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)
1305 * If the task is part of a workload that spans multiple NUMA nodes,
1306 * and is migrating into one of the workload's active nodes, remember
1307 * this node as the task's preferred numa node, so the workload can
1309 * A task that migrated to a second choice node will be better off
1310 * trying for a better one later. Do not set the preferred node here.
1312 if (p->numa_group && node_isset(env.dst_nid, p->numa_group->active_nodes))
1313 sched_setnuma(p, env.dst_nid);
1316 * Reset the scan period if the task is being rescheduled on an
1317 * alternative node to recheck if the tasks is now properly placed.
1319 p->numa_scan_period = task_scan_min(p);
1321 if (env.best_task == NULL) {
1322 ret = migrate_task_to(p, env.best_cpu);
1324 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1328 ret = migrate_swap(p, env.best_task);
1330 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1331 put_task_struct(env.best_task);
1335 /* Attempt to migrate a task to a CPU on the preferred node. */
1336 static void numa_migrate_preferred(struct task_struct *p)
1338 unsigned long interval = HZ;
1340 /* This task has no NUMA fault statistics yet */
1341 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults_memory))
1344 /* Periodically retry migrating the task to the preferred node */
1345 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1346 p->numa_migrate_retry = jiffies + interval;
1348 /* Success if task is already running on preferred CPU */
1349 if (task_node(p) == p->numa_preferred_nid)
1352 /* Otherwise, try migrate to a CPU on the preferred node */
1353 task_numa_migrate(p);
1357 * Find the nodes on which the workload is actively running. We do this by
1358 * tracking the nodes from which NUMA hinting faults are triggered. This can
1359 * be different from the set of nodes where the workload's memory is currently
1362 * The bitmask is used to make smarter decisions on when to do NUMA page
1363 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1364 * are added when they cause over 6/16 of the maximum number of faults, but
1365 * only removed when they drop below 3/16.
1367 static void update_numa_active_node_mask(struct numa_group *numa_group)
1369 unsigned long faults, max_faults = 0;
1372 for_each_online_node(nid) {
1373 faults = group_faults_cpu(numa_group, nid);
1374 if (faults > max_faults)
1375 max_faults = faults;
1378 for_each_online_node(nid) {
1379 faults = group_faults_cpu(numa_group, nid);
1380 if (!node_isset(nid, numa_group->active_nodes)) {
1381 if (faults > max_faults * 6 / 16)
1382 node_set(nid, numa_group->active_nodes);
1383 } else if (faults < max_faults * 3 / 16)
1384 node_clear(nid, numa_group->active_nodes);
1389 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1390 * increments. The more local the fault statistics are, the higher the scan
1391 * period will be for the next scan window. If local/remote ratio is below
1392 * NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) the
1393 * scan period will decrease
1395 #define NUMA_PERIOD_SLOTS 10
1396 #define NUMA_PERIOD_THRESHOLD 3
1399 * Increase the scan period (slow down scanning) if the majority of
1400 * our memory is already on our local node, or if the majority of
1401 * the page accesses are shared with other processes.
1402 * Otherwise, decrease the scan period.
1404 static void update_task_scan_period(struct task_struct *p,
1405 unsigned long shared, unsigned long private)
1407 unsigned int period_slot;
1411 unsigned long remote = p->numa_faults_locality[0];
1412 unsigned long local = p->numa_faults_locality[1];
1415 * If there were no record hinting faults then either the task is
1416 * completely idle or all activity is areas that are not of interest
1417 * to automatic numa balancing. Scan slower
1419 if (local + shared == 0) {
1420 p->numa_scan_period = min(p->numa_scan_period_max,
1421 p->numa_scan_period << 1);
1423 p->mm->numa_next_scan = jiffies +
1424 msecs_to_jiffies(p->numa_scan_period);
1430 * Prepare to scale scan period relative to the current period.
1431 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1432 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1433 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1435 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1436 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1437 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1438 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1441 diff = slot * period_slot;
1443 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1446 * Scale scan rate increases based on sharing. There is an
1447 * inverse relationship between the degree of sharing and
1448 * the adjustment made to the scanning period. Broadly
1449 * speaking the intent is that there is little point
1450 * scanning faster if shared accesses dominate as it may
1451 * simply bounce migrations uselessly
1453 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared));
1454 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1457 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1458 task_scan_min(p), task_scan_max(p));
1459 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1463 * Get the fraction of time the task has been running since the last
1464 * NUMA placement cycle. The scheduler keeps similar statistics, but
1465 * decays those on a 32ms period, which is orders of magnitude off
1466 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1467 * stats only if the task is so new there are no NUMA statistics yet.
1469 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1471 u64 runtime, delta, now;
1472 /* Use the start of this time slice to avoid calculations. */
1473 now = p->se.exec_start;
1474 runtime = p->se.sum_exec_runtime;
1476 if (p->last_task_numa_placement) {
1477 delta = runtime - p->last_sum_exec_runtime;
1478 *period = now - p->last_task_numa_placement;
1480 delta = p->se.avg.runnable_avg_sum;
1481 *period = p->se.avg.runnable_avg_period;
1484 p->last_sum_exec_runtime = runtime;
1485 p->last_task_numa_placement = now;
1490 static void task_numa_placement(struct task_struct *p)
1492 int seq, nid, max_nid = -1, max_group_nid = -1;
1493 unsigned long max_faults = 0, max_group_faults = 0;
1494 unsigned long fault_types[2] = { 0, 0 };
1495 unsigned long total_faults;
1496 u64 runtime, period;
1497 spinlock_t *group_lock = NULL;
1499 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1500 if (p->numa_scan_seq == seq)
1502 p->numa_scan_seq = seq;
1503 p->numa_scan_period_max = task_scan_max(p);
1505 total_faults = p->numa_faults_locality[0] +
1506 p->numa_faults_locality[1];
1507 runtime = numa_get_avg_runtime(p, &period);
1509 /* If the task is part of a group prevent parallel updates to group stats */
1510 if (p->numa_group) {
1511 group_lock = &p->numa_group->lock;
1512 spin_lock_irq(group_lock);
1515 /* Find the node with the highest number of faults */
1516 for_each_online_node(nid) {
1517 unsigned long faults = 0, group_faults = 0;
1520 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1521 long diff, f_diff, f_weight;
1523 i = task_faults_idx(nid, priv);
1525 /* Decay existing window, copy faults since last scan */
1526 diff = p->numa_faults_buffer_memory[i] - p->numa_faults_memory[i] / 2;
1527 fault_types[priv] += p->numa_faults_buffer_memory[i];
1528 p->numa_faults_buffer_memory[i] = 0;
1531 * Normalize the faults_from, so all tasks in a group
1532 * count according to CPU use, instead of by the raw
1533 * number of faults. Tasks with little runtime have
1534 * little over-all impact on throughput, and thus their
1535 * faults are less important.
1537 f_weight = div64_u64(runtime << 16, period + 1);
1538 f_weight = (f_weight * p->numa_faults_buffer_cpu[i]) /
1540 f_diff = f_weight - p->numa_faults_cpu[i] / 2;
1541 p->numa_faults_buffer_cpu[i] = 0;
1543 p->numa_faults_memory[i] += diff;
1544 p->numa_faults_cpu[i] += f_diff;
1545 faults += p->numa_faults_memory[i];
1546 p->total_numa_faults += diff;
1547 if (p->numa_group) {
1548 /* safe because we can only change our own group */
1549 p->numa_group->faults[i] += diff;
1550 p->numa_group->faults_cpu[i] += f_diff;
1551 p->numa_group->total_faults += diff;
1552 group_faults += p->numa_group->faults[i];
1556 if (faults > max_faults) {
1557 max_faults = faults;
1561 if (group_faults > max_group_faults) {
1562 max_group_faults = group_faults;
1563 max_group_nid = nid;
1567 update_task_scan_period(p, fault_types[0], fault_types[1]);
1569 if (p->numa_group) {
1570 update_numa_active_node_mask(p->numa_group);
1572 * If the preferred task and group nids are different,
1573 * iterate over the nodes again to find the best place.
1575 if (max_nid != max_group_nid) {
1576 unsigned long weight, max_weight = 0;
1578 for_each_online_node(nid) {
1579 weight = task_weight(p, nid) + group_weight(p, nid);
1580 if (weight > max_weight) {
1581 max_weight = weight;
1587 spin_unlock_irq(group_lock);
1590 /* Preferred node as the node with the most faults */
1591 if (max_faults && max_nid != p->numa_preferred_nid) {
1592 /* Update the preferred nid and migrate task if possible */
1593 sched_setnuma(p, max_nid);
1594 numa_migrate_preferred(p);
1598 static inline int get_numa_group(struct numa_group *grp)
1600 return atomic_inc_not_zero(&grp->refcount);
1603 static inline void put_numa_group(struct numa_group *grp)
1605 if (atomic_dec_and_test(&grp->refcount))
1606 kfree_rcu(grp, rcu);
1609 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1612 struct numa_group *grp, *my_grp;
1613 struct task_struct *tsk;
1615 int cpu = cpupid_to_cpu(cpupid);
1618 if (unlikely(!p->numa_group)) {
1619 unsigned int size = sizeof(struct numa_group) +
1620 4*nr_node_ids*sizeof(unsigned long);
1622 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1626 atomic_set(&grp->refcount, 1);
1627 spin_lock_init(&grp->lock);
1628 INIT_LIST_HEAD(&grp->task_list);
1630 /* Second half of the array tracks nids where faults happen */
1631 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1634 node_set(task_node(current), grp->active_nodes);
1636 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1637 grp->faults[i] = p->numa_faults_memory[i];
1639 grp->total_faults = p->total_numa_faults;
1641 list_add(&p->numa_entry, &grp->task_list);
1643 rcu_assign_pointer(p->numa_group, grp);
1647 tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1649 if (!cpupid_match_pid(tsk, cpupid))
1652 grp = rcu_dereference(tsk->numa_group);
1656 my_grp = p->numa_group;
1661 * Only join the other group if its bigger; if we're the bigger group,
1662 * the other task will join us.
1664 if (my_grp->nr_tasks > grp->nr_tasks)
1668 * Tie-break on the grp address.
1670 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1673 /* Always join threads in the same process. */
1674 if (tsk->mm == current->mm)
1677 /* Simple filter to avoid false positives due to PID collisions */
1678 if (flags & TNF_SHARED)
1681 /* Update priv based on whether false sharing was detected */
1684 if (join && !get_numa_group(grp))
1692 BUG_ON(irqs_disabled());
1693 double_lock_irq(&my_grp->lock, &grp->lock);
1695 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
1696 my_grp->faults[i] -= p->numa_faults_memory[i];
1697 grp->faults[i] += p->numa_faults_memory[i];
1699 my_grp->total_faults -= p->total_numa_faults;
1700 grp->total_faults += p->total_numa_faults;
1702 list_move(&p->numa_entry, &grp->task_list);
1706 spin_unlock(&my_grp->lock);
1707 spin_unlock_irq(&grp->lock);
1709 rcu_assign_pointer(p->numa_group, grp);
1711 put_numa_group(my_grp);
1719 void task_numa_free(struct task_struct *p)
1721 struct numa_group *grp = p->numa_group;
1723 void *numa_faults = p->numa_faults_memory;
1726 spin_lock_irq(&grp->lock);
1727 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1728 grp->faults[i] -= p->numa_faults_memory[i];
1729 grp->total_faults -= p->total_numa_faults;
1731 list_del(&p->numa_entry);
1733 spin_unlock_irq(&grp->lock);
1734 rcu_assign_pointer(p->numa_group, NULL);
1735 put_numa_group(grp);
1738 p->numa_faults_memory = NULL;
1739 p->numa_faults_buffer_memory = NULL;
1740 p->numa_faults_cpu= NULL;
1741 p->numa_faults_buffer_cpu = NULL;
1746 * Got a PROT_NONE fault for a page on @node.
