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 /* This task has no NUMA fault statistics yet */
1330 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults_memory))
1333 /* Periodically retry migrating the task to the preferred node */
1334 p->numa_migrate_retry = jiffies + HZ;
1336 /* Success if task is already running on preferred CPU */
1337 if (task_node(p) == p->numa_preferred_nid)
1340 /* Otherwise, try migrate to a CPU on the preferred node */
1341 task_numa_migrate(p);
1345 * Find the nodes on which the workload is actively running. We do this by
1346 * tracking the nodes from which NUMA hinting faults are triggered. This can
1347 * be different from the set of nodes where the workload's memory is currently
1350 * The bitmask is used to make smarter decisions on when to do NUMA page
1351 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1352 * are added when they cause over 6/16 of the maximum number of faults, but
1353 * only removed when they drop below 3/16.
1355 static void update_numa_active_node_mask(struct numa_group *numa_group)
1357 unsigned long faults, max_faults = 0;
1360 for_each_online_node(nid) {
1361 faults = group_faults_cpu(numa_group, nid);
1362 if (faults > max_faults)
1363 max_faults = faults;
1366 for_each_online_node(nid) {
1367 faults = group_faults_cpu(numa_group, nid);
1368 if (!node_isset(nid, numa_group->active_nodes)) {
1369 if (faults > max_faults * 6 / 16)
1370 node_set(nid, numa_group->active_nodes);
1371 } else if (faults < max_faults * 3 / 16)
1372 node_clear(nid, numa_group->active_nodes);
1377 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1378 * increments. The more local the fault statistics are, the higher the scan
1379 * period will be for the next scan window. If local/remote ratio is below
1380 * NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) the
1381 * scan period will decrease
1383 #define NUMA_PERIOD_SLOTS 10
1384 #define NUMA_PERIOD_THRESHOLD 3
1387 * Increase the scan period (slow down scanning) if the majority of
1388 * our memory is already on our local node, or if the majority of
1389 * the page accesses are shared with other processes.
1390 * Otherwise, decrease the scan period.
1392 static void update_task_scan_period(struct task_struct *p,
1393 unsigned long shared, unsigned long private)
1395 unsigned int period_slot;
1399 unsigned long remote = p->numa_faults_locality[0];
1400 unsigned long local = p->numa_faults_locality[1];
1403 * If there were no record hinting faults then either the task is
1404 * completely idle or all activity is areas that are not of interest
1405 * to automatic numa balancing. Scan slower
1407 if (local + shared == 0) {
1408 p->numa_scan_period = min(p->numa_scan_period_max,
1409 p->numa_scan_period << 1);
1411 p->mm->numa_next_scan = jiffies +
1412 msecs_to_jiffies(p->numa_scan_period);
1418 * Prepare to scale scan period relative to the current period.
1419 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1420 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1421 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1423 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1424 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1425 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1426 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1429 diff = slot * period_slot;
1431 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1434 * Scale scan rate increases based on sharing. There is an
1435 * inverse relationship between the degree of sharing and
1436 * the adjustment made to the scanning period. Broadly
1437 * speaking the intent is that there is little point
1438 * scanning faster if shared accesses dominate as it may
1439 * simply bounce migrations uselessly
1441 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared));
1442 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1445 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1446 task_scan_min(p), task_scan_max(p));
1447 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1451 * Get the fraction of time the task has been running since the last
1452 * NUMA placement cycle. The scheduler keeps similar statistics, but
1453 * decays those on a 32ms period, which is orders of magnitude off
1454 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1455 * stats only if the task is so new there are no NUMA statistics yet.
1457 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1459 u64 runtime, delta, now;
1460 /* Use the start of this time slice to avoid calculations. */
1461 now = p->se.exec_start;
1462 runtime = p->se.sum_exec_runtime;
1464 if (p->last_task_numa_placement) {
1465 delta = runtime - p->last_sum_exec_runtime;
1466 *period = now - p->last_task_numa_placement;
1468 delta = p->se.avg.runnable_avg_sum;
1469 *period = p->se.avg.runnable_avg_period;
1472 p->last_sum_exec_runtime = runtime;
1473 p->last_task_numa_placement = now;
1478 static void task_numa_placement(struct task_struct *p)
1480 int seq, nid, max_nid = -1, max_group_nid = -1;
1481 unsigned long max_faults = 0, max_group_faults = 0;
1482 unsigned long fault_types[2] = { 0, 0 };
1483 unsigned long total_faults;
1484 u64 runtime, period;
1485 spinlock_t *group_lock = NULL;
1487 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1488 if (p->numa_scan_seq == seq)
1490 p->numa_scan_seq = seq;
1491 p->numa_scan_period_max = task_scan_max(p);
1493 total_faults = p->numa_faults_locality[0] +
1494 p->numa_faults_locality[1];
1495 runtime = numa_get_avg_runtime(p, &period);
1497 /* If the task is part of a group prevent parallel updates to group stats */
1498 if (p->numa_group) {
1499 group_lock = &p->numa_group->lock;
1500 spin_lock(group_lock);
1503 /* Find the node with the highest number of faults */
1504 for_each_online_node(nid) {
1505 unsigned long faults = 0, group_faults = 0;
1508 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1509 long diff, f_diff, f_weight;
1511 i = task_faults_idx(nid, priv);
1513 /* Decay existing window, copy faults since last scan */
1514 diff = p->numa_faults_buffer_memory[i] - p->numa_faults_memory[i] / 2;
1515 fault_types[priv] += p->numa_faults_buffer_memory[i];
1516 p->numa_faults_buffer_memory[i] = 0;
1519 * Normalize the faults_from, so all tasks in a group
1520 * count according to CPU use, instead of by the raw
1521 * number of faults. Tasks with little runtime have
1522 * little over-all impact on throughput, and thus their
1523 * faults are less important.
1525 f_weight = div64_u64(runtime << 16, period + 1);
1526 f_weight = (f_weight * p->numa_faults_buffer_cpu[i]) /
1528 f_diff = f_weight - p->numa_faults_cpu[i] / 2;
1529 p->numa_faults_buffer_cpu[i] = 0;
1531 p->numa_faults_memory[i] += diff;
1532 p->numa_faults_cpu[i] += f_diff;
1533 faults += p->numa_faults_memory[i];
1534 p->total_numa_faults += diff;
1535 if (p->numa_group) {
1536 /* safe because we can only change our own group */
1537 p->numa_group->faults[i] += diff;
1538 p->numa_group->faults_cpu[i] += f_diff;
1539 p->numa_group->total_faults += diff;
1540 group_faults += p->numa_group->faults[i];
1544 if (faults > max_faults) {
1545 max_faults = faults;
1549 if (group_faults > max_group_faults) {
1550 max_group_faults = group_faults;
1551 max_group_nid = nid;
1555 update_task_scan_period(p, fault_types[0], fault_types[1]);
1557 if (p->numa_group) {
1558 update_numa_active_node_mask(p->numa_group);
1560 * If the preferred task and group nids are different,
1561 * iterate over the nodes again to find the best place.
1563 if (max_nid != max_group_nid) {
1564 unsigned long weight, max_weight = 0;
1566 for_each_online_node(nid) {
1567 weight = task_weight(p, nid) + group_weight(p, nid);
1568 if (weight > max_weight) {
1569 max_weight = weight;
1575 spin_unlock(group_lock);
1578 /* Preferred node as the node with the most faults */
1579 if (max_faults && max_nid != p->numa_preferred_nid) {
1580 /* Update the preferred nid and migrate task if possible */
1581 sched_setnuma(p, max_nid);
1582 numa_migrate_preferred(p);
1586 static inline int get_numa_group(struct numa_group *grp)
1588 return atomic_inc_not_zero(&grp->refcount);
1591 static inline void put_numa_group(struct numa_group *grp)
1593 if (atomic_dec_and_test(&grp->refcount))
1594 kfree_rcu(grp, rcu);
1597 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1600 struct numa_group *grp, *my_grp;
1601 struct task_struct *tsk;
1603 int cpu = cpupid_to_cpu(cpupid);
1606 if (unlikely(!p->numa_group)) {
1607 unsigned int size = sizeof(struct numa_group) +
1608 4*nr_node_ids*sizeof(unsigned long);
1610 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1614 atomic_set(&grp->refcount, 1);
1615 spin_lock_init(&grp->lock);
1616 INIT_LIST_HEAD(&grp->task_list);
1618 /* Second half of the array tracks nids where faults happen */
1619 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1622 node_set(task_node(current), grp->active_nodes);
1624 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1625 grp->faults[i] = p->numa_faults_memory[i];
1627 grp->total_faults = p->total_numa_faults;
1629 list_add(&p->numa_entry, &grp->task_list);
1631 rcu_assign_pointer(p->numa_group, grp);
1635 tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1637 if (!cpupid_match_pid(tsk, cpupid))
1640 grp = rcu_dereference(tsk->numa_group);
1644 my_grp = p->numa_group;
1649 * Only join the other group if its bigger; if we're the bigger group,
1650 * the other task will join us.
1652 if (my_grp->nr_tasks > grp->nr_tasks)
1656 * Tie-break on the grp address.
1658 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1661 /* Always join threads in the same process. */
1662 if (tsk->mm == current->mm)
1665 /* Simple filter to avoid false positives due to PID collisions */
1666 if (flags & TNF_SHARED)
1669 /* Update priv based on whether false sharing was detected */
1672 if (join && !get_numa_group(grp))
1680 double_lock(&my_grp->lock, &grp->lock);
1682 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
1683 my_grp->faults[i] -= p->numa_faults_memory[i];
1684 grp->faults[i] += p->numa_faults_memory[i];
1686 my_grp->total_faults -= p->total_numa_faults;
1687 grp->total_faults += p->total_numa_faults;
1689 list_move(&p->numa_entry, &grp->task_list);
1693 spin_unlock(&my_grp->lock);
1694 spin_unlock(&grp->lock);
1696 rcu_assign_pointer(p->numa_group, grp);
1698 put_numa_group(my_grp);
1706 void task_numa_free(struct task_struct *p)
1708 struct numa_group *grp = p->numa_group;
1710 void *numa_faults = p->numa_faults_memory;
1713 spin_lock(&grp->lock);
1714 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1715 grp->faults[i] -= p->numa_faults_memory[i];
1716 grp->total_faults -= p->total_numa_faults;
1718 list_del(&p->numa_entry);
1720 spin_unlock(&grp->lock);
1721 rcu_assign_pointer(p->numa_group, NULL);
1722 put_numa_group(grp);
1725 p->numa_faults_memory = NULL;
1726 p->numa_faults_buffer_memory = NULL;
1727 p->numa_faults_cpu= NULL;
1728 p->numa_faults_buffer_cpu = NULL;
1733 * Got a PROT_NONE fault for a page on @node.
1735 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
1737 struct task_struct *p = current;
1738 bool migrated = flags & TNF_MIGRATED;
1739 int cpu_node = task_node(current);
1742 if (!numabalancing_enabled)
1745 /* for example, ksmd faulting in a user's mm */
1749 /* Do not worry about placement if exiting */
1750 if (p->state == TASK_DEAD)
1753 /* Allocate buffer to track faults on a per-node basis */
1754 if (unlikely(!p->numa_faults_memory)) {
1755 int size = sizeof(*p->numa_faults_memory) *
1756 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
1758 p->numa_faults_memory = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
1759 if (!p->numa_faults_memory)
1762 BUG_ON(p->numa_faults_buffer_memory);
1764 * The averaged statistics, shared & private, memory & cpu,
1765 * occupy the first half of the array. The second half of the
1766 * array is for current counters, which are averaged into the
1767 * first set by task_numa_placement.
1769 p->numa_faults_cpu = p->numa_faults_memory + (2 * nr_node_ids);
1770 p->numa_faults_buffer_memory = p->numa_faults_memory + (4 * nr_node_ids);
1771 p->numa_faults_buffer_cpu = p->numa_faults_memory + (6 * nr_node_ids);
1772 p->total_numa_faults = 0;
1773 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1777 * First accesses are treated as private, otherwise consider accesses
1778 * to be private if the accessing pid has not changed
1780 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
1783 priv = cpupid_match_pid(p, last_cpupid);
1784 if (!priv && !(flags & TNF_NO_GROUP))
1785 task_numa_group(p, last_cpupid, flags, &priv);
1788 task_numa_placement(p);
1791 * Retry task to preferred node migration periodically, in case it
1792 * case it previously failed, or the scheduler moved us.
1794 if (time_after(jiffies, p->numa_migrate_retry))
1795 numa_migrate_preferred(p);
1798 p->numa_pages_migrated += pages;
1800 p->numa_faults_buffer_memory[task_faults_idx(mem_node, priv)] += pages;
1801 p->numa_faults_buffer_cpu[task_faults_idx(cpu_node, priv)] += pages;
1802 p->numa_faults_locality[!!(flags & TNF_FAULT_LOCAL)] += pages;
1805 static void reset_ptenuma_scan(struct task_struct *p)
1807 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1808 p->mm->numa_scan_offset = 0;
1812 * The expensive part of numa migration is done from task_work context.
1813 * Triggered from task_tick_numa().
1815 void task_numa_work(struct callback_head *work)
1817 unsigned long migrate, next_scan, now = jiffies;
1818 struct task_struct *p = current;
1819 struct mm_struct *mm = p->mm;
1820 struct vm_area_struct *vma;
1821 unsigned long start, end;
1822 unsigned long nr_pte_updates = 0;
1825 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1827 work->next = work; /* protect against double add */
1829 * Who cares about NUMA placement when they're dying.
1831 * NOTE: make sure not to dereference p->mm before this check,
1832 * exit_task_work() happens _after_ exit_mm() so we could be called
1833 * without p->mm even though we still had it when we enqueued this
1836 if (p->flags & PF_EXITING)
1839 if (!mm->numa_next_scan) {
1840 mm->numa_next_scan = now +
1841 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1845 * Enforce maximal scan/migration frequency..
1847 migrate = mm->numa_next_scan;
1848 if (time_before(now, migrate))
1851 if (p->numa_scan_period == 0) {
1852 p->numa_scan_period_max = task_scan_max(p);
1853 p->numa_scan_period = task_scan_min(p);
1856 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1857 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1861 * Delay this task enough that another task of this mm will likely win
1862 * the next time around.
1864 p->node_stamp += 2 * TICK_NSEC;
1866 start = mm->numa_scan_offset;
1867 pages = sysctl_numa_balancing_scan_size;
1868 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1872 down_read(&mm->mmap_sem);
1873 vma = find_vma(mm, start);
1875 reset_ptenuma_scan(p);
1879 for (; vma; vma = vma->vm_next) {
1880 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1884 * Shared library pages mapped by multiple processes are not
1885 * migrated as it is expected they are cache replicated. Avoid
1886 * hinting faults in read-only file-backed mappings or the vdso
1887 * as migrating the pages will be of marginal benefit.
1890 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1894 * Skip inaccessible VMAs to avoid any confusion between
1895 * PROT_NONE and NUMA hinting ptes
1897 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
1901 start = max(start, vma->vm_start);
1902 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1903 end = min(end, vma->vm_end);
1904 nr_pte_updates += change_prot_numa(vma, start, end);
1907 * Scan sysctl_numa_balancing_scan_size but ensure that
1908 * at least one PTE is updated so that unused virtual
1909 * address space is quickly skipped.
1912 pages -= (end - start) >> PAGE_SHIFT;
1919 } while (end != vma->vm_end);
1924 * It is possible to reach the end of the VMA list but the last few
1925 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1926 * would find the !migratable VMA on the next scan but not reset the
1927 * scanner to the start so check it now.
1930 mm->numa_scan_offset = start;
1932 reset_ptenuma_scan(p);
1933 up_read(&mm->mmap_sem);
1937 * Drive the periodic memory faults..
1939 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1941 struct callback_head *work = &curr->numa_work;
1945 * We don't care about NUMA placement if we don't have memory.
1947 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1951 * Using runtime rather than walltime has the dual advantage that
1952 * we (mostly) drive the selection from busy threads and that the
1953 * task needs to have done some actual work before we bother with
1956 now = curr->se.sum_exec_runtime;
1957 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1959 if (now - curr->node_stamp > period) {
1960 if (!curr->node_stamp)
1961 curr->numa_scan_period = task_scan_min(curr);
1962 curr->node_stamp += period;
1964 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1965 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1966 task_work_add(curr, work, true);
1971 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1975 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1979 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1982 #endif /* CONFIG_NUMA_BALANCING */
1985 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1987 update_load_add(&cfs_rq->load, se->load.weight);
1988 if (!parent_entity(se))
1989 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1991 if (entity_is_task(se)) {
1992 struct rq *rq = rq_of(cfs_rq);
1994 account_numa_enqueue(rq, task_of(se));
1995 list_add(&se->group_node, &rq->cfs_tasks);
1998 cfs_rq->nr_running++;
2002 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2004 update_load_sub(&cfs_rq->load, se->load.weight);
2005 if (!parent_entity(se))
2006 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2007 if (entity_is_task(se)) {
2008 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2009 list_del_init(&se->group_node);
2011 cfs_rq->nr_running--;
2014 #ifdef CONFIG_FAIR_GROUP_SCHED
2016 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2021 * Use this CPU's actual weight instead of the last load_contribution
2022 * to gain a more accurate current total weight. See
2023 * update_cfs_rq_load_contribution().
2025 tg_weight = atomic_long_read(&tg->load_avg);
2026 tg_weight -= cfs_rq->tg_load_contrib;
2027 tg_weight += cfs_rq->load.weight;
2032 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2034 long tg_weight, load, shares;
2036 tg_weight = calc_tg_weight(tg, cfs_rq);
2037 load = cfs_rq->load.weight;
2039 shares = (tg->shares * load);
2041 shares /= tg_weight;
2043 if (shares < MIN_SHARES)
2044 shares = MIN_SHARES;
2045 if (shares > tg->shares)
2046 shares = tg->shares;
2050 # else /* CONFIG_SMP */
2051 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2055 # endif /* CONFIG_SMP */
2056 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2057 unsigned long weight)
2060 /* commit outstanding execution time */
2061 if (cfs_rq->curr == se)
2062 update_curr(cfs_rq);
2063 account_entity_dequeue(cfs_rq, se);
2066 update_load_set(&se->load, weight);
2069 account_entity_enqueue(cfs_rq, se);
2072 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2074 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2076 struct task_group *tg;
2077 struct sched_entity *se;
2081 se = tg->se[cpu_of(rq_of(cfs_rq))];
2082 if (!se || throttled_hierarchy(cfs_rq))
2085 if (likely(se->load.weight == tg->shares))
2088 shares = calc_cfs_shares(cfs_rq, tg);
2090 reweight_entity(cfs_rq_of(se), se, shares);
2092 #else /* CONFIG_FAIR_GROUP_SCHED */
2093 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2096 #endif /* CONFIG_FAIR_GROUP_SCHED */
2100 * We choose a half-life close to 1 scheduling period.
