is cache affine and idle, don't be stupid.
4388 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4392 * Otherwise, iterate the domains and find an elegible idle cpu.
4394 sd = rcu_dereference(per_cpu(sd_llc, target));
4395 for_each_lower_domain(sd) {
4398 if (!cpumask_intersects(sched_group_cpus(sg),
4399 tsk_cpus_allowed(p)))
4402 for_each_cpu(i, sched_group_cpus(sg)) {
4403 if (i == target || !idle_cpu(i))
4407 target = cpumask_first_and(sched_group_cpus(sg),
4408 tsk_cpus_allowed(p));
4412 } while (sg != sd->groups);
4419 * select_task_rq_fair: Select target runqueue for the waking task in domains
4420 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4421 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4423 * Balances load by selecting the idlest cpu in the idlest group, or under
4424 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4426 * Returns the target cpu number.
4428 * preempt must be disabled.
4431 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4433 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4434 int cpu = smp_processor_id();
4436 int want_affine = 0;
4437 int sync = wake_flags & WF_SYNC;
4439 if (p->nr_cpus_allowed == 1)
4442 if (sd_flag & SD_BALANCE_WAKE) {
4443 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
4449 for_each_domain(cpu, tmp) {
4450 if (!(tmp->flags & SD_LOAD_BALANCE))
4454 * If both cpu and prev_cpu are part of this domain,
4455 * cpu is a valid SD_WAKE_AFFINE target.
4457 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4458 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4463 if (tmp->flags & sd_flag)
4468 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4471 new_cpu = select_idle_sibling(p, prev_cpu);
4476 struct sched_group *group;
4479 if (!(sd->flags & sd_flag)) {
4484 group = find_idlest_group(sd, p, cpu, sd_flag);
4490 new_cpu = find_idlest_cpu(group, p, cpu);
4491 if (new_cpu == -1 || new_cpu == cpu) {
4492 /* Now try balancing at a lower domain level of cpu */
4497 /* Now try balancing at a lower domain level of new_cpu */
4499 weight = sd->span_weight;
4501 for_each_domain(cpu, tmp) {
4502 if (weight <= tmp->span_weight)
4504 if (tmp->flags & sd_flag)
4507 /* while loop will break here if sd == NULL */
4516 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4517 * cfs_rq_of(p) references at time of call are still valid and identify the
4518 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4519 * other assumptions, including the state of rq->lock, should be made.
4522 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4524 struct sched_entity *se = &p->se;
4525 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4528 * Load tracking: accumulate removed load so that it can be processed
4529 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4530 * to blocked load iff they have a positive decay-count. It can never
4531 * be negative here since on-rq tasks have decay-count == 0.
4533 if (se->avg.decay_count) {
4534 se->avg.decay_count = -__synchronize_entity_decay(se);
4535 atomic_long_add(se->avg.load_avg_contrib,
4536 &cfs_rq->removed_load);
4539 #endif /* CONFIG_SMP */
4541 static unsigned long
4542 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4544 unsigned long gran = sysctl_sched_wakeup_granularity;
4547 * Since its curr running now, convert the gran from real-time
4548 * to virtual-time in his units.
4550 * By using 'se' instead of 'curr' we penalize light tasks, so
4551 * they get preempted easier. That is, if 'se' < 'curr' then
4552 * the resulting gran will be larger, therefore penalizing the
4553 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4554 * be smaller, again penalizing the lighter task.
4556 * This is especially important for buddies when the leftmost
4557 * task is higher priority than the buddy.
4559 return calc_delta_fair(gran, se);
4563 * Should 'se' preempt 'curr'.
4577 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4579 s64 gran, vdiff = curr->vruntime - se->vruntime;
4584 gran = wakeup_gran(curr, se);
4591 static void set_last_buddy(struct sched_entity *se)
4593 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4596 for_each_sched_entity(se)
4597 cfs_rq_of(se)->last = se;
4600 static void set_next_buddy(struct sched_entity *se)
4602 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4605 for_each_sched_entity(se)
4606 cfs_rq_of(se)->next = se;
4609 static void set_skip_buddy(struct sched_entity *se)
4611 for_each_sched_entity(se)
4612 cfs_rq_of(se)->skip = se;
4616 * Preempt the current task with a newly woken task if needed:
4618 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4620 struct task_struct *curr = rq->curr;
4621 struct sched_entity *se = &curr->se, *pse = &p->se;
4622 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4623 int scale = cfs_rq->nr_running >= sched_nr_latency;
4624 int next_buddy_marked = 0;
4626 if (unlikely(se == pse))
4630 * This is possible from callers such as move_task(), in which we
4631 * unconditionally check_prempt_curr() after an enqueue (which may have
4632 * lead to a throttle). This both saves work and prevents false
4633 * next-buddy nomination below.
4635 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4638 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4639 set_next_buddy(pse);
4640 next_buddy_marked = 1;
4644 * We can come here with TIF_NEED_RESCHED already set from new task
4647 * Note: this also catches the edge-case of curr being in a throttled
4648 * group (e.g. via set_curr_task), since update_curr() (in the
4649 * enqueue of curr) will have resulted in resched being set. This
4650 * prevents us from potentially nominating it as a false LAST_BUDDY
4653 if (test_tsk_need_resched(curr))
4656 /* Idle tasks are by definition preempted by non-idle tasks. */
4657 if (unlikely(curr->policy == SCHED_IDLE) &&
4658 likely(p->policy != SCHED_IDLE))
4662 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4663 * is driven by the tick):
4665 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4668 find_matching_se(&se, &pse);
4669 update_curr(cfs_rq_of(se));
4671 if (wakeup_preempt_entity(se, pse) == 1) {
4673 * Bias pick_next to pick the sched entity that is
4674 * triggering this preemption.
4676 if (!next_buddy_marked)
4677 set_next_buddy(pse);
4686 * Only set the backward buddy when the current task is still
4687 * on the rq. This can happen when a wakeup gets interleaved
4688 * with schedule on the ->pre_schedule() or idle_balance()
4689 * point, either of which can * drop the rq lock.
4691 * Also, during early boot the idle thread is in the fair class,
4692 * for obvious reasons its a bad idea to schedule back to it.
4694 if (unlikely(!se->on_rq || curr == rq->idle))
4697 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4701 static struct task_struct *
4702 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
4704 struct cfs_rq *cfs_rq = &rq->cfs;
4705 struct sched_entity *se;
4706 struct task_struct *p;
4710 #ifdef CONFIG_FAIR_GROUP_SCHED
4711 if (!cfs_rq->nr_running)
4714 if (prev->sched_class != &fair_sched_class)
4718 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
4719 * likely that a next task is from the same cgroup as the current.
4721 * Therefore attempt to avoid putting and setting the entire cgroup
4722 * hierarchy, only change the part that actually changes.
4726 struct sched_entity *curr = cfs_rq->curr;
4729 * Since we got here without doing put_prev_entity() we also
4730 * have to consider cfs_rq->curr. If it is still a runnable
4731 * entity, update_curr() will update its vruntime, otherwise
4732 * forget we've ever seen it.
4734 if (curr && curr->on_rq)
4735 update_curr(cfs_rq);
4740 * This call to check_cfs_rq_runtime() will do the throttle and
4741 * dequeue its entity in the parent(s). Therefore the 'simple'
4742 * nr_running test will indeed be correct.
4744 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
4747 se = pick_next_entity(cfs_rq, curr);
4748 cfs_rq = group_cfs_rq(se);
4754 * Since we haven't yet done put_prev_entity and if the selected task
4755 * is a different task than we started out with, try and touch the
4756 * least amount of cfs_rqs.
4759 struct sched_entity *pse = &prev->se;
4761 while (!(cfs_rq = is_same_group(se, pse))) {
4762 int se_depth = se->depth;
4763 int pse_depth = pse->depth;
4765 if (se_depth <= pse_depth) {
4766 put_prev_entity(cfs_rq_of(pse), pse);
4767 pse = parent_entity(pse);
4769 if (se_depth >= pse_depth) {
4770 set_next_entity(cfs_rq_of(se), se);
4771 se = parent_entity(se);
4775 put_prev_entity(cfs_rq, pse);
4776 set_next_entity(cfs_rq, se);
4779 if (hrtick_enabled(rq))
4780 hrtick_start_fair(rq, p);
4787 if (!cfs_rq->nr_running)
4790 put_prev_task(rq, prev);
4793 se = pick_next_entity(cfs_rq, NULL);
4794 set_next_entity(cfs_rq, se);
4795 cfs_rq = group_cfs_rq(se);
4800 if (hrtick_enabled(rq))
4801 hrtick_start_fair(rq, p);
4806 new_tasks = idle_balance(rq);
4808 * Because idle_balance() releases (and re-acquires) rq->lock, it is
4809 * possible for any higher priority task to appear. In that case we
4810 * must re-start the pick_next_entity() loop.
4822 * Account for a descheduled task:
4824 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4826 struct sched_entity *se = &prev->se;
4827 struct cfs_rq *cfs_rq;
4829 for_each_sched_entity(se) {
4830 cfs_rq = cfs_rq_of(se);
4831 put_prev_entity(cfs_rq, se);
4836 * sched_yield() is very simple
4838 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4840 static void yield_task_fair(struct rq *rq)
4842 struct task_struct *curr = rq->curr;
4843 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4844 struct sched_entity *se = &curr->se;
4847 * Are we the only task in the tree?
4849 if (unlikely(rq->nr_running == 1))
4852 clear_buddies(cfs_rq, se);
4854 if (curr->policy != SCHED_BATCH) {
4855 update_rq_clock(rq);
4857 * Update run-time statistics of the 'current'.