1748 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
1750 struct task_struct *p = current;
1751 bool migrated = flags & TNF_MIGRATED;
1752 int cpu_node = task_node(current);
1753 int local = !!(flags & TNF_FAULT_LOCAL);
1756 if (!numabalancing_enabled)
1759 /* for example, ksmd faulting in a user's mm */
1763 /* Do not worry about placement if exiting */
1764 if (p->state == TASK_DEAD)
1767 /* Allocate buffer to track faults on a per-node basis */
1768 if (unlikely(!p->numa_faults_memory)) {
1769 int size = sizeof(*p->numa_faults_memory) *
1770 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
1772 p->numa_faults_memory = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
1773 if (!p->numa_faults_memory)
1776 BUG_ON(p->numa_faults_buffer_memory);
1778 * The averaged statistics, shared & private, memory & cpu,
1779 * occupy the first half of the array. The second half of the
1780 * array is for current counters, which are averaged into the
1781 * first set by task_numa_placement.
1783 p->numa_faults_cpu = p->numa_faults_memory + (2 * nr_node_ids);
1784 p->numa_faults_buffer_memory = p->numa_faults_memory + (4 * nr_node_ids);
1785 p->numa_faults_buffer_cpu = p->numa_faults_memory + (6 * nr_node_ids);
1786 p->total_numa_faults = 0;
1787 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1791 * First accesses are treated as private, otherwise consider accesses
1792 * to be private if the accessing pid has not changed
1794 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
1797 priv = cpupid_match_pid(p, last_cpupid);
1798 if (!priv && !(flags & TNF_NO_GROUP))
1799 task_numa_group(p, last_cpupid, flags, &priv);
1803 * If a workload spans multiple NUMA nodes, a shared fault that
1804 * occurs wholly within the set of nodes that the workload is
1805 * actively using should be counted as local. This allows the
1806 * scan rate to slow down when a workload has settled down.
1808 if (!priv && !local && p->numa_group &&
1809 node_isset(cpu_node, p->numa_group->active_nodes) &&
1810 node_isset(mem_node, p->numa_group->active_nodes))
1813 task_numa_placement(p);
1816 * Retry task to preferred node migration periodically, in case it
1817 * case it previously failed, or the scheduler moved us.
1819 if (time_after(jiffies, p->numa_migrate_retry))
1820 numa_migrate_preferred(p);
1823 p->numa_pages_migrated += pages;
1825 p->numa_faults_buffer_memory[task_faults_idx(mem_node, priv)] += pages;
1826 p->numa_faults_buffer_cpu[task_faults_idx(cpu_node, priv)] += pages;
1827 p->numa_faults_locality[local] += pages;
1830 static void reset_ptenuma_scan(struct task_struct *p)
1832 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1833 p->mm->numa_scan_offset = 0;
1837 * The expensive part of numa migration is done from task_work context.
1838 * Triggered from task_tick_numa().
1840 void task_numa_work(struct callback_head *work)
1842 unsigned long migrate, next_scan, now = jiffies;
1843 struct task_struct *p = current;
1844 struct mm_struct *mm = p->mm;
1845 struct vm_area_struct *vma;
1846 unsigned long start, end;
1847 unsigned long nr_pte_updates = 0;
1850 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1852 work->next = work; /* protect against double add */
1854 * Who cares about NUMA placement when they're dying.
1856 * NOTE: make sure not to dereference p->mm before this check,
1857 * exit_task_work() happens _after_ exit_mm() so we could be called
1858 * without p->mm even though we still had it when we enqueued this
1861 if (p->flags & PF_EXITING)
1864 if (!mm->numa_next_scan) {
1865 mm->numa_next_scan = now +
1866 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1870 * Enforce maximal scan/migration frequency..
1872 migrate = mm->numa_next_scan;
1873 if (time_before(now, migrate))
1876 if (p->numa_scan_period == 0) {
1877 p->numa_scan_period_max = task_scan_max(p);
1878 p->numa_scan_period = task_scan_min(p);
1881 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1882 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1886 * Delay this task enough that another task of this mm will likely win
1887 * the next time around.
1889 p->node_stamp += 2 * TICK_NSEC;
1891 start = mm->numa_scan_offset;
1892 pages = sysctl_numa_balancing_scan_size;
1893 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1897 down_read(&mm->mmap_sem);
1898 vma = find_vma(mm, start);
1900 reset_ptenuma_scan(p);
1904 for (; vma; vma = vma->vm_next) {
1905 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1909 * Shared library pages mapped by multiple processes are not
1910 * migrated as it is expected they are cache replicated. Avoid
1911 * hinting faults in read-only file-backed mappings or the vdso
1912 * as migrating the pages will be of marginal benefit.
1915 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1919 * Skip inaccessible VMAs to avoid any confusion between
1920 * PROT_NONE and NUMA hinting ptes
1922 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
1926 start = max(start, vma->vm_start);
1927 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1928 end = min(end, vma->vm_end);
1929 nr_pte_updates += change_prot_numa(vma, start, end);
1932 * Scan sysctl_numa_balancing_scan_size but ensure that
1933 * at least one PTE is updated so that unused virtual
1934 * address space is quickly skipped.
1937 pages -= (end - start) >> PAGE_SHIFT;
1944 } while (end != vma->vm_end);
1949 * It is possible to reach the end of the VMA list but the last few
1950 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1951 * would find the !migratable VMA on the next scan but not reset the
1952 * scanner to the start so check it now.
1955 mm->numa_scan_offset = start;
1957 reset_ptenuma_scan(p);
1958 up_read(&mm->mmap_sem);
1962 * Drive the periodic memory faults..
1964 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1966 struct callback_head *work = &curr->numa_work;
1970 * We don't care about NUMA placement if we don't have memory.
1972 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1976 * Using runtime rather than walltime has the dual advantage that
1977 * we (mostly) drive the selection from busy threads and that the
1978 * task needs to have done some actual work before we bother with
1981 now = curr->se.sum_exec_runtime;
1982 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1984 if (now - curr->node_stamp > period) {
1985 if (!curr->node_stamp)
1986 curr->numa_scan_period = task_scan_min(curr);
1987 curr->node_stamp += period;
1989 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1990 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1991 task_work_add(curr, work, true);
1996 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2000 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2004 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2007 #endif /* CONFIG_NUMA_BALANCING */
2010 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2012 update_load_add(&cfs_rq->load, se->load.weight);
2013 if (!parent_entity(se))
2014 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2016 if (entity_is_task(se)) {
2017 struct rq *rq = rq_of(cfs_rq);
2019 account_numa_enqueue(rq, task_of(se));
2020 list_add(&se->group_node, &rq->cfs_tasks);
2023 cfs_rq->nr_running++;
2027 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2029 update_load_sub(&cfs_rq->load, se->load.weight);
2030 if (!parent_entity(se))
2031 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2032 if (entity_is_task(se)) {
2033 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2034 list_del_init(&se->group_node);
2036 cfs_rq->nr_running--;
2039 #ifdef CONFIG_FAIR_GROUP_SCHED
2041 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2046 * Use this CPU's actual weight instead of the last load_contribution
2047 * to gain a more accurate current total weight. See
2048 * update_cfs_rq_load_contribution().
2050 tg_weight = atomic_long_read(&tg->load_avg);
2051 tg_weight -= cfs_rq->tg_load_contrib;
2052 tg_weight += cfs_rq->load.weight;
2057 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2059 long tg_weight, load, shares;
2061 tg_weight = calc_tg_weight(tg, cfs_rq);
2062 load = cfs_rq->load.weight;
2064 shares = (tg->shares * load);
2066 shares /= tg_weight;
2068 if (shares < MIN_SHARES)
2069 shares = MIN_SHARES;
2070 if (shares > tg->shares)
2071 shares = tg->shares;
2075 # else /* CONFIG_SMP */
2076 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2080 # endif /* CONFIG_SMP */
2081 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2082 unsigned long weight)
2085 /* commit outstanding execution time */
2086 if (cfs_rq->curr == se)
2087 update_curr(cfs_rq);
2088 account_entity_dequeue(cfs_rq, se);
2091 update_load_set(&se->load, weight);
2094 account_entity_enqueue(cfs_rq, se);
2097 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2099 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2101 struct task_group *tg;
2102 struct sched_entity *se;
2106 se = tg->se[cpu_of(rq_of(cfs_rq))];
2107 if (!se || throttled_hierarchy(cfs_rq))
2110 if (likely(se->load.weight == tg->shares))
2113 shares = calc_cfs_shares(cfs_rq, tg);
2115 reweight_entity(cfs_rq_of(se), se, shares);
2117 #else /* CONFIG_FAIR_GROUP_SCHED */
2118 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2121 #endif /* CONFIG_FAIR_GROUP_SCHED */
2125 * We choose a half-life close to 1 scheduling period.
2126 * Note: The tables below are dependent on this value.
2128 #define LOAD_AVG_PERIOD 32
2129 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2130 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2132 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2133 static const u32 runnable_avg_yN_inv[] = {
2134 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2135 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2136 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2137 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2138 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2139 0x85aac367, 0x82cd8698,
2143 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2144 * over-estimates when re-combining.
2146 static const u32 runnable_avg_yN_sum[] = {
2147 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2148 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2149 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2154 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2156 static __always_inline u64 decay_load(u64 val, u64 n)
2158 unsigned int local_n;
2162 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2165 /* after bounds checking we can collapse to 32-bit */
2169 * As y^PERIOD = 1/2, we can combine
2170 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
2171 * With a look-up table which covers k^n (n<PERIOD)
2173 * To achieve constant time decay_load.
2175 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2176 val >>= local_n / LOAD_AVG_PERIOD;
2177 local_n %= LOAD_AVG_PERIOD;
2180 val *= runnable_avg_yN_inv[local_n];
2181 /* We don't use SRR here since we always want to round down. */
2186 * For updates fully spanning n periods, the contribution to runnable
2187 * average will be: \Sum 1024*y^n
2189 * We can compute this reasonably efficiently by combining:
2190 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2192 static u32 __compute_runnable_contrib(u64 n)
2196 if (likely(n <= LOAD_AVG_PERIOD))
2197 return runnable_avg_yN_sum[n];
2198 else if (unlikely(n >= LOAD_AVG_MAX_N))
2199 return LOAD_AVG_MAX;
2201 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2203 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2204 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2206 n -= LOAD_AVG_PERIOD;
2207 } while (n > LOAD_AVG_PERIOD);
2209 contrib = decay_load(contrib, n);
2210 return contrib + runnable_avg_yN_sum[n];
2214 * We can represent the historical contribution to runnable average as the
2215 * coefficients of a geometric series. To do this we sub-divide our runnable
2216 * history into segments of approximately 1ms (1024us); label the segment that
2217 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2219 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2221 * (now) (~1ms ago) (~2ms ago)
2223 * Let u_i denote the fraction of p_i that the entity was runnable.
2225 * We then designate the fractions u_i as our co-efficients, yielding the
2226 * following representation of historical load:
2227 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2229 * We choose y based on the with of a reasonably scheduling period, fixing:
2232 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2233 * approximately half as much as the contribution to load within the last ms
2236 * When a period "rolls over" and we have new u_0`, multiplying the previous
2237 * sum again by y is sufficient to update:
2238 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2239 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2241 static __always_inline int __update_entity_runnable_avg(u64 now,
2242 struct sched_avg *sa,
2246 u32 runnable_contrib;
2247 int delta_w, decayed = 0;
2249 delta = now - sa->last_runnable_update;
2251 * This should only happen when time goes backwards, which it
2252 * unfortunately does during sched clock init when we swap over to TSC.
2254 if ((s64)delta < 0) {
2255 sa->last_runnable_update = now;
2260 * Use 1024ns as the unit of measurement since it's a reasonable
2261 * approximation of 1us and fast to compute.
2266 sa->last_runnable_update = now;
2268 /* delta_w is the amount already accumulated against our next period */
2269 delta_w = sa->runnable_avg_period % 1024;
2270 if (delta + delta_w >= 1024) {
2271 /* period roll-over */
2275 * Now that we know we're crossing a period boundary, figure
2276 * out how much from delta we need to complete the current
2277 * period and accrue it.