2101 * Note: The tables below are dependent on this value.
2103 #define LOAD_AVG_PERIOD 32
2104 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2105 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2107 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2108 static const u32 runnable_avg_yN_inv[] = {
2109 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2110 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2111 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2112 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2113 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2114 0x85aac367, 0x82cd8698,
2118 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2119 * over-estimates when re-combining.
2121 static const u32 runnable_avg_yN_sum[] = {
2122 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2123 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2124 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2129 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2131 static __always_inline u64 decay_load(u64 val, u64 n)
2133 unsigned int local_n;
2137 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2140 /* after bounds checking we can collapse to 32-bit */
2144 * As y^PERIOD = 1/2, we can combine
2145 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
2146 * With a look-up table which covers k^n (n<PERIOD)
2148 * To achieve constant time decay_load.
2150 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2151 val >>= local_n / LOAD_AVG_PERIOD;
2152 local_n %= LOAD_AVG_PERIOD;
2155 val *= runnable_avg_yN_inv[local_n];
2156 /* We don't use SRR here since we always want to round down. */
2161 * For updates fully spanning n periods, the contribution to runnable
2162 * average will be: \Sum 1024*y^n
2164 * We can compute this reasonably efficiently by combining:
2165 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2167 static u32 __compute_runnable_contrib(u64 n)
2171 if (likely(n <= LOAD_AVG_PERIOD))
2172 return runnable_avg_yN_sum[n];
2173 else if (unlikely(n >= LOAD_AVG_MAX_N))
2174 return LOAD_AVG_MAX;
2176 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2178 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2179 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2181 n -= LOAD_AVG_PERIOD;
2182 } while (n > LOAD_AVG_PERIOD);
2184 contrib = decay_load(contrib, n);
2185 return contrib + runnable_avg_yN_sum[n];
2189 * We can represent the historical contribution to runnable average as the
2190 * coefficients of a geometric series. To do this we sub-divide our runnable
2191 * history into segments of approximately 1ms (1024us); label the segment that
2192 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2194 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2196 * (now) (~1ms ago) (~2ms ago)
2198 * Let u_i denote the fraction of p_i that the entity was runnable.
2200 * We then designate the fractions u_i as our co-efficients, yielding the
2201 * following representation of historical load:
2202 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2204 * We choose y based on the with of a reasonably scheduling period, fixing:
2207 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2208 * approximately half as much as the contribution to load within the last ms
2211 * When a period "rolls over" and we have new u_0`, multiplying the previous
2212 * sum again by y is sufficient to update:
2213 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2214 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2216 static __always_inline int __update_entity_runnable_avg(u64 now,
2217 struct sched_avg *sa,
2221 u32 runnable_contrib;
2222 int delta_w, decayed = 0;
2224 delta = now - sa->last_runnable_update;
2226 * This should only happen when time goes backwards, which it
2227 * unfortunately does during sched clock init when we swap over to TSC.
2229 if ((s64)delta < 0) {
2230 sa->last_runnable_update = now;
2235 * Use 1024ns as the unit of measurement since it's a reasonable
2236 * approximation of 1us and fast to compute.
2241 sa->last_runnable_update = now;
2243 /* delta_w is the amount already accumulated against our next period */
2244 delta_w = sa->runnable_avg_period % 1024;
2245 if (delta + delta_w >= 1024) {
2246 /* period roll-over */
2250 * Now that we know we're crossing a period boundary, figure
2251 * out how much from delta we need to complete the current
2252 * period and accrue it.
2254 delta_w = 1024 - delta_w;
2256 sa->runnable_avg_sum += delta_w;
2257 sa->runnable_avg_period += delta_w;
2261 /* Figure out how many additional periods this update spans */
2262 periods = delta / 1024;
2265 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2267 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
2270 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2271 runnable_contrib = __compute_runnable_contrib(periods);
2273 sa->runnable_avg_sum += runnable_contrib;
2274 sa->runnable_avg_period += runnable_contrib;
2277 /* Remainder of delta accrued against u_0` */
2279 sa->runnable_avg_sum += delta;
2280 sa->runnable_avg_period += delta;
2285 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2286 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2288 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2289 u64 decays = atomic64_read(&cfs_rq->decay_counter);
2291 decays -= se->avg.decay_count;
2295 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2296 se->avg.decay_count = 0;
2301 #ifdef CONFIG_FAIR_GROUP_SCHED
2302 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2305 struct task_group *tg = cfs_rq->tg;
2308 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2309 tg_contrib -= cfs_rq->tg_load_contrib;
2311 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2312 atomic_long_add(tg_contrib, &tg->load_avg);
2313 cfs_rq->tg_load_contrib += tg_contrib;
2318 * Aggregate cfs_rq runnable averages into an equivalent task_group
2319 * representation for computing load contributions.
2321 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2322 struct cfs_rq *cfs_rq)
2324 struct task_group *tg = cfs_rq->tg;
2327 /* The fraction of a cpu used by this cfs_rq */
2328 contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT,
2329 sa->runnable_avg_period + 1);
2330 contrib -= cfs_rq->tg_runnable_contrib;
2332 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2333 atomic_add(contrib, &tg->runnable_avg);
2334 cfs_rq->tg_runnable_contrib += contrib;
2338 static inline void __update_group_entity_contrib(struct sched_entity *se)
2340 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2341 struct task_group *tg = cfs_rq->tg;
2346 contrib = cfs_rq->tg_load_contrib * tg->shares;
2347 se->avg.load_avg_contrib = div_u64(contrib,
2348 atomic_long_read(&tg->load_avg) + 1);
2351 * For group entities we need to compute a correction term in the case
2352 * that they are consuming <1 cpu so that we would contribute the same
2353 * load as a task of equal weight.
2355 * Explicitly co-ordinating this measurement would be expensive, but
2356 * fortunately the sum of each cpus contribution forms a usable
2357 * lower-bound on the true value.
2359 * Consider the aggregate of 2 contributions. Either they are disjoint
2360 * (and the sum represents true value) or they are disjoint and we are
2361 * understating by the aggregate of their overlap.
2363 * Extending this to N cpus, for a given overlap, the maximum amount we
2364 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2365 * cpus that overlap for this interval and w_i is the interval width.
2367 * On a small machine; the first term is well-bounded which bounds the
2368 * total error since w_i is a subset of the period. Whereas on a
2369 * larger machine, while this first term can be larger, if w_i is the
2370 * of consequential size guaranteed to see n_i*w_i quickly converge to
2371 * our upper bound of 1-cpu.
2373 runnable_avg = atomic_read(&tg->runnable_avg);
2374 if (runnable_avg < NICE_0_LOAD) {
2375 se->avg.load_avg_contrib *= runnable_avg;
2376 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2380 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2382 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
2383 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2385 #else /* CONFIG_FAIR_GROUP_SCHED */
2386 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2387 int force_update) {}
2388 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2389 struct cfs_rq *cfs_rq) {}
2390 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2391 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2392 #endif /* CONFIG_FAIR_GROUP_SCHED */
2394 static inline void __update_task_entity_contrib(struct sched_entity *se)
2398 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2399 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2400 contrib /= (se->avg.runnable_avg_period + 1);
2401 se->avg.load_avg_contrib = scale_load(contrib);
2404 /* Compute the current contribution to load_avg by se, return any delta */
2405 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2407 long old_contrib = se->avg.load_avg_contrib;
2409 if (entity_is_task(se)) {
2410 __update_task_entity_contrib(se);
2412 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2413 __update_group_entity_contrib(se);
2416 return se->avg.load_avg_contrib - old_contrib;
2419 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2422 if (likely(load_contrib < cfs_rq->blocked_load_avg))
2423 cfs_rq->blocked_load_avg -= load_contrib;
2425 cfs_rq->blocked_load_avg = 0;
2428 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2430 /* Update a sched_entity's runnable average */
2431 static inline void update_entity_load_avg(struct sched_entity *se,
2434 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2439 * For a group entity we need to use their owned cfs_rq_clock_task() in
2440 * case they are the parent of a throttled hierarchy.
2442 if (entity_is_task(se))
2443 now = cfs_rq_clock_task(cfs_rq);
2445 now = cfs_rq_clock_task(group_cfs_rq(se));
2447 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
2450 contrib_delta = __update_entity_load_avg_contrib(se);
2456 cfs_rq->runnable_load_avg += contrib_delta;
2458 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2462 * Decay the load contributed by all blocked children and account this so that
2463 * their contribution may appropriately discounted when they wake up.
2465 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2467 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2470 decays = now - cfs_rq->last_decay;
2471 if (!decays && !force_update)
2474 if (atomic_long_read(&cfs_rq->removed_load)) {
2475 unsigned long removed_load;
2476 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2477 subtract_blocked_load_contrib(cfs_rq, removed_load);
2481 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2483 atomic64_add(decays, &cfs_rq->decay_counter);
2484 cfs_rq->last_decay = now;
2487 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2490 /* Add the load generated by se into cfs_rq's child load-average */
2491 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2492 struct sched_entity *se,
2496 * We track migrations using entity decay_count <= 0, on a wake-up
2497 * migration we use a negative decay count to track the remote decays
2498 * accumulated while sleeping.
2500 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2501 * are seen by enqueue_entity_load_avg() as a migration with an already
2502 * constructed load_avg_contrib.
2504 if (unlikely(se->avg.decay_count <= 0)) {
2505 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2506 if (se->avg.decay_count) {
2508 * In a wake-up migration we have to approximate the
2509 * time sleeping. This is because we can't synchronize
2510 * clock_task between the two cpus, and it is not
2511 * guaranteed to be read-safe. Instead, we can
2512 * approximate this using our carried decays, which are
2513 * explicitly atomically readable.
2515 se->avg.last_runnable_update -= (-se->avg.decay_count)
2517 update_entity_load_avg(se, 0);
2518 /* Indicate that we're now synchronized and on-rq */
2519 se->avg.decay_count = 0;
2523 __synchronize_entity_decay(se);
2526 /* migrated tasks did not contribute to our blocked load */
2528 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2529 update_entity_load_avg(se, 0);
2532 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2533 /* we force update consideration on load-balancer moves */
2534 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2538 * Remove se's load from this cfs_rq child load-average, if the entity is
2539 * transitioning to a blocked state we track its projected decay using
2542 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2543 struct sched_entity *se,
2546 update_entity_load_avg(se, 1);
2547 /* we force update consideration on load-balancer moves */
2548 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2550 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2552 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2553 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2554 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2558 * Update the rq's load with the elapsed running time before entering
2559 * idle. if the last scheduled task is not a CFS task, idle_enter will
2560 * be the only way to update the runnable statistic.
2562 void idle_enter_fair(struct rq *this_rq)
2564 update_rq_runnable_avg(this_rq, 1);
2568 * Update the rq's load with the elapsed idle time before a task is
2569 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2570 * be the only way to update the runnable statistic.
2572 void idle_exit_fair(struct rq *this_rq)
2574 update_rq_runnable_avg(this_rq, 0);
2577 static int idle_balance(struct rq *this_rq);
2579 #else /* CONFIG_SMP */
2581 static inline void update_entity_load_avg(struct sched_entity *se,
2582 int update_cfs_rq) {}
2583 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2584 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2585 struct sched_entity *se,
2587 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2588 struct sched_entity *se,
2590 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2591 int force_update) {}
2593 static inline int idle_balance(struct rq *rq)
2598 #endif /* CONFIG_SMP */
2600 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2602 #ifdef CONFIG_SCHEDSTATS
2603 struct task_struct *tsk = NULL;
2605 if (entity_is_task(se))
2608 if (se->statistics.sleep_start) {
2609 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2614 if (unlikely(delta > se->statistics.sleep_max))
2615 se->statistics.sleep_max = delta;
2617 se->statistics.sleep_start = 0;
2618 se->statistics.sum_sleep_runtime += delta;
2621 account_scheduler_latency(tsk, delta >> 10, 1);
2622 trace_sched_stat_sleep(tsk, delta);
2625 if (se->statistics.block_start) {
2626 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2631 if (unlikely(delta > se->statistics.block_max))
2632 se->statistics.block_max = delta;
2634 se->statistics.block_start = 0;
2635 se->statistics.sum_sleep_runtime += delta;
2638 if (tsk->in_iowait) {
2639 se->statistics.iowait_sum += delta;
2640 se->statistics.iowait_count++;
2641 trace_sched_stat_iowait(tsk, delta);
2644 trace_sched_stat_blocked(tsk, delta);
2647 * Blocking time is in units of nanosecs, so shift by
2648 * 20 to get a milliseconds-range estimation of the
2649 * amount of time that the task spent sleeping:
2651 if (unlikely(prof_on == SLEEP_PROFILING)) {
2652 profile_hits(SLEEP_PROFILING,
2653 (void *)get_wchan(tsk),
2656 account_scheduler_latency(tsk, delta >> 10, 0);
2662 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2664 #ifdef CONFIG_SCHED_DEBUG
2665 s64 d = se->vruntime - cfs_rq->min_vruntime;
2670 if (d > 3*sysctl_sched_latency)
2671 schedstat_inc(cfs_rq, nr_spread_over);
2676 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2678 u64 vruntime = cfs_rq->min_vruntime;
2681 * The 'current' period is already promised to the current tasks,
2682 * however the extra weight of the new task will slow them down a
2683 * little, place the new task so that it fits in the slot that
2684 * stays open at the end.
2686 if (initial && sched_feat(START_DEBIT))
2687 vruntime += sched_vslice(cfs_rq, se);
2689 /* sleeps up to a single latency don't count. */
2691 unsigned long thresh = sysctl_sched_latency;
2694 * Halve their sleep time's effect, to allow
2695 * for a gentler effect of sleepers:
2697 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2703 /* ensure we never gain time by being placed backwards. */
2704 se->vruntime = max_vruntime(se->vruntime, vruntime);
2707 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2710 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2713 * Update the normalized vruntime before updating min_vruntime
2714 * through calling update_curr().
2716 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2717 se->vruntime += cfs_rq->min_vruntime;
2720 * Update run-time statistics of the 'current'.
2722 update_curr(cfs_rq);
2723 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2724 account_entity_enqueue(cfs_rq, se);
2725 update_cfs_shares(cfs_rq);
2727 if (flags & ENQUEUE_WAKEUP) {
2728 place_entity(cfs_rq, se, 0);
2729 enqueue_sleeper(cfs_rq, se);
2732 update_stats_enqueue(cfs_rq, se);
2733 check_spread(cfs_rq, se);
2734 if (se != cfs_rq->curr)
2735 __enqueue_entity(cfs_rq, se);
2738 if (cfs_rq->nr_running == 1) {
2739 list_add_leaf_cfs_rq(cfs_rq);
2740 check_enqueue_throttle(cfs_rq);
2744 static void __clear_buddies_last(struct sched_entity *se)
2746 for_each_sched_entity(se) {
2747 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2748 if (cfs_rq->last != se)
2751 cfs_rq->last = NULL;
2755 static void __clear_buddies_next(struct sched_entity *se)
2757 for_each_sched_entity(se) {
2758 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2759 if (cfs_rq->next != se)
2762 cfs_rq->next = NULL;
2766 static void __clear_buddies_skip(struct sched_entity *se)
2768 for_each_sched_entity(se) {
2769 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2770 if (cfs_rq->skip != se)
2773 cfs_rq->skip = NULL;
2777 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2779 if (cfs_rq->last == se)
2780 __clear_buddies_last(se);
2782 if (cfs_rq->next == se)
2783 __clear_buddies_next(se);
2785 if (cfs_rq->skip == se)
2786 __clear_buddies_skip(se);
2789 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2792 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2795 * Update run-time statistics of the 'current'.
2797 update_curr(cfs_rq);
2798 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2800 update_stats_dequeue(cfs_rq, se);
2801 if (flags & DEQUEUE_SLEEP) {
2802 #ifdef CONFIG_SCHEDSTATS
2803 if (entity_is_task(se)) {
2804 struct task_struct *tsk = task_of(se);
2806 if (tsk->state & TASK_INTERRUPTIBLE)
2807 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2808 if (tsk->state & TASK_UNINTERRUPTIBLE)
2809 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2814 clear_buddies(cfs_rq, se);
2816 if (se != cfs_rq->curr)
2817 __dequeue_entity(cfs_rq, se);
2819 account_entity_dequeue(cfs_rq, se);
2822 * Normalize the entity after updating the min_vruntime because the
2823 * update can refer to the ->curr item and we need to reflect this
2824 * movement in our normalized position.
2826 if (!(flags & DEQUEUE_SLEEP))
2827 se->vruntime -= cfs_rq->min_vruntime;
2829 /* return excess runtime on last dequeue */
2830 return_cfs_rq_runtime(cfs_rq);
2832 update_min_vruntime(cfs_rq);
2833 update_cfs_shares(cfs_rq);
2837 * Preempt the current task with a newly woken task if needed:
2840 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2842 unsigned long ideal_runtime, delta_exec;
2843 struct sched_entity *se;
2846 ideal_runtime = sched_slice(cfs_rq, curr);
2847 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2848 if (delta_exec > ideal_runtime) {
2849 resched_task(rq_of(cfs_rq)->curr);
2851 * The current task ran long enough, ensure it doesn't get
2852 * re-elected due to buddy favours.
2854 clear_buddies(cfs_rq, curr);
2859 * Ensure that a task that missed wakeup preemption by a
2860 * narrow margin doesn't have to wait for a full slice.
2861 * This also mitigates buddy induced latencies under load.