4859 update_curr(cfs_rq);
4861 * Tell update_rq_clock() that we've just updated,
4862 * so we don't do microscopic update in schedule()
4863 * and double the fastpath cost.
4865 rq->skip_clock_update = 1;
4871 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4873 struct sched_entity *se = &p->se;
4875 /* throttled hierarchies are not runnable */
4876 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4879 /* Tell the scheduler that we'd really like pse to run next. */
4882 yield_task_fair(rq);
4888 /**************************************************
4889 * Fair scheduling class load-balancing methods.
4893 * The purpose of load-balancing is to achieve the same basic fairness the
4894 * per-cpu scheduler provides, namely provide a proportional amount of compute
4895 * time to each task. This is expressed in the following equation:
4897 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4899 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4900 * W_i,0 is defined as:
4902 * W_i,0 = \Sum_j w_i,j (2)
4904 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4905 * is derived from the nice value as per prio_to_weight[].
4907 * The weight average is an exponential decay average of the instantaneous
4910 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4912 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4913 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4914 * can also include other factors [XXX].
4916 * To achieve this balance we define a measure of imbalance which follows
4917 * directly from (1):
4919 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4921 * We them move tasks around to minimize the imbalance. In the continuous
4922 * function space it is obvious this converges, in the discrete case we get
4923 * a few fun cases generally called infeasible weight scenarios.
4926 * - infeasible weights;
4927 * - local vs global optima in the discrete case. ]
4932 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4933 * for all i,j solution, we create a tree of cpus that follows the hardware
4934 * topology where each level pairs two lower groups (or better). This results
4935 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4936 * tree to only the first of the previous level and we decrease the frequency
4937 * of load-balance at each level inv. proportional to the number of cpus in
4943 * \Sum { --- * --- * 2^i } = O(n) (5)
4945 * `- size of each group
4946 * | | `- number of cpus doing load-balance
4948 * `- sum over all levels
4950 * Coupled with a limit on how many tasks we can migrate every balance pass,
4951 * this makes (5) the runtime complexity of the balancer.
4953 * An important property here is that each CPU is still (indirectly) connected
4954 * to every other cpu in at most O(log n) steps:
4956 * The adjacency matrix of the resulting graph is given by:
4959 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4962 * And you'll find that:
4964 * A^(log_2 n)_i,j != 0 for all i,j (7)
4966 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4967 * The task movement gives a factor of O(m), giving a convergence complexity
4970 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4975 * In order to avoid CPUs going idle while there's still work to do, new idle
4976 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4977 * tree itself instead of relying on other CPUs to bring it work.
4979 * This adds some complexity to both (5) and (8) but it reduces the total idle
4987 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4990 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4995 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4997 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4999 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5002 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5003 * rewrite all of this once again.]
5006 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5008 enum fbq_type { regular, remote, all };
5010 #define LBF_ALL_PINNED 0x01
5011 #define LBF_NEED_BREAK 0x02
5012 #define LBF_DST_PINNED 0x04
5013 #define LBF_SOME_PINNED 0x08
5016 struct sched_domain *sd;
5024 struct cpumask *dst_grpmask;
5026 enum cpu_idle_type idle;
5028 /* The set of CPUs under consideration for load-balancing */
5029 struct cpumask *cpus;
5034 unsigned int loop_break;
5035 unsigned int loop_max;
5037 enum fbq_type fbq_type;
5041 * move_task - move a task from one runqueue to another runqueue.
5042 * Both runqueues must be locked.
5044 static void move_task(struct task_struct *p, struct lb_env *env)
5046 deactivate_task(env->src_rq, p, 0);
5047 set_task_cpu(p, env->dst_cpu);
5048 activate_task(env->dst_rq, p, 0);
5049 check_preempt_curr(env->dst_rq, p, 0);
5053 * Is this task likely cache-hot:
5056 task_hot(struct task_struct *p, u64 now)
5060 if (p->sched_class != &fair_sched_class)
5063 if (unlikely(p->policy == SCHED_IDLE))
5067 * Buddy candidates are cache hot:
5069 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
5070 (&p->se == cfs_rq_of(&p->se)->next ||
5071 &p->se == cfs_rq_of(&p->se)->last))
5074 if (sysctl_sched_migration_cost == -1)
5076 if (sysctl_sched_migration_cost == 0)
5079 delta = now - p->se.exec_start;
5081 return delta < (s64)sysctl_sched_migration_cost;
5084 #ifdef CONFIG_NUMA_BALANCING
5085 /* Returns true if the destination node has incurred more faults */
5086 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
5088 int src_nid, dst_nid;
5090 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults_memory ||
5091 !(env->sd->flags & SD_NUMA)) {
5095 src_nid = cpu_to_node(env->src_cpu);
5096 dst_nid = cpu_to_node(env->dst_cpu);
5098 if (src_nid == dst_nid)
5101 /* Always encourage migration to the preferred node. */
5102 if (dst_nid == p->numa_preferred_nid)
5105 /* If both task and group weight improve, this move is a winner. */
5106 if (task_weight(p, dst_nid) > task_weight(p, src_nid) &&
5107 group_weight(p, dst_nid) > group_weight(p, src_nid))
5114 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5116 int src_nid, dst_nid;
5118 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
5121 if (!p->numa_faults_memory || !(env->sd->flags & SD_NUMA))
5124 src_nid = cpu_to_node(env->src_cpu);
5125 dst_nid = cpu_to_node(env->dst_cpu);
5127 if (src_nid == dst_nid)
5130 /* Migrating away from the preferred node is always bad. */
5131 if (src_nid == p->numa_preferred_nid)
5134 /* If either task or group weight get worse, don't do it. */
5135 if (task_weight(p, dst_nid) < task_weight(p, src_nid) ||
5136 group_weight(p, dst_nid) < group_weight(p, src_nid))
5143 static inline bool migrate_improves_locality(struct task_struct *p,
5149 static inline bool migrate_degrades_locality(struct task_struct *p,
5157 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5160 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5162 int tsk_cache_hot = 0;
5164 * We do not migrate tasks that are:
5165 * 1) throttled_lb_pair, or
5166 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5167 * 3) running (obviously), or
5168 * 4) are cache-hot on their current CPU.
5170 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5173 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5176 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5178 env->flags |= LBF_SOME_PINNED;
5181 * Remember if this task can be migrated to any other cpu in
5182 * our sched_group. We may want to revisit it if we couldn't
5183 * meet load balance goals by pulling other tasks on src_cpu.
5185 * Also avoid computing new_dst_cpu if we have already computed
5186 * one in current iteration.
5188 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5191 /* Prevent to re-select dst_cpu via env's cpus */
5192 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5193 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5194 env->flags |= LBF_DST_PINNED;
5195 env->new_dst_cpu = cpu;
5203 /* Record that we found atleast one task that could run on dst_cpu */
5204 env->flags &= ~LBF_ALL_PINNED;
5206 if (task_running(env->src_rq, p)) {
5207 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5212 * Aggressive migration if:
5213 * 1) destination numa is preferred
5214 * 2) task is cache cold, or
5215 * 3) too many balance attempts have failed.
5217 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq));
5219 tsk_cache_hot = migrate_degrades_locality(p, env);
5221 if (migrate_improves_locality(p, env)) {
5222 #ifdef CONFIG_SCHEDSTATS
5223 if (tsk_cache_hot) {
5224 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5225 schedstat_inc(p, se.statistics.nr_forced_migrations);
5231 if (!tsk_cache_hot ||
5232 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5234 if (tsk_cache_hot) {
5235 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5236 schedstat_inc(p, se.statistics.nr_forced_migrations);
5242 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5247 * move_one_task tries to move exactly one task from busiest to this_rq, as
5248 * part of active balancing operations within "domain".
5249 * Returns 1 if successful and 0 otherwise.
5251 * Called with both runqueues locked.
5253 static int move_one_task(struct lb_env *env)
5255 struct task_struct *p, *n;
5257 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5258 if (!can_migrate_task(p, env))
5263 * Right now, this is only the second place move_task()
5264 * is called, so we can safely collect move_task()
5265 * stats here rather than inside move_task().
5267 schedstat_inc(env->sd, lb_gained[env->idle]);
5273 static const unsigned int sched_nr_migrate_break = 32;
5276 * move_tasks tries to move up to imbalance weighted load from busiest to
5277 * this_rq, as part of a balancing operation within domain "sd".
5278 * Returns 1 if successful and 0 otherwise.
5280 * Called with both runqueues locked.
5282 static int move_tasks(struct lb_env *env)
5284 struct list_head *tasks = &env->src_rq->cfs_tasks;
5285 struct task_struct *p;
5289 if (env->imbalance <= 0)
5292 while (!list_empty(tasks)) {
5293 p = list_first_entry(tasks, struct task_struct, se.group_node);
5296 /* We've more or less seen every task there is, call it quits */
5297 if (env->loop > env->loop_max)
5300 /* take a breather every nr_migrate tasks */
5301 if (env->loop > env->loop_break) {
5302 env->loop_break += sched_nr_migrate_break;
5303 env->flags |= LBF_NEED_BREAK;
5307 if (!can_migrate_task(p, env))
5310 load = task_h_load(p);
5312 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5315 if ((load / 2) > env->imbalance)
5320 env->imbalance -= load;
5322 #ifdef CONFIG_PREEMPT
5324 * NEWIDLE balancing is a source of latency, so preemptible
5325 * kernels will stop after the first task is pulled to minimize
5326 * the critical section.
5328 if (env->idle == CPU_NEWLY_IDLE)
5333 * We only want to steal up to the prescribed amount of
5336 if (env->imbalance <= 0)
5341 list_move_tail(&p->se.group_node, tasks);
5345 * Right now, this is one of only two places move_task() is called,
5346 * so we can safely collect move_task() stats here rather than
5347 * inside move_task().