2279 delta_w = 1024 - delta_w;
2281 sa->runnable_avg_sum += delta_w;
2282 sa->runnable_avg_period += delta_w;
2286 /* Figure out how many additional periods this update spans */
2287 periods = delta / 1024;
2290 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2292 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
2295 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2296 runnable_contrib = __compute_runnable_contrib(periods);
2298 sa->runnable_avg_sum += runnable_contrib;
2299 sa->runnable_avg_period += runnable_contrib;
2302 /* Remainder of delta accrued against u_0` */
2304 sa->runnable_avg_sum += delta;
2305 sa->runnable_avg_period += delta;
2310 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2311 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2313 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2314 u64 decays = atomic64_read(&cfs_rq->decay_counter);
2316 decays -= se->avg.decay_count;
2320 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2321 se->avg.decay_count = 0;
2326 #ifdef CONFIG_FAIR_GROUP_SCHED
2327 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2330 struct task_group *tg = cfs_rq->tg;
2333 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2334 tg_contrib -= cfs_rq->tg_load_contrib;
2336 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2337 atomic_long_add(tg_contrib, &tg->load_avg);
2338 cfs_rq->tg_load_contrib += tg_contrib;
2343 * Aggregate cfs_rq runnable averages into an equivalent task_group
2344 * representation for computing load contributions.
2346 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2347 struct cfs_rq *cfs_rq)
2349 struct task_group *tg = cfs_rq->tg;
2352 /* The fraction of a cpu used by this cfs_rq */
2353 contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT,
2354 sa->runnable_avg_period + 1);
2355 contrib -= cfs_rq->tg_runnable_contrib;
2357 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2358 atomic_add(contrib, &tg->runnable_avg);
2359 cfs_rq->tg_runnable_contrib += contrib;
2363 static inline void __update_group_entity_contrib(struct sched_entity *se)
2365 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2366 struct task_group *tg = cfs_rq->tg;
2371 contrib = cfs_rq->tg_load_contrib * tg->shares;
2372 se->avg.load_avg_contrib = div_u64(contrib,
2373 atomic_long_read(&tg->load_avg) + 1);
2376 * For group entities we need to compute a correction term in the case
2377 * that they are consuming <1 cpu so that we would contribute the same
2378 * load as a task of equal weight.
2380 * Explicitly co-ordinating this measurement would be expensive, but
2381 * fortunately the sum of each cpus contribution forms a usable
2382 * lower-bound on the true value.
2384 * Consider the aggregate of 2 contributions. Either they are disjoint
2385 * (and the sum represents true value) or they are disjoint and we are
2386 * understating by the aggregate of their overlap.
2388 * Extending this to N cpus, for a given overlap, the maximum amount we
2389 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2390 * cpus that overlap for this interval and w_i is the interval width.
2392 * On a small machine; the first term is well-bounded which bounds the
2393 * total error since w_i is a subset of the period. Whereas on a
2394 * larger machine, while this first term can be larger, if w_i is the
2395 * of consequential size guaranteed to see n_i*w_i quickly converge to
2396 * our upper bound of 1-cpu.
2398 runnable_avg = atomic_read(&tg->runnable_avg);
2399 if (runnable_avg < NICE_0_LOAD) {
2400 se->avg.load_avg_contrib *= runnable_avg;
2401 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2405 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2407 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
2408 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2410 #else /* CONFIG_FAIR_GROUP_SCHED */
2411 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2412 int force_update) {}
2413 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2414 struct cfs_rq *cfs_rq) {}
2415 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2416 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2417 #endif /* CONFIG_FAIR_GROUP_SCHED */
2419 static inline void __update_task_entity_contrib(struct sched_entity *se)
2423 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2424 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2425 contrib /= (se->avg.runnable_avg_period + 1);
2426 se->avg.load_avg_contrib = scale_load(contrib);
2429 /* Compute the current contribution to load_avg by se, return any delta */
2430 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2432 long old_contrib = se->avg.load_avg_contrib;
2434 if (entity_is_task(se)) {
2435 __update_task_entity_contrib(se);
2437 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2438 __update_group_entity_contrib(se);
2441 return se->avg.load_avg_contrib - old_contrib;
2444 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2447 if (likely(load_contrib < cfs_rq->blocked_load_avg))
2448 cfs_rq->blocked_load_avg -= load_contrib;
2450 cfs_rq->blocked_load_avg = 0;
2453 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2455 /* Update a sched_entity's runnable average */
2456 static inline void update_entity_load_avg(struct sched_entity *se,
2459 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2464 * For a group entity we need to use their owned cfs_rq_clock_task() in
2465 * case they are the parent of a throttled hierarchy.
2467 if (entity_is_task(se))
2468 now = cfs_rq_clock_task(cfs_rq);
2470 now = cfs_rq_clock_task(group_cfs_rq(se));
2472 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
2475 contrib_delta = __update_entity_load_avg_contrib(se);
2481 cfs_rq->runnable_load_avg += contrib_delta;
2483 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2487 * Decay the load contributed by all blocked children and account this so that
2488 * their contribution may appropriately discounted when they wake up.
2490 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2492 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2495 decays = now - cfs_rq->last_decay;
2496 if (!decays && !force_update)
2499 if (atomic_long_read(&cfs_rq->removed_load)) {
2500 unsigned long removed_load;
2501 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2502 subtract_blocked_load_contrib(cfs_rq, removed_load);
2506 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2508 atomic64_add(decays, &cfs_rq->decay_counter);
2509 cfs_rq->last_decay = now;
2512 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2515 /* Add the load generated by se into cfs_rq's child load-average */
2516 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2517 struct sched_entity *se,
2521 * We track migrations using entity decay_count <= 0, on a wake-up
2522 * migration we use a negative decay count to track the remote decays
2523 * accumulated while sleeping.
2525 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2526 * are seen by enqueue_entity_load_avg() as a migration with an already
2527 * constructed load_avg_contrib.
2529 if (unlikely(se->avg.decay_count <= 0)) {
2530 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2531 if (se->avg.decay_count) {
2533 * In a wake-up migration we have to approximate the
2534 * time sleeping. This is because we can't synchronize
2535 * clock_task between the two cpus, and it is not
2536 * guaranteed to be read-safe. Instead, we can
2537 * approximate this using our carried decays, which are
2538 * explicitly atomically readable.
2540 se->avg.last_runnable_update -= (-se->avg.decay_count)
2542 update_entity_load_avg(se, 0);
2543 /* Indicate that we're now synchronized and on-rq */
2544 se->avg.decay_count = 0;
2548 __synchronize_entity_decay(se);
2551 /* migrated tasks did not contribute to our blocked load */
2553 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2554 update_entity_load_avg(se, 0);
2557 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2558 /* we force update consideration on load-balancer moves */
2559 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2563 * Remove se's load from this cfs_rq child load-average, if the entity is
2564 * transitioning to a blocked state we track its projected decay using
2567 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2568 struct sched_entity *se,
2571 update_entity_load_avg(se, 1);
2572 /* we force update consideration on load-balancer moves */
2573 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2575 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2577 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2578 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2579 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2583 * Update the rq's load with the elapsed running time before entering
2584 * idle. if the last scheduled task is not a CFS task, idle_enter will
2585 * be the only way to update the runnable statistic.
2587 void idle_enter_fair(struct rq *this_rq)
2589 update_rq_runnable_avg(this_rq, 1);
2593 * Update the rq's load with the elapsed idle time before a task is
2594 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2595 * be the only way to update the runnable statistic.
2597 void idle_exit_fair(struct rq *this_rq)
2599 update_rq_runnable_avg(this_rq, 0);
2602 static int idle_balance(struct rq *this_rq);
2604 #else /* CONFIG_SMP */
2606 static inline void update_entity_load_avg(struct sched_entity *se,
2607 int update_cfs_rq) {}
2608 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2609 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2610 struct sched_entity *se,
2612 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2613 struct sched_entity *se,
2615 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2616 int force_update) {}
2618 static inline int idle_balance(struct rq *rq)
2623 #endif /* CONFIG_SMP */
2625 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2627 #ifdef CONFIG_SCHEDSTATS
2628 struct task_struct *tsk = NULL;
2630 if (entity_is_task(se))
2633 if (se->statistics.sleep_start) {
2634 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2639 if (unlikely(delta > se->statistics.sleep_max))
2640 se->statistics.sleep_max = delta;
2642 se->statistics.sleep_start = 0;
2643 se->statistics.sum_sleep_runtime += delta;
2646 account_scheduler_latency(tsk, delta >> 10, 1);
2647 trace_sched_stat_sleep(tsk, delta);
2650 if (se->statistics.block_start) {
2651 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2656 if (unlikely(delta > se->statistics.block_max))
2657 se->statistics.block_max = delta;
2659 se->statistics.block_start = 0;
2660 se->statistics.sum_sleep_runtime += delta;
2663 if (tsk->in_iowait) {
2664 se->statistics.iowait_sum += delta;
2665 se->statistics.iowait_count++;
2666 trace_sched_stat_iowait(tsk, delta);
2669 trace_sched_stat_blocked(tsk, delta);
2672 * Blocking time is in units of nanosecs, so shift by
2673 * 20 to get a milliseconds-range estimation of the
2674 * amount of time that the task spent sleeping:
2676 if (unlikely(prof_on == SLEEP_PROFILING)) {
2677 profile_hits(SLEEP_PROFILING,
2678 (void *)get_wchan(tsk),
2681 account_scheduler_latency(tsk, delta >> 10, 0);
2687 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2689 #ifdef CONFIG_SCHED_DEBUG
2690 s64 d = se->vruntime - cfs_rq->min_vruntime;
2695 if (d > 3*sysctl_sched_latency)
2696 schedstat_inc(cfs_rq, nr_spread_over);
2701 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2703 u64 vruntime = cfs_rq->min_vruntime;
2706 * The 'current' period is already promised to the current tasks,
2707 * however the extra weight of the new task will slow them down a
2708 * little, place the new task so that it fits in the slot that
2709 * stays open at the end.
2711 if (initial && sched_feat(START_DEBIT))
2712 vruntime += sched_vslice(cfs_rq, se);
2714 /* sleeps up to a single latency don't count. */
2716 unsigned long thresh = sysctl_sched_latency;
2719 * Halve their sleep time's effect, to allow
2720 * for a gentler effect of sleepers:
2722 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2728 /* ensure we never gain time by being placed backwards. */
2729 se->vruntime = max_vruntime(se->vruntime, vruntime);
2732 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2735 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2738 * Update the normalized vruntime before updating min_vruntime
2739 * through calling update_curr().
2741 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2742 se->vruntime += cfs_rq->min_vruntime;
2745 * Update run-time statistics of the 'current'.
2747 update_curr(cfs_rq);
2748 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2749 account_entity_enqueue(cfs_rq, se);
2750 update_cfs_shares(cfs_rq);
2752 if (flags & ENQUEUE_WAKEUP) {
2753 place_entity(cfs_rq, se, 0);
2754 enqueue_sleeper(cfs_rq, se);
2757 update_stats_enqueue(cfs_rq, se);
2758 check_spread(cfs_rq, se);
2759 if (se != cfs_rq->curr)
2760 __enqueue_entity(cfs_rq, se);
2763 if (cfs_rq->nr_running == 1) {
2764 list_add_leaf_cfs_rq(cfs_rq);
2765 check_enqueue_throttle(cfs_rq);
2769 static void __clear_buddies_last(struct sched_entity *se)
2771 for_each_sched_entity(se) {
2772 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2773 if (cfs_rq->last != se)
2776 cfs_rq->last = NULL;
2780 static void __clear_buddies_next(struct sched_entity *se)
2782 for_each_sched_entity(se) {
2783 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2784 if (cfs_rq->next != se)
2787 cfs_rq->next = NULL;
2791 static void __clear_buddies_skip(struct sched_entity *se)
2793 for_each_sched_entity(se) {
2794 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2795 if (cfs_rq->skip != se)
2798 cfs_rq->skip = NULL;
2802 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2804 if (cfs_rq->last == se)
2805 __clear_buddies_last(se);
2807 if (cfs_rq->next == se)
2808 __clear_buddies_next(se);
2810 if (cfs_rq->skip == se)
2811 __clear_buddies_skip(se);
2814 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2817 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2820 * Update run-time statistics of the 'current'.
2822 update_curr(cfs_rq);
2823 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2825 update_stats_dequeue(cfs_rq, se);
2826 if (flags & DEQUEUE_SLEEP) {
2827 #ifdef CONFIG_SCHEDSTATS
2828 if (entity_is_task(se)) {
2829 struct task_struct *tsk = task_of(se);
2831 if (tsk->state & TASK_INTERRUPTIBLE)
2832 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2833 if (tsk->state & TASK_UNINTERRUPTIBLE)
2834 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2839 clear_buddies(cfs_rq, se);
2841 if (se != cfs_rq->curr)
2842 __dequeue_entity(cfs_rq, se);
2844 account_entity_dequeue(cfs_rq, se);
2847 * Normalize the entity after updating the min_vruntime because the
2848 * update can refer to the ->curr item and we need to reflect this
2849 * movement in our normalized position.