2863 if (delta_exec < sysctl_sched_min_granularity)
2866 se = __pick_first_entity(cfs_rq);
2867 delta = curr->vruntime - se->vruntime;
2872 if (delta > ideal_runtime)
2873 resched_task(rq_of(cfs_rq)->curr);
2877 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2879 /* 'current' is not kept within the tree. */
2882 * Any task has to be enqueued before it get to execute on
2883 * a CPU. So account for the time it spent waiting on the
2886 update_stats_wait_end(cfs_rq, se);
2887 __dequeue_entity(cfs_rq, se);
2890 update_stats_curr_start(cfs_rq, se);
2892 #ifdef CONFIG_SCHEDSTATS
2894 * Track our maximum slice length, if the CPU's load is at
2895 * least twice that of our own weight (i.e. dont track it
2896 * when there are only lesser-weight tasks around):
2898 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2899 se->statistics.slice_max = max(se->statistics.slice_max,
2900 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2903 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2907 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2910 * Pick the next process, keeping these things in mind, in this order:
2911 * 1) keep things fair between processes/task groups
2912 * 2) pick the "next" process, since someone really wants that to run
2913 * 3) pick the "last" process, for cache locality
2914 * 4) do not run the "skip" process, if something else is available
2916 static struct sched_entity *
2917 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2919 struct sched_entity *left = __pick_first_entity(cfs_rq);
2920 struct sched_entity *se;
2923 * If curr is set we have to see if its left of the leftmost entity
2924 * still in the tree, provided there was anything in the tree at all.
2926 if (!left || (curr && entity_before(curr, left)))
2929 se = left; /* ideally we run the leftmost entity */
2932 * Avoid running the skip buddy, if running something else can
2933 * be done without getting too unfair.
2935 if (cfs_rq->skip == se) {
2936 struct sched_entity *second;
2939 second = __pick_first_entity(cfs_rq);
2941 second = __pick_next_entity(se);
2942 if (!second || (curr && entity_before(curr, second)))
2946 if (second && wakeup_preempt_entity(second, left) < 1)
2951 * Prefer last buddy, try to return the CPU to a preempted task.
2953 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2957 * Someone really wants this to run. If it's not unfair, run it.
2959 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2962 clear_buddies(cfs_rq, se);
2967 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2969 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2972 * If still on the runqueue then deactivate_task()
2973 * was not called and update_curr() has to be done:
2976 update_curr(cfs_rq);
2978 /* throttle cfs_rqs exceeding runtime */
2979 check_cfs_rq_runtime(cfs_rq);
2981 check_spread(cfs_rq, prev);
2983 update_stats_wait_start(cfs_rq, prev);
2984 /* Put 'current' back into the tree. */
2985 __enqueue_entity(cfs_rq, prev);
2986 /* in !on_rq case, update occurred at dequeue */
2987 update_entity_load_avg(prev, 1);
2989 cfs_rq->curr = NULL;
2993 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2996 * Update run-time statistics of the 'current'.
2998 update_curr(cfs_rq);
3001 * Ensure that runnable average is periodically updated.
3003 update_entity_load_avg(curr, 1);
3004 update_cfs_rq_blocked_load(cfs_rq, 1);
3005 update_cfs_shares(cfs_rq);
3007 #ifdef CONFIG_SCHED_HRTICK
3009 * queued ticks are scheduled to match the slice, so don't bother
3010 * validating it and just reschedule.
3013 resched_task(rq_of(cfs_rq)->curr);
3017 * don't let the period tick interfere with the hrtick preemption
3019 if (!sched_feat(DOUBLE_TICK) &&
3020 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3024 if (cfs_rq->nr_running > 1)
3025 check_preempt_tick(cfs_rq, curr);
3029 /**************************************************
3030 * CFS bandwidth control machinery
3033 #ifdef CONFIG_CFS_BANDWIDTH
3035 #ifdef HAVE_JUMP_LABEL
3036 static struct static_key __cfs_bandwidth_used;
3038 static inline bool cfs_bandwidth_used(void)
3040 return static_key_false(&__cfs_bandwidth_used);
3043 void cfs_bandwidth_usage_inc(void)
3045 static_key_slow_inc(&__cfs_bandwidth_used);
3048 void cfs_bandwidth_usage_dec(void)
3050 static_key_slow_dec(&__cfs_bandwidth_used);
3052 #else /* HAVE_JUMP_LABEL */
3053 static bool cfs_bandwidth_used(void)
3058 void cfs_bandwidth_usage_inc(void) {}
3059 void cfs_bandwidth_usage_dec(void) {}
3060 #endif /* HAVE_JUMP_LABEL */
3063 * default period for cfs group bandwidth.
3064 * default: 0.1s, units: nanoseconds
3066 static inline u64 default_cfs_period(void)
3068 return 100000000ULL;
3071 static inline u64 sched_cfs_bandwidth_slice(void)
3073 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3077 * Replenish runtime according to assigned quota and update expiration time.
3078 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3079 * additional synchronization around rq->lock.
3081 * requires cfs_b->lock
3083 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3087 if (cfs_b->quota == RUNTIME_INF)
3090 now = sched_clock_cpu(smp_processor_id());
3091 cfs_b->runtime = cfs_b->quota;
3092 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3095 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3097 return &tg->cfs_bandwidth;
3100 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3101 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3103 if (unlikely(cfs_rq->throttle_count))
3104 return cfs_rq->throttled_clock_task;
3106 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3109 /* returns 0 on failure to allocate runtime */
3110 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3112 struct task_group *tg = cfs_rq->tg;
3113 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3114 u64 amount = 0, min_amount, expires;
3116 /* note: this is a positive sum as runtime_remaining <= 0 */
3117 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3119 raw_spin_lock(&cfs_b->lock);
3120 if (cfs_b->quota == RUNTIME_INF)
3121 amount = min_amount;
3124 * If the bandwidth pool has become inactive, then at least one
3125 * period must have elapsed since the last consumption.
3126 * Refresh the global state and ensure bandwidth timer becomes
3129 if (!cfs_b->timer_active) {
3130 __refill_cfs_bandwidth_runtime(cfs_b);
3131 __start_cfs_bandwidth(cfs_b);
3134 if (cfs_b->runtime > 0) {
3135 amount = min(cfs_b->runtime, min_amount);
3136 cfs_b->runtime -= amount;
3140 expires = cfs_b->runtime_expires;
3141 raw_spin_unlock(&cfs_b->lock);
3143 cfs_rq->runtime_remaining += amount;
3145 * we may have advanced our local expiration to account for allowed
3146 * spread between our sched_clock and the one on which runtime was
3149 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3150 cfs_rq->runtime_expires = expires;
3152 return cfs_rq->runtime_remaining > 0;
3156 * Note: This depends on the synchronization provided by sched_clock and the
3157 * fact that rq->clock snapshots this value.
3159 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3161 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3163 /* if the deadline is ahead of our clock, nothing to do */
3164 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3167 if (cfs_rq->runtime_remaining < 0)
3171 * If the local deadline has passed we have to consider the
3172 * possibility that our sched_clock is 'fast' and the global deadline
3173 * has not truly expired.
3175 * Fortunately we can check determine whether this the case by checking
3176 * whether the global deadline has advanced.
3179 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
3180 /* extend local deadline, drift is bounded above by 2 ticks */
3181 cfs_rq->runtime_expires += TICK_NSEC;
3183 /* global deadline is ahead, expiration has passed */
3184 cfs_rq->runtime_remaining = 0;
3188 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3190 /* dock delta_exec before expiring quota (as it could span periods) */
3191 cfs_rq->runtime_remaining -= delta_exec;
3192 expire_cfs_rq_runtime(cfs_rq);
3194 if (likely(cfs_rq->runtime_remaining > 0))
3198 * if we're unable to extend our runtime we resched so that the active
3199 * hierarchy can be throttled
3201 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3202 resched_task(rq_of(cfs_rq)->curr);
3205 static __always_inline
3206 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3208 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3211 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3214 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3216 return cfs_bandwidth_used() && cfs_rq->throttled;
3219 /* check whether cfs_rq, or any parent, is throttled */
3220 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3222 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3226 * Ensure that neither of the group entities corresponding to src_cpu or
3227 * dest_cpu are members of a throttled hierarchy when performing group
3228 * load-balance operations.
3230 static inline int throttled_lb_pair(struct task_group *tg,
3231 int src_cpu, int dest_cpu)
3233 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3235 src_cfs_rq = tg->cfs_rq[src_cpu];
3236 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3238 return throttled_hierarchy(src_cfs_rq) ||
3239 throttled_hierarchy(dest_cfs_rq);
3242 /* updated child weight may affect parent so we have to do this bottom up */
3243 static int tg_unthrottle_up(struct task_group *tg, void *data)
3245 struct rq *rq = data;
3246 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3248 cfs_rq->throttle_count--;
3250 if (!cfs_rq->throttle_count) {
3251 /* adjust cfs_rq_clock_task() */
3252 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3253 cfs_rq->throttled_clock_task;
3260 static int tg_throttle_down(struct task_group *tg, void *data)
3262 struct rq *rq = data;
3263 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3265 /* group is entering throttled state, stop time */
3266 if (!cfs_rq->throttle_count)
3267 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3268 cfs_rq->throttle_count++;
3273 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3275 struct rq *rq = rq_of(cfs_rq);
3276 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3277 struct sched_entity *se;
3278 long task_delta, dequeue = 1;
3280 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3282 /* freeze hierarchy runnable averages while throttled */
3284 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3287 task_delta = cfs_rq->h_nr_running;
3288 for_each_sched_entity(se) {
3289 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3290 /* throttled entity or throttle-on-deactivate */
3295 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3296 qcfs_rq->h_nr_running -= task_delta;
3298 if (qcfs_rq->load.weight)
3303 rq->nr_running -= task_delta;
3305 cfs_rq->throttled = 1;
3306 cfs_rq->throttled_clock = rq_clock(rq);
3307 raw_spin_lock(&cfs_b->lock);
3308 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3309 if (!cfs_b->timer_active)
3310 __start_cfs_bandwidth(cfs_b);
3311 raw_spin_unlock(&cfs_b->lock);
3314 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3316 struct rq *rq = rq_of(cfs_rq);
3317 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3318 struct sched_entity *se;
3322 se = cfs_rq->tg->se[cpu_of(rq)];
3324 cfs_rq->throttled = 0;
3326 update_rq_clock(rq);
3328 raw_spin_lock(&cfs_b->lock);
3329 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3330 list_del_rcu(&cfs_rq->throttled_list);
3331 raw_spin_unlock(&cfs_b->lock);
3333 /* update hierarchical throttle state */
3334 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3336 if (!cfs_rq->load.weight)
3339 task_delta = cfs_rq->h_nr_running;
3340 for_each_sched_entity(se) {
3344 cfs_rq = cfs_rq_of(se);
3346 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3347 cfs_rq->h_nr_running += task_delta;
3349 if (cfs_rq_throttled(cfs_rq))
3354 rq->nr_running += task_delta;
3356 /* determine whether we need to wake up potentially idle cpu */
3357 if (rq->curr == rq->idle && rq->cfs.nr_running)
3358 resched_task(rq->curr);
3361 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3362 u64 remaining, u64 expires)
3364 struct cfs_rq *cfs_rq;
3365 u64 runtime = remaining;
3368 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3370 struct rq *rq = rq_of(cfs_rq);
3372 raw_spin_lock(&rq->lock);
3373 if (!cfs_rq_throttled(cfs_rq))
3376 runtime = -cfs_rq->runtime_remaining + 1;
3377 if (runtime > remaining)
3378 runtime = remaining;
3379 remaining -= runtime;
3381 cfs_rq->runtime_remaining += runtime;
3382 cfs_rq->runtime_expires = expires;
3384 /* we check whether we're throttled above */
3385 if (cfs_rq->runtime_remaining > 0)
3386 unthrottle_cfs_rq(cfs_rq);
3389 raw_spin_unlock(&rq->lock);
3400 * Responsible for refilling a task_group's bandwidth and unthrottling its
3401 * cfs_rqs as appropriate. If there has been no activity within the last
3402 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3403 * used to track this state.
3405 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3407 u64 runtime, runtime_expires;
3408 int idle = 1, throttled;
3410 raw_spin_lock(&cfs_b->lock);
3411 /* no need to continue the timer with no bandwidth constraint */
3412 if (cfs_b->quota == RUNTIME_INF)
3415 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3416 /* idle depends on !throttled (for the case of a large deficit) */
3417 idle = cfs_b->idle && !throttled;
3418 cfs_b->nr_periods += overrun;
3420 /* if we're going inactive then everything else can be deferred */
3425 * if we have relooped after returning idle once, we need to update our
3426 * status as actually running, so that other cpus doing
3427 * __start_cfs_bandwidth will stop trying to cancel us.
3429 cfs_b->timer_active = 1;
3431 __refill_cfs_bandwidth_runtime(cfs_b);
3434 /* mark as potentially idle for the upcoming period */
3439 /* account preceding periods in which throttling occurred */
3440 cfs_b->nr_throttled += overrun;
3443 * There are throttled entities so we must first use the new bandwidth
3444 * to unthrottle them before making it generally available. This
3445 * ensures that all existing debts will be paid before a new cfs_rq is
3448 runtime = cfs_b->runtime;
3449 runtime_expires = cfs_b->runtime_expires;
3453 * This check is repeated as we are holding onto the new bandwidth
3454 * while we unthrottle. This can potentially race with an unthrottled
3455 * group trying to acquire new bandwidth from the global pool.
3457 while (throttled && runtime > 0) {
3458 raw_spin_unlock(&cfs_b->lock);
3459 /* we can't nest cfs_b->lock while distributing bandwidth */
3460 runtime = distribute_cfs_runtime(cfs_b, runtime,
3462 raw_spin_lock(&cfs_b->lock);
3464 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3467 /* return (any) remaining runtime */
3468 cfs_b->runtime = runtime;
3470 * While we are ensured activity in the period following an
3471 * unthrottle, this also covers the case in which the new bandwidth is
3472 * insufficient to cover the existing bandwidth deficit. (Forcing the
3473 * timer to remain active while there are any throttled entities.)
3478 cfs_b->timer_active = 0;
3479 raw_spin_unlock(&cfs_b->lock);
3484 /* a cfs_rq won't donate quota below this amount */
3485 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3486 /* minimum remaining period time to redistribute slack quota */
3487 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3488 /* how long we wait to gather additional slack before distributing */
3489 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3492 * Are we near the end of the current quota period?
3494 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3495 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3496 * migrate_hrtimers, base is never cleared, so we are fine.
3498 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3500 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3503 /* if the call-back is running a quota refresh is already occurring */
3504 if (hrtimer_callback_running(refresh_timer))
3507 /* is a quota refresh about to occur? */
3508 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3509 if (remaining < min_expire)
3515 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3517 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3519 /* if there's a quota refresh soon don't bother with slack */
3520 if (runtime_refresh_within(cfs_b, min_left))
3523 start_bandwidth_timer(&cfs_b->slack_timer,
3524 ns_to_ktime(cfs_bandwidth_slack_period));
3527 /* we know any runtime found here is valid as update_curr() precedes return */
3528 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3530 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3531 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3533 if (slack_runtime <= 0)
3536 raw_spin_lock(&cfs_b->lock);
3537 if (cfs_b->quota != RUNTIME_INF &&
3538 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3539 cfs_b->runtime += slack_runtime;
3541 /* we are under rq->lock, defer unthrottling using a timer */
3542 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3543 !list_empty(&cfs_b->throttled_cfs_rq))
3544 start_cfs_slack_bandwidth(cfs_b);
3546 raw_spin_unlock(&cfs_b->lock);
3548 /* even if it's not valid for return we don't want to try again */
3549 cfs_rq->runtime_remaining -= slack_runtime;
3552 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3554 if (!cfs_bandwidth_used())
3557 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3560 __return_cfs_rq_runtime(cfs_rq);
3564 * This is done with a timer (instead of inline with bandwidth return) since
3565 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3567 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3569 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3572 /* confirm we're still not at a refresh boundary */
3573 raw_spin_lock(&cfs_b->lock);
3574 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3575 raw_spin_unlock(&cfs_b->lock);
3579 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
3580 runtime = cfs_b->runtime;
3583 expires = cfs_b->runtime_expires;
3584 raw_spin_unlock(&cfs_b->lock);
3589 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3591 raw_spin_lock(&cfs_b->lock);
3592 if (expires == cfs_b->runtime_expires)
3593 cfs_b->runtime = runtime;
3594 raw_spin_unlock(&cfs_b->lock);
3598 * When a group wakes up we want to make sure that its quota is not already
3599 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3600 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3602 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3604 if (!cfs_bandwidth_used())
3607 /* an active group must be handled by the update_curr()->put() path */
3608 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3611 /* ensure the group is not already throttled */
3612 if (cfs_rq_throttled(cfs_rq))
3615 /* update runtime allocation */
3616 account_cfs_rq_runtime(cfs_rq, 0);
3617 if (cfs_rq->runtime_remaining <= 0)
3618 throttle_cfs_rq(cfs_rq);
3621 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3622 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3624 if (!cfs_bandwidth_used())
3627 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3631 * it's possible for a throttled entity to be forced into a running
3632 * state (e.g. set_curr_task), in this case we're finished.