5349 schedstat_add(env->sd, lb_gained[env->idle], pulled);
5354 #ifdef CONFIG_FAIR_GROUP_SCHED
5356 * update tg->load_weight by folding this cpu's load_avg
5358 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5360 struct sched_entity *se = tg->se[cpu];
5361 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5363 /* throttled entities do not contribute to load */
5364 if (throttled_hierarchy(cfs_rq))
5367 update_cfs_rq_blocked_load(cfs_rq, 1);
5370 update_entity_load_avg(se, 1);
5372 * We pivot on our runnable average having decayed to zero for
5373 * list removal. This generally implies that all our children
5374 * have also been removed (modulo rounding error or bandwidth
5375 * control); however, such cases are rare and we can fix these
5378 * TODO: fix up out-of-order children on enqueue.
5380 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5381 list_del_leaf_cfs_rq(cfs_rq);
5383 struct rq *rq = rq_of(cfs_rq);
5384 update_rq_runnable_avg(rq, rq->nr_running);
5388 static void update_blocked_averages(int cpu)
5390 struct rq *rq = cpu_rq(cpu);
5391 struct cfs_rq *cfs_rq;
5392 unsigned long flags;
5394 raw_spin_lock_irqsave(&rq->lock, flags);
5395 update_rq_clock(rq);
5397 * Iterates the task_group tree in a bottom up fashion, see
5398 * list_add_leaf_cfs_rq() for details.
5400 for_each_leaf_cfs_rq(rq, cfs_rq) {
5402 * Note: We may want to consider periodically releasing
5403 * rq->lock about these updates so that creating many task
5404 * groups does not result in continually extending hold time.
5406 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5409 raw_spin_unlock_irqrestore(&rq->lock, flags);
5413 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5414 * This needs to be done in a top-down fashion because the load of a child
5415 * group is a fraction of its parents load.
5417 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5419 struct rq *rq = rq_of(cfs_rq);
5420 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5421 unsigned long now = jiffies;
5424 if (cfs_rq->last_h_load_update == now)
5427 cfs_rq->h_load_next = NULL;
5428 for_each_sched_entity(se) {
5429 cfs_rq = cfs_rq_of(se);
5430 cfs_rq->h_load_next = se;
5431 if (cfs_rq->last_h_load_update == now)
5436 cfs_rq->h_load = cfs_rq->runnable_load_avg;
5437 cfs_rq->last_h_load_update = now;
5440 while ((se = cfs_rq->h_load_next) != NULL) {
5441 load = cfs_rq->h_load;
5442 load = div64_ul(load * se->avg.load_avg_contrib,
5443 cfs_rq->runnable_load_avg + 1);
5444 cfs_rq = group_cfs_rq(se);
5445 cfs_rq->h_load = load;
5446 cfs_rq->last_h_load_update = now;
5450 static unsigned long task_h_load(struct task_struct *p)
5452 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5454 update_cfs_rq_h_load(cfs_rq);
5455 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
5456 cfs_rq->runnable_load_avg + 1);
5459 static inline void update_blocked_averages(int cpu)
5463 static unsigned long task_h_load(struct task_struct *p)
5465 return p->se.avg.load_avg_contrib;
5469 /********** Helpers for find_busiest_group ************************/
5471 * sg_lb_stats - stats of a sched_group required for load_balancing
5473 struct sg_lb_stats {
5474 unsigned long avg_load; /*Avg load across the CPUs of the group */
5475 unsigned long group_load; /* Total load over the CPUs of the group */
5476 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5477 unsigned long load_per_task;
5478 unsigned long group_power;
5479 unsigned int sum_nr_running; /* Nr tasks running in the group */
5480 unsigned int group_capacity;
5481 unsigned int idle_cpus;
5482 unsigned int group_weight;
5483 int group_imb; /* Is there an imbalance in the group ? */
5484 int group_has_capacity; /* Is there extra capacity in the group? */
5485 #ifdef CONFIG_NUMA_BALANCING
5486 unsigned int nr_numa_running;
5487 unsigned int nr_preferred_running;
5492 * sd_lb_stats - Structure to store the statistics of a sched_domain
5493 * during load balancing.
5495 struct sd_lb_stats {
5496 struct sched_group *busiest; /* Busiest group in this sd */
5497 struct sched_group *local; /* Local group in this sd */
5498 unsigned long total_load; /* Total load of all groups in sd */
5499 unsigned long total_pwr; /* Total power of all groups in sd */
5500 unsigned long avg_load; /* Average load across all groups in sd */
5502 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5503 struct sg_lb_stats local_stat; /* Statistics of the local group */
5506 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5509 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5510 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5511 * We must however clear busiest_stat::avg_load because
5512 * update_sd_pick_busiest() reads this before assignment.
5514 *sds = (struct sd_lb_stats){
5526 * get_sd_load_idx - Obtain the load index for a given sched domain.
5527 * @sd: The sched_domain whose load_idx is to be obtained.
5528 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5530 * Return: The load index.
5532 static inline int get_sd_load_idx(struct sched_domain *sd,
5533 enum cpu_idle_type idle)
5539 load_idx = sd->busy_idx;
5542 case CPU_NEWLY_IDLE:
5543 load_idx = sd->newidle_idx;
5546 load_idx = sd->idle_idx;
5553 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
5555 return SCHED_POWER_SCALE;
5558 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
5560 return default_scale_freq_power(sd, cpu);
5563 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
5565 unsigned long weight = sd->span_weight;
5566 unsigned long smt_gain = sd->smt_gain;
5573 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
5575 return default_scale_smt_power(sd, cpu);
5578 static unsigned long scale_rt_power(int cpu)
5580 struct rq *rq = cpu_rq(cpu);
5581 u64 total, available, age_stamp, avg;
5585 * Since we're reading these variables without serialization make sure
5586 * we read them once before doing sanity checks on them.
5588 age_stamp = ACCESS_ONCE(rq->age_stamp);
5589 avg = ACCESS_ONCE(rq->rt_avg);
5591 delta = rq_clock(rq) - age_stamp;
5592 if (unlikely(delta < 0))
5595 total = sched_avg_period() + delta;
5597 if (unlikely(total < avg)) {
5598 /* Ensures that power won't end up being negative */
5601 available = total - avg;
5604 if (unlikely((s64)total < SCHED_POWER_SCALE))
5605 total = SCHED_POWER_SCALE;
5607 total >>= SCHED_POWER_SHIFT;
5609 return div_u64(available, total);
5612 static void update_cpu_power(struct sched_domain *sd, int cpu)
5614 unsigned long weight = sd->span_weight;
5615 unsigned long power = SCHED_POWER_SCALE;
5616 struct sched_group *sdg = sd->groups;
5618 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
5619 if (sched_feat(ARCH_POWER))
5620 power *= arch_scale_smt_power(sd, cpu);
5622 power *= default_scale_smt_power(sd, cpu);
5624 power >>= SCHED_POWER_SHIFT;
5627 sdg->sgp->power_orig = power;
5629 if (sched_feat(ARCH_POWER))
5630 power *= arch_scale_freq_power(sd, cpu);
5632 power *= default_scale_freq_power(sd, cpu);
5634 power >>= SCHED_POWER_SHIFT;
5636 power *= scale_rt_power(cpu);
5637 power >>= SCHED_POWER_SHIFT;
5642 cpu_rq(cpu)->cpu_power = power;
5643 sdg->sgp->power = power;
5646 void update_group_power(struct sched_domain *sd, int cpu)
5648 struct sched_domain *child = sd->child;
5649 struct sched_group *group, *sdg = sd->groups;
5650 unsigned long power, power_orig;
5651 unsigned long interval;
5653 interval = msecs_to_jiffies(sd->balance_interval);
5654 interval = clamp(interval, 1UL, max_load_balance_interval);
5655 sdg->sgp->next_update = jiffies + interval;
5658 update_cpu_power(sd, cpu);
5662 power_orig = power = 0;
5664 if (child->flags & SD_OVERLAP) {
5666 * SD_OVERLAP domains cannot assume that child groups
5667 * span the current group.
5670 for_each_cpu(cpu, sched_group_cpus(sdg)) {
5671 struct sched_group_power *sgp;
5672 struct rq *rq = cpu_rq(cpu);
5675 * build_sched_domains() -> init_sched_groups_power()
5676 * gets here before we've attached the domains to the
5679 * Use power_of(), which is set irrespective of domains
5680 * in update_cpu_power().
5682 * This avoids power/power_orig from being 0 and
5683 * causing divide-by-zero issues on boot.
5685 * Runtime updates will correct power_orig.
5687 if (unlikely(!rq->sd)) {
5688 power_orig += power_of(cpu);
5689 power += power_of(cpu);
5693 sgp = rq->sd->groups->sgp;
5694 power_orig += sgp->power_orig;
5695 power += sgp->power;
5699 * !SD_OVERLAP domains can assume that child groups
5700 * span the current group.
5703 group = child->groups;
5705 power_orig += group->sgp->power_orig;
5706 power += group->sgp->power;
5707 group = group->next;
5708 } while (group != child->groups);
5711 sdg->sgp->power_orig = power_orig;
5712 sdg->sgp->power = power;
5716 * Try and fix up capacity for tiny siblings, this is needed when
5717 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5718 * which on its own isn't powerful enough.
5720 * See update_sd_pick_busiest() and check_asym_packing().
5723 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5726 * Only siblings can have significantly less than SCHED_POWER_SCALE
5728 if (!(sd->flags & SD_SHARE_CPUPOWER))
5732 * If ~90% of the cpu_power is still there, we're good.
5734 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
5741 * Group imbalance indicates (and tries to solve) the problem where balancing
5742 * groups is inadequate due to tsk_cpus_allowed() constraints.