2851 if (!(flags & DEQUEUE_SLEEP))
2852 se->vruntime -= cfs_rq->min_vruntime;
2854 /* return excess runtime on last dequeue */
2855 return_cfs_rq_runtime(cfs_rq);
2857 update_min_vruntime(cfs_rq);
2858 update_cfs_shares(cfs_rq);
2862 * Preempt the current task with a newly woken task if needed:
2865 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2867 unsigned long ideal_runtime, delta_exec;
2868 struct sched_entity *se;
2871 ideal_runtime = sched_slice(cfs_rq, curr);
2872 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2873 if (delta_exec > ideal_runtime) {
2874 resched_task(rq_of(cfs_rq)->curr);
2876 * The current task ran long enough, ensure it doesn't get
2877 * re-elected due to buddy favours.
2879 clear_buddies(cfs_rq, curr);
2884 * Ensure that a task that missed wakeup preemption by a
2885 * narrow margin doesn't have to wait for a full slice.
2886 * This also mitigates buddy induced latencies under load.
2888 if (delta_exec < sysctl_sched_min_granularity)
2891 se = __pick_first_entity(cfs_rq);
2892 delta = curr->vruntime - se->vruntime;
2897 if (delta > ideal_runtime)
2898 resched_task(rq_of(cfs_rq)->curr);
2902 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2904 /* 'current' is not kept within the tree. */
2907 * Any task has to be enqueued before it get to execute on
2908 * a CPU. So account for the time it spent waiting on the
2911 update_stats_wait_end(cfs_rq, se);
2912 __dequeue_entity(cfs_rq, se);
2915 update_stats_curr_start(cfs_rq, se);
2917 #ifdef CONFIG_SCHEDSTATS
2919 * Track our maximum slice length, if the CPU's load is at
2920 * least twice that of our own weight (i.e. dont track it
2921 * when there are only lesser-weight tasks around):
2923 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2924 se->statistics.slice_max = max(se->statistics.slice_max,
2925 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2928 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2932 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2935 * Pick the next process, keeping these things in mind, in this order:
2936 * 1) keep things fair between processes/task groups
2937 * 2) pick the "next" process, since someone really wants that to run
2938 * 3) pick the "last" process, for cache locality
2939 * 4) do not run the "skip" process, if something else is available
2941 static struct sched_entity *
2942 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2944 struct sched_entity *left = __pick_first_entity(cfs_rq);
2945 struct sched_entity *se;
2948 * If curr is set we have to see if its left of the leftmost entity
2949 * still in the tree, provided there was anything in the tree at all.
2951 if (!left || (curr && entity_before(curr, left)))
2954 se = left; /* ideally we run the leftmost entity */
2957 * Avoid running the skip buddy, if running something else can
2958 * be done without getting too unfair.
2960 if (cfs_rq->skip == se) {
2961 struct sched_entity *second;
2964 second = __pick_first_entity(cfs_rq);
2966 second = __pick_next_entity(se);
2967 if (!second || (curr && entity_before(curr, second)))
2971 if (second && wakeup_preempt_entity(second, left) < 1)
2976 * Prefer last buddy, try to return the CPU to a preempted task.
2978 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2982 * Someone really wants this to run. If it's not unfair, run it.
2984 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2987 clear_buddies(cfs_rq, se);
2992 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2994 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2997 * If still on the runqueue then deactivate_task()
2998 * was not called and update_curr() has to be done:
3001 update_curr(cfs_rq);
3003 /* throttle cfs_rqs exceeding runtime */
3004 check_cfs_rq_runtime(cfs_rq);
3006 check_spread(cfs_rq, prev);
3008 update_stats_wait_start(cfs_rq, prev);
3009 /* Put 'current' back into the tree. */
3010 __enqueue_entity(cfs_rq, prev);
3011 /* in !on_rq case, update occurred at dequeue */
3012 update_entity_load_avg(prev, 1);
3014 cfs_rq->curr = NULL;
3018 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3021 * Update run-time statistics of the 'current'.
3023 update_curr(cfs_rq);
3026 * Ensure that runnable average is periodically updated.
3028 update_entity_load_avg(curr, 1);
3029 update_cfs_rq_blocked_load(cfs_rq, 1);
3030 update_cfs_shares(cfs_rq);
3032 #ifdef CONFIG_SCHED_HRTICK
3034 * queued ticks are scheduled to match the slice, so don't bother
3035 * validating it and just reschedule.
3038 resched_task(rq_of(cfs_rq)->curr);
3042 * don't let the period tick interfere with the hrtick preemption
3044 if (!sched_feat(DOUBLE_TICK) &&
3045 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3049 if (cfs_rq->nr_running > 1)
3050 check_preempt_tick(cfs_rq, curr);
3054 /**************************************************
3055 * CFS bandwidth control machinery
3058 #ifdef CONFIG_CFS_BANDWIDTH
3060 #ifdef HAVE_JUMP_LABEL
3061 static struct static_key __cfs_bandwidth_used;
3063 static inline bool cfs_bandwidth_used(void)
3065 return static_key_false(&__cfs_bandwidth_used);
3068 void cfs_bandwidth_usage_inc(void)
3070 static_key_slow_inc(&__cfs_bandwidth_used);
3073 void cfs_bandwidth_usage_dec(void)
3075 static_key_slow_dec(&__cfs_bandwidth_used);
3077 #else /* HAVE_JUMP_LABEL */
3078 static bool cfs_bandwidth_used(void)
3083 void cfs_bandwidth_usage_inc(void) {}
3084 void cfs_bandwidth_usage_dec(void) {}
3085 #endif /* HAVE_JUMP_LABEL */
3088 * default period for cfs group bandwidth.
3089 * default: 0.1s, units: nanoseconds
3091 static inline u64 default_cfs_period(void)
3093 return 100000000ULL;
3096 static inline u64 sched_cfs_bandwidth_slice(void)
3098 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3102 * Replenish runtime according to assigned quota and update expiration time.
3103 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3104 * additional synchronization around rq->lock.
3106 * requires cfs_b->lock
3108 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3112 if (cfs_b->quota == RUNTIME_INF)
3115 now = sched_clock_cpu(smp_processor_id());
3116 cfs_b->runtime = cfs_b->quota;
3117 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3120 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3122 return &tg->cfs_bandwidth;
3125 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3126 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3128 if (unlikely(cfs_rq->throttle_count))
3129 return cfs_rq->throttled_clock_task;
3131 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3134 /* returns 0 on failure to allocate runtime */
3135 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3137 struct task_group *tg = cfs_rq->tg;
3138 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3139 u64 amount = 0, min_amount, expires;
3141 /* note: this is a positive sum as runtime_remaining <= 0 */
3142 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3144 raw_spin_lock(&cfs_b->lock);
3145 if (cfs_b->quota == RUNTIME_INF)
3146 amount = min_amount;
3149 * If the bandwidth pool has become inactive, then at least one
3150 * period must have elapsed since the last consumption.
3151 * Refresh the global state and ensure bandwidth timer becomes
3154 if (!cfs_b->timer_active) {
3155 __refill_cfs_bandwidth_runtime(cfs_b);
3156 __start_cfs_bandwidth(cfs_b);
3159 if (cfs_b->runtime > 0) {
3160 amount = min(cfs_b->runtime, min_amount);
3161 cfs_b->runtime -= amount;
3165 expires = cfs_b->runtime_expires;
3166 raw_spin_unlock(&cfs_b->lock);
3168 cfs_rq->runtime_remaining += amount;
3170 * we may have advanced our local expiration to account for allowed
3171 * spread between our sched_clock and the one on which runtime was
3174 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3175 cfs_rq->runtime_expires = expires;
3177 return cfs_rq->runtime_remaining > 0;
3181 * Note: This depends on the synchronization provided by sched_clock and the
3182 * fact that rq->clock snapshots this value.
3184 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3186 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3188 /* if the deadline is ahead of our clock, nothing to do */
3189 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3192 if (cfs_rq->runtime_remaining < 0)
3196 * If the local deadline has passed we have to consider the
3197 * possibility that our sched_clock is 'fast' and the global deadline
3198 * has not truly expired.
3200 * Fortunately we can check determine whether this the case by checking
3201 * whether the global deadline has advanced.
3204 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
3205 /* extend local deadline, drift is bounded above by 2 ticks */
3206 cfs_rq->runtime_expires += TICK_NSEC;
3208 /* global deadline is ahead, expiration has passed */
3209 cfs_rq->runtime_remaining = 0;
3213 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3215 /* dock delta_exec before expiring quota (as it could span periods) */
3216 cfs_rq->runtime_remaining -= delta_exec;
3217 expire_cfs_rq_runtime(cfs_rq);
3219 if (likely(cfs_rq->runtime_remaining > 0))
3223 * if we're unable to extend our runtime we resched so that the active
3224 * hierarchy can be throttled
3226 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3227 resched_task(rq_of(cfs_rq)->curr);
3230 static __always_inline
3231 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3233 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3236 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3239 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3241 return cfs_bandwidth_used() && cfs_rq->throttled;
3244 /* check whether cfs_rq, or any parent, is throttled */
3245 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3247 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3251 * Ensure that neither of the group entities corresponding to src_cpu or
3252 * dest_cpu are members of a throttled hierarchy when performing group
3253 * load-balance operations.
3255 static inline int throttled_lb_pair(struct task_group *tg,
3256 int src_cpu, int dest_cpu)
3258 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3260 src_cfs_rq = tg->cfs_rq[src_cpu];
3261 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3263 return throttled_hierarchy(src_cfs_rq) ||
3264 throttled_hierarchy(dest_cfs_rq);
3267 /* updated child weight may affect parent so we have to do this bottom up */
3268 static int tg_unthrottle_up(struct task_group *tg, void *data)
3270 struct rq *rq = data;
3271 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3273 cfs_rq->throttle_count--;
3275 if (!cfs_rq->throttle_count) {
3276 /* adjust cfs_rq_clock_task() */
3277 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3278 cfs_rq->throttled_clock_task;
3285 static int tg_throttle_down(struct task_group *tg, void *data)
3287 struct rq *rq = data;
3288 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3290 /* group is entering throttled state, stop time */
3291 if (!cfs_rq->throttle_count)
3292 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3293 cfs_rq->throttle_count++;
3298 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3300 struct rq *rq = rq_of(cfs_rq);
3301 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3302 struct sched_entity *se;
3303 long task_delta, dequeue = 1;
3305 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3307 /* freeze hierarchy runnable averages while throttled */
3309 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3312 task_delta = cfs_rq->h_nr_running;
3313 for_each_sched_entity(se) {
3314 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3315 /* throttled entity or throttle-on-deactivate */
3320 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3321 qcfs_rq->h_nr_running -= task_delta;
3323 if (qcfs_rq->load.weight)
3328 rq->nr_running -= task_delta;
3330 cfs_rq->throttled = 1;
3331 cfs_rq->throttled_clock = rq_clock(rq);
3332 raw_spin_lock(&cfs_b->lock);
3333 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3334 if (!cfs_b->timer_active)
3335 __start_cfs_bandwidth(cfs_b);
3336 raw_spin_unlock(&cfs_b->lock);
3339 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3341 struct rq *rq = rq_of(cfs_rq);
3342 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3343 struct sched_entity *se;
3347 se = cfs_rq->tg->se[cpu_of(rq)];
3349 cfs_rq->throttled = 0;
3351 update_rq_clock(rq);
3353 raw_spin_lock(&cfs_b->lock);
3354 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3355 list_del_rcu(&cfs_rq->throttled_list);
3356 raw_spin_unlock(&cfs_b->lock);
3358 /* update hierarchical throttle state */
3359 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3361 if (!cfs_rq->load.weight)
3364 task_delta = cfs_rq->h_nr_running;
3365 for_each_sched_entity(se) {
3369 cfs_rq = cfs_rq_of(se);
3371 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3372 cfs_rq->h_nr_running += task_delta;
3374 if (cfs_rq_throttled(cfs_rq))
3379 rq->nr_running += task_delta;
3381 /* determine whether we need to wake up potentially idle cpu */
3382 if (rq->curr == rq->idle && rq->cfs.nr_running)
3383 resched_task(rq->curr);
3386 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3387 u64 remaining, u64 expires)
3389 struct cfs_rq *cfs_rq;
3390 u64 runtime = remaining;
3393 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3395 struct rq *rq = rq_of(cfs_rq);
3397 raw_spin_lock(&rq->lock);
3398 if (!cfs_rq_throttled(cfs_rq))
3401 runtime = -cfs_rq->runtime_remaining + 1;
3402 if (runtime > remaining)
3403 runtime = remaining;
3404 remaining -= runtime;
3406 cfs_rq->runtime_remaining += runtime;
3407 cfs_rq->runtime_expires = expires;
3409 /* we check whether we're throttled above */
3410 if (cfs_rq->runtime_remaining > 0)
3411 unthrottle_cfs_rq(cfs_rq);
3414 raw_spin_unlock(&rq->lock);
3425 * Responsible for refilling a task_group's bandwidth and unthrottling its
3426 * cfs_rqs as appropriate. If there has been no activity within the last
3427 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3428 * used to track this state.