3634 if (cfs_rq_throttled(cfs_rq))
3637 throttle_cfs_rq(cfs_rq);
3641 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3643 struct cfs_bandwidth *cfs_b =
3644 container_of(timer, struct cfs_bandwidth, slack_timer);
3645 do_sched_cfs_slack_timer(cfs_b);
3647 return HRTIMER_NORESTART;
3650 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3652 struct cfs_bandwidth *cfs_b =
3653 container_of(timer, struct cfs_bandwidth, period_timer);
3659 now = hrtimer_cb_get_time(timer);
3660 overrun = hrtimer_forward(timer, now, cfs_b->period);
3665 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3668 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3671 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3673 raw_spin_lock_init(&cfs_b->lock);
3675 cfs_b->quota = RUNTIME_INF;
3676 cfs_b->period = ns_to_ktime(default_cfs_period());
3678 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3679 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3680 cfs_b->period_timer.function = sched_cfs_period_timer;
3681 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3682 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3685 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3687 cfs_rq->runtime_enabled = 0;
3688 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3691 /* requires cfs_b->lock, may release to reprogram timer */
3692 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3695 * The timer may be active because we're trying to set a new bandwidth
3696 * period or because we're racing with the tear-down path
3697 * (timer_active==0 becomes visible before the hrtimer call-back
3698 * terminates). In either case we ensure that it's re-programmed
3700 while (unlikely(hrtimer_active(&cfs_b->period_timer)) &&
3701 hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) {
3702 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3703 raw_spin_unlock(&cfs_b->lock);
3705 raw_spin_lock(&cfs_b->lock);
3706 /* if someone else restarted the timer then we're done */
3707 if (cfs_b->timer_active)
3711 cfs_b->timer_active = 1;
3712 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3715 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3717 hrtimer_cancel(&cfs_b->period_timer);
3718 hrtimer_cancel(&cfs_b->slack_timer);
3721 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3723 struct cfs_rq *cfs_rq;
3725 for_each_leaf_cfs_rq(rq, cfs_rq) {
3726 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3728 if (!cfs_rq->runtime_enabled)
3732 * clock_task is not advancing so we just need to make sure
3733 * there's some valid quota amount
3735 cfs_rq->runtime_remaining = cfs_b->quota;
3736 if (cfs_rq_throttled(cfs_rq))
3737 unthrottle_cfs_rq(cfs_rq);
3741 #else /* CONFIG_CFS_BANDWIDTH */
3742 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3744 return rq_clock_task(rq_of(cfs_rq));
3747 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
3748 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
3749 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3750 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3752 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3757 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3762 static inline int throttled_lb_pair(struct task_group *tg,
3763 int src_cpu, int dest_cpu)
3768 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3770 #ifdef CONFIG_FAIR_GROUP_SCHED
3771 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3774 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3778 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3779 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3781 #endif /* CONFIG_CFS_BANDWIDTH */
3783 /**************************************************
3784 * CFS operations on tasks:
3787 #ifdef CONFIG_SCHED_HRTICK
3788 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3790 struct sched_entity *se = &p->se;
3791 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3793 WARN_ON(task_rq(p) != rq);
3795 if (cfs_rq->nr_running > 1) {
3796 u64 slice = sched_slice(cfs_rq, se);
3797 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3798 s64 delta = slice - ran;
3807 * Don't schedule slices shorter than 10000ns, that just
3808 * doesn't make sense. Rely on vruntime for fairness.
3811 delta = max_t(s64, 10000LL, delta);
3813 hrtick_start(rq, delta);
3818 * called from enqueue/dequeue and updates the hrtick when the
3819 * current task is from our class and nr_running is low enough
3822 static void hrtick_update(struct rq *rq)
3824 struct task_struct *curr = rq->curr;
3826 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3829 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3830 hrtick_start_fair(rq, curr);
3832 #else /* !CONFIG_SCHED_HRTICK */
3834 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3838 static inline void hrtick_update(struct rq *rq)
3844 * The enqueue_task method is called before nr_running is
3845 * increased. Here we update the fair scheduling stats and
3846 * then put the task into the rbtree:
3849 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3851 struct cfs_rq *cfs_rq;
3852 struct sched_entity *se = &p->se;
3854 for_each_sched_entity(se) {
3857 cfs_rq = cfs_rq_of(se);
3858 enqueue_entity(cfs_rq, se, flags);
3861 * end evaluation on encountering a throttled cfs_rq
3863 * note: in the case of encountering a throttled cfs_rq we will
3864 * post the final h_nr_running increment below.
3866 if (cfs_rq_throttled(cfs_rq))
3868 cfs_rq->h_nr_running++;
3870 flags = ENQUEUE_WAKEUP;
3873 for_each_sched_entity(se) {
3874 cfs_rq = cfs_rq_of(se);
3875 cfs_rq->h_nr_running++;
3877 if (cfs_rq_throttled(cfs_rq))
3880 update_cfs_shares(cfs_rq);
3881 update_entity_load_avg(se, 1);
3885 update_rq_runnable_avg(rq, rq->nr_running);
3891 static void set_next_buddy(struct sched_entity *se);
3894 * The dequeue_task method is called before nr_running is
3895 * decreased. We remove the task from the rbtree and
3896 * update the fair scheduling stats:
3898 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3900 struct cfs_rq *cfs_rq;
3901 struct sched_entity *se = &p->se;
3902 int task_sleep = flags & DEQUEUE_SLEEP;
3904 for_each_sched_entity(se) {
3905 cfs_rq = cfs_rq_of(se);
3906 dequeue_entity(cfs_rq, se, flags);
3909 * end evaluation on encountering a throttled cfs_rq
3911 * note: in the case of encountering a throttled cfs_rq we will
3912 * post the final h_nr_running decrement below.
3914 if (cfs_rq_throttled(cfs_rq))
3916 cfs_rq->h_nr_running--;
3918 /* Don't dequeue parent if it has other entities besides us */
3919 if (cfs_rq->load.weight) {
3921 * Bias pick_next to pick a task from this cfs_rq, as
3922 * p is sleeping when it is within its sched_slice.
3924 if (task_sleep && parent_entity(se))
3925 set_next_buddy(parent_entity(se));
3927 /* avoid re-evaluating load for this entity */
3928 se = parent_entity(se);
3931 flags |= DEQUEUE_SLEEP;
3934 for_each_sched_entity(se) {
3935 cfs_rq = cfs_rq_of(se);
3936 cfs_rq->h_nr_running--;
3938 if (cfs_rq_throttled(cfs_rq))
3941 update_cfs_shares(cfs_rq);
3942 update_entity_load_avg(se, 1);
3947 update_rq_runnable_avg(rq, 1);
3953 /* Used instead of source_load when we know the type == 0 */
3954 static unsigned long weighted_cpuload(const int cpu)
3956 return cpu_rq(cpu)->cfs.runnable_load_avg;
3960 * Return a low guess at the load of a migration-source cpu weighted
3961 * according to the scheduling class and "nice" value.
3963 * We want to under-estimate the load of migration sources, to
3964 * balance conservatively.
3966 static unsigned long source_load(int cpu, int type)
3968 struct rq *rq = cpu_rq(cpu);
3969 unsigned long total = weighted_cpuload(cpu);
3971 if (type == 0 || !sched_feat(LB_BIAS))
3974 return min(rq->cpu_load[type-1], total);
3978 * Return a high guess at the load of a migration-target cpu weighted
3979 * according to the scheduling class and "nice" value.
3981 static unsigned long target_load(int cpu, int type)
3983 struct rq *rq = cpu_rq(cpu);
3984 unsigned long total = weighted_cpuload(cpu);
3986 if (type == 0 || !sched_feat(LB_BIAS))
3989 return max(rq->cpu_load[type-1], total);
3992 static unsigned long power_of(int cpu)
3994 return cpu_rq(cpu)->cpu_power;
3997 static unsigned long cpu_avg_load_per_task(int cpu)
3999 struct rq *rq = cpu_rq(cpu);
4000 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
4001 unsigned long load_avg = rq->cfs.runnable_load_avg;
4004 return load_avg / nr_running;
4009 static void record_wakee(struct task_struct *p)
4012 * Rough decay (wiping) for cost saving, don't worry
4013 * about the boundary, really active task won't care
4016 if (jiffies > current->wakee_flip_decay_ts + HZ) {
4017 current->wakee_flips = 0;
4018 current->wakee_flip_decay_ts = jiffies;
4021 if (current->last_wakee != p) {
4022 current->last_wakee = p;
4023 current->wakee_flips++;
4027 static void task_waking_fair(struct task_struct *p)
4029 struct sched_entity *se = &p->se;
4030 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4033 #ifndef CONFIG_64BIT
4034 u64 min_vruntime_copy;
4037 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4039 min_vruntime = cfs_rq->min_vruntime;
4040 } while (min_vruntime != min_vruntime_copy);
4042 min_vruntime = cfs_rq->min_vruntime;
4045 se->vruntime -= min_vruntime;
4049 #ifdef CONFIG_FAIR_GROUP_SCHED
4051 * effective_load() calculates the load change as seen from the root_task_group
4053 * Adding load to a group doesn't make a group heavier, but can cause movement
4054 * of group shares between cpus. Assuming the shares were perfectly aligned one
4055 * can calculate the shift in shares.
4057 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4058 * on this @cpu and results in a total addition (subtraction) of @wg to the
4059 * total group weight.
4061 * Given a runqueue weight distribution (rw_i) we can compute a shares
4062 * distribution (s_i) using:
4064 * s_i = rw_i / \Sum rw_j (1)
4066 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4067 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4068 * shares distribution (s_i):
4070 * rw_i = { 2, 4, 1, 0 }
4071 * s_i = { 2/7, 4/7, 1/7, 0 }
4073 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4074 * task used to run on and the CPU the waker is running on), we need to
4075 * compute the effect of waking a task on either CPU and, in case of a sync
4076 * wakeup, compute the effect of the current task going to sleep.
4078 * So for a change of @wl to the local @cpu with an overall group weight change
4079 * of @wl we can compute the new shares distribution (s'_i) using:
4081 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4083 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4084 * differences in waking a task to CPU 0. The additional task changes the
4085 * weight and shares distributions like:
4087 * rw'_i = { 3, 4, 1, 0 }
4088 * s'_i = { 3/8, 4/8, 1/8, 0 }
4090 * We can then compute the difference in effective weight by using:
4092 * dw_i = S * (s'_i - s_i) (3)
4094 * Where 'S' is the group weight as seen by its parent.
4096 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4097 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4098 * 4/7) times the weight of the group.
4100 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4102 struct sched_entity *se = tg->se[cpu];
4104 if (!tg->parent) /* the trivial, non-cgroup case */
4107 for_each_sched_entity(se) {
4113 * W = @wg + \Sum rw_j
4115 W = wg + calc_tg_weight(tg, se->my_q);
4120 w = se->my_q->load.weight + wl;
4123 * wl = S * s'_i; see (2)
4126 wl = (w * tg->shares) / W;
4131 * Per the above, wl is the new se->load.weight value; since
4132 * those are clipped to [MIN_SHARES, ...) do so now. See
4133 * calc_cfs_shares().
4135 if (wl < MIN_SHARES)
4139 * wl = dw_i = S * (s'_i - s_i); see (3)
4141 wl -= se->load.weight;
4144 * Recursively apply this logic to all parent groups to compute
4145 * the final effective load change on the root group. Since
4146 * only the @tg group gets extra weight, all parent groups can
4147 * only redistribute existing shares. @wl is the shift in shares
4148 * resulting from this level per the above.
4157 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4164 static int wake_wide(struct task_struct *p)
4166 int factor = this_cpu_read(sd_llc_size);
4169 * Yeah, it's the switching-frequency, could means many wakee or
4170 * rapidly switch, use factor here will just help to automatically
4171 * adjust the loose-degree, so bigger node will lead to more pull.
4173 if (p->wakee_flips > factor) {
4175 * wakee is somewhat hot, it needs certain amount of cpu
4176 * resource, so if waker is far more hot, prefer to leave
4179 if (current->wakee_flips > (factor * p->wakee_flips))
4186 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4188 s64 this_load, load;
4189 int idx, this_cpu, prev_cpu;
4190 unsigned long tl_per_task;
4191 struct task_group *tg;
4192 unsigned long weight;
4196 * If we wake multiple tasks be careful to not bounce
4197 * ourselves around too much.
4203 this_cpu = smp_processor_id();
4204 prev_cpu = task_cpu(p);
4205 load = source_load(prev_cpu, idx);
4206 this_load = target_load(this_cpu, idx);
4209 * If sync wakeup then subtract the (maximum possible)
4210 * effect of the currently running task from the load
4211 * of the current CPU:
4214 tg = task_group(current);
4215 weight = current->se.load.weight;
4217 this_load += effective_load(tg, this_cpu, -weight, -weight);
4218 load += effective_load(tg, prev_cpu, 0, -weight);
4222 weight = p->se.load.weight;
4225 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4226 * due to the sync cause above having dropped this_load to 0, we'll
4227 * always have an imbalance, but there's really nothing you can do
4228 * about that, so that's good too.
4230 * Otherwise check if either cpus are near enough in load to allow this
4231 * task to be woken on this_cpu.
4233 if (this_load > 0) {
4234 s64 this_eff_load, prev_eff_load;
4236 this_eff_load = 100;
4237 this_eff_load *= power_of(prev_cpu);
4238 this_eff_load *= this_load +
4239 effective_load(tg, this_cpu, weight, weight);
4241 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4242 prev_eff_load *= power_of(this_cpu);
4243 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4245 balanced = this_eff_load <= prev_eff_load;
4250 * If the currently running task will sleep within
4251 * a reasonable amount of time then attract this newly
4254 if (sync && balanced)
4257 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4258 tl_per_task = cpu_avg_load_per_task(this_cpu);
4261 (this_load <= load &&
4262 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
4264 * This domain has SD_WAKE_AFFINE and
4265 * p is cache cold in this domain, and
4266 * there is no bad imbalance.
4268 schedstat_inc(sd, ttwu_move_affine);
4269 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4277 * find_idlest_group finds and returns the least busy CPU group within the
4280 static struct sched_group *
4281 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4282 int this_cpu, int sd_flag)
4284 struct sched_group *idlest = NULL, *group = sd->groups;
4285 unsigned long min_load = ULONG_MAX, this_load = 0;
4286 int load_idx = sd->forkexec_idx;
4287 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4289 if (sd_flag & SD_BALANCE_WAKE)
4290 load_idx = sd->wake_idx;
4293 unsigned long load, avg_load;
4297 /* Skip over this group if it has no CPUs allowed */
4298 if (!cpumask_intersects(sched_group_cpus(group),
4299 tsk_cpus_allowed(p)))
4302 local_group = cpumask_test_cpu(this_cpu,
4303 sched_group_cpus(group));
4305 /* Tally up the load of all CPUs in the group */
4308 for_each_cpu(i, sched_group_cpus(group)) {
4309 /* Bias balancing toward cpus of our domain */
4311 load = source_load(i, load_idx);
4313 load = target_load(i, load_idx);
4318 /* Adjust by relative CPU power of the group */
4319 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
4322 this_load = avg_load;
4323 } else if (avg_load < min_load) {
4324 min_load = avg_load;
4327 } while (group = group->next, group != sd->groups);
4329 if (!idlest || 100*this_load < imbalance*min_load)
4335 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4338 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4340 unsigned long load, min_load = ULONG_MAX;
4344 /* Traverse only the allowed CPUs */
4345 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4346 load = weighted_cpuload(i);
4348 if (load < min_load || (load == min_load && i == this_cpu)) {
4358 * Try and locate an idle CPU in the sched_domain.
4360 static int select_idle_sibling(struct task_struct *p, int target)
4362 struct sched_domain *sd;
4363 struct sched_group *sg;
4364 int i = task_cpu(p);
4366 if (idle_cpu(target))
4370 * If the prevous cpu is cache affine and idle, don't be stupid.
4372 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4376 * Otherwise, iterate the domains and find an elegible idle cpu.
4378 sd = rcu_dereference(per_cpu(sd_llc, target));
4379 for_each_lower_domain(sd) {
4382 if (!cpumask_intersects(sched_group_cpus(sg),
4383 tsk_cpus_allowed(p)))
4386 for_each_cpu(i, sched_group_cpus(sg)) {
4387 if (i == target || !idle_cpu(i))
4391 target = cpumask_first_and(sched_group_cpus(sg),
4392 tsk_cpus_allowed(p));
4396 } while (sg != sd->groups);
4403 * select_task_rq_fair: Select target runqueue for the waking task in domains
4404 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4405 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4407 * Balances load by selecting the idlest cpu in the idlest group, or under
4408 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4410 * Returns the target cpu number.
4412 * preempt must be disabled.
4415 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4417 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4418 int cpu = smp_processor_id();
4420 int want_affine = 0;
4421 int sync = wake_flags & WF_SYNC;
4423 if (p->nr_cpus_allowed == 1)
4426 if (sd_flag & SD_BALANCE_WAKE) {
4427 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
4433 for_each_domain(cpu, tmp) {
4434 if (!(tmp->flags & SD_LOAD_BALANCE))
4438 * If both cpu and prev_cpu are part of this domain,
4439 * cpu is a valid SD_WAKE_AFFINE target.
4441 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4442 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4447 if (tmp->flags & sd_flag)
4452 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4455 new_cpu = select_idle_sibling(p, prev_cpu);
4460 struct sched_group *group;
4463 if (!(sd->flags & sd_flag)) {
4468 group = find_idlest_group(sd, p, cpu, sd_flag);
4474 new_cpu = find_idlest_cpu(group, p, cpu);
4475 if (new_cpu == -1 || new_cpu == cpu) {
4476 /* Now try balancing at a lower domain level of cpu */
4481 /* Now try balancing at a lower domain level of new_cpu */
4483 weight = sd->span_weight;
4485 for_each_domain(cpu, tmp) {
4486 if (weight <= tmp->span_weight)
4488 if (tmp->flags & sd_flag)
4491 /* while loop will break here if sd == NULL */
4500 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4501 * cfs_rq_of(p) references at time of call are still valid and identify the
4502 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4503 * other assumptions, including the state of rq->lock, should be made.
4506 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4508 struct sched_entity *se = &p->se;
4509 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4512 * Load tracking: accumulate removed load so that it can be processed
4513 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4514 * to blocked load iff they have a positive decay-count. It can never
4515 * be negative here since on-rq tasks have decay-count == 0.
4517 if (se->avg.decay_count) {
4518 se->avg.decay_count = -__synchronize_entity_decay(se);
4519 atomic_long_add(se->avg.load_avg_contrib,
4520 &cfs_rq->removed_load);
4523 #endif /* CONFIG_SMP */
4525 static unsigned long
4526 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4528 unsigned long gran = sysctl_sched_wakeup_granularity;
4531 * Since its curr running now, convert the gran from real-time
4532 * to virtual-time in his units.
4534 * By using 'se' instead of 'curr' we penalize light tasks, so
4535 * they get preempted easier. That is, if 'se' < 'curr' then
4536 * the resulting gran will be larger, therefore penalizing the
4537 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4538 * be smaller, again penalizing the lighter task.
4540 * This is especially important for buddies when the leftmost
4541 * task is higher priority than the buddy.
4543 return calc_delta_fair(gran, se);
4547 * Should 'se' preempt 'curr'.