5744 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5745 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5748 * { 0 1 2 3 } { 4 5 6 7 }
5751 * If we were to balance group-wise we'd place two tasks in the first group and
5752 * two tasks in the second group. Clearly this is undesired as it will overload
5753 * cpu 3 and leave one of the cpus in the second group unused.
5755 * The current solution to this issue is detecting the skew in the first group
5756 * by noticing the lower domain failed to reach balance and had difficulty
5757 * moving tasks due to affinity constraints.
5759 * When this is so detected; this group becomes a candidate for busiest; see
5760 * update_sd_pick_busiest(). And calculate_imbalance() and
5761 * find_busiest_group() avoid some of the usual balance conditions to allow it
5762 * to create an effective group imbalance.
5764 * This is a somewhat tricky proposition since the next run might not find the
5765 * group imbalance and decide the groups need to be balanced again. A most
5766 * subtle and fragile situation.
5769 static inline int sg_imbalanced(struct sched_group *group)
5771 return group->sgp->imbalance;
5775 * Compute the group capacity.
5777 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
5778 * first dividing out the smt factor and computing the actual number of cores
5779 * and limit power unit capacity with that.
5781 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
5783 unsigned int capacity, smt, cpus;
5784 unsigned int power, power_orig;
5786 power = group->sgp->power;
5787 power_orig = group->sgp->power_orig;
5788 cpus = group->group_weight;
5790 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
5791 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
5792 capacity = cpus / smt; /* cores */
5794 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
5796 capacity = fix_small_capacity(env->sd, group);
5802 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5803 * @env: The load balancing environment.
5804 * @group: sched_group whose statistics are to be updated.
5805 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5806 * @local_group: Does group contain this_cpu.
5807 * @sgs: variable to hold the statistics for this group.
5809 static inline void update_sg_lb_stats(struct lb_env *env,
5810 struct sched_group *group, int load_idx,
5811 int local_group, struct sg_lb_stats *sgs)
5816 memset(sgs, 0, sizeof(*sgs));
5818 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5819 struct rq *rq = cpu_rq(i);
5821 /* Bias balancing toward cpus of our domain */
5823 load = target_load(i, load_idx);
5825 load = source_load(i, load_idx);
5827 sgs->group_load += load;
5828 sgs->sum_nr_running += rq->nr_running;
5829 #ifdef CONFIG_NUMA_BALANCING
5830 sgs->nr_numa_running += rq->nr_numa_running;
5831 sgs->nr_preferred_running += rq->nr_preferred_running;
5833 sgs->sum_weighted_load += weighted_cpuload(i);
5838 /* Adjust by relative CPU power of the group */
5839 sgs->group_power = group->sgp->power;
5840 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
5842 if (sgs->sum_nr_running)
5843 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
5845 sgs->group_weight = group->group_weight;
5847 sgs->group_imb = sg_imbalanced(group);
5848 sgs->group_capacity = sg_capacity(env, group);
5850 if (sgs->group_capacity > sgs->sum_nr_running)
5851 sgs->group_has_capacity = 1;
5855 * update_sd_pick_busiest - return 1 on busiest group
5856 * @env: The load balancing environment.
5857 * @sds: sched_domain statistics
5858 * @sg: sched_group candidate to be checked for being the busiest
5859 * @sgs: sched_group statistics
5861 * Determine if @sg is a busier group than the previously selected
5864 * Return: %true if @sg is a busier group than the previously selected
5865 * busiest group. %false otherwise.
5867 static bool update_sd_pick_busiest(struct lb_env *env,
5868 struct sd_lb_stats *sds,
5869 struct sched_group *sg,
5870 struct sg_lb_stats *sgs)
5872 if (sgs->avg_load <= sds->busiest_stat.avg_load)
5875 if (sgs->sum_nr_running > sgs->group_capacity)
5882 * ASYM_PACKING needs to move all the work to the lowest
5883 * numbered CPUs in the group, therefore mark all groups
5884 * higher than ourself as busy.
5886 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5887 env->dst_cpu < group_first_cpu(sg)) {
5891 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5898 #ifdef CONFIG_NUMA_BALANCING
5899 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5901 if (sgs->sum_nr_running > sgs->nr_numa_running)
5903 if (sgs->sum_nr_running > sgs->nr_preferred_running)
5908 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5910 if (rq->nr_running > rq->nr_numa_running)
5912 if (rq->nr_running > rq->nr_preferred_running)
5917 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5922 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5926 #endif /* CONFIG_NUMA_BALANCING */
5929 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5930 * @env: The load balancing environment.
5931 * @sds: variable to hold the statistics for this sched_domain.
5933 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
5935 struct sched_domain *child = env->sd->child;
5936 struct sched_group *sg = env->sd->groups;
5937 struct sg_lb_stats tmp_sgs;
5938 int load_idx, prefer_sibling = 0;
5940 if (child && child->flags & SD_PREFER_SIBLING)
5943 load_idx = get_sd_load_idx(env->sd, env->idle);
5946 struct sg_lb_stats *sgs = &tmp_sgs;
5949 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
5952 sgs = &sds->local_stat;
5954 if (env->idle != CPU_NEWLY_IDLE ||
5955 time_after_eq(jiffies, sg->sgp->next_update))
5956 update_group_power(env->sd, env->dst_cpu);
5959 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
5965 * In case the child domain prefers tasks go to siblings
5966 * first, lower the sg capacity to one so that we'll try
5967 * and move all the excess tasks away. We lower the capacity
5968 * of a group only if the local group has the capacity to fit
5969 * these excess tasks, i.e. nr_running < group_capacity. The
5970 * extra check prevents the case where you always pull from the
5971 * heaviest group when it is already under-utilized (possible
5972 * with a large weight task outweighs the tasks on the system).
5974 if (prefer_sibling && sds->local &&
5975 sds->local_stat.group_has_capacity)
5976 sgs->group_capacity = min(sgs->group_capacity, 1U);
5978 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
5980 sds->busiest_stat = *sgs;
5984 /* Now, start updating sd_lb_stats */
5985 sds->total_load += sgs->group_load;
5986 sds->total_pwr += sgs->group_power;
5989 } while (sg != env->sd->groups);
5991 if (env->sd->flags & SD_NUMA)
5992 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
5996 * check_asym_packing - Check to see if the group is packed into the
5999 * This is primarily intended to used at the sibling level. Some
6000 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6001 * case of POWER7, it can move to lower SMT modes only when higher
6002 * threads are idle. When in lower SMT modes, the threads will
6003 * perform better since they share less core resources. Hence when we
6004 * have idle threads, we want them to be the higher ones.
6006 * This packing function is run on idle threads. It checks to see if
6007 * the busiest CPU in this domain (core in the P7 case) has a higher
6008 * CPU number than the packing function is being run on. Here we are
6009 * assuming lower CPU number will be equivalent to lower a SMT thread
6012 * Return: 1 when packing is required and a task should be moved to
6013 * this CPU. The amount of the imbalance is returned in *imbalance.
6015 * @env: The load balancing environment.
6016 * @sds: Statistics of the sched_domain which is to be packed
6018 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6022 if (!(env->sd->flags & SD_ASYM_PACKING))
6028 busiest_cpu = group_first_cpu(sds->busiest);
6029 if (env->dst_cpu > busiest_cpu)
6032 env->imbalance = DIV_ROUND_CLOSEST(
6033 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
6040 * fix_small_imbalance - Calculate the minor imbalance that exists
6041 * amongst the groups of a sched_domain, during
6043 * @env: The load balancing environment.
6044 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6047 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6049 unsigned long tmp, pwr_now = 0, pwr_move = 0;
6050 unsigned int imbn = 2;
6051 unsigned long scaled_busy_load_per_task;
6052 struct sg_lb_stats *local, *busiest;
6054 local = &sds->local_stat;
6055 busiest = &sds->busiest_stat;
6057 if (!local->sum_nr_running)
6058 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6059 else if (busiest->load_per_task > local->load_per_task)
6062 scaled_busy_load_per_task =
6063 (busiest->load_per_task * SCHED_POWER_SCALE) /
6064 busiest->group_power;
6066 if (busiest->avg_load + scaled_busy_load_per_task >=
6067 local->avg_load + (scaled_busy_load_per_task * imbn)) {
6068 env->imbalance = busiest->load_per_task;
6073 * OK, we don't have enough imbalance to justify moving tasks,
6074 * however we may be able to increase total CPU power used by
6078 pwr_now += busiest->group_power *
6079 min(busiest->load_per_task, busiest->avg_load);
6080 pwr_now += local->group_power *
6081 min(local->load_per_task, local->avg_load);
6082 pwr_now /= SCHED_POWER_SCALE;
6084 /* Amount of load we'd subtract */
6085 if (busiest->avg_load > scaled_busy_load_per_task) {
6086 pwr_move += busiest->group_power *
6087 min(busiest->load_per_task,
6088 busiest->avg_load - scaled_busy_load_per_task);
6091 /* Amount of load we'd add */
6092 if (busiest->avg_load * busiest->group_power <
6093 busiest->load_per_task * SCHED_POWER_SCALE) {
6094 tmp = (busiest->avg_load * busiest->group_power) /
6097 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
6100 pwr_move += local->group_power *
6101 min(local->load_per_task, local->avg_load + tmp);
6102 pwr_move /= SCHED_POWER_SCALE;
6104 /* Move if we gain throughput */
6105 if (pwr_move > pwr_now)
6106 env->imbalance = busiest->load_per_task;
6110 * calculate_imbalance - Calculate the amount of imbalance present within the
6111 * groups of a given sched_domain during load balance.