3430 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3432 u64 runtime, runtime_expires;
3433 int idle = 1, throttled;
3435 raw_spin_lock(&cfs_b->lock);
3436 /* no need to continue the timer with no bandwidth constraint */
3437 if (cfs_b->quota == RUNTIME_INF)
3440 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3441 /* idle depends on !throttled (for the case of a large deficit) */
3442 idle = cfs_b->idle && !throttled;
3443 cfs_b->nr_periods += overrun;
3445 /* if we're going inactive then everything else can be deferred */
3450 * if we have relooped after returning idle once, we need to update our
3451 * status as actually running, so that other cpus doing
3452 * __start_cfs_bandwidth will stop trying to cancel us.
3454 cfs_b->timer_active = 1;
3456 __refill_cfs_bandwidth_runtime(cfs_b);
3459 /* mark as potentially idle for the upcoming period */
3464 /* account preceding periods in which throttling occurred */
3465 cfs_b->nr_throttled += overrun;
3468 * There are throttled entities so we must first use the new bandwidth
3469 * to unthrottle them before making it generally available. This
3470 * ensures that all existing debts will be paid before a new cfs_rq is
3473 runtime = cfs_b->runtime;
3474 runtime_expires = cfs_b->runtime_expires;
3478 * This check is repeated as we are holding onto the new bandwidth
3479 * while we unthrottle. This can potentially race with an unthrottled
3480 * group trying to acquire new bandwidth from the global pool.
3482 while (throttled && runtime > 0) {
3483 raw_spin_unlock(&cfs_b->lock);
3484 /* we can't nest cfs_b->lock while distributing bandwidth */
3485 runtime = distribute_cfs_runtime(cfs_b, runtime,
3487 raw_spin_lock(&cfs_b->lock);
3489 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3492 /* return (any) remaining runtime */
3493 cfs_b->runtime = runtime;
3495 * While we are ensured activity in the period following an
3496 * unthrottle, this also covers the case in which the new bandwidth is
3497 * insufficient to cover the existing bandwidth deficit. (Forcing the
3498 * timer to remain active while there are any throttled entities.)
3503 cfs_b->timer_active = 0;
3504 raw_spin_unlock(&cfs_b->lock);
3509 /* a cfs_rq won't donate quota below this amount */
3510 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3511 /* minimum remaining period time to redistribute slack quota */
3512 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3513 /* how long we wait to gather additional slack before distributing */
3514 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3517 * Are we near the end of the current quota period?
3519 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3520 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3521 * migrate_hrtimers, base is never cleared, so we are fine.
3523 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3525 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3528 /* if the call-back is running a quota refresh is already occurring */
3529 if (hrtimer_callback_running(refresh_timer))
3532 /* is a quota refresh about to occur? */
3533 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3534 if (remaining < min_expire)
3540 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3542 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3544 /* if there's a quota refresh soon don't bother with slack */
3545 if (runtime_refresh_within(cfs_b, min_left))
3548 start_bandwidth_timer(&cfs_b->slack_timer,
3549 ns_to_ktime(cfs_bandwidth_slack_period));
3552 /* we know any runtime found here is valid as update_curr() precedes return */
3553 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3555 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3556 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3558 if (slack_runtime <= 0)
3561 raw_spin_lock(&cfs_b->lock);
3562 if (cfs_b->quota != RUNTIME_INF &&
3563 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3564 cfs_b->runtime += slack_runtime;
3566 /* we are under rq->lock, defer unthrottling using a timer */
3567 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3568 !list_empty(&cfs_b->throttled_cfs_rq))
3569 start_cfs_slack_bandwidth(cfs_b);
3571 raw_spin_unlock(&cfs_b->lock);
3573 /* even if it's not valid for return we don't want to try again */
3574 cfs_rq->runtime_remaining -= slack_runtime;
3577 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3579 if (!cfs_bandwidth_used())
3582 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3585 __return_cfs_rq_runtime(cfs_rq);
3589 * This is done with a timer (instead of inline with bandwidth return) since
3590 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3592 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3594 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3597 /* confirm we're still not at a refresh boundary */
3598 raw_spin_lock(&cfs_b->lock);
3599 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3600 raw_spin_unlock(&cfs_b->lock);
3604 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
3605 runtime = cfs_b->runtime;
3608 expires = cfs_b->runtime_expires;
3609 raw_spin_unlock(&cfs_b->lock);
3614 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3616 raw_spin_lock(&cfs_b->lock);
3617 if (expires == cfs_b->runtime_expires)
3618 cfs_b->runtime = runtime;
3619 raw_spin_unlock(&cfs_b->lock);
3623 * When a group wakes up we want to make sure that its quota is not already
3624 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3625 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3627 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3629 if (!cfs_bandwidth_used())
3632 /* an active group must be handled by the update_curr()->put() path */
3633 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3636 /* ensure the group is not already throttled */
3637 if (cfs_rq_throttled(cfs_rq))
3640 /* update runtime allocation */
3641 account_cfs_rq_runtime(cfs_rq, 0);
3642 if (cfs_rq->runtime_remaining <= 0)
3643 throttle_cfs_rq(cfs_rq);
3646 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3647 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3649 if (!cfs_bandwidth_used())
3652 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3656 * it's possible for a throttled entity to be forced into a running
3657 * state (e.g. set_curr_task), in this case we're finished.
3659 if (cfs_rq_throttled(cfs_rq))
3662 throttle_cfs_rq(cfs_rq);
3666 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3668 struct cfs_bandwidth *cfs_b =
3669 container_of(timer, struct cfs_bandwidth, slack_timer);
3670 do_sched_cfs_slack_timer(cfs_b);
3672 return HRTIMER_NORESTART;
3675 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3677 struct cfs_bandwidth *cfs_b =
3678 container_of(timer, struct cfs_bandwidth, period_timer);
3684 now = hrtimer_cb_get_time(timer);
3685 overrun = hrtimer_forward(timer, now, cfs_b->period);
3690 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3693 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3696 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3698 raw_spin_lock_init(&cfs_b->lock);
3700 cfs_b->quota = RUNTIME_INF;
3701 cfs_b->period = ns_to_ktime(default_cfs_period());
3703 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3704 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3705 cfs_b->period_timer.function = sched_cfs_period_timer;
3706 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3707 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3710 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3712 cfs_rq->runtime_enabled = 0;
3713 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3716 /* requires cfs_b->lock, may release to reprogram timer */
3717 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3720 * The timer may be active because we're trying to set a new bandwidth
3721 * period or because we're racing with the tear-down path
3722 * (timer_active==0 becomes visible before the hrtimer call-back
3723 * terminates). In either case we ensure that it's re-programmed
3725 while (unlikely(hrtimer_active(&cfs_b->period_timer)) &&
3726 hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) {
3727 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3728 raw_spin_unlock(&cfs_b->lock);
3730 raw_spin_lock(&cfs_b->lock);
3731 /* if someone else restarted the timer then we're done */
3732 if (cfs_b->timer_active)
3736 cfs_b->timer_active = 1;
3737 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3740 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3742 hrtimer_cancel(&cfs_b->period_timer);
3743 hrtimer_cancel(&cfs_b->slack_timer);
3746 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3748 struct cfs_rq *cfs_rq;
3750 for_each_leaf_cfs_rq(rq, cfs_rq) {
3751 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3753 if (!cfs_rq->runtime_enabled)
3757 * clock_task is not advancing so we just need to make sure
3758 * there's some valid quota amount
3760 cfs_rq->runtime_remaining = cfs_b->quota;
3761 if (cfs_rq_throttled(cfs_rq))
3762 unthrottle_cfs_rq(cfs_rq);
3766 #else /* CONFIG_CFS_BANDWIDTH */
3767 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3769 return rq_clock_task(rq_of(cfs_rq));
3772 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
3773 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
3774 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3775 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3777 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3782 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3787 static inline int throttled_lb_pair(struct task_group *tg,
3788 int src_cpu, int dest_cpu)
3793 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3795 #ifdef CONFIG_FAIR_GROUP_SCHED
3796 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3799 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3803 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3804 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3806 #endif /* CONFIG_CFS_BANDWIDTH */
3808 /**************************************************
3809 * CFS operations on tasks:
3812 #ifdef CONFIG_SCHED_HRTICK
3813 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3815 struct sched_entity *se = &p->se;
3816 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3818 WARN_ON(task_rq(p) != rq);
3820 if (cfs_rq->nr_running > 1) {
3821 u64 slice = sched_slice(cfs_rq, se);
3822 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3823 s64 delta = slice - ran;
3832 * Don't schedule slices shorter than 10000ns, that just
3833 * doesn't make sense. Rely on vruntime for fairness.
3836 delta = max_t(s64, 10000LL, delta);
3838 hrtick_start(rq, delta);
3843 * called from enqueue/dequeue and updates the hrtick when the
3844 * current task is from our class and nr_running is low enough
3847 static void hrtick_update(struct rq *rq)
3849 struct task_struct *curr = rq->curr;
3851 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3854 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3855 hrtick_start_fair(rq, curr);
3857 #else /* !CONFIG_SCHED_HRTICK */
3859 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3863 static inline void hrtick_update(struct rq *rq)
3869 * The enqueue_task method is called before nr_running is
3870 * increased. Here we update the fair scheduling stats and
3871 * then put the task into the rbtree:
3874 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3876 struct cfs_rq *cfs_rq;
3877 struct sched_entity *se = &p->se;
3879 for_each_sched_entity(se) {
3882 cfs_rq = cfs_rq_of(se);
3883 enqueue_entity(cfs_rq, se, flags);
3886 * end evaluation on encountering a throttled cfs_rq
3888 * note: in the case of encountering a throttled cfs_rq we will
3889 * post the final h_nr_running increment below.
3891 if (cfs_rq_throttled(cfs_rq))
3893 cfs_rq->h_nr_running++;
3895 flags = ENQUEUE_WAKEUP;
3898 for_each_sched_entity(se) {
3899 cfs_rq = cfs_rq_of(se);
3900 cfs_rq->h_nr_running++;
3902 if (cfs_rq_throttled(cfs_rq))
3905 update_cfs_shares(cfs_rq);
3906 update_entity_load_avg(se, 1);
3910 update_rq_runnable_avg(rq, rq->nr_running);
3916 static void set_next_buddy(struct sched_entity *se);
3919 * The dequeue_task method is called before nr_running is
3920 * decreased. We remove the task from the rbtree and
3921 * update the fair scheduling stats:
3923 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3925 struct cfs_rq *cfs_rq;
3926 struct sched_entity *se = &p->se;
3927 int task_sleep = flags & DEQUEUE_SLEEP;
3929 for_each_sched_entity(se) {
3930 cfs_rq = cfs_rq_of(se);
3931 dequeue_entity(cfs_rq, se, flags);
3934 * end evaluation on encountering a throttled cfs_rq
3936 * note: in the case of encountering a throttled cfs_rq we will
3937 * post the final h_nr_running decrement below.