4561 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4563 s64 gran, vdiff = curr->vruntime - se->vruntime;
4568 gran = wakeup_gran(curr, se);
4575 static void set_last_buddy(struct sched_entity *se)
4577 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4580 for_each_sched_entity(se)
4581 cfs_rq_of(se)->last = se;
4584 static void set_next_buddy(struct sched_entity *se)
4586 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4589 for_each_sched_entity(se)
4590 cfs_rq_of(se)->next = se;
4593 static void set_skip_buddy(struct sched_entity *se)
4595 for_each_sched_entity(se)
4596 cfs_rq_of(se)->skip = se;
4600 * Preempt the current task with a newly woken task if needed:
4602 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4604 struct task_struct *curr = rq->curr;
4605 struct sched_entity *se = &curr->se, *pse = &p->se;
4606 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4607 int scale = cfs_rq->nr_running >= sched_nr_latency;
4608 int next_buddy_marked = 0;
4610 if (unlikely(se == pse))
4614 * This is possible from callers such as move_task(), in which we
4615 * unconditionally check_prempt_curr() after an enqueue (which may have
4616 * lead to a throttle). This both saves work and prevents false
4617 * next-buddy nomination below.
4619 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4622 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4623 set_next_buddy(pse);
4624 next_buddy_marked = 1;
4628 * We can come here with TIF_NEED_RESCHED already set from new task
4631 * Note: this also catches the edge-case of curr being in a throttled
4632 * group (e.g. via set_curr_task), since update_curr() (in the
4633 * enqueue of curr) will have resulted in resched being set. This
4634 * prevents us from potentially nominating it as a false LAST_BUDDY
4637 if (test_tsk_need_resched(curr))
4640 /* Idle tasks are by definition preempted by non-idle tasks. */
4641 if (unlikely(curr->policy == SCHED_IDLE) &&
4642 likely(p->policy != SCHED_IDLE))
4646 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4647 * is driven by the tick):
4649 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4652 find_matching_se(&se, &pse);
4653 update_curr(cfs_rq_of(se));
4655 if (wakeup_preempt_entity(se, pse) == 1) {
4657 * Bias pick_next to pick the sched entity that is
4658 * triggering this preemption.
4660 if (!next_buddy_marked)
4661 set_next_buddy(pse);
4670 * Only set the backward buddy when the current task is still
4671 * on the rq. This can happen when a wakeup gets interleaved
4672 * with schedule on the ->pre_schedule() or idle_balance()
4673 * point, either of which can * drop the rq lock.
4675 * Also, during early boot the idle thread is in the fair class,
4676 * for obvious reasons its a bad idea to schedule back to it.
4678 if (unlikely(!se->on_rq || curr == rq->idle))
4681 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4685 static struct task_struct *
4686 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
4688 struct cfs_rq *cfs_rq = &rq->cfs;
4689 struct sched_entity *se;
4690 struct task_struct *p;
4694 #ifdef CONFIG_FAIR_GROUP_SCHED
4695 if (!cfs_rq->nr_running)
4698 if (prev->sched_class != &fair_sched_class)
4702 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
4703 * likely that a next task is from the same cgroup as the current.
4705 * Therefore attempt to avoid putting and setting the entire cgroup
4706 * hierarchy, only change the part that actually changes.
4710 struct sched_entity *curr = cfs_rq->curr;
4713 * Since we got here without doing put_prev_entity() we also
4714 * have to consider cfs_rq->curr. If it is still a runnable
4715 * entity, update_curr() will update its vruntime, otherwise
4716 * forget we've ever seen it.
4718 if (curr && curr->on_rq)
4719 update_curr(cfs_rq);
4724 * This call to check_cfs_rq_runtime() will do the throttle and
4725 * dequeue its entity in the parent(s). Therefore the 'simple'
4726 * nr_running test will indeed be correct.
4728 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
4731 se = pick_next_entity(cfs_rq, curr);
4732 cfs_rq = group_cfs_rq(se);
4738 * Since we haven't yet done put_prev_entity and if the selected task
4739 * is a different task than we started out with, try and touch the
4740 * least amount of cfs_rqs.
4743 struct sched_entity *pse = &prev->se;
4745 while (!(cfs_rq = is_same_group(se, pse))) {
4746 int se_depth = se->depth;
4747 int pse_depth = pse->depth;
4749 if (se_depth <= pse_depth) {
4750 put_prev_entity(cfs_rq_of(pse), pse);
4751 pse = parent_entity(pse);
4753 if (se_depth >= pse_depth) {
4754 set_next_entity(cfs_rq_of(se), se);
4755 se = parent_entity(se);
4759 put_prev_entity(cfs_rq, pse);
4760 set_next_entity(cfs_rq, se);
4763 if (hrtick_enabled(rq))
4764 hrtick_start_fair(rq, p);
4771 if (!cfs_rq->nr_running)
4774 put_prev_task(rq, prev);
4777 se = pick_next_entity(cfs_rq, NULL);
4778 set_next_entity(cfs_rq, se);
4779 cfs_rq = group_cfs_rq(se);
4784 if (hrtick_enabled(rq))
4785 hrtick_start_fair(rq, p);
4791 * Because idle_balance() releases (and re-acquires) rq->lock, it is
4792 * possible for any higher priority task to appear. In that case we
4793 * must re-start the pick_next_entity() loop.
4795 new_tasks = idle_balance(rq);
4797 if (rq->nr_running != rq->cfs.h_nr_running)
4807 * Account for a descheduled task:
4809 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4811 struct sched_entity *se = &prev->se;
4812 struct cfs_rq *cfs_rq;
4814 for_each_sched_entity(se) {
4815 cfs_rq = cfs_rq_of(se);
4816 put_prev_entity(cfs_rq, se);
4821 * sched_yield() is very simple
4823 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4825 static void yield_task_fair(struct rq *rq)
4827 struct task_struct *curr = rq->curr;
4828 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4829 struct sched_entity *se = &curr->se;
4832 * Are we the only task in the tree?
4834 if (unlikely(rq->nr_running == 1))
4837 clear_buddies(cfs_rq, se);
4839 if (curr->policy != SCHED_BATCH) {
4840 update_rq_clock(rq);
4842 * Update run-time statistics of the 'current'.
4844 update_curr(cfs_rq);
4846 * Tell update_rq_clock() that we've just updated,
4847 * so we don't do microscopic update in schedule()
4848 * and double the fastpath cost.
4850 rq->skip_clock_update = 1;
4856 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4858 struct sched_entity *se = &p->se;
4860 /* throttled hierarchies are not runnable */
4861 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4864 /* Tell the scheduler that we'd really like pse to run next. */
4867 yield_task_fair(rq);
4873 /**************************************************
4874 * Fair scheduling class load-balancing methods.
4878 * The purpose of load-balancing is to achieve the same basic fairness the
4879 * per-cpu scheduler provides, namely provide a proportional amount of compute
4880 * time to each task. This is expressed in the following equation:
4882 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4884 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4885 * W_i,0 is defined as:
4887 * W_i,0 = \Sum_j w_i,j (2)
4889 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4890 * is derived from the nice value as per prio_to_weight[].
4892 * The weight average is an exponential decay average of the instantaneous
4895 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4897 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4898 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4899 * can also include other factors [XXX].
4901 * To achieve this balance we define a measure of imbalance which follows
4902 * directly from (1):
4904 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4906 * We them move tasks around to minimize the imbalance. In the continuous
4907 * function space it is obvious this converges, in the discrete case we get
4908 * a few fun cases generally called infeasible weight scenarios.
4911 * - infeasible weights;
4912 * - local vs global optima in the discrete case. ]
4917 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4918 * for all i,j solution, we create a tree of cpus that follows the hardware
4919 * topology where each level pairs two lower groups (or better). This results
4920 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4921 * tree to only the first of the previous level and we decrease the frequency
4922 * of load-balance at each level inv. proportional to the number of cpus in
4928 * \Sum { --- * --- * 2^i } = O(n) (5)
4930 * `- size of each group
4931 * | | `- number of cpus doing load-balance
4933 * `- sum over all levels
4935 * Coupled with a limit on how many tasks we can migrate every balance pass,
4936 * this makes (5) the runtime complexity of the balancer.
4938 * An important property here is that each CPU is still (indirectly) connected
4939 * to every other cpu in at most O(log n) steps:
4941 * The adjacency matrix of the resulting graph is given by:
4944 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4947 * And you'll find that:
4949 * A^(log_2 n)_i,j != 0 for all i,j (7)
4951 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4952 * The task movement gives a factor of O(m), giving a convergence complexity
4955 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4960 * In order to avoid CPUs going idle while there's still work to do, new idle
4961 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4962 * tree itself instead of relying on other CPUs to bring it work.
4964 * This adds some complexity to both (5) and (8) but it reduces the total idle
4972 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4975 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4980 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4982 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4984 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4987 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4988 * rewrite all of this once again.]
4991 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4993 enum fbq_type { regular, remote, all };
4995 #define LBF_ALL_PINNED 0x01
4996 #define LBF_NEED_BREAK 0x02
4997 #define LBF_DST_PINNED 0x04
4998 #define LBF_SOME_PINNED 0x08
5001 struct sched_domain *sd;
5009 struct cpumask *dst_grpmask;
5011 enum cpu_idle_type idle;
5013 /* The set of CPUs under consideration for load-balancing */
5014 struct cpumask *cpus;
5019 unsigned int loop_break;
5020 unsigned int loop_max;
5022 enum fbq_type fbq_type;
5026 * move_task - move a task from one runqueue to another runqueue.
5027 * Both runqueues must be locked.
5029 static void move_task(struct task_struct *p, struct lb_env *env)
5031 deactivate_task(env->src_rq, p, 0);
5032 set_task_cpu(p, env->dst_cpu);
5033 activate_task(env->dst_rq, p, 0);
5034 check_preempt_curr(env->dst_rq, p, 0);
5038 * Is this task likely cache-hot:
5041 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
5045 if (p->sched_class != &fair_sched_class)
5048 if (unlikely(p->policy == SCHED_IDLE))
5052 * Buddy candidates are cache hot:
5054 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
5055 (&p->se == cfs_rq_of(&p->se)->next ||
5056 &p->se == cfs_rq_of(&p->se)->last))
5059 if (sysctl_sched_migration_cost == -1)
5061 if (sysctl_sched_migration_cost == 0)
5064 delta = now - p->se.exec_start;
5066 return delta < (s64)sysctl_sched_migration_cost;
5069 #ifdef CONFIG_NUMA_BALANCING
5070 /* Returns true if the destination node has incurred more faults */
5071 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
5073 int src_nid, dst_nid;
5075 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults_memory ||
5076 !(env->sd->flags & SD_NUMA)) {
5080 src_nid = cpu_to_node(env->src_cpu);
5081 dst_nid = cpu_to_node(env->dst_cpu);
5083 if (src_nid == dst_nid)
5086 /* Always encourage migration to the preferred node. */
5087 if (dst_nid == p->numa_preferred_nid)
5090 /* If both task and group weight improve, this move is a winner. */
5091 if (task_weight(p, dst_nid) > task_weight(p, src_nid) &&
5092 group_weight(p, dst_nid) > group_weight(p, src_nid))
5099 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5101 int src_nid, dst_nid;
5103 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
5106 if (!p->numa_faults_memory || !(env->sd->flags & SD_NUMA))
5109 src_nid = cpu_to_node(env->src_cpu);
5110 dst_nid = cpu_to_node(env->dst_cpu);
5112 if (src_nid == dst_nid)
5115 /* Migrating away from the preferred node is always bad. */
5116 if (src_nid == p->numa_preferred_nid)
5119 /* If either task or group weight get worse, don't do it. */
5120 if (task_weight(p, dst_nid) < task_weight(p, src_nid) ||
5121 group_weight(p, dst_nid) < group_weight(p, src_nid))
5128 static inline bool migrate_improves_locality(struct task_struct *p,
5134 static inline bool migrate_degrades_locality(struct task_struct *p,
5142 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5145 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5147 int tsk_cache_hot = 0;
5149 * We do not migrate tasks that are:
5150 * 1) throttled_lb_pair, or
5151 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5152 * 3) running (obviously), or
5153 * 4) are cache-hot on their current CPU.
5155 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5158 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5161 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5163 env->flags |= LBF_SOME_PINNED;
5166 * Remember if this task can be migrated to any other cpu in
5167 * our sched_group. We may want to revisit it if we couldn't
5168 * meet load balance goals by pulling other tasks on src_cpu.
5170 * Also avoid computing new_dst_cpu if we have already computed
5171 * one in current iteration.
5173 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5176 /* Prevent to re-select dst_cpu via env's cpus */
5177 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5178 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5179 env->flags |= LBF_DST_PINNED;
5180 env->new_dst_cpu = cpu;
5188 /* Record that we found atleast one task that could run on dst_cpu */
5189 env->flags &= ~LBF_ALL_PINNED;
5191 if (task_running(env->src_rq, p)) {
5192 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5197 * Aggressive migration if:
5198 * 1) destination numa is preferred
5199 * 2) task is cache cold, or
5200 * 3) too many balance attempts have failed.
5202 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
5204 tsk_cache_hot = migrate_degrades_locality(p, env);
5206 if (migrate_improves_locality(p, env)) {
5207 #ifdef CONFIG_SCHEDSTATS
5208 if (tsk_cache_hot) {
5209 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5210 schedstat_inc(p, se.statistics.nr_forced_migrations);
5216 if (!tsk_cache_hot ||
5217 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5219 if (tsk_cache_hot) {
5220 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5221 schedstat_inc(p, se.statistics.nr_forced_migrations);
5227 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5232 * move_one_task tries to move exactly one task from busiest to this_rq, as
5233 * part of active balancing operations within "domain".
5234 * Returns 1 if successful and 0 otherwise.
5236 * Called with both runqueues locked.
5238 static int move_one_task(struct lb_env *env)
5240 struct task_struct *p, *n;
5242 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5243 if (!can_migrate_task(p, env))
5248 * Right now, this is only the second place move_task()
5249 * is called, so we can safely collect move_task()
5250 * stats here rather than inside move_task().
5252 schedstat_inc(env->sd, lb_gained[env->idle]);
5258 static const unsigned int sched_nr_migrate_break = 32;
5261 * move_tasks tries to move up to imbalance weighted load from busiest to
5262 * this_rq, as part of a balancing operation within domain "sd".
5263 * Returns 1 if successful and 0 otherwise.
5265 * Called with both runqueues locked.
5267 static int move_tasks(struct lb_env *env)
5269 struct list_head *tasks = &env->src_rq->cfs_tasks;
5270 struct task_struct *p;
5274 if (env->imbalance <= 0)
5277 while (!list_empty(tasks)) {
5278 p = list_first_entry(tasks, struct task_struct, se.group_node);
5281 /* We've more or less seen every task there is, call it quits */
5282 if (env->loop > env->loop_max)
5285 /* take a breather every nr_migrate tasks */
5286 if (env->loop > env->loop_break) {
5287 env->loop_break += sched_nr_migrate_break;
5288 env->flags |= LBF_NEED_BREAK;
5292 if (!can_migrate_task(p, env))
5295 load = task_h_load(p);
5297 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5300 if ((load / 2) > env->imbalance)
5305 env->imbalance -= load;
5307 #ifdef CONFIG_PREEMPT
5309 * NEWIDLE balancing is a source of latency, so preemptible
5310 * kernels will stop after the first task is pulled to minimize
5311 * the critical section.
5313 if (env->idle == CPU_NEWLY_IDLE)
5318 * We only want to steal up to the prescribed amount of
5321 if (env->imbalance <= 0)
5326 list_move_tail(&p->se.group_node, tasks);
5330 * Right now, this is one of only two places move_task() is called,
5331 * so we can safely collect move_task() stats here rather than
5332 * inside move_task().
5334 schedstat_add(env->sd, lb_gained[env->idle], pulled);
5339 #ifdef CONFIG_FAIR_GROUP_SCHED
5341 * update tg->load_weight by folding this cpu's load_avg
5343 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5345 struct sched_entity *se = tg->se[cpu];
5346 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5348 /* throttled entities do not contribute to load */
5349 if (throttled_hierarchy(cfs_rq))
5352 update_cfs_rq_blocked_load(cfs_rq, 1);
5355 update_entity_load_avg(se, 1);
5357 * We pivot on our runnable average having decayed to zero for
5358 * list removal. This generally implies that all our children
5359 * have also been removed (modulo rounding error or bandwidth
5360 * control); however, such cases are rare and we can fix these
5363 * TODO: fix up out-of-order children on enqueue.
5365 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5366 list_del_leaf_cfs_rq(cfs_rq);
5368 struct rq *rq = rq_of(cfs_rq);
5369 update_rq_runnable_avg(rq, rq->nr_running);
5373 static void update_blocked_averages(int cpu)
5375 struct rq *rq = cpu_rq(cpu);
5376 struct cfs_rq *cfs_rq;
5377 unsigned long flags;
5379 raw_spin_lock_irqsave(&rq->lock, flags);
5380 update_rq_clock(rq);
5382 * Iterates the task_group tree in a bottom up fashion, see
5383 * list_add_leaf_cfs_rq() for details.
5385 for_each_leaf_cfs_rq(rq, cfs_rq) {
5387 * Note: We may want to consider periodically releasing
5388 * rq->lock about these updates so that creating many task
5389 * groups does not result in continually extending hold time.
5391 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5394 raw_spin_unlock_irqrestore(&rq->lock, flags);
5398 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5399 * This needs to be done in a top-down fashion because the load of a child
5400 * group is a fraction of its parents load.