6112 * @env: load balance environment
6113 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6115 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6117 unsigned long max_pull, load_above_capacity = ~0UL;
6118 struct sg_lb_stats *local, *busiest;
6120 local = &sds->local_stat;
6121 busiest = &sds->busiest_stat;
6123 if (busiest->group_imb) {
6125 * In the group_imb case we cannot rely on group-wide averages
6126 * to ensure cpu-load equilibrium, look at wider averages. XXX
6128 busiest->load_per_task =
6129 min(busiest->load_per_task, sds->avg_load);
6133 * In the presence of smp nice balancing, certain scenarios can have
6134 * max load less than avg load(as we skip the groups at or below
6135 * its cpu_power, while calculating max_load..)
6137 if (busiest->avg_load <= sds->avg_load ||
6138 local->avg_load >= sds->avg_load) {
6140 return fix_small_imbalance(env, sds);
6143 if (!busiest->group_imb) {
6145 * Don't want to pull so many tasks that a group would go idle.
6146 * Except of course for the group_imb case, since then we might
6147 * have to drop below capacity to reach cpu-load equilibrium.
6149 load_above_capacity =
6150 (busiest->sum_nr_running - busiest->group_capacity);
6152 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
6153 load_above_capacity /= busiest->group_power;
6157 * We're trying to get all the cpus to the average_load, so we don't
6158 * want to push ourselves above the average load, nor do we wish to
6159 * reduce the max loaded cpu below the average load. At the same time,
6160 * we also don't want to reduce the group load below the group capacity
6161 * (so that we can implement power-savings policies etc). Thus we look
6162 * for the minimum possible imbalance.
6164 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6166 /* How much load to actually move to equalise the imbalance */
6167 env->imbalance = min(
6168 max_pull * busiest->group_power,
6169 (sds->avg_load - local->avg_load) * local->group_power
6170 ) / SCHED_POWER_SCALE;
6173 * if *imbalance is less than the average load per runnable task
6174 * there is no guarantee that any tasks will be moved so we'll have
6175 * a think about bumping its value to force at least one task to be
6178 if (env->imbalance < busiest->load_per_task)
6179 return fix_small_imbalance(env, sds);
6182 /******* find_busiest_group() helpers end here *********************/
6185 * find_busiest_group - Returns the busiest group within the sched_domain
6186 * if there is an imbalance. If there isn't an imbalance, and
6187 * the user has opted for power-savings, it returns a group whose
6188 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6189 * such a group exists.
6191 * Also calculates the amount of weighted load which should be moved
6192 * to restore balance.
6194 * @env: The load balancing environment.
6196 * Return: - The busiest group if imbalance exists.
6197 * - If no imbalance and user has opted for power-savings balance,
6198 * return the least loaded group whose CPUs can be
6199 * put to idle by rebalancing its tasks onto our group.
6201 static struct sched_group *find_busiest_group(struct lb_env *env)
6203 struct sg_lb_stats *local, *busiest;
6204 struct sd_lb_stats sds;
6206 init_sd_lb_stats(&sds);
6209 * Compute the various statistics relavent for load balancing at
6212 update_sd_lb_stats(env, &sds);
6213 local = &sds.local_stat;
6214 busiest = &sds.busiest_stat;
6216 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6217 check_asym_packing(env, &sds))
6220 /* There is no busy sibling group to pull tasks from */
6221 if (!sds.busiest || busiest->sum_nr_running == 0)
6224 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
6227 * If the busiest group is imbalanced the below checks don't
6228 * work because they assume all things are equal, which typically
6229 * isn't true due to cpus_allowed constraints and the like.
6231 if (busiest->group_imb)
6234 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6235 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
6236 !busiest->group_has_capacity)
6240 * If the local group is more busy than the selected busiest group
6241 * don't try and pull any tasks.
6243 if (local->avg_load >= busiest->avg_load)
6247 * Don't pull any tasks if this group is already above the domain
6250 if (local->avg_load >= sds.avg_load)
6253 if (env->idle == CPU_IDLE) {
6255 * This cpu is idle. If the busiest group load doesn't
6256 * have more tasks than the number of available cpu's and
6257 * there is no imbalance between this and busiest group
6258 * wrt to idle cpu's, it is balanced.
6260 if ((local->idle_cpus < busiest->idle_cpus) &&
6261 busiest->sum_nr_running <= busiest->group_weight)
6265 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6266 * imbalance_pct to be conservative.
6268 if (100 * busiest->avg_load <=
6269 env->sd->imbalance_pct * local->avg_load)
6274 /* Looks like there is an imbalance. Compute it */
6275 calculate_imbalance(env, &sds);
6284 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6286 static struct rq *find_busiest_queue(struct lb_env *env,
6287 struct sched_group *group)
6289 struct rq *busiest = NULL, *rq;
6290 unsigned long busiest_load = 0, busiest_power = 1;
6293 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6294 unsigned long power, capacity, wl;
6298 rt = fbq_classify_rq(rq);
6301 * We classify groups/runqueues into three groups:
6302 * - regular: there are !numa tasks
6303 * - remote: there are numa tasks that run on the 'wrong' node
6304 * - all: there is no distinction
6306 * In order to avoid migrating ideally placed numa tasks,
6307 * ignore those when there's better options.
6309 * If we ignore the actual busiest queue to migrate another
6310 * task, the next balance pass can still reduce the busiest
6311 * queue by moving tasks around inside the node.
6313 * If we cannot move enough load due to this classification
6314 * the next pass will adjust the group classification and
6315 * allow migration of more tasks.
6317 * Both cases only affect the total convergence complexity.
6319 if (rt > env->fbq_type)
6322 power = power_of(i);
6323 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
6325 capacity = fix_small_capacity(env->sd, group);
6327 wl = weighted_cpuload(i);
6330 * When comparing with imbalance, use weighted_cpuload()
6331 * which is not scaled with the cpu power.
6333 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
6337 * For the load comparisons with the other cpu's, consider
6338 * the weighted_cpuload() scaled with the cpu power, so that
6339 * the load can be moved away from the cpu that is potentially
6340 * running at a lower capacity.
6342 * Thus we're looking for max(wl_i / power_i), crosswise
6343 * multiplication to rid ourselves of the division works out
6344 * to: wl_i * power_j > wl_j * power_i; where j is our
6347 if (wl * busiest_power > busiest_load * power) {
6349 busiest_power = power;
6358 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6359 * so long as it is large enough.
6361 #define MAX_PINNED_INTERVAL 512
6363 /* Working cpumask for load_balance and load_balance_newidle. */
6364 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6366 static int need_active_balance(struct lb_env *env)
6368 struct sched_domain *sd = env->sd;
6370 if (env->idle == CPU_NEWLY_IDLE) {
6373 * ASYM_PACKING needs to force migrate tasks from busy but
6374 * higher numbered CPUs in order to pack all tasks in the
6375 * lowest numbered CPUs.
6377 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6381 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6384 static int active_load_balance_cpu_stop(void *data);
6386 static int should_we_balance(struct lb_env *env)
6388 struct sched_group *sg = env->sd->groups;
6389 struct cpumask *sg_cpus, *sg_mask;
6390 int cpu, balance_cpu = -1;
6393 * In the newly idle case, we will allow all the cpu's
6394 * to do the newly idle load balance.
6396 if (env->idle == CPU_NEWLY_IDLE)
6399 sg_cpus = sched_group_cpus(sg);
6400 sg_mask = sched_group_mask(sg);
6401 /* Try to find first idle cpu */
6402 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6403 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6410 if (balance_cpu == -1)
6411 balance_cpu = group_balance_cpu(sg);
6414 * First idle cpu or the first cpu(busiest) in this sched group
6415 * is eligible for doing load balancing at this and above domains.
6417 return balance_cpu == env->dst_cpu;
6421 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6422 * tasks if there is an imbalance.
6424 static int load_balance(int this_cpu, struct rq *this_rq,
6425 struct sched_domain *sd, enum cpu_idle_type idle,
6426 int *continue_balancing)
6428 int ld_moved, cur_ld_moved, active_balance = 0;
6429 struct sched_domain *sd_parent = sd->parent;
6430 struct sched_group *group;
6432 unsigned long flags;
6433 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
6435 struct lb_env env = {
6437 .dst_cpu = this_cpu,
6439 .dst_grpmask = sched_group_cpus(sd->groups),
6441 .loop_break = sched_nr_migrate_break,
6447 * For NEWLY_IDLE load_balancing, we don't need to consider
6448 * other cpus in our group
6450 if (idle == CPU_NEWLY_IDLE)
6451 env.dst_grpmask = NULL;
6453 cpumask_copy(cpus, cpu_active_mask);
6455 schedstat_inc(sd, lb_count[idle]);
6458 if (!should_we_balance(&env)) {
6459 *continue_balancing = 0;
6463 group = find_busiest_group(&env);
6465 schedstat_inc(sd, lb_nobusyg[idle]);
6469 busiest = find_busiest_queue(&env, group);
6471 schedstat_inc(sd, lb_nobusyq[idle]);
6475 BUG_ON(busiest == env.dst_rq);
6477 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6480 if (busiest->nr_running > 1) {
6482 * Attempt to move tasks. If find_busiest_group has found
6483 * an imbalance but busiest->nr_running <= 1, the group is
6484 * still unbalanced. ld_moved simply stays zero, so it is
6485 * correctly treated as an imbalance.
6487 env.flags |= LBF_ALL_PINNED;
6488 env.src_cpu = busiest->cpu;
6489 env.src_rq = busiest;
6490 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
6493 local_irq_save(flags);
6494 double_rq_lock(env.dst_rq, busiest);
6497 * cur_ld_moved - load moved in current iteration
6498 * ld_moved - cumulative load moved across iterations
6500 cur_ld_moved = move_tasks(&env);
6501 ld_moved += cur_ld_moved;
6502 double_rq_unlock(env.dst_rq, busiest);
6503 local_irq_restore(flags);
6506 * some other cpu did the load balance for us.