3939 if (cfs_rq_throttled(cfs_rq))
3941 cfs_rq->h_nr_running--;
3943 /* Don't dequeue parent if it has other entities besides us */
3944 if (cfs_rq->load.weight) {
3946 * Bias pick_next to pick a task from this cfs_rq, as
3947 * p is sleeping when it is within its sched_slice.
3949 if (task_sleep && parent_entity(se))
3950 set_next_buddy(parent_entity(se));
3952 /* avoid re-evaluating load for this entity */
3953 se = parent_entity(se);
3956 flags |= DEQUEUE_SLEEP;
3959 for_each_sched_entity(se) {
3960 cfs_rq = cfs_rq_of(se);
3961 cfs_rq->h_nr_running--;
3963 if (cfs_rq_throttled(cfs_rq))
3966 update_cfs_shares(cfs_rq);
3967 update_entity_load_avg(se, 1);
3972 update_rq_runnable_avg(rq, 1);
3978 /* Used instead of source_load when we know the type == 0 */
3979 static unsigned long weighted_cpuload(const int cpu)
3981 return cpu_rq(cpu)->cfs.runnable_load_avg;
3985 * Return a low guess at the load of a migration-source cpu weighted
3986 * according to the scheduling class and "nice" value.
3988 * We want to under-estimate the load of migration sources, to
3989 * balance conservatively.
3991 static unsigned long source_load(int cpu, int type)
3993 struct rq *rq = cpu_rq(cpu);
3994 unsigned long total = weighted_cpuload(cpu);
3996 if (type == 0 || !sched_feat(LB_BIAS))
3999 return min(rq->cpu_load[type-1], total);
4003 * Return a high guess at the load of a migration-target cpu weighted
4004 * according to the scheduling class and "nice" value.
4006 static unsigned long target_load(int cpu, int type)
4008 struct rq *rq = cpu_rq(cpu);
4009 unsigned long total = weighted_cpuload(cpu);
4011 if (type == 0 || !sched_feat(LB_BIAS))
4014 return max(rq->cpu_load[type-1], total);
4017 static unsigned long power_of(int cpu)
4019 return cpu_rq(cpu)->cpu_power;
4022 static unsigned long cpu_avg_load_per_task(int cpu)
4024 struct rq *rq = cpu_rq(cpu);
4025 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
4026 unsigned long load_avg = rq->cfs.runnable_load_avg;
4029 return load_avg / nr_running;
4034 static void record_wakee(struct task_struct *p)
4037 * Rough decay (wiping) for cost saving, don't worry
4038 * about the boundary, really active task won't care
4041 if (jiffies > current->wakee_flip_decay_ts + HZ) {
4042 current->wakee_flips = 0;
4043 current->wakee_flip_decay_ts = jiffies;
4046 if (current->last_wakee != p) {
4047 current->last_wakee = p;
4048 current->wakee_flips++;
4052 static void task_waking_fair(struct task_struct *p)
4054 struct sched_entity *se = &p->se;
4055 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4058 #ifndef CONFIG_64BIT
4059 u64 min_vruntime_copy;
4062 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4064 min_vruntime = cfs_rq->min_vruntime;
4065 } while (min_vruntime != min_vruntime_copy);
4067 min_vruntime = cfs_rq->min_vruntime;
4070 se->vruntime -= min_vruntime;
4074 #ifdef CONFIG_FAIR_GROUP_SCHED
4076 * effective_load() calculates the load change as seen from the root_task_group
4078 * Adding load to a group doesn't make a group heavier, but can cause movement
4079 * of group shares between cpus. Assuming the shares were perfectly aligned one
4080 * can calculate the shift in shares.
4082 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4083 * on this @cpu and results in a total addition (subtraction) of @wg to the
4084 * total group weight.
4086 * Given a runqueue weight distribution (rw_i) we can compute a shares
4087 * distribution (s_i) using:
4089 * s_i = rw_i / \Sum rw_j (1)
4091 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4092 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4093 * shares distribution (s_i):
4095 * rw_i = { 2, 4, 1, 0 }
4096 * s_i = { 2/7, 4/7, 1/7, 0 }
4098 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4099 * task used to run on and the CPU the waker is running on), we need to
4100 * compute the effect of waking a task on either CPU and, in case of a sync
4101 * wakeup, compute the effect of the current task going to sleep.
4103 * So for a change of @wl to the local @cpu with an overall group weight change
4104 * of @wl we can compute the new shares distribution (s'_i) using:
4106 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4108 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4109 * differences in waking a task to CPU 0. The additional task changes the
4110 * weight and shares distributions like:
4112 * rw'_i = { 3, 4, 1, 0 }
4113 * s'_i = { 3/8, 4/8, 1/8, 0 }
4115 * We can then compute the difference in effective weight by using:
4117 * dw_i = S * (s'_i - s_i) (3)
4119 * Where 'S' is the group weight as seen by its parent.
4121 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4122 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4123 * 4/7) times the weight of the group.
4125 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4127 struct sched_entity *se = tg->se[cpu];
4129 if (!tg->parent) /* the trivial, non-cgroup case */
4132 for_each_sched_entity(se) {
4138 * W = @wg + \Sum rw_j
4140 W = wg + calc_tg_weight(tg, se->my_q);
4145 w = se->my_q->load.weight + wl;
4148 * wl = S * s'_i; see (2)
4151 wl = (w * tg->shares) / W;
4156 * Per the above, wl is the new se->load.weight value; since
4157 * those are clipped to [MIN_SHARES, ...) do so now. See
4158 * calc_cfs_shares().
4160 if (wl < MIN_SHARES)
4164 * wl = dw_i = S * (s'_i - s_i); see (3)
4166 wl -= se->load.weight;
4169 * Recursively apply this logic to all parent groups to compute
4170 * the final effective load change on the root group. Since
4171 * only the @tg group gets extra weight, all parent groups can
4172 * only redistribute existing shares. @wl is the shift in shares
4173 * resulting from this level per the above.
4182 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4189 static int wake_wide(struct task_struct *p)
4191 int factor = this_cpu_read(sd_llc_size);
4194 * Yeah, it's the switching-frequency, could means many wakee or
4195 * rapidly switch, use factor here will just help to automatically
4196 * adjust the loose-degree, so bigger node will lead to more pull.
4198 if (p->wakee_flips > factor) {
4200 * wakee is somewhat hot, it needs certain amount of cpu
4201 * resource, so if waker is far more hot, prefer to leave
4204 if (current->wakee_flips > (factor * p->wakee_flips))
4211 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4213 s64 this_load, load;
4214 int idx, this_cpu, prev_cpu;
4215 unsigned long tl_per_task;
4216 struct task_group *tg;
4217 unsigned long weight;
4221 * If we wake multiple tasks be careful to not bounce
4222 * ourselves around too much.
4228 this_cpu = smp_processor_id();
4229 prev_cpu = task_cpu(p);
4230 load = source_load(prev_cpu, idx);
4231 this_load = target_load(this_cpu, idx);
4234 * If sync wakeup then subtract the (maximum possible)
4235 * effect of the currently running task from the load
4236 * of the current CPU:
4239 tg = task_group(current);
4240 weight = current->se.load.weight;
4242 this_load += effective_load(tg, this_cpu, -weight, -weight);
4243 load += effective_load(tg, prev_cpu, 0, -weight);
4247 weight = p->se.load.weight;
4250 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4251 * due to the sync cause above having dropped this_load to 0, we'll
4252 * always have an imbalance, but there's really nothing you can do
4253 * about that, so that's good too.
4255 * Otherwise check if either cpus are near enough in load to allow this
4256 * task to be woken on this_cpu.
4258 if (this_load > 0) {
4259 s64 this_eff_load, prev_eff_load;
4261 this_eff_load = 100;
4262 this_eff_load *= power_of(prev_cpu);
4263 this_eff_load *= this_load +
4264 effective_load(tg, this_cpu, weight, weight);
4266 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4267 prev_eff_load *= power_of(this_cpu);
4268 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4270 balanced = this_eff_load <= prev_eff_load;
4275 * If the currently running task will sleep within
4276 * a reasonable amount of time then attract this newly
4279 if (sync && balanced)
4282 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4283 tl_per_task = cpu_avg_load_per_task(this_cpu);
4286 (this_load <= load &&
4287 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
4289 * This domain has SD_WAKE_AFFINE and
4290 * p is cache cold in this domain, and
4291 * there is no bad imbalance.
4293 schedstat_inc(sd, ttwu_move_affine);
4294 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4302 * find_idlest_group finds and returns the least busy CPU group within the
4305 static struct sched_group *
4306 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4307 int this_cpu, int sd_flag)
4309 struct sched_group *idlest = NULL, *group = sd->groups;
4310 unsigned long min_load = ULONG_MAX, this_load = 0;
4311 int load_idx = sd->forkexec_idx;
4312 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4314 if (sd_flag & SD_BALANCE_WAKE)
4315 load_idx = sd->wake_idx;
4318 unsigned long load, avg_load;
4322 /* Skip over this group if it has no CPUs allowed */
4323 if (!cpumask_intersects(sched_group_cpus(group),
4324 tsk_cpus_allowed(p)))
4327 local_group = cpumask_test_cpu(this_cpu,
4328 sched_group_cpus(group));
4330 /* Tally up the load of all CPUs in the group */
4333 for_each_cpu(i, sched_group_cpus(group)) {
4334 /* Bias balancing toward cpus of our domain */
4336 load = source_load(i, load_idx);
4338 load = target_load(i, load_idx);
4343 /* Adjust by relative CPU power of the group */
4344 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
4347 this_load = avg_load;
4348 } else if (avg_load < min_load) {
4349 min_load = avg_load;
4352 } while (group = group->next, group != sd->groups);
4354 if (!idlest || 100*this_load < imbalance*min_load)
4360 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4363 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4365 unsigned long load, min_load = ULONG_MAX;
4369 /* Traverse only the allowed CPUs */
4370 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4371 load = weighted_cpuload(i);
4373 if (load < min_load || (load == min_load && i == this_cpu)) {
4383 * Try and locate an idle CPU in the sched_domain.
4385 static int select_idle_sibling(struct task_struct *p, int target)
4387 struct sched_domain *sd;
4388 struct sched_group *sg;
4389 int i = task_cpu(p);
4391 if (idle_cpu(target))
4395 * If the prevous cpu is cache affine and idle, don't be stupid.
4397 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4401 * Otherwise, iterate the domains and find an elegible idle cpu.
4403 sd = rcu_dereference(per_cpu(sd_llc, target));
4404 for_each_lower_domain(sd) {
4407 if (!cpumask_intersects(sched_group_cpus(sg),
4408 tsk_cpus_allowed(p)))
4411 for_each_cpu(i, sched_group_cpus(sg)) {
4412 if (i == target || !idle_cpu(i))
4416 target = cpumask_first_and(sched_group_cpus(sg),
4417 tsk_cpus_allowed(p));
4421 } while (sg != sd->groups);
4428 * select_task_rq_fair: Select target runqueue for the waking task in domains
4429 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4430 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4432 * Balances load by selecting the idlest cpu in the idlest group, or under
4433 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4435 * Returns the target cpu number.
4437 * preempt must be disabled.
4440 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4442 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4443 int cpu = smp_processor_id();
4445 int want_affine = 0;
4446 int sync = wake_flags & WF_SYNC;
4448 if (p->nr_cpus_allowed == 1)
4451 if (sd_flag & SD_BALANCE_WAKE) {
4452 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
4458 for_each_domain(cpu, tmp) {
4459 if (!(tmp->flags & SD_LOAD_BALANCE))
4463 * If both cpu and prev_cpu are part of this domain,
4464 * cpu is a valid SD_WAKE_AFFINE target.
4466 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4467 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4472 if (tmp->flags & sd_flag)
4477 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4480 new_cpu = select_idle_sibling(p, prev_cpu);
4485 struct sched_group *group;
4488 if (!(sd->flags & sd_flag)) {
4493 group = find_idlest_group(sd, p, cpu, sd_flag);
4499 new_cpu = find_idlest_cpu(group, p, cpu);
4500 if (new_cpu == -1 || new_cpu == cpu) {
4501 /* Now try balancing at a lower domain level of cpu */
4506 /* Now try balancing at a lower domain level of new_cpu */
4508 weight = sd->span_weight;
4510 for_each_domain(cpu, tmp) {
4511 if (weight <= tmp->span_weight)
4513 if (tmp->flags & sd_flag)
4516 /* while loop will break here if sd == NULL */
4525 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4526 * cfs_rq_of(p) references at time of call are still valid and identify the
4527 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4528 * other assumptions, including the state of rq->lock, should be made.