5402 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5404 struct rq *rq = rq_of(cfs_rq);
5405 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5406 unsigned long now = jiffies;
5409 if (cfs_rq->last_h_load_update == now)
5412 cfs_rq->h_load_next = NULL;
5413 for_each_sched_entity(se) {
5414 cfs_rq = cfs_rq_of(se);
5415 cfs_rq->h_load_next = se;
5416 if (cfs_rq->last_h_load_update == now)
5421 cfs_rq->h_load = cfs_rq->runnable_load_avg;
5422 cfs_rq->last_h_load_update = now;
5425 while ((se = cfs_rq->h_load_next) != NULL) {
5426 load = cfs_rq->h_load;
5427 load = div64_ul(load * se->avg.load_avg_contrib,
5428 cfs_rq->runnable_load_avg + 1);
5429 cfs_rq = group_cfs_rq(se);
5430 cfs_rq->h_load = load;
5431 cfs_rq->last_h_load_update = now;
5435 static unsigned long task_h_load(struct task_struct *p)
5437 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5439 update_cfs_rq_h_load(cfs_rq);
5440 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
5441 cfs_rq->runnable_load_avg + 1);
5444 static inline void update_blocked_averages(int cpu)
5448 static unsigned long task_h_load(struct task_struct *p)
5450 return p->se.avg.load_avg_contrib;
5454 /********** Helpers for find_busiest_group ************************/
5456 * sg_lb_stats - stats of a sched_group required for load_balancing
5458 struct sg_lb_stats {
5459 unsigned long avg_load; /*Avg load across the CPUs of the group */
5460 unsigned long group_load; /* Total load over the CPUs of the group */
5461 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5462 unsigned long load_per_task;
5463 unsigned long group_power;
5464 unsigned int sum_nr_running; /* Nr tasks running in the group */
5465 unsigned int group_capacity;
5466 unsigned int idle_cpus;
5467 unsigned int group_weight;
5468 int group_imb; /* Is there an imbalance in the group ? */
5469 int group_has_capacity; /* Is there extra capacity in the group? */
5470 #ifdef CONFIG_NUMA_BALANCING
5471 unsigned int nr_numa_running;
5472 unsigned int nr_preferred_running;
5477 * sd_lb_stats - Structure to store the statistics of a sched_domain
5478 * during load balancing.
5480 struct sd_lb_stats {
5481 struct sched_group *busiest; /* Busiest group in this sd */
5482 struct sched_group *local; /* Local group in this sd */
5483 unsigned long total_load; /* Total load of all groups in sd */
5484 unsigned long total_pwr; /* Total power of all groups in sd */
5485 unsigned long avg_load; /* Average load across all groups in sd */
5487 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5488 struct sg_lb_stats local_stat; /* Statistics of the local group */
5491 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5494 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5495 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5496 * We must however clear busiest_stat::avg_load because
5497 * update_sd_pick_busiest() reads this before assignment.
5499 *sds = (struct sd_lb_stats){
5511 * get_sd_load_idx - Obtain the load index for a given sched domain.
5512 * @sd: The sched_domain whose load_idx is to be obtained.
5513 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5515 * Return: The load index.
5517 static inline int get_sd_load_idx(struct sched_domain *sd,
5518 enum cpu_idle_type idle)
5524 load_idx = sd->busy_idx;
5527 case CPU_NEWLY_IDLE:
5528 load_idx = sd->newidle_idx;
5531 load_idx = sd->idle_idx;
5538 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
5540 return SCHED_POWER_SCALE;
5543 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
5545 return default_scale_freq_power(sd, cpu);
5548 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
5550 unsigned long weight = sd->span_weight;
5551 unsigned long smt_gain = sd->smt_gain;
5558 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
5560 return default_scale_smt_power(sd, cpu);
5563 static unsigned long scale_rt_power(int cpu)
5565 struct rq *rq = cpu_rq(cpu);
5566 u64 total, available, age_stamp, avg;
5569 * Since we're reading these variables without serialization make sure
5570 * we read them once before doing sanity checks on them.
5572 age_stamp = ACCESS_ONCE(rq->age_stamp);
5573 avg = ACCESS_ONCE(rq->rt_avg);
5575 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
5577 if (unlikely(total < avg)) {
5578 /* Ensures that power won't end up being negative */
5581 available = total - avg;
5584 if (unlikely((s64)total < SCHED_POWER_SCALE))
5585 total = SCHED_POWER_SCALE;
5587 total >>= SCHED_POWER_SHIFT;
5589 return div_u64(available, total);
5592 static void update_cpu_power(struct sched_domain *sd, int cpu)
5594 unsigned long weight = sd->span_weight;
5595 unsigned long power = SCHED_POWER_SCALE;
5596 struct sched_group *sdg = sd->groups;
5598 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
5599 if (sched_feat(ARCH_POWER))
5600 power *= arch_scale_smt_power(sd, cpu);
5602 power *= default_scale_smt_power(sd, cpu);
5604 power >>= SCHED_POWER_SHIFT;
5607 sdg->sgp->power_orig = power;
5609 if (sched_feat(ARCH_POWER))
5610 power *= arch_scale_freq_power(sd, cpu);
5612 power *= default_scale_freq_power(sd, cpu);
5614 power >>= SCHED_POWER_SHIFT;
5616 power *= scale_rt_power(cpu);
5617 power >>= SCHED_POWER_SHIFT;
5622 cpu_rq(cpu)->cpu_power = power;
5623 sdg->sgp->power = power;
5626 void update_group_power(struct sched_domain *sd, int cpu)
5628 struct sched_domain *child = sd->child;
5629 struct sched_group *group, *sdg = sd->groups;
5630 unsigned long power, power_orig;
5631 unsigned long interval;
5633 interval = msecs_to_jiffies(sd->balance_interval);
5634 interval = clamp(interval, 1UL, max_load_balance_interval);
5635 sdg->sgp->next_update = jiffies + interval;
5638 update_cpu_power(sd, cpu);
5642 power_orig = power = 0;
5644 if (child->flags & SD_OVERLAP) {
5646 * SD_OVERLAP domains cannot assume that child groups
5647 * span the current group.
5650 for_each_cpu(cpu, sched_group_cpus(sdg)) {
5651 struct sched_group_power *sgp;
5652 struct rq *rq = cpu_rq(cpu);
5655 * build_sched_domains() -> init_sched_groups_power()
5656 * gets here before we've attached the domains to the
5659 * Use power_of(), which is set irrespective of domains
5660 * in update_cpu_power().
5662 * This avoids power/power_orig from being 0 and
5663 * causing divide-by-zero issues on boot.
5665 * Runtime updates will correct power_orig.
5667 if (unlikely(!rq->sd)) {
5668 power_orig += power_of(cpu);
5669 power += power_of(cpu);
5673 sgp = rq->sd->groups->sgp;
5674 power_orig += sgp->power_orig;
5675 power += sgp->power;
5679 * !SD_OVERLAP domains can assume that child groups
5680 * span the current group.
5683 group = child->groups;
5685 power_orig += group->sgp->power_orig;
5686 power += group->sgp->power;
5687 group = group->next;
5688 } while (group != child->groups);
5691 sdg->sgp->power_orig = power_orig;
5692 sdg->sgp->power = power;
5696 * Try and fix up capacity for tiny siblings, this is needed when
5697 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5698 * which on its own isn't powerful enough.
5700 * See update_sd_pick_busiest() and check_asym_packing().
5703 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5706 * Only siblings can have significantly less than SCHED_POWER_SCALE
5708 if (!(sd->flags & SD_SHARE_CPUPOWER))
5712 * If ~90% of the cpu_power is still there, we're good.
5714 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
5721 * Group imbalance indicates (and tries to solve) the problem where balancing
5722 * groups is inadequate due to tsk_cpus_allowed() constraints.
5724 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5725 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5728 * { 0 1 2 3 } { 4 5 6 7 }
5731 * If we were to balance group-wise we'd place two tasks in the first group and
5732 * two tasks in the second group. Clearly this is undesired as it will overload
5733 * cpu 3 and leave one of the cpus in the second group unused.
5735 * The current solution to this issue is detecting the skew in the first group
5736 * by noticing the lower domain failed to reach balance and had difficulty
5737 * moving tasks due to affinity constraints.
5739 * When this is so detected; this group becomes a candidate for busiest; see
5740 * update_sd_pick_busiest(). And calculate_imbalance() and
5741 * find_busiest_group() avoid some of the usual balance conditions to allow it
5742 * to create an effective group imbalance.
5744 * This is a somewhat tricky proposition since the next run might not find the
5745 * group imbalance and decide the groups need to be balanced again. A most
5746 * subtle and fragile situation.
5749 static inline int sg_imbalanced(struct sched_group *group)
5751 return group->sgp->imbalance;
5755 * Compute the group capacity.
5757 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
5758 * first dividing out the smt factor and computing the actual number of cores
5759 * and limit power unit capacity with that.
5761 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
5763 unsigned int capacity, smt, cpus;
5764 unsigned int power, power_orig;
5766 power = group->sgp->power;
5767 power_orig = group->sgp->power_orig;
5768 cpus = group->group_weight;
5770 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
5771 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
5772 capacity = cpus / smt; /* cores */
5774 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
5776 capacity = fix_small_capacity(env->sd, group);
5782 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5783 * @env: The load balancing environment.
5784 * @group: sched_group whose statistics are to be updated.
5785 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5786 * @local_group: Does group contain this_cpu.
5787 * @sgs: variable to hold the statistics for this group.
5789 static inline void update_sg_lb_stats(struct lb_env *env,
5790 struct sched_group *group, int load_idx,
5791 int local_group, struct sg_lb_stats *sgs)
5796 memset(sgs, 0, sizeof(*sgs));
5798 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5799 struct rq *rq = cpu_rq(i);
5801 /* Bias balancing toward cpus of our domain */
5803 load = target_load(i, load_idx);
5805 load = source_load(i, load_idx);
5807 sgs->group_load += load;
5808 sgs->sum_nr_running += rq->nr_running;
5809 #ifdef CONFIG_NUMA_BALANCING
5810 sgs->nr_numa_running += rq->nr_numa_running;
5811 sgs->nr_preferred_running += rq->nr_preferred_running;
5813 sgs->sum_weighted_load += weighted_cpuload(i);
5818 /* Adjust by relative CPU power of the group */
5819 sgs->group_power = group->sgp->power;
5820 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
5822 if (sgs->sum_nr_running)
5823 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
5825 sgs->group_weight = group->group_weight;
5827 sgs->group_imb = sg_imbalanced(group);
5828 sgs->group_capacity = sg_capacity(env, group);
5830 if (sgs->group_capacity > sgs->sum_nr_running)
5831 sgs->group_has_capacity = 1;
5835 * update_sd_pick_busiest - return 1 on busiest group
5836 * @env: The load balancing environment.
5837 * @sds: sched_domain statistics
5838 * @sg: sched_group candidate to be checked for being the busiest
5839 * @sgs: sched_group statistics
5841 * Determine if @sg is a busier group than the previously selected
5844 * Return: %true if @sg is a busier group than the previously selected
5845 * busiest group. %false otherwise.
5847 static bool update_sd_pick_busiest(struct lb_env *env,
5848 struct sd_lb_stats *sds,
5849 struct sched_group *sg,
5850 struct sg_lb_stats *sgs)
5852 if (sgs->avg_load <= sds->busiest_stat.avg_load)
5855 if (sgs->sum_nr_running > sgs->group_capacity)
5862 * ASYM_PACKING needs to move all the work to the lowest
5863 * numbered CPUs in the group, therefore mark all groups
5864 * higher than ourself as busy.
5866 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5867 env->dst_cpu < group_first_cpu(sg)) {
5871 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5878 #ifdef CONFIG_NUMA_BALANCING
5879 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5881 if (sgs->sum_nr_running > sgs->nr_numa_running)
5883 if (sgs->sum_nr_running > sgs->nr_preferred_running)
5888 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5890 if (rq->nr_running > rq->nr_numa_running)
5892 if (rq->nr_running > rq->nr_preferred_running)
5897 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5902 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5906 #endif /* CONFIG_NUMA_BALANCING */
5909 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5910 * @env: The load balancing environment.
5911 * @sds: variable to hold the statistics for this sched_domain.
5913 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
5915 struct sched_domain *child = env->sd->child;
5916 struct sched_group *sg = env->sd->groups;
5917 struct sg_lb_stats tmp_sgs;
5918 int load_idx, prefer_sibling = 0;
5920 if (child && child->flags & SD_PREFER_SIBLING)
5923 load_idx = get_sd_load_idx(env->sd, env->idle);
5926 struct sg_lb_stats *sgs = &tmp_sgs;
5929 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
5932 sgs = &sds->local_stat;
5934 if (env->idle != CPU_NEWLY_IDLE ||
5935 time_after_eq(jiffies, sg->sgp->next_update))
5936 update_group_power(env->sd, env->dst_cpu);
5939 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
5945 * In case the child domain prefers tasks go to siblings
5946 * first, lower the sg capacity to one so that we'll try
5947 * and move all the excess tasks away. We lower the capacity
5948 * of a group only if the local group has the capacity to fit
5949 * these excess tasks, i.e. nr_running < group_capacity. The
5950 * extra check prevents the case where you always pull from the
5951 * heaviest group when it is already under-utilized (possible
5952 * with a large weight task outweighs the tasks on the system).
5954 if (prefer_sibling && sds->local &&
5955 sds->local_stat.group_has_capacity)
5956 sgs->group_capacity = min(sgs->group_capacity, 1U);
5958 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
5960 sds->busiest_stat = *sgs;
5964 /* Now, start updating sd_lb_stats */
5965 sds->total_load += sgs->group_load;
5966 sds->total_pwr += sgs->group_power;
5969 } while (sg != env->sd->groups);
5971 if (env->sd->flags & SD_NUMA)
5972 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
5976 * check_asym_packing - Check to see if the group is packed into the
5979 * This is primarily intended to used at the sibling level. Some
5980 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5981 * case of POWER7, it can move to lower SMT modes only when higher
5982 * threads are idle. When in lower SMT modes, the threads will
5983 * perform better since they share less core resources. Hence when we
5984 * have idle threads, we want them to be the higher ones.
5986 * This packing function is run on idle threads. It checks to see if
5987 * the busiest CPU in this domain (core in the P7 case) has a higher
5988 * CPU number than the packing function is being run on. Here we are
5989 * assuming lower CPU number will be equivalent to lower a SMT thread
5992 * Return: 1 when packing is required and a task should be moved to
5993 * this CPU. The amount of the imbalance is returned in *imbalance.
5995 * @env: The load balancing environment.
5996 * @sds: Statistics of the sched_domain which is to be packed
5998 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6002 if (!(env->sd->flags & SD_ASYM_PACKING))
6008 busiest_cpu = group_first_cpu(sds->busiest);
6009 if (env->dst_cpu > busiest_cpu)
6012 env->imbalance = DIV_ROUND_CLOSEST(
6013 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
6020 * fix_small_imbalance - Calculate the minor imbalance that exists
6021 * amongst the groups of a sched_domain, during
6023 * @env: The load balancing environment.
6024 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6027 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6029 unsigned long tmp, pwr_now = 0, pwr_move = 0;
6030 unsigned int imbn = 2;
6031 unsigned long scaled_busy_load_per_task;
6032 struct sg_lb_stats *local, *busiest;
6034 local = &sds->local_stat;
6035 busiest = &sds->busiest_stat;
6037 if (!local->sum_nr_running)
6038 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6039 else if (busiest->load_per_task > local->load_per_task)
6042 scaled_busy_load_per_task =
6043 (busiest->load_per_task * SCHED_POWER_SCALE) /
6044 busiest->group_power;
6046 if (busiest->avg_load + scaled_busy_load_per_task >=
6047 local->avg_load + (scaled_busy_load_per_task * imbn)) {
6048 env->imbalance = busiest->load_per_task;
6053 * OK, we don't have enough imbalance to justify moving tasks,
6054 * however we may be able to increase total CPU power used by
6058 pwr_now += busiest->group_power *
6059 min(busiest->load_per_task, busiest->avg_load);
6060 pwr_now += local->group_power *
6061 min(local->load_per_task, local->avg_load);
6062 pwr_now /= SCHED_POWER_SCALE;
6064 /* Amount of load we'd subtract */
6065 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
6066 busiest->group_power;
6067 if (busiest->avg_load > tmp) {
6068 pwr_move += busiest->group_power *
6069 min(busiest->load_per_task,
6070 busiest->avg_load - tmp);
6073 /* Amount of load we'd add */
6074 if (busiest->avg_load * busiest->group_power <
6075 busiest->load_per_task * SCHED_POWER_SCALE) {
6076 tmp = (busiest->avg_load * busiest->group_power) /
6079 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
6082 pwr_move += local->group_power *
6083 min(local->load_per_task, local->avg_load + tmp);
6084 pwr_move /= SCHED_POWER_SCALE;
6086 /* Move if we gain throughput */
6087 if (pwr_move > pwr_now)
6088 env->imbalance = busiest->load_per_task;
6092 * calculate_imbalance - Calculate the amount of imbalance present within the
6093 * groups of a given sched_domain during load balance.
6094 * @env: load balance environment
6095 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6097 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6099 unsigned long max_pull, load_above_capacity = ~0UL;
6100 struct sg_lb_stats *local, *busiest;
6102 local = &sds->local_stat;
6103 busiest = &sds->busiest_stat;
6105 if (busiest->group_imb) {
6107 * In the group_imb case we cannot rely on group-wide averages
6108 * to ensure cpu-load equilibrium, look at wider averages. XXX
6110 busiest->load_per_task =
6111 min(busiest->load_per_task, sds->avg_load);
6115 * In the presence of smp nice balancing, certain scenarios can have
6116 * max load less than avg load(as we skip the groups at or below
6117 * its cpu_power, while calculating max_load..)
6119 if (busiest->avg_load <= sds->avg_load ||
6120 local->avg_load >= sds->avg_load) {
6122 return fix_small_imbalance(env, sds);
6125 if (!busiest->group_imb) {
6127 * Don't want to pull so many tasks that a group would go idle.
6128 * Except of course for the group_imb case, since then we might
6129 * have to drop below capacity to reach cpu-load equilibrium.
6131 load_above_capacity =
6132 (busiest->sum_nr_running - busiest->group_capacity);
6134 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
6135 load_above_capacity /= busiest->group_power;
6139 * We're trying to get all the cpus to the average_load, so we don't
6140 * want to push ourselves above the average load, nor do we wish to
6141 * reduce the max loaded cpu below the average load. At the same time,
6142 * we also don't want to reduce the group load below the group capacity
6143 * (so that we can implement power-savings policies etc). Thus we look
6144 * for the minimum possible imbalance.