6508 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
6509 resched_cpu(env.dst_cpu);
6511 if (env.flags & LBF_NEED_BREAK) {
6512 env.flags &= ~LBF_NEED_BREAK;
6517 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6518 * us and move them to an alternate dst_cpu in our sched_group
6519 * where they can run. The upper limit on how many times we
6520 * iterate on same src_cpu is dependent on number of cpus in our
6523 * This changes load balance semantics a bit on who can move
6524 * load to a given_cpu. In addition to the given_cpu itself
6525 * (or a ilb_cpu acting on its behalf where given_cpu is
6526 * nohz-idle), we now have balance_cpu in a position to move
6527 * load to given_cpu. In rare situations, this may cause
6528 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6529 * _independently_ and at _same_ time to move some load to
6530 * given_cpu) causing exceess load to be moved to given_cpu.
6531 * This however should not happen so much in practice and
6532 * moreover subsequent load balance cycles should correct the
6533 * excess load moved.
6535 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
6537 /* Prevent to re-select dst_cpu via env's cpus */
6538 cpumask_clear_cpu(env.dst_cpu, env.cpus);
6540 env.dst_rq = cpu_rq(env.new_dst_cpu);
6541 env.dst_cpu = env.new_dst_cpu;
6542 env.flags &= ~LBF_DST_PINNED;
6544 env.loop_break = sched_nr_migrate_break;
6547 * Go back to "more_balance" rather than "redo" since we
6548 * need to continue with same src_cpu.
6554 * We failed to reach balance because of affinity.
6557 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
6559 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
6560 *group_imbalance = 1;
6561 } else if (*group_imbalance)
6562 *group_imbalance = 0;
6565 /* All tasks on this runqueue were pinned by CPU affinity */
6566 if (unlikely(env.flags & LBF_ALL_PINNED)) {
6567 cpumask_clear_cpu(cpu_of(busiest), cpus);
6568 if (!cpumask_empty(cpus)) {
6570 env.loop_break = sched_nr_migrate_break;
6578 schedstat_inc(sd, lb_failed[idle]);
6580 * Increment the failure counter only on periodic balance.
6581 * We do not want newidle balance, which can be very
6582 * frequent, pollute the failure counter causing
6583 * excessive cache_hot migrations and active balances.
6585 if (idle != CPU_NEWLY_IDLE)
6586 sd->nr_balance_failed++;
6588 if (need_active_balance(&env)) {
6589 raw_spin_lock_irqsave(&busiest->lock, flags);
6591 /* don't kick the active_load_balance_cpu_stop,
6592 * if the curr task on busiest cpu can't be
6595 if (!cpumask_test_cpu(this_cpu,
6596 tsk_cpus_allowed(busiest->curr))) {
6597 raw_spin_unlock_irqrestore(&busiest->lock,
6599 env.flags |= LBF_ALL_PINNED;
6600 goto out_one_pinned;
6604 * ->active_balance synchronizes accesses to
6605 * ->active_balance_work. Once set, it's cleared
6606 * only after active load balance is finished.
6608 if (!busiest->active_balance) {
6609 busiest->active_balance = 1;
6610 busiest->push_cpu = this_cpu;
6613 raw_spin_unlock_irqrestore(&busiest->lock, flags);
6615 if (active_balance) {
6616 stop_one_cpu_nowait(cpu_of(busiest),
6617 active_load_balance_cpu_stop, busiest,
6618 &busiest->active_balance_work);
6622 * We've kicked active balancing, reset the failure
6625 sd->nr_balance_failed = sd->cache_nice_tries+1;
6628 sd->nr_balance_failed = 0;
6630 if (likely(!active_balance)) {
6631 /* We were unbalanced, so reset the balancing interval */
6632 sd->balance_interval = sd->min_interval;
6635 * If we've begun active balancing, start to back off. This
6636 * case may not be covered by the all_pinned logic if there
6637 * is only 1 task on the busy runqueue (because we don't call
6640 if (sd->balance_interval < sd->max_interval)
6641 sd->balance_interval *= 2;
6647 schedstat_inc(sd, lb_balanced[idle]);
6649 sd->nr_balance_failed = 0;
6652 /* tune up the balancing interval */
6653 if (((env.flags & LBF_ALL_PINNED) &&
6654 sd->balance_interval < MAX_PINNED_INTERVAL) ||
6655 (sd->balance_interval < sd->max_interval))
6656 sd->balance_interval *= 2;
6664 * idle_balance is called by schedule() if this_cpu is about to become
6665 * idle. Attempts to pull tasks from other CPUs.
6667 static int idle_balance(struct rq *this_rq)
6669 struct sched_domain *sd;
6670 int pulled_task = 0;
6671 unsigned long next_balance = jiffies + HZ;
6673 int this_cpu = this_rq->cpu;
6675 idle_enter_fair(this_rq);
6678 * We must set idle_stamp _before_ calling idle_balance(), such that we
6679 * measure the duration of idle_balance() as idle time.
6681 this_rq->idle_stamp = rq_clock(this_rq);
6683 if (this_rq->avg_idle < sysctl_sched_migration_cost)
6687 * Drop the rq->lock, but keep IRQ/preempt disabled.
6689 raw_spin_unlock(&this_rq->lock);
6691 update_blocked_averages(this_cpu);
6693 for_each_domain(this_cpu, sd) {
6694 unsigned long interval;
6695 int continue_balancing = 1;
6696 u64 t0, domain_cost;
6698 if (!(sd->flags & SD_LOAD_BALANCE))
6701 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
6704 if (sd->flags & SD_BALANCE_NEWIDLE) {
6705 t0 = sched_clock_cpu(this_cpu);
6707 /* If we've pulled tasks over stop searching: */
6708 pulled_task = load_balance(this_cpu, this_rq,
6710 &continue_balancing);
6712 domain_cost = sched_clock_cpu(this_cpu) - t0;
6713 if (domain_cost > sd->max_newidle_lb_cost)
6714 sd->max_newidle_lb_cost = domain_cost;
6716 curr_cost += domain_cost;
6719 interval = msecs_to_jiffies(sd->balance_interval);
6720 if (time_after(next_balance, sd->last_balance + interval))
6721 next_balance = sd->last_balance + interval;
6727 raw_spin_lock(&this_rq->lock);
6729 if (curr_cost > this_rq->max_idle_balance_cost)
6730 this_rq->max_idle_balance_cost = curr_cost;
6733 * While browsing the domains, we released the rq lock, a task could
6734 * have been enqueued in the meantime. Since we're not going idle,
6735 * pretend we pulled a task.
6737 if (this_rq->cfs.h_nr_running && !pulled_task)
6740 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
6742 * We are going idle. next_balance may be set based on
6743 * a busy processor. So reset next_balance.
6745 this_rq->next_balance = next_balance;
6749 /* Is there a task of a high priority class? */
6750 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
6754 idle_exit_fair(this_rq);
6755 this_rq->idle_stamp = 0;
6762 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6763 * running tasks off the busiest CPU onto idle CPUs. It requires at
6764 * least 1 task to be running on each physical CPU where possible, and
6765 * avoids physical / logical imbalances.
6767 static int active_load_balance_cpu_stop(void *data)
6769 struct rq *busiest_rq = data;
6770 int busiest_cpu = cpu_of(busiest_rq);
6771 int target_cpu = busiest_rq->push_cpu;
6772 struct rq *target_rq = cpu_rq(target_cpu);
6773 struct sched_domain *sd;
6775 raw_spin_lock_irq(&busiest_rq->lock);
6777 /* make sure the requested cpu hasn't gone down in the meantime */
6778 if (unlikely(busiest_cpu != smp_processor_id() ||
6779 !busiest_rq->active_balance))
6782 /* Is there any task to move? */
6783 if (busiest_rq->nr_running <= 1)
6787 * This condition is "impossible", if it occurs
6788 * we need to fix it. Originally reported by
6789 * Bjorn Helgaas on a 128-cpu setup.
6791 BUG_ON(busiest_rq == target_rq);
6793 /* move a task from busiest_rq to target_rq */
6794 double_lock_balance(busiest_rq, target_rq);
6796 /* Search for an sd spanning us and the target CPU. */
6798 for_each_domain(target_cpu, sd) {
6799 if ((sd->flags & SD_LOAD_BALANCE) &&
6800 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6805 struct lb_env env = {
6807 .dst_cpu = target_cpu,
6808 .dst_rq = target_rq,
6809 .src_cpu = busiest_rq->cpu,
6810 .src_rq = busiest_rq,
6814 schedstat_inc(sd, alb_count);
6816 if (move_one_task(&env))
6817 schedstat_inc(sd, alb_pushed);
6819 schedstat_inc(sd, alb_failed);
6822 double_unlock_balance(busiest_rq, target_rq);
6824 busiest_rq->active_balance = 0;
6825 raw_spin_unlock_irq(&busiest_rq->lock);
6829 static inline int on_null_domain(struct rq *rq)
6831 return unlikely(!rcu_dereference_sched(rq->sd));
6834 #ifdef CONFIG_NO_HZ_COMMON
6836 * idle load balancing details
6837 * - When one of the busy CPUs notice that there may be an idle rebalancing
6838 * needed, they will kick the idle load balancer, which then does idle
6839 * load balancing for all the idle CPUs.
6842 cpumask_var_t idle_cpus_mask;
6844 unsigned long next_balance; /* in jiffy units */
6845 } nohz ____cacheline_aligned;
6847 static inline int find_new_ilb(void)
6849 int ilb = cpumask_first(nohz.idle_cpus_mask);
6851 if (ilb < nr_cpu_ids && idle_cpu(ilb))
6858 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6859 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6860 * CPU (if there is one).