4531 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4533 struct sched_entity *se = &p->se;
4534 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4537 * Load tracking: accumulate removed load so that it can be processed
4538 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4539 * to blocked load iff they have a positive decay-count. It can never
4540 * be negative here since on-rq tasks have decay-count == 0.
4542 if (se->avg.decay_count) {
4543 se->avg.decay_count = -__synchronize_entity_decay(se);
4544 atomic_long_add(se->avg.load_avg_contrib,
4545 &cfs_rq->removed_load);
4548 #endif /* CONFIG_SMP */
4550 static unsigned long
4551 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4553 unsigned long gran = sysctl_sched_wakeup_granularity;
4556 * Since its curr running now, convert the gran from real-time
4557 * to virtual-time in his units.
4559 * By using 'se' instead of 'curr' we penalize light tasks, so
4560 * they get preempted easier. That is, if 'se' < 'curr' then
4561 * the resulting gran will be larger, therefore penalizing the
4562 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4563 * be smaller, again penalizing the lighter task.
4565 * This is especially important for buddies when the leftmost
4566 * task is higher priority than the buddy.
4568 return calc_delta_fair(gran, se);
4572 * Should 'se' preempt 'curr'.
4586 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4588 s64 gran, vdiff = curr->vruntime - se->vruntime;
4593 gran = wakeup_gran(curr, se);
4600 static void set_last_buddy(struct sched_entity *se)
4602 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4605 for_each_sched_entity(se)
4606 cfs_rq_of(se)->last = se;
4609 static void set_next_buddy(struct sched_entity *se)
4611 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4614 for_each_sched_entity(se)
4615 cfs_rq_of(se)->next = se;
4618 static void set_skip_buddy(struct sched_entity *se)
4620 for_each_sched_entity(se)
4621 cfs_rq_of(se)->skip = se;
4625 * Preempt the current task with a newly woken task if needed:
4627 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4629 struct task_struct *curr = rq->curr;
4630 struct sched_entity *se = &curr->se, *pse = &p->se;
4631 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4632 int scale = cfs_rq->nr_running >= sched_nr_latency;
4633 int next_buddy_marked = 0;
4635 if (unlikely(se == pse))
4639 * This is possible from callers such as move_task(), in which we
4640 * unconditionally check_prempt_curr() after an enqueue (which may have
4641 * lead to a throttle). This both saves work and prevents false
4642 * next-buddy nomination below.
4644 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4647 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4648 set_next_buddy(pse);
4649 next_buddy_marked = 1;
4653 * We can come here with TIF_NEED_RESCHED already set from new task
4656 * Note: this also catches the edge-case of curr being in a throttled
4657 * group (e.g. via set_curr_task), since update_curr() (in the
4658 * enqueue of curr) will have resulted in resched being set. This
4659 * prevents us from potentially nominating it as a false LAST_BUDDY
4662 if (test_tsk_need_resched(curr))
4665 /* Idle tasks are by definition preempted by non-idle tasks. */
4666 if (unlikely(curr->policy == SCHED_IDLE) &&
4667 likely(p->policy != SCHED_IDLE))
4671 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4672 * is driven by the tick):
4674 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4677 find_matching_se(&se, &pse);
4678 update_curr(cfs_rq_of(se));
4680 if (wakeup_preempt_entity(se, pse) == 1) {
4682 * Bias pick_next to pick the sched entity that is
4683 * triggering this preemption.
4685 if (!next_buddy_marked)
4686 set_next_buddy(pse);
4695 * Only set the backward buddy when the current task is still
4696 * on the rq. This can happen when a wakeup gets interleaved
4697 * with schedule on the ->pre_schedule() or idle_balance()
4698 * point, either of which can * drop the rq lock.
4700 * Also, during early boot the idle thread is in the fair class,
4701 * for obvious reasons its a bad idea to schedule back to it.
4703 if (unlikely(!se->on_rq || curr == rq->idle))
4706 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4710 static struct task_struct *
4711 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
4713 struct cfs_rq *cfs_rq = &rq->cfs;
4714 struct sched_entity *se;
4715 struct task_struct *p;
4719 #ifdef CONFIG_FAIR_GROUP_SCHED
4720 if (!cfs_rq->nr_running)
4723 if (prev->sched_class != &fair_sched_class)
4727 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
4728 * likely that a next task is from the same cgroup as the current.
4730 * Therefore attempt to avoid putting and setting the entire cgroup
4731 * hierarchy, only change the part that actually changes.
4735 struct sched_entity *curr = cfs_rq->curr;
4738 * Since we got here without doing put_prev_entity() we also
4739 * have to consider cfs_rq->curr. If it is still a runnable
4740 * entity, update_curr() will update its vruntime, otherwise
4741 * forget we've ever seen it.
4743 if (curr && curr->on_rq)
4744 update_curr(cfs_rq);
4749 * This call to check_cfs_rq_runtime() will do the throttle and
4750 * dequeue its entity in the parent(s). Therefore the 'simple'
4751 * nr_running test will indeed be correct.
4753 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
4756 se = pick_next_entity(cfs_rq, curr);
4757 cfs_rq = group_cfs_rq(se);
4763 * Since we haven't yet done put_prev_entity and if the selected task
4764 * is a different task than we started out with, try and touch the
4765 * least amount of cfs_rqs.
4768 struct sched_entity *pse = &prev->se;
4770 while (!(cfs_rq = is_same_group(se, pse))) {
4771 int se_depth = se->depth;
4772 int pse_depth = pse->depth;
4774 if (se_depth <= pse_depth) {
4775 put_prev_entity(cfs_rq_of(pse), pse);
4776 pse = parent_entity(pse);
4778 if (se_depth >= pse_depth) {
4779 set_next_entity(cfs_rq_of(se), se);
4780 se = parent_entity(se);
4784 put_prev_entity(cfs_rq, pse);
4785 set_next_entity(cfs_rq, se);
4788 if (hrtick_enabled(rq))
4789 hrtick_start_fair(rq, p);
4796 if (!cfs_rq->nr_running)
4799 put_prev_task(rq, prev);
4802 se = pick_next_entity(cfs_rq, NULL);
4803 set_next_entity(cfs_rq, se);
4804 cfs_rq = group_cfs_rq(se);
4809 if (hrtick_enabled(rq))
4810 hrtick_start_fair(rq, p);
4815 new_tasks = idle_balance(rq);
4817 * Because idle_balance() releases (and re-acquires) rq->lock, it is
4818 * possible for any higher priority task to appear. In that case we
4819 * must re-start the pick_next_entity() loop.
4831 * Account for a descheduled task:
4833 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4835 struct sched_entity *se = &prev->se;
4836 struct cfs_rq *cfs_rq;
4838 for_each_sched_entity(se) {
4839 cfs_rq = cfs_rq_of(se);
4840 put_prev_entity(cfs_rq, se);
4845 * sched_yield() is very simple
4847 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4849 static void yield_task_fair(struct rq *rq)
4851 struct task_struct *curr = rq->curr;
4852 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4853 struct sched_entity *se = &curr->se;
4856 * Are we the only task in the tree?
4858 if (unlikely(rq->nr_running == 1))
4861 clear_buddies(cfs_rq, se);
4863 if (curr->policy != SCHED_BATCH) {
4864 update_rq_clock(rq);
4866 * Update run-time statistics of the 'current'.
4868 update_curr(cfs_rq);
4870 * Tell update_rq_clock() that we've just updated,
4871 * so we don't do microscopic update in schedule()
4872 * and double the fastpath cost.
4874 rq->skip_clock_update = 1;
4880 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4882 struct sched_entity *se = &p->se;
4884 /* throttled hierarchies are not runnable */
4885 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4888 /* Tell the scheduler that we'd really like pse to run next. */
4891 yield_task_fair(rq);
4897 /**************************************************
4898 * Fair scheduling class load-balancing methods.
4902 * The purpose of load-balancing is to achieve the same basic fairness the
4903 * per-cpu scheduler provides, namely provide a proportional amount of compute
4904 * time to each task. This is expressed in the following equation:
4906 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4908 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4909 * W_i,0 is defined as:
4911 * W_i,0 = \Sum_j w_i,j (2)
4913 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4914 * is derived from the nice value as per prio_to_weight[].
4916 * The weight average is an exponential decay average of the instantaneous
4919 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4921 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4922 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4923 * can also include other factors [XXX].
4925 * To achieve this balance we define a measure of imbalance which follows
4926 * directly from (1):
4928 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4930 * We them move tasks around to minimize the imbalance. In the continuous
4931 * function space it is obvious this converges, in the discrete case we get
4932 * a few fun cases generally called infeasible weight scenarios.
4935 * - infeasible weights;
4936 * - local vs global optima in the discrete case. ]
4941 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4942 * for all i,j solution, we create a tree of cpus that follows the hardware
4943 * topology where each level pairs two lower groups (or better). This results
4944 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4945 * tree to only the first of the previous level and we decrease the frequency
4946 * of load-balance at each level inv. proportional to the number of cpus in
4952 * \Sum { --- * --- * 2^i } = O(n) (5)
4954 * `- size of each group
4955 * | | `- number of cpus doing load-balance
4957 * `- sum over all levels
4959 * Coupled with a limit on how many tasks we can migrate every balance pass,
4960 * this makes (5) the runtime complexity of the balancer.
4962 * An important property here is that each CPU is still (indirectly) connected
4963 * to every other cpu in at most O(log n) steps:
4965 * The adjacency matrix of the resulting graph is given by:
4968 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4971 * And you'll find that:
4973 * A^(log_2 n)_i,j != 0 for all i,j (7)
4975 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4976 * The task movement gives a factor of O(m), giving a convergence complexity
4979 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4984 * In order to avoid CPUs going idle while there's still work to do, new idle
4985 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4986 * tree itself instead of relying on other CPUs to bring it work.
4988 * This adds some complexity to both (5) and (8) but it reduces the total idle
4996 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4999 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5004 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5006 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5008 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5011 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5012 * rewrite all of this once again.]
5015 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5017 enum fbq_type { regular, remote, all };
5019 #define LBF_ALL_PINNED 0x01
5020 #define LBF_NEED_BREAK 0x02
5021 #define LBF_DST_PINNED 0x04
5022 #define LBF_SOME_PINNED 0x08
5025 struct sched_domain *sd;
5033 struct cpumask *dst_grpmask;
5035 enum cpu_idle_type idle;
5037 /* The set of CPUs under consideration for load-balancing */
5038 struct cpumask *cpus;
5043 unsigned int loop_break;
5044 unsigned int loop_max;
5046 enum fbq_type fbq_type;
5050 * move_task - move a task from one runqueue to another runqueue.
5051 * Both runqueues must be locked.