6146 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6148 /* How much load to actually move to equalise the imbalance */
6149 env->imbalance = min(
6150 max_pull * busiest->group_power,
6151 (sds->avg_load - local->avg_load) * local->group_power
6152 ) / SCHED_POWER_SCALE;
6155 * if *imbalance is less than the average load per runnable task
6156 * there is no guarantee that any tasks will be moved so we'll have
6157 * a think about bumping its value to force at least one task to be
6160 if (env->imbalance < busiest->load_per_task)
6161 return fix_small_imbalance(env, sds);
6164 /******* find_busiest_group() helpers end here *********************/
6167 * find_busiest_group - Returns the busiest group within the sched_domain
6168 * if there is an imbalance. If there isn't an imbalance, and
6169 * the user has opted for power-savings, it returns a group whose
6170 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6171 * such a group exists.
6173 * Also calculates the amount of weighted load which should be moved
6174 * to restore balance.
6176 * @env: The load balancing environment.
6178 * Return: - The busiest group if imbalance exists.
6179 * - If no imbalance and user has opted for power-savings balance,
6180 * return the least loaded group whose CPUs can be
6181 * put to idle by rebalancing its tasks onto our group.
6183 static struct sched_group *find_busiest_group(struct lb_env *env)
6185 struct sg_lb_stats *local, *busiest;
6186 struct sd_lb_stats sds;
6188 init_sd_lb_stats(&sds);
6191 * Compute the various statistics relavent for load balancing at
6194 update_sd_lb_stats(env, &sds);
6195 local = &sds.local_stat;
6196 busiest = &sds.busiest_stat;
6198 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6199 check_asym_packing(env, &sds))
6202 /* There is no busy sibling group to pull tasks from */
6203 if (!sds.busiest || busiest->sum_nr_running == 0)
6206 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
6209 * If the busiest group is imbalanced the below checks don't
6210 * work because they assume all things are equal, which typically
6211 * isn't true due to cpus_allowed constraints and the like.
6213 if (busiest->group_imb)
6216 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6217 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
6218 !busiest->group_has_capacity)
6222 * If the local group is more busy than the selected busiest group
6223 * don't try and pull any tasks.
6225 if (local->avg_load >= busiest->avg_load)
6229 * Don't pull any tasks if this group is already above the domain
6232 if (local->avg_load >= sds.avg_load)
6235 if (env->idle == CPU_IDLE) {
6237 * This cpu is idle. If the busiest group load doesn't
6238 * have more tasks than the number of available cpu's and
6239 * there is no imbalance between this and busiest group
6240 * wrt to idle cpu's, it is balanced.
6242 if ((local->idle_cpus < busiest->idle_cpus) &&
6243 busiest->sum_nr_running <= busiest->group_weight)
6247 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6248 * imbalance_pct to be conservative.
6250 if (100 * busiest->avg_load <=
6251 env->sd->imbalance_pct * local->avg_load)
6256 /* Looks like there is an imbalance. Compute it */
6257 calculate_imbalance(env, &sds);
6266 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6268 static struct rq *find_busiest_queue(struct lb_env *env,
6269 struct sched_group *group)
6271 struct rq *busiest = NULL, *rq;
6272 unsigned long busiest_load = 0, busiest_power = 1;
6275 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6276 unsigned long power, capacity, wl;
6280 rt = fbq_classify_rq(rq);
6283 * We classify groups/runqueues into three groups:
6284 * - regular: there are !numa tasks
6285 * - remote: there are numa tasks that run on the 'wrong' node
6286 * - all: there is no distinction
6288 * In order to avoid migrating ideally placed numa tasks,
6289 * ignore those when there's better options.
6291 * If we ignore the actual busiest queue to migrate another
6292 * task, the next balance pass can still reduce the busiest
6293 * queue by moving tasks around inside the node.
6295 * If we cannot move enough load due to this classification
6296 * the next pass will adjust the group classification and
6297 * allow migration of more tasks.
6299 * Both cases only affect the total convergence complexity.
6301 if (rt > env->fbq_type)
6304 power = power_of(i);
6305 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
6307 capacity = fix_small_capacity(env->sd, group);
6309 wl = weighted_cpuload(i);
6312 * When comparing with imbalance, use weighted_cpuload()
6313 * which is not scaled with the cpu power.
6315 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
6319 * For the load comparisons with the other cpu's, consider
6320 * the weighted_cpuload() scaled with the cpu power, so that
6321 * the load can be moved away from the cpu that is potentially
6322 * running at a lower capacity.
6324 * Thus we're looking for max(wl_i / power_i), crosswise
6325 * multiplication to rid ourselves of the division works out
6326 * to: wl_i * power_j > wl_j * power_i; where j is our
6329 if (wl * busiest_power > busiest_load * power) {
6331 busiest_power = power;
6340 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6341 * so long as it is large enough.
6343 #define MAX_PINNED_INTERVAL 512
6345 /* Working cpumask for load_balance and load_balance_newidle. */
6346 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6348 static int need_active_balance(struct lb_env *env)
6350 struct sched_domain *sd = env->sd;
6352 if (env->idle == CPU_NEWLY_IDLE) {
6355 * ASYM_PACKING needs to force migrate tasks from busy but
6356 * higher numbered CPUs in order to pack all tasks in the
6357 * lowest numbered CPUs.
6359 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6363 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6366 static int active_load_balance_cpu_stop(void *data);
6368 static int should_we_balance(struct lb_env *env)
6370 struct sched_group *sg = env->sd->groups;
6371 struct cpumask *sg_cpus, *sg_mask;
6372 int cpu, balance_cpu = -1;
6375 * In the newly idle case, we will allow all the cpu's
6376 * to do the newly idle load balance.
6378 if (env->idle == CPU_NEWLY_IDLE)
6381 sg_cpus = sched_group_cpus(sg);
6382 sg_mask = sched_group_mask(sg);
6383 /* Try to find first idle cpu */
6384 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6385 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6392 if (balance_cpu == -1)
6393 balance_cpu = group_balance_cpu(sg);
6396 * First idle cpu or the first cpu(busiest) in this sched group
6397 * is eligible for doing load balancing at this and above domains.
6399 return balance_cpu == env->dst_cpu;
6403 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6404 * tasks if there is an imbalance.
6406 static int load_balance(int this_cpu, struct rq *this_rq,
6407 struct sched_domain *sd, enum cpu_idle_type idle,
6408 int *continue_balancing)
6410 int ld_moved, cur_ld_moved, active_balance = 0;
6411 struct sched_domain *sd_parent = sd->parent;
6412 struct sched_group *group;
6414 unsigned long flags;
6415 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
6417 struct lb_env env = {
6419 .dst_cpu = this_cpu,
6421 .dst_grpmask = sched_group_cpus(sd->groups),
6423 .loop_break = sched_nr_migrate_break,
6429 * For NEWLY_IDLE load_balancing, we don't need to consider
6430 * other cpus in our group
6432 if (idle == CPU_NEWLY_IDLE)
6433 env.dst_grpmask = NULL;
6435 cpumask_copy(cpus, cpu_active_mask);
6437 schedstat_inc(sd, lb_count[idle]);
6440 if (!should_we_balance(&env)) {
6441 *continue_balancing = 0;
6445 group = find_busiest_group(&env);
6447 schedstat_inc(sd, lb_nobusyg[idle]);
6451 busiest = find_busiest_queue(&env, group);
6453 schedstat_inc(sd, lb_nobusyq[idle]);
6457 BUG_ON(busiest == env.dst_rq);
6459 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6462 if (busiest->nr_running > 1) {
6464 * Attempt to move tasks. If find_busiest_group has found
6465 * an imbalance but busiest->nr_running <= 1, the group is
6466 * still unbalanced. ld_moved simply stays zero, so it is
6467 * correctly treated as an imbalance.
6469 env.flags |= LBF_ALL_PINNED;
6470 env.src_cpu = busiest->cpu;
6471 env.src_rq = busiest;
6472 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
6475 local_irq_save(flags);
6476 double_rq_lock(env.dst_rq, busiest);
6479 * cur_ld_moved - load moved in current iteration
6480 * ld_moved - cumulative load moved across iterations
6482 cur_ld_moved = move_tasks(&env);
6483 ld_moved += cur_ld_moved;
6484 double_rq_unlock(env.dst_rq, busiest);
6485 local_irq_restore(flags);
6488 * some other cpu did the load balance for us.
6490 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
6491 resched_cpu(env.dst_cpu);
6493 if (env.flags & LBF_NEED_BREAK) {
6494 env.flags &= ~LBF_NEED_BREAK;
6499 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6500 * us and move them to an alternate dst_cpu in our sched_group
6501 * where they can run. The upper limit on how many times we
6502 * iterate on same src_cpu is dependent on number of cpus in our
6505 * This changes load balance semantics a bit on who can move
6506 * load to a given_cpu. In addition to the given_cpu itself
6507 * (or a ilb_cpu acting on its behalf where given_cpu is
6508 * nohz-idle), we now have balance_cpu in a position to move
6509 * load to given_cpu. In rare situations, this may cause
6510 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6511 * _independently_ and at _same_ time to move some load to
6512 * given_cpu) causing exceess load to be moved to given_cpu.
6513 * This however should not happen so much in practice and
6514 * moreover subsequent load balance cycles should correct the
6515 * excess load moved.
6517 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
6519 /* Prevent to re-select dst_cpu via env's cpus */
6520 cpumask_clear_cpu(env.dst_cpu, env.cpus);
6522 env.dst_rq = cpu_rq(env.new_dst_cpu);
6523 env.dst_cpu = env.new_dst_cpu;
6524 env.flags &= ~LBF_DST_PINNED;
6526 env.loop_break = sched_nr_migrate_break;
6529 * Go back to "more_balance" rather than "redo" since we
6530 * need to continue with same src_cpu.
6536 * We failed to reach balance because of affinity.
6539 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
6541 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
6542 *group_imbalance = 1;
6543 } else if (*group_imbalance)
6544 *group_imbalance = 0;
6547 /* All tasks on this runqueue were pinned by CPU affinity */
6548 if (unlikely(env.flags & LBF_ALL_PINNED)) {
6549 cpumask_clear_cpu(cpu_of(busiest), cpus);
6550 if (!cpumask_empty(cpus)) {
6552 env.loop_break = sched_nr_migrate_break;
6560 schedstat_inc(sd, lb_failed[idle]);
6562 * Increment the failure counter only on periodic balance.
6563 * We do not want newidle balance, which can be very
6564 * frequent, pollute the failure counter causing
6565 * excessive cache_hot migrations and active balances.
6567 if (idle != CPU_NEWLY_IDLE)
6568 sd->nr_balance_failed++;
6570 if (need_active_balance(&env)) {
6571 raw_spin_lock_irqsave(&busiest->lock, flags);
6573 /* don't kick the active_load_balance_cpu_stop,
6574 * if the curr task on busiest cpu can't be
6577 if (!cpumask_test_cpu(this_cpu,
6578 tsk_cpus_allowed(busiest->curr))) {
6579 raw_spin_unlock_irqrestore(&busiest->lock,
6581 env.flags |= LBF_ALL_PINNED;
6582 goto out_one_pinned;
6586 * ->active_balance synchronizes accesses to
6587 * ->active_balance_work. Once set, it's cleared
6588 * only after active load balance is finished.
6590 if (!busiest->active_balance) {
6591 busiest->active_balance = 1;
6592 busiest->push_cpu = this_cpu;
6595 raw_spin_unlock_irqrestore(&busiest->lock, flags);
6597 if (active_balance) {
6598 stop_one_cpu_nowait(cpu_of(busiest),
6599 active_load_balance_cpu_stop, busiest,
6600 &busiest->active_balance_work);
6604 * We've kicked active balancing, reset the failure
6607 sd->nr_balance_failed = sd->cache_nice_tries+1;
6610 sd->nr_balance_failed = 0;
6612 if (likely(!active_balance)) {
6613 /* We were unbalanced, so reset the balancing interval */
6614 sd->balance_interval = sd->min_interval;
6617 * If we've begun active balancing, start to back off. This
6618 * case may not be covered by the all_pinned logic if there
6619 * is only 1 task on the busy runqueue (because we don't call
6622 if (sd->balance_interval < sd->max_interval)
6623 sd->balance_interval *= 2;
6629 schedstat_inc(sd, lb_balanced[idle]);
6631 sd->nr_balance_failed = 0;
6634 /* tune up the balancing interval */
6635 if (((env.flags & LBF_ALL_PINNED) &&
6636 sd->balance_interval < MAX_PINNED_INTERVAL) ||
6637 (sd->balance_interval < sd->max_interval))
6638 sd->balance_interval *= 2;
6646 * idle_balance is called by schedule() if this_cpu is about to become
6647 * idle. Attempts to pull tasks from other CPUs.
6649 static int idle_balance(struct rq *this_rq)
6651 struct sched_domain *sd;
6652 int pulled_task = 0;
6653 unsigned long next_balance = jiffies + HZ;
6655 int this_cpu = this_rq->cpu;
6657 idle_enter_fair(this_rq);
6659 * We must set idle_stamp _before_ calling idle_balance(), such that we
6660 * measure the duration of idle_balance() as idle time.
6662 this_rq->idle_stamp = rq_clock(this_rq);
6664 if (this_rq->avg_idle < sysctl_sched_migration_cost)
6668 * Drop the rq->lock, but keep IRQ/preempt disabled.
6670 raw_spin_unlock(&this_rq->lock);
6672 update_blocked_averages(this_cpu);
6674 for_each_domain(this_cpu, sd) {
6675 unsigned long interval;
6676 int continue_balancing = 1;
6677 u64 t0, domain_cost;
6679 if (!(sd->flags & SD_LOAD_BALANCE))
6682 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
6685 if (sd->flags & SD_BALANCE_NEWIDLE) {
6686 t0 = sched_clock_cpu(this_cpu);
6688 /* If we've pulled tasks over stop searching: */
6689 pulled_task = load_balance(this_cpu, this_rq,
6691 &continue_balancing);
6693 domain_cost = sched_clock_cpu(this_cpu) - t0;
6694 if (domain_cost > sd->max_newidle_lb_cost)
6695 sd->max_newidle_lb_cost = domain_cost;
6697 curr_cost += domain_cost;
6700 interval = msecs_to_jiffies(sd->balance_interval);
6701 if (time_after(next_balance, sd->last_balance + interval))
6702 next_balance = sd->last_balance + interval;
6708 raw_spin_lock(&this_rq->lock);
6711 * While browsing the domains, we released the rq lock.
6712 * A task could have be enqueued in the meantime
6714 if (this_rq->nr_running && !pulled_task) {
6719 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
6721 * We are going idle. next_balance may be set based on
6722 * a busy processor. So reset next_balance.
6724 this_rq->next_balance = next_balance;
6727 if (curr_cost > this_rq->max_idle_balance_cost)
6728 this_rq->max_idle_balance_cost = curr_cost;
6732 this_rq->idle_stamp = 0;
6738 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6739 * running tasks off the busiest CPU onto idle CPUs. It requires at
6740 * least 1 task to be running on each physical CPU where possible, and
6741 * avoids physical / logical imbalances.
6743 static int active_load_balance_cpu_stop(void *data)
6745 struct rq *busiest_rq = data;
6746 int busiest_cpu = cpu_of(busiest_rq);
6747 int target_cpu = busiest_rq->push_cpu;
6748 struct rq *target_rq = cpu_rq(target_cpu);
6749 struct sched_domain *sd;
6751 raw_spin_lock_irq(&busiest_rq->lock);
6753 /* make sure the requested cpu hasn't gone down in the meantime */
6754 if (unlikely(busiest_cpu != smp_processor_id() ||
6755 !busiest_rq->active_balance))
6758 /* Is there any task to move? */
6759 if (busiest_rq->nr_running <= 1)
6763 * This condition is "impossible", if it occurs
6764 * we need to fix it. Originally reported by
6765 * Bjorn Helgaas on a 128-cpu setup.
6767 BUG_ON(busiest_rq == target_rq);
6769 /* move a task from busiest_rq to target_rq */
6770 double_lock_balance(busiest_rq, target_rq);
6772 /* Search for an sd spanning us and the target CPU. */
6774 for_each_domain(target_cpu, sd) {
6775 if ((sd->flags & SD_LOAD_BALANCE) &&
6776 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6781 struct lb_env env = {
6783 .dst_cpu = target_cpu,
6784 .dst_rq = target_rq,
6785 .src_cpu = busiest_rq->cpu,
6786 .src_rq = busiest_rq,
6790 schedstat_inc(sd, alb_count);
6792 if (move_one_task(&env))
6793 schedstat_inc(sd, alb_pushed);
6795 schedstat_inc(sd, alb_failed);
6798 double_unlock_balance(busiest_rq, target_rq);
6800 busiest_rq->active_balance = 0;
6801 raw_spin_unlock_irq(&busiest_rq->lock);
6805 static inline int on_null_domain(struct rq *rq)
6807 return unlikely(!rcu_dereference_sched(rq->sd));
6810 #ifdef CONFIG_NO_HZ_COMMON
6812 * idle load balancing details
6813 * - When one of the busy CPUs notice that there may be an idle rebalancing
6814 * needed, they will kick the idle load balancer, which then does idle
6815 * load balancing for all the idle CPUs.
6818 cpumask_var_t idle_cpus_mask;
6820 unsigned long next_balance; /* in jiffy units */
6821 } nohz ____cacheline_aligned;
6823 static inline int find_new_ilb(void)
6825 int ilb = cpumask_first(nohz.idle_cpus_mask);
6827 if (ilb < nr_cpu_ids && idle_cpu(ilb))
6834 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6835 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6836 * CPU (if there is one).
6838 static void nohz_balancer_kick(void)
6842 nohz.next_balance++;
6844 ilb_cpu = find_new_ilb();
6846 if (ilb_cpu >= nr_cpu_ids)
6849 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
6852 * Use smp_send_reschedule() instead of resched_cpu().
6853 * This way we generate a sched IPI on the target cpu which
6854 * is idle. And the softirq performing nohz idle load balance
6855 * will be run before returning from the IPI.
6857 smp_send_reschedule(ilb_cpu);
6861 static inline void nohz_balance_exit_idle(int cpu)
6863 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
6865 * Completely isolated CPUs don't ever set, so we must test.