6862 static void nohz_balancer_kick(void)
6866 nohz.next_balance++;
6868 ilb_cpu = find_new_ilb();
6870 if (ilb_cpu >= nr_cpu_ids)
6873 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
6876 * Use smp_send_reschedule() instead of resched_cpu().
6877 * This way we generate a sched IPI on the target cpu which
6878 * is idle. And the softirq performing nohz idle load balance
6879 * will be run before returning from the IPI.
6881 smp_send_reschedule(ilb_cpu);
6885 static inline void nohz_balance_exit_idle(int cpu)
6887 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
6889 * Completely isolated CPUs don't ever set, so we must test.
6891 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
6892 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
6893 atomic_dec(&nohz.nr_cpus);
6895 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6899 static inline void set_cpu_sd_state_busy(void)
6901 struct sched_domain *sd;
6902 int cpu = smp_processor_id();
6905 sd = rcu_dereference(per_cpu(sd_busy, cpu));
6907 if (!sd || !sd->nohz_idle)
6911 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
6916 void set_cpu_sd_state_idle(void)
6918 struct sched_domain *sd;
6919 int cpu = smp_processor_id();
6922 sd = rcu_dereference(per_cpu(sd_busy, cpu));
6924 if (!sd || sd->nohz_idle)
6928 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
6934 * This routine will record that the cpu is going idle with tick stopped.
6935 * This info will be used in performing idle load balancing in the future.
6937 void nohz_balance_enter_idle(int cpu)
6940 * If this cpu is going down, then nothing needs to be done.
6942 if (!cpu_active(cpu))
6945 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
6949 * If we're a completely isolated CPU, we don't play.
6951 if (on_null_domain(cpu_rq(cpu)))
6954 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
6955 atomic_inc(&nohz.nr_cpus);
6956 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6959 static int sched_ilb_notifier(struct notifier_block *nfb,
6960 unsigned long action, void *hcpu)
6962 switch (action & ~CPU_TASKS_FROZEN) {
6964 nohz_balance_exit_idle(smp_processor_id());
6972 static DEFINE_SPINLOCK(balancing);
6975 * Scale the max load_balance interval with the number of CPUs in the system.
6976 * This trades load-balance latency on larger machines for less cross talk.
6978 void update_max_interval(void)
6980 max_load_balance_interval = HZ*num_online_cpus()/10;
6984 * It checks each scheduling domain to see if it is due to be balanced,
6985 * and initiates a balancing operation if so.
6987 * Balancing parameters are set up in init_sched_domains.
6989 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
6991 int continue_balancing = 1;
6993 unsigned long interval;
6994 struct sched_domain *sd;
6995 /* Earliest time when we have to do rebalance again */
6996 unsigned long next_balance = jiffies + 60*HZ;
6997 int update_next_balance = 0;
6998 int need_serialize, need_decay = 0;
7001 update_blocked_averages(cpu);
7004 for_each_domain(cpu, sd) {
7006 * Decay the newidle max times here because this is a regular
7007 * visit to all the domains. Decay ~1% per second.
7009 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7010 sd->max_newidle_lb_cost =
7011 (sd->max_newidle_lb_cost * 253) / 256;
7012 sd->next_decay_max_lb_cost = jiffies + HZ;
7015 max_cost += sd->max_newidle_lb_cost;
7017 if (!(sd->flags & SD_LOAD_BALANCE))
7021 * Stop the load balance at this level. There is another
7022 * CPU in our sched group which is doing load balancing more
7025 if (!continue_balancing) {
7031 interval = sd->balance_interval;
7032 if (idle != CPU_IDLE)
7033 interval *= sd->busy_factor;
7035 /* scale ms to jiffies */
7036 interval = msecs_to_jiffies(interval);
7037 interval = clamp(interval, 1UL, max_load_balance_interval);
7039 need_serialize = sd->flags & SD_SERIALIZE;
7041 if (need_serialize) {
7042 if (!spin_trylock(&balancing))
7046 if (time_after_eq(jiffies, sd->last_balance + interval)) {
7047 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7049 * The LBF_DST_PINNED logic could have changed
7050 * env->dst_cpu, so we can't know our idle
7051 * state even if we migrated tasks. Update it.
7053 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7055 sd->last_balance = jiffies;
7058 spin_unlock(&balancing);
7060 if (time_after(next_balance, sd->last_balance + interval)) {
7061 next_balance = sd->last_balance + interval;
7062 update_next_balance = 1;
7067 * Ensure the rq-wide value also decays but keep it at a
7068 * reasonable floor to avoid funnies with rq->avg_idle.
7070 rq->max_idle_balance_cost =
7071 max((u64)sysctl_sched_migration_cost, max_cost);
7076 * next_balance will be updated only when there is a need.
7077 * When the cpu is attached to null domain for ex, it will not be
7080 if (likely(update_next_balance))
7081 rq->next_balance = next_balance;
7084 #ifdef CONFIG_NO_HZ_COMMON
7086 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7087 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7089 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7091 int this_cpu = this_rq->cpu;
7095 if (idle != CPU_IDLE ||
7096 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7099 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7100 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7104 * If this cpu gets work to do, stop the load balancing
7105 * work being done for other cpus. Next load
7106 * balancing owner will pick it up.
7111 rq = cpu_rq(balance_cpu);
7113 raw_spin_lock_irq(&rq->lock);
7114 update_rq_clock(rq);
7115 update_idle_cpu_load(rq);
7116 raw_spin_unlock_irq(&rq->lock);
7118 rebalance_domains(rq, CPU_IDLE);
7120 if (time_after(this_rq->next_balance, rq->next_balance))
7121 this_rq->next_balance = rq->next_balance;
7123 nohz.next_balance = this_rq->next_balance;
7125 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7129 * Current heuristic for kicking the idle load balancer in the presence
7130 * of an idle cpu is the system.
7131 * - This rq has more than one task.
7132 * - At any scheduler domain level, this cpu's scheduler group has multiple
7133 * busy cpu's exceeding the group's power.
7134 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7135 * domain span are idle.
7137 static inline int nohz_kick_needed(struct rq *rq)
7139 unsigned long now = jiffies;
7140 struct sched_domain *sd;
7141 struct sched_group_power *sgp;
7142 int nr_busy, cpu = rq->cpu;
7144 if (unlikely(rq->idle_balance))
7148 * We may be recently in ticked or tickless idle mode. At the first
7149 * busy tick after returning from idle, we will update the busy stats.
7151 set_cpu_sd_state_busy();
7152 nohz_balance_exit_idle(cpu);
7155 * None are in tickless mode and hence no need for NOHZ idle load
7158 if (likely(!atomic_read(&nohz.nr_cpus)))
7161 if (time_before(now, nohz.next_balance))
7164 if (rq->nr_running >= 2)
7168 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7171 sgp = sd->groups->sgp;
7172 nr_busy = atomic_read(&sgp->nr_busy_cpus);
7175 goto need_kick_unlock;
7178 sd = rcu_dereference(per_cpu(sd_asym, cpu));
7180 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7181 sched_domain_span(sd)) < cpu))
7182 goto need_kick_unlock;
7193 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7197 * run_rebalance_domains is triggered when needed from the scheduler tick.
7198 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7200 static void run_rebalance_domains(struct softirq_action *h)
7202 struct rq *this_rq = this_rq();
7203 enum cpu_idle_type idle = this_rq->idle_balance ?
7204 CPU_IDLE : CPU_NOT_IDLE;
7206 rebalance_domains(this_rq, idle);
7209 * If this cpu has a pending nohz_balance_kick, then do the
7210 * balancing on behalf of the other idle cpus whose ticks are
7213 nohz_idle_balance(this_rq, idle);
7217 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7219 void trigger_load_balance(struct rq *rq)
7221 /* Don't need to rebalance while attached to NULL domain */
7222 if (unlikely(on_null_domain(rq)))
7225 if (time_after_eq(jiffies, rq->next_balance))
7226 raise_softirq(SCHED_SOFTIRQ);
7227 #ifdef CONFIG_NO_HZ_COMMON
7228 if (nohz_kick_needed(rq))
7229 nohz_balancer_kick();
7233 static void rq_online_fair(struct rq *rq)
7238 static void rq_offline_fair(struct rq *rq)
7242 /* Ensure any throttled groups are reachable by pick_next_task */
7243 unthrottle_offline_cfs_rqs(rq);
7246 #endif /* CONFIG_SMP */
7249 * scheduler tick hitting a task of our scheduling class:
7251 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7253 struct cfs_rq *cfs_rq;
7254 struct sched_entity *se = &curr->se;
7256 for_each_sched_entity(se) {
7257 cfs_rq = cfs_rq_of(se);
7258 entity_tick(cfs_rq, se, queued);
7261 if (numabalancing_enabled)
7262 task_tick_numa(rq, curr);
7264 update_rq_runnable_avg(rq, 1);
7268 * called on fork with the child task as argument from the parent's context
7269 * - child not yet on the tasklist
7270 * - preemption disabled
7272 static void task_fork_fair(struct task_struct *p)
7274 struct cfs_rq *cfs_rq;
7275 struct sched_entity *se = &p->se, *curr;
7276 int this_cpu = smp_processor_id();
7277 struct rq *rq = this_rq();
7278 unsigned long flags;
7280 raw_spin_lock_irqsave(&rq->lock, flags);
7282 update_rq_clock(rq);
7284 cfs_rq = task_cfs_rq(current);
7285 curr = cfs_rq->curr;
7288 * Not only the cpu but also the task_group of the parent might have
7289 * been changed after parent->se.parent,cfs_rq were copied to
7290 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7291 * of child point to valid ones.