6867 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
6868 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
6869 atomic_dec(&nohz.nr_cpus);
6871 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6875 static inline void set_cpu_sd_state_busy(void)
6877 struct sched_domain *sd;
6878 int cpu = smp_processor_id();
6881 sd = rcu_dereference(per_cpu(sd_busy, cpu));
6883 if (!sd || !sd->nohz_idle)
6887 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
6892 void set_cpu_sd_state_idle(void)
6894 struct sched_domain *sd;
6895 int cpu = smp_processor_id();
6898 sd = rcu_dereference(per_cpu(sd_busy, cpu));
6900 if (!sd || sd->nohz_idle)
6904 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
6910 * This routine will record that the cpu is going idle with tick stopped.
6911 * This info will be used in performing idle load balancing in the future.
6913 void nohz_balance_enter_idle(int cpu)
6916 * If this cpu is going down, then nothing needs to be done.
6918 if (!cpu_active(cpu))
6921 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
6925 * If we're a completely isolated CPU, we don't play.
6927 if (on_null_domain(cpu_rq(cpu)))
6930 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
6931 atomic_inc(&nohz.nr_cpus);
6932 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6935 static int sched_ilb_notifier(struct notifier_block *nfb,
6936 unsigned long action, void *hcpu)
6938 switch (action & ~CPU_TASKS_FROZEN) {
6940 nohz_balance_exit_idle(smp_processor_id());
6948 static DEFINE_SPINLOCK(balancing);
6951 * Scale the max load_balance interval with the number of CPUs in the system.
6952 * This trades load-balance latency on larger machines for less cross talk.
6954 void update_max_interval(void)
6956 max_load_balance_interval = HZ*num_online_cpus()/10;
6960 * It checks each scheduling domain to see if it is due to be balanced,
6961 * and initiates a balancing operation if so.
6963 * Balancing parameters are set up in init_sched_domains.
6965 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
6967 int continue_balancing = 1;
6969 unsigned long interval;
6970 struct sched_domain *sd;
6971 /* Earliest time when we have to do rebalance again */
6972 unsigned long next_balance = jiffies + 60*HZ;
6973 int update_next_balance = 0;
6974 int need_serialize, need_decay = 0;
6977 update_blocked_averages(cpu);
6980 for_each_domain(cpu, sd) {
6982 * Decay the newidle max times here because this is a regular
6983 * visit to all the domains. Decay ~1% per second.
6985 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
6986 sd->max_newidle_lb_cost =
6987 (sd->max_newidle_lb_cost * 253) / 256;
6988 sd->next_decay_max_lb_cost = jiffies + HZ;
6991 max_cost += sd->max_newidle_lb_cost;
6993 if (!(sd->flags & SD_LOAD_BALANCE))
6997 * Stop the load balance at this level. There is another
6998 * CPU in our sched group which is doing load balancing more
7001 if (!continue_balancing) {
7007 interval = sd->balance_interval;
7008 if (idle != CPU_IDLE)
7009 interval *= sd->busy_factor;
7011 /* scale ms to jiffies */
7012 interval = msecs_to_jiffies(interval);
7013 interval = clamp(interval, 1UL, max_load_balance_interval);
7015 need_serialize = sd->flags & SD_SERIALIZE;
7017 if (need_serialize) {
7018 if (!spin_trylock(&balancing))
7022 if (time_after_eq(jiffies, sd->last_balance + interval)) {
7023 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7025 * The LBF_DST_PINNED logic could have changed
7026 * env->dst_cpu, so we can't know our idle
7027 * state even if we migrated tasks. Update it.
7029 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7031 sd->last_balance = jiffies;
7034 spin_unlock(&balancing);
7036 if (time_after(next_balance, sd->last_balance + interval)) {
7037 next_balance = sd->last_balance + interval;
7038 update_next_balance = 1;
7043 * Ensure the rq-wide value also decays but keep it at a
7044 * reasonable floor to avoid funnies with rq->avg_idle.
7046 rq->max_idle_balance_cost =
7047 max((u64)sysctl_sched_migration_cost, max_cost);
7052 * next_balance will be updated only when there is a need.
7053 * When the cpu is attached to null domain for ex, it will not be
7056 if (likely(update_next_balance))
7057 rq->next_balance = next_balance;
7060 #ifdef CONFIG_NO_HZ_COMMON
7062 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7063 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7065 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7067 int this_cpu = this_rq->cpu;
7071 if (idle != CPU_IDLE ||
7072 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7075 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7076 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7080 * If this cpu gets work to do, stop the load balancing
7081 * work being done for other cpus. Next load
7082 * balancing owner will pick it up.
7087 rq = cpu_rq(balance_cpu);
7089 raw_spin_lock_irq(&rq->lock);
7090 update_rq_clock(rq);
7091 update_idle_cpu_load(rq);
7092 raw_spin_unlock_irq(&rq->lock);
7094 rebalance_domains(rq, CPU_IDLE);
7096 if (time_after(this_rq->next_balance, rq->next_balance))
7097 this_rq->next_balance = rq->next_balance;
7099 nohz.next_balance = this_rq->next_balance;
7101 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7105 * Current heuristic for kicking the idle load balancer in the presence
7106 * of an idle cpu is the system.
7107 * - This rq has more than one task.
7108 * - At any scheduler domain level, this cpu's scheduler group has multiple
7109 * busy cpu's exceeding the group's power.
7110 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7111 * domain span are idle.
7113 static inline int nohz_kick_needed(struct rq *rq)
7115 unsigned long now = jiffies;
7116 struct sched_domain *sd;
7117 struct sched_group_power *sgp;
7118 int nr_busy, cpu = rq->cpu;
7120 if (unlikely(rq->idle_balance))
7124 * We may be recently in ticked or tickless idle mode. At the first
7125 * busy tick after returning from idle, we will update the busy stats.
7127 set_cpu_sd_state_busy();
7128 nohz_balance_exit_idle(cpu);
7131 * None are in tickless mode and hence no need for NOHZ idle load
7134 if (likely(!atomic_read(&nohz.nr_cpus)))
7137 if (time_before(now, nohz.next_balance))
7140 if (rq->nr_running >= 2)
7144 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7147 sgp = sd->groups->sgp;
7148 nr_busy = atomic_read(&sgp->nr_busy_cpus);
7151 goto need_kick_unlock;
7154 sd = rcu_dereference(per_cpu(sd_asym, cpu));
7156 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7157 sched_domain_span(sd)) < cpu))
7158 goto need_kick_unlock;
7169 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7173 * run_rebalance_domains is triggered when needed from the scheduler tick.
7174 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7176 static void run_rebalance_domains(struct softirq_action *h)
7178 struct rq *this_rq = this_rq();
7179 enum cpu_idle_type idle = this_rq->idle_balance ?
7180 CPU_IDLE : CPU_NOT_IDLE;
7182 rebalance_domains(this_rq, idle);
7185 * If this cpu has a pending nohz_balance_kick, then do the
7186 * balancing on behalf of the other idle cpus whose ticks are
7189 nohz_idle_balance(this_rq, idle);
7193 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7195 void trigger_load_balance(struct rq *rq)
7197 /* Don't need to rebalance while attached to NULL domain */
7198 if (unlikely(on_null_domain(rq)))
7201 if (time_after_eq(jiffies, rq->next_balance))
7202 raise_softirq(SCHED_SOFTIRQ);
7203 #ifdef CONFIG_NO_HZ_COMMON
7204 if (nohz_kick_needed(rq))
7205 nohz_balancer_kick();
7209 static void rq_online_fair(struct rq *rq)
7214 static void rq_offline_fair(struct rq *rq)
7218 /* Ensure any throttled groups are reachable by pick_next_task */
7219 unthrottle_offline_cfs_rqs(rq);
7222 #endif /* CONFIG_SMP */
7225 * scheduler tick hitting a task of our scheduling class:
7227 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7229 struct cfs_rq *cfs_rq;
7230 struct sched_entity *se = &curr->se;
7232 for_each_sched_entity(se) {
7233 cfs_rq = cfs_rq_of(se);
7234 entity_tick(cfs_rq, se, queued);
7237 if (numabalancing_enabled)
7238 task_tick_numa(rq, curr);
7240 update_rq_runnable_avg(rq, 1);
7244 * called on fork with the child task as argument from the parent's context
7245 * - child not yet on the tasklist
7246 * - preemption disabled
7248 static void task_fork_fair(struct task_struct *p)
7250 struct cfs_rq *cfs_rq;
7251 struct sched_entity *se = &p->se, *curr;
7252 int this_cpu = smp_processor_id();
7253 struct rq *rq = this_rq();
7254 unsigned long flags;
7256 raw_spin_lock_irqsave(&rq->lock, flags);
7258 update_rq_clock(rq);
7260 cfs_rq = task_cfs_rq(current);
7261 curr = cfs_rq->curr;
7264 * Not only the cpu but also the task_group of the parent might have
7265 * been changed after parent->se.parent,cfs_rq were copied to
7266 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7267 * of child point to valid ones.
7270 __set_task_cpu(p, this_cpu);
7273 update_curr(cfs_rq);
7276 se->vruntime = curr->vruntime;
7277 place_entity(cfs_rq, se, 1);
7279 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
7281 * Upon rescheduling, sched_class::put_prev_task() will place
7282 * 'current' within the tree based on its new key value.
7284 swap(curr->vruntime, se->vruntime);
7285 resched_task(rq->curr);
7288 se->vruntime -= cfs_rq->min_vruntime;
7290 raw_spin_unlock_irqrestore(&rq->lock, flags);
7294 * Priority of the task has changed. Check to see if we preempt
7298 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
7304 * Reschedule if we are currently running on this runqueue and
7305 * our priority decreased, or if we are not currently running on
7306 * this runqueue and our priority is higher than the current's
7308 if (rq->curr == p) {
7309 if (p->prio > oldprio)
7310 resched_task(rq->curr);
7312 check_preempt_curr(rq, p, 0);
7315 static void switched_from_fair(struct rq *rq, struct task_struct *p)
7317 struct sched_entity *se = &p->se;
7318 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7321 * Ensure the task's vruntime is normalized, so that when it's
7322 * switched back to the fair class the enqueue_entity(.flags=0) will
7323 * do the right thing.
7325 * If it's on_rq, then the dequeue_entity(.flags=0) will already
7326 * have normalized the vruntime, if it's !on_rq, then only when
7327 * the task is sleeping will it still have non-normalized vruntime.
7329 if (!p->on_rq && p->state != TASK_RUNNING) {
7331 * Fix up our vruntime so that the current sleep doesn't
7332 * cause 'unlimited' sleep bonus.
7334 place_entity(cfs_rq, se, 0);
7335 se->vruntime -= cfs_rq->min_vruntime;
7340 * Remove our load from contribution when we leave sched_fair
7341 * and ensure we don't carry in an old decay_count if we
7344 if (se->avg.decay_count) {
7345 __synchronize_entity_decay(se);
7346 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
7352 * We switched to the sched_fair class.
7354 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7356 struct sched_entity *se = &p->se;
7357 #ifdef CONFIG_FAIR_GROUP_SCHED
7359 * Since the real-depth could have been changed (only FAIR
7360 * class maintain depth value), reset depth properly.
7362 se->depth = se->parent ? se->parent->depth + 1 : 0;
7368 * We were most likely switched from sched_rt, so
7369 * kick off the schedule if running, otherwise just see
7370 * if we can still preempt the current task.
7373 resched_task(rq->curr);
7375 check_preempt_curr(rq, p, 0);
7378 /* Account for a task changing its policy or group.
7380 * This routine is mostly called to set cfs_rq->curr field when a task
7381 * migrates between groups/classes.
7383 static void set_curr_task_fair(struct rq *rq)
7385 struct sched_entity *se = &rq->curr->se;
7387 for_each_sched_entity(se) {
7388 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7390 set_next_entity(cfs_rq, se);
7391 /* ensure bandwidth has been allocated on our new cfs_rq */
7392 account_cfs_rq_runtime(cfs_rq, 0);
7396 void init_cfs_rq(struct cfs_rq *cfs_rq)
7398 cfs_rq->tasks_timeline = RB_ROOT;
7399 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7400 #ifndef CONFIG_64BIT
7401 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7404 atomic64_set(&cfs_rq->decay_counter, 1);
7405 atomic_long_set(&cfs_rq->removed_load, 0);
7409 #ifdef CONFIG_FAIR_GROUP_SCHED
7410 static void task_move_group_fair(struct task_struct *p, int on_rq)
7412 struct sched_entity *se = &p->se;
7413 struct cfs_rq *cfs_rq;
7416 * If the task was not on the rq at the time of this cgroup movement
7417 * it must have been asleep, sleeping tasks keep their ->vruntime
7418 * absolute on their old rq until wakeup (needed for the fair sleeper
7419 * bonus in place_entity()).
7421 * If it was on the rq, we've just 'preempted' it, which does convert
7422 * ->vruntime to a relative base.
7424 * Make sure both cases convert their relative position when migrating
7425 * to another cgroup's rq. This does somewhat interfere with the
7426 * fair sleeper stuff for the first placement, but who cares.
7429 * When !on_rq, vruntime of the task has usually NOT been normalized.
7430 * But there are some cases where it has already been normalized:
7432 * - Moving a forked child which is waiting for being woken up by
7433 * wake_up_new_task().
7434 * - Moving a task which has been woken up by try_to_wake_up() and
7435 * waiting for actually being woken up by sched_ttwu_pending().
7437 * To prevent boost or penalty in the new cfs_rq caused by delta
7438 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7440 if (!on_rq && (!se->sum_exec_runtime || p->state == TASK_WAKING))
7444 se->vruntime -= cfs_rq_of(se)->min_vruntime;
7445 set_task_rq(p, task_cpu(p));
7446 se->depth = se->parent ? se->parent->depth + 1 : 0;
7448 cfs_rq = cfs_rq_of(se);
7449 se->vruntime += cfs_rq->min_vruntime;
7452 * migrate_task_rq_fair() will have removed our previous
7453 * contribution, but we must synchronize for ongoing future
7456 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7457 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
7462 void free_fair_sched_group(struct task_group *tg)
7466 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7468 for_each_possible_cpu(i) {
7470 kfree(tg->cfs_rq[i]);
7479 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7481 struct cfs_rq *cfs_rq;
7482 struct sched_entity *se;
7485 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7488 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7492 tg->shares = NICE_0_LOAD;
7494 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7496 for_each_possible_cpu(i) {
7497 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7498 GFP_KERNEL, cpu_to_node(i));
7502 se = kzalloc_node(sizeof(struct sched_entity),
7503 GFP_KERNEL, cpu_to_node(i));
7507 init_cfs_rq(cfs_rq);
7508 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7519 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7521 struct rq *rq = cpu_rq(cpu);
7522 unsigned long flags;
7525 * Only empty task groups can be destroyed; so we can speculatively
7526 * check on_list without danger of it being re-added.
7528 if (!tg->cfs_rq[cpu]->on_list)
7531 raw_spin_lock_irqsave(&rq->lock, flags);
7532 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7533 raw_spin_unlock_irqrestore(&rq->lock, flags);
7536 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7537 struct sched_entity *se, int cpu,
7538 struct sched_entity *parent)
7540 struct rq *rq = cpu_rq(cpu);
7544 init_cfs_rq_runtime(cfs_rq);
7546 tg->cfs_rq[cpu] = cfs_rq;
7549 /* se could be NULL for root_task_group */
7554 se->cfs_rq = &rq->cfs;
7557 se->cfs_rq = parent->my_q;
7558 se->depth = parent->depth + 1;
7562 /* guarantee group entities always have weight */
7563 update_load_set(&se->load, NICE_0_LOAD);
7564 se->parent = parent;
7567 static DEFINE_MUTEX(shares_mutex);
7569 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7572 unsigned long flags;
7575 * We can't change the weight of the root cgroup.
7580 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7582 mutex_lock(&shares_mutex);
7583 if (tg->shares == shares)
7586 tg->shares = shares;
7587 for_each_possible_cpu(i) {
7588 struct rq *rq = cpu_rq(i);
7589 struct sched_entity *se;
7592 /* Propagate contribution to hierarchy */
7593 raw_spin_lock_irqsave(&rq->lock, flags);
7595 /* Possible calls to update_curr() need rq clock */
7596 update_rq_clock(rq);
7597 for_each_sched_entity(se)
7598 update_cfs_shares(group_cfs_rq(se));
7599 raw_spin_unlock_irqrestore(&rq->lock, flags);
7603 mutex_unlock(&shares_mutex);
7606 #else /* CONFIG_FAIR_GROUP_SCHED */
7608 void free_fair_sched_group(struct task_group *tg) { }
7610 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7615 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7617 #endif /* CONFIG_FAIR_GROUP_SCHED */
7620 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
7622 struct sched_entity *se = &task->se;
7623 unsigned int rr_interval = 0;
7626 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7629 if (rq->cfs.load.weight)
7630 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
7636 * All the scheduling class methods:
7638 const struct sched_class fair_sched_class = {
7639 .next = &idle_sched_class,
7640 .enqueue_task = enqueue_task_fair,
7641 .dequeue_task = dequeue_task_fair,
7642 .yield_task = yield_task_fair,
7643 .yield_to_task = yield_to_task_fair,
7645 .check_preempt_curr = check_preempt_wakeup,
7647 .pick_next_task = pick_next_task_fair,
7648 .put_prev_task = put_prev_task_fair,
7651 .select_task_rq = select_task_rq_fair,
7652 .migrate_task_rq = migrate_task_rq_fair,
7654 .rq_online = rq_online_fair,
7655 .rq_offline = rq_offline_fair,
7657 .task_waking = task_waking_fair,
7660 .set_curr_task = set_curr_task_fair,
7661 .task_tick = task_tick_fair,
7662 .task_fork = task_fork_fair,
7664 .prio_changed = prio_changed_fair,
7665 .switched_from = switched_from_fair,
7666 .switched_to = switched_to_fair,
7668 .get_rr_interval = get_rr_interval_fair,
7670 #ifdef CONFIG_FAIR_GROUP_SCHED
7671 .task_move_group = task_move_group_fair,
7675 #ifdef CONFIG_SCHED_DEBUG
7676 void print_cfs_stats(struct seq_file *m, int cpu)
7678 struct cfs_rq *cfs_rq;
7681 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7682 print_cfs_rq(m, cpu, cfs_rq);
7687 __init void init_sched_fair_class(void)
7690 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7692 #ifdef CONFIG_NO_HZ_COMMON
7693 nohz.next_balance = jiffies;
7694 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7695 cpu_notifier(sched_ilb_notifier, 0);