7294 __set_task_cpu(p, this_cpu);
7297 update_curr(cfs_rq);
7300 se->vruntime = curr->vruntime;
7301 place_entity(cfs_rq, se, 1);
7303 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
7305 * Upon rescheduling, sched_class::put_prev_task() will place
7306 * 'current' within the tree based on its new key value.
7308 swap(curr->vruntime, se->vruntime);
7309 resched_task(rq->curr);
7312 se->vruntime -= cfs_rq->min_vruntime;
7314 raw_spin_unlock_irqrestore(&rq->lock, flags);
7318 * Priority of the task has changed. Check to see if we preempt
7322 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
7328 * Reschedule if we are currently running on this runqueue and
7329 * our priority decreased, or if we are not currently running on
7330 * this runqueue and our priority is higher than the current's
7332 if (rq->curr == p) {
7333 if (p->prio > oldprio)
7334 resched_task(rq->curr);
7336 check_preempt_curr(rq, p, 0);
7339 static void switched_from_fair(struct rq *rq, struct task_struct *p)
7341 struct sched_entity *se = &p->se;
7342 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7345 * Ensure the task's vruntime is normalized, so that when it's
7346 * switched back to the fair class the enqueue_entity(.flags=0) will
7347 * do the right thing.
7349 * If it's on_rq, then the dequeue_entity(.flags=0) will already
7350 * have normalized the vruntime, if it's !on_rq, then only when
7351 * the task is sleeping will it still have non-normalized vruntime.
7353 if (!p->on_rq && p->state != TASK_RUNNING) {
7355 * Fix up our vruntime so that the current sleep doesn't
7356 * cause 'unlimited' sleep bonus.
7358 place_entity(cfs_rq, se, 0);
7359 se->vruntime -= cfs_rq->min_vruntime;
7364 * Remove our load from contribution when we leave sched_fair
7365 * and ensure we don't carry in an old decay_count if we
7368 if (se->avg.decay_count) {
7369 __synchronize_entity_decay(se);
7370 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
7376 * We switched to the sched_fair class.
7378 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7380 struct sched_entity *se = &p->se;
7381 #ifdef CONFIG_FAIR_GROUP_SCHED
7383 * Since the real-depth could have been changed (only FAIR
7384 * class maintain depth value), reset depth properly.
7386 se->depth = se->parent ? se->parent->depth + 1 : 0;
7392 * We were most likely switched from sched_rt, so
7393 * kick off the schedule if running, otherwise just see
7394 * if we can still preempt the current task.
7397 resched_task(rq->curr);
7399 check_preempt_curr(rq, p, 0);
7402 /* Account for a task changing its policy or group.
7404 * This routine is mostly called to set cfs_rq->curr field when a task
7405 * migrates between groups/classes.
7407 static void set_curr_task_fair(struct rq *rq)
7409 struct sched_entity *se = &rq->curr->se;
7411 for_each_sched_entity(se) {
7412 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7414 set_next_entity(cfs_rq, se);
7415 /* ensure bandwidth has been allocated on our new cfs_rq */
7416 account_cfs_rq_runtime(cfs_rq, 0);
7420 void init_cfs_rq(struct cfs_rq *cfs_rq)
7422 cfs_rq->tasks_timeline = RB_ROOT;
7423 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7424 #ifndef CONFIG_64BIT
7425 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7428 atomic64_set(&cfs_rq->decay_counter, 1);
7429 atomic_long_set(&cfs_rq->removed_load, 0);
7433 #ifdef CONFIG_FAIR_GROUP_SCHED
7434 static void task_move_group_fair(struct task_struct *p, int on_rq)
7436 struct sched_entity *se = &p->se;
7437 struct cfs_rq *cfs_rq;
7440 * If the task was not on the rq at the time of this cgroup movement
7441 * it must have been asleep, sleeping tasks keep their ->vruntime
7442 * absolute on their old rq until wakeup (needed for the fair sleeper
7443 * bonus in place_entity()).
7445 * If it was on the rq, we've just 'preempted' it, which does convert
7446 * ->vruntime to a relative base.
7448 * Make sure both cases convert their relative position when migrating
7449 * to another cgroup's rq. This does somewhat interfere with the
7450 * fair sleeper stuff for the first placement, but who cares.
7453 * When !on_rq, vruntime of the task has usually NOT been normalized.
7454 * But there are some cases where it has already been normalized:
7456 * - Moving a forked child which is waiting for being woken up by
7457 * wake_up_new_task().
7458 * - Moving a task which has been woken up by try_to_wake_up() and
7459 * waiting for actually being woken up by sched_ttwu_pending().
7461 * To prevent boost or penalty in the new cfs_rq caused by delta
7462 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7464 if (!on_rq && (!se->sum_exec_runtime || p->state == TASK_WAKING))
7468 se->vruntime -= cfs_rq_of(se)->min_vruntime;
7469 set_task_rq(p, task_cpu(p));
7470 se->depth = se->parent ? se->parent->depth + 1 : 0;
7472 cfs_rq = cfs_rq_of(se);
7473 se->vruntime += cfs_rq->min_vruntime;
7476 * migrate_task_rq_fair() will have removed our previous
7477 * contribution, but we must synchronize for ongoing future
7480 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7481 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
7486 void free_fair_sched_group(struct task_group *tg)
7490 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7492 for_each_possible_cpu(i) {
7494 kfree(tg->cfs_rq[i]);
7503 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7505 struct cfs_rq *cfs_rq;
7506 struct sched_entity *se;
7509 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7512 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7516 tg->shares = NICE_0_LOAD;
7518 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7520 for_each_possible_cpu(i) {
7521 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7522 GFP_KERNEL, cpu_to_node(i));
7526 se = kzalloc_node(sizeof(struct sched_entity),
7527 GFP_KERNEL, cpu_to_node(i));
7531 init_cfs_rq(cfs_rq);
7532 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7543 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7545 struct rq *rq = cpu_rq(cpu);
7546 unsigned long flags;
7549 * Only empty task groups can be destroyed; so we can speculatively
7550 * check on_list without danger of it being re-added.
7552 if (!tg->cfs_rq[cpu]->on_list)
7555 raw_spin_lock_irqsave(&rq->lock, flags);
7556 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7557 raw_spin_unlock_irqrestore(&rq->lock, flags);
7560 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7561 struct sched_entity *se, int cpu,
7562 struct sched_entity *parent)
7564 struct rq *rq = cpu_rq(cpu);
7568 init_cfs_rq_runtime(cfs_rq);
7570 tg->cfs_rq[cpu] = cfs_rq;
7573 /* se could be NULL for root_task_group */
7578 se->cfs_rq = &rq->cfs;
7581 se->cfs_rq = parent->my_q;
7582 se->depth = parent->depth + 1;
7586 /* guarantee group entities always have weight */
7587 update_load_set(&se->load, NICE_0_LOAD);
7588 se->parent = parent;
7591 static DEFINE_MUTEX(shares_mutex);
7593 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7596 unsigned long flags;
7599 * We can't change the weight of the root cgroup.
7604 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7606 mutex_lock(&shares_mutex);
7607 if (tg->shares == shares)
7610 tg->shares = shares;
7611 for_each_possible_cpu(i) {
7612 struct rq *rq = cpu_rq(i);
7613 struct sched_entity *se;
7616 /* Propagate contribution to hierarchy */
7617 raw_spin_lock_irqsave(&rq->lock, flags);
7619 /* Possible calls to update_curr() need rq clock */
7620 update_rq_clock(rq);
7621 for_each_sched_entity(se)
7622 update_cfs_shares(group_cfs_rq(se));
7623 raw_spin_unlock_irqrestore(&rq->lock, flags);
7627 mutex_unlock(&shares_mutex);
7630 #else /* CONFIG_FAIR_GROUP_SCHED */
7632 void free_fair_sched_group(struct task_group *tg) { }
7634 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7639 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7641 #endif /* CONFIG_FAIR_GROUP_SCHED */
7644 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
7646 struct sched_entity *se = &task->se;
7647 unsigned int rr_interval = 0;
7650 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7653 if (rq->cfs.load.weight)
7654 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
7660 * All the scheduling class methods:
7662 const struct sched_class fair_sched_class = {
7663 .next = &idle_sched_class,
7664 .enqueue_task = enqueue_task_fair,
7665 .dequeue_task = dequeue_task_fair,
7666 .yield_task = yield_task_fair,
7667 .yield_to_task = yield_to_task_fair,
7669 .check_preempt_curr = check_preempt_wakeup,
7671 .pick_next_task = pick_next_task_fair,
7672 .put_prev_task = put_prev_task_fair,
7675 .select_task_rq = select_task_rq_fair,
7676 .migrate_task_rq = migrate_task_rq_fair,
7678 .rq_online = rq_online_fair,
7679 .rq_offline = rq_offline_fair,
7681 .task_waking = task_waking_fair,
7684 .set_curr_task = set_curr_task_fair,
7685 .task_tick = task_tick_fair,
7686 .task_fork = task_fork_fair,
7688 .prio_changed = prio_changed_fair,
7689 .switched_from = switched_from_fair,
7690 .switched_to = switched_to_fair,
7692 .get_rr_interval = get_rr_interval_fair,
7694 #ifdef CONFIG_FAIR_GROUP_SCHED
7695 .task_move_group = task_move_group_fair,
7699 #ifdef CONFIG_SCHED_DEBUG
7700 void print_cfs_stats(struct seq_file *m, int cpu)
7702 struct cfs_rq *cfs_rq;
7705 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7706 print_cfs_rq(m, cpu, cfs_rq);
7711 __init void init_sched_fair_class(void)
7714 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7716 #ifdef CONFIG_NO_HZ_COMMON
7717 nohz.next_balance = jiffies;
7718 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7719 cpu_notifier(sched_ilb_notifier, 0);