5053 static void move_task(struct task_struct *p, struct lb_env *env)
5055 deactivate_task(env->src_rq, p, 0);
5056 set_task_cpu(p, env->dst_cpu);
5057 activate_task(env->dst_rq, p, 0);
5058 check_preempt_curr(env->dst_rq, p, 0);
5062 * Is this task likely cache-hot:
5065 task_hot(struct task_struct *p, u64 now)
5069 if (p->sched_class != &fair_sched_class)
5072 if (unlikely(p->policy == SCHED_IDLE))
5076 * Buddy candidates are cache hot:
5078 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
5079 (&p->se == cfs_rq_of(&p->se)->next ||
5080 &p->se == cfs_rq_of(&p->se)->last))
5083 if (sysctl_sched_migration_cost == -1)
5085 if (sysctl_sched_migration_cost == 0)
5088 delta = now - p->se.exec_start;
5090 return delta < (s64)sysctl_sched_migration_cost;
5093 #ifdef CONFIG_NUMA_BALANCING
5094 /* Returns true if the destination node has incurred more faults */
5095 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
5097 int src_nid, dst_nid;
5099 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults_memory ||
5100 !(env->sd->flags & SD_NUMA)) {
5104 src_nid = cpu_to_node(env->src_cpu);
5105 dst_nid = cpu_to_node(env->dst_cpu);
5107 if (src_nid == dst_nid)
5110 /* Always encourage migration to the preferred node. */
5111 if (dst_nid == p->numa_preferred_nid)
5114 /* If both task and group weight improve, this move is a winner. */
5115 if (task_weight(p, dst_nid) > task_weight(p, src_nid) &&
5116 group_weight(p, dst_nid) > group_weight(p, src_nid))
5123 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5125 int src_nid, dst_nid;
5127 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
5130 if (!p->numa_faults_memory || !(env->sd->flags & SD_NUMA))
5133 src_nid = cpu_to_node(env->src_cpu);
5134 dst_nid = cpu_to_node(env->dst_cpu);
5136 if (src_nid == dst_nid)
5139 /* Migrating away from the preferred node is always bad. */
5140 if (src_nid == p->numa_preferred_nid)
5143 /* If either task or group weight get worse, don't do it. */
5144 if (task_weight(p, dst_nid) < task_weight(p, src_nid) ||
5145 group_weight(p, dst_nid) < group_weight(p, src_nid))
5152 static inline bool migrate_improves_locality(struct task_struct *p,
5158 static inline bool migrate_degrades_locality(struct task_struct *p,
5166 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5169 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5171 int tsk_cache_hot = 0;
5173 * We do not migrate tasks that are:
5174 * 1) throttled_lb_pair, or
5175 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5176 * 3) running (obviously), or
5177 * 4) are cache-hot on their current CPU.
5179 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5182 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5185 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5187 env->flags |= LBF_SOME_PINNED;
5190 * Remember if this task can be migrated to any other cpu in
5191 * our sched_group. We may want to revisit it if we couldn't
5192 * meet load balance goals by pulling other tasks on src_cpu.
5194 * Also avoid computing new_dst_cpu if we have already computed
5195 * one in current iteration.
5197 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5200 /* Prevent to re-select dst_cpu via env's cpus */
5201 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5202 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5203 env->flags |= LBF_DST_PINNED;
5204 env->new_dst_cpu = cpu;
5212 /* Record that we found atleast one task that could run on dst_cpu */
5213 env->flags &= ~LBF_ALL_PINNED;
5215 if (task_running(env->src_rq, p)) {
5216 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5221 * Aggressive migration if:
5222 * 1) destination numa is preferred
5223 * 2) task is cache cold, or
5224 * 3) too many balance attempts have failed.
5226 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq));
5228 tsk_cache_hot = migrate_degrades_locality(p, env);
5230 if (migrate_improves_locality(p, env)) {
5231 #ifdef CONFIG_SCHEDSTATS
5232 if (tsk_cache_hot) {
5233 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5234 schedstat_inc(p, se.statistics.nr_forced_migrations);
5240 if (!tsk_cache_hot ||
5241 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5243 if (tsk_cache_hot) {
5244 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5245 schedstat_inc(p, se.statistics.nr_forced_migrations);
5251 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5256 * move_one_task tries to move exactly one task from busiest to this_rq, as
5257 * part of active balancing operations within "domain".
5258 * Returns 1 if successful and 0 otherwise.
5260 * Called with both runqueues locked.
5262 static int move_one_task(struct lb_env *env)
5264 struct task_struct *p, *n;
5266 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5267 if (!can_migrate_task(p, env))
5272 * Right now, this is only the second place move_task()
5273 * is called, so we can safely collect move_task()
5274 * stats here rather than inside move_task().
5276 schedstat_inc(env->sd, lb_gained[env->idle]);
5282 static const unsigned int sched_nr_migrate_break = 32;
5285 * move_tasks tries to move up to imbalance weighted load from busiest to
5286 * this_rq, as part of a balancing operation within domain "sd".
5287 * Returns 1 if successful and 0 otherwise.
5289 * Called with both runqueues locked.
5291 static int move_tasks(struct lb_env *env)
5293 struct list_head *tasks = &env->src_rq->cfs_tasks;
5294 struct task_struct *p;
5298 if (env->imbalance <= 0)
5301 while (!list_empty(tasks)) {
5302 p = list_first_entry(tasks, struct task_struct, se.group_node);
5305 /* We've more or less seen every task there is, call it quits */
5306 if (env->loop > env->loop_max)
5309 /* take a breather every nr_migrate tasks */
5310 if (env->loop > env->loop_break) {
5311 env->loop_break += sched_nr_migrate_break;
5312 env->flags |= LBF_NEED_BREAK;
5316 if (!can_migrate_task(p, env))
5319 load = task_h_load(p);
5321 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5324 if ((load / 2) > env->imbalance)
5329 env->imbalance -= load;
5331 #ifdef CONFIG_PREEMPT
5333 * NEWIDLE balancing is a source of latency, so preemptible
5334 * kernels will stop after the first task is pulled to minimize
5335 * the critical section.
5337 if (env->idle == CPU_NEWLY_IDLE)
5342 * We only want to steal up to the prescribed amount of
5345 if (env->imbalance <= 0)
5350 list_move_tail(&p->se.group_node, tasks);
5354 * Right now, this is one of only two places move_task() is called,
5355 * so we can safely collect move_task() stats here rather than
5356 * inside move_task().
5358 schedstat_add(env->sd, lb_gained[env->idle], pulled);
5363 #ifdef CONFIG_FAIR_GROUP_SCHED
5365 * update tg->load_weight by folding this cpu's load_avg
5367 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5369 struct sched_entity *se = tg->se[cpu];
5370 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5372 /* throttled entities do not contribute to load */
5373 if (throttled_hierarchy(cfs_rq))
5376 update_cfs_rq_blocked_load(cfs_rq, 1);
5379 update_entity_load_avg(se, 1);
5381 * We pivot on our runnable average having decayed to zero for
5382 * list removal. This generally implies that all our children
5383 * have also been removed (modulo rounding error or bandwidth
5384 * control); however, such cases are rare and we can fix these
5387 * TODO: fix up out-of-order children on enqueue.
5389 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5390 list_del_leaf_cfs_rq(cfs_rq);
5392 struct rq *rq = rq_of(cfs_rq);
5393 update_rq_runnable_avg(rq, rq->nr_running);
5397 static void update_blocked_averages(int cpu)
5399 struct rq *rq = cpu_rq(cpu);
5400 struct cfs_rq *cfs_rq;
5401 unsigned long flags;
5403 raw_spin_lock_irqsave(&rq->lock, flags);
5404 update_rq_clock(rq);
5406 * Iterates the task_group tree in a bottom up fashion, see
5407 * list_add_leaf_cfs_rq() for details.
5409 for_each_leaf_cfs_rq(rq, cfs_rq) {
5411 * Note: We may want to consider periodically releasing
5412 * rq->lock about these updates so that creating many task
5413 * groups does not result in continually extending hold time.
5415 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5418 raw_spin_unlock_irqrestore(&rq->lock, flags);
5422 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5423 * This needs to be done in a top-down fashion because the load of a child
5424 * group is a fraction of its parents load.
5426 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5428 struct rq *rq = rq_of(cfs_rq);
5429 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5430 unsigned long now = jiffies;
5433 if (cfs_rq->last_h_load_update == now)
5436 cfs_rq->h_load_next = NULL;
5437 for_each_sched_entity(se) {
5438 cfs_rq = cfs_rq_of(se);
5439 cfs_rq->h_load_next = se;
5440 if (cfs_rq->last_h_load_update == now)
5445 cfs_rq->h_load = cfs_rq->runnable_load_avg;
5446 cfs_rq->last_h_load_update = now;
5449 while ((se = cfs_rq->h_load_next) != NULL) {
5450 load = cfs_rq->h_load;
5451 load = div64_ul(load * se->avg.load_avg_contrib,
5452 cfs_rq->runnable_load_avg + 1);
5453 cfs_rq = group_cfs_rq(se);
5454 cfs_rq->h_load = load;
5455 cfs_rq->last_h_load_update = now;
5459 static unsigned long task_h_load(struct task_struct *p)
5461 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5463 update_cfs_rq_h_load(cfs_rq);
5464 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
5465 cfs_rq->runnable_load_avg + 1);
5468 static inline void update_blocked_averages(int cpu)
5472 static unsigned long task_h_load(struct task_struct *p)
5474 return p->se.avg.load_avg_contrib;
5478 /********** Helpers for find_busiest_group ************************/
5480 * sg_lb_stats - stats of a sched_group required for load_balancing
5482 struct sg_lb_stats {
5483 unsigned long avg_load; /*Avg load across the CPUs of the group */
5484 unsigned long group_load; /* Total load over the CPUs of the group */
5485 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5486 unsigned long load_per_task;
5487 unsigned long group_power;
5488 unsigned int sum_nr_running; /* Nr tasks running in the group */
5489 unsigned int group_capacity;
5490 unsigned int idle_cpus;
5491 unsigned int group_weight;
5492 int group_imb; /* Is there an imbalance in the group ? */
5493 int group_has_capacity; /* Is there extra capacity in the group? */
5494 #ifdef CONFIG_NUMA_BALANCING
5495 unsigned int nr_numa_running;
5496 unsigned int nr_preferred_running;
5501 * sd_lb_stats - Structure to store the statistics of a sched_domain
5502 * during load balancing.
5504 struct sd_lb_stats {
5505 struct sched_group *busiest; /* Busiest group in this sd */
5506 struct sched_group *local; /* Local group in this sd */
5507 unsigned long total_load; /* Total load of all groups in sd */
5508 unsigned long total_pwr; /* Total power of all groups in sd */
5509 unsigned long avg_load; /* Average load across all groups in sd */
5511 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5512 struct sg_lb_stats local_stat; /* Statistics of the local group */
5515 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5518 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5519 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5520 * We must however clear busiest_stat::avg_load because
5521 * update_sd_pick_busiest() reads this before assignment.
5523 *sds = (struct sd_lb_stats){
5535 * get_sd_load_idx - Obtain the load index for a given sched domain.
5536 * @sd: The sched_domain whose load_idx is to be obtained.
5537 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5539 * Return: The load index.
5541 static inline int get_sd_load_idx(struct sched_domain *sd,
5542 enum cpu_idle_type idle)
5548 load_idx = sd->busy_idx;
5551 case CPU_NEWLY_IDLE:
5552 load_idx = sd->newidle_idx;
5555 load_idx = sd->idle_idx;
5562 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
5564 return SCHED_POWER_SCALE;
5567 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
5569 return default_scale_freq_power(sd, cpu);
5572 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
5574 unsigned long weight = sd->span_weight;
5575 unsigned long smt_gain = sd->smt_gain;
5582 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
5584 return default_scale_smt_power(sd, cpu);
5587 static unsigned long scale_rt_power(int cpu)
5589 struct rq *rq = cpu_rq(cpu);
5590 u64 total, available, age_stamp, avg;
5594 * Since we're reading these variables without serialization make sure
5595 * we read them once before doing sanity checks on them.
5597 age_stamp = ACCESS_ONCE(rq->age_stamp);
5598 avg = ACCESS_ONCE(rq->rt_avg);
5600 delta = rq_clock(rq) - age_stamp;
5601 if (unlikely(delta < 0))
5604 total = sched_avg_period() + delta;
5606 if (unlikely(total < avg)) {
5607 /* Ensures that power won't end up being negative */
5610 available = total - avg;
5613 if (unlikely((s64)total < SCHED_POWER_SCALE))
5614 total = SCHED_POWER_SCALE;
5616 total >>= SCHED_POWER_SHIFT;
5618 return div_u64(available, total);
5621 static void update_cpu_power(struct sched_domain *sd, int cpu)
5623 unsigned long weight = sd->span_weight;
5624 unsigned long power = SCHED_POWER_SCALE;
5625 struct sched_group *sdg = sd->groups;
5627 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
5628 if (sched_feat(ARCH_POWER))
5629 power *= arch_scale_smt_power(sd, cpu);
5631 power *= default_scale_smt_power(sd, cpu);
5633 power >>= SCHED_POWER_SHIFT;
5636 sdg->sgp->power_orig = power;
5638 if (sched_feat(ARCH_POWER))
5639 power *= arch_scale_freq_power(sd, cpu);
5641 power *= default_scale_freq_power(sd, cpu);
5643 power >>= SCHED_POWER_SHIFT;
5645 power *= scale_rt_power(cpu);
5646 power >>= SCHED_POWER_SHIFT;
5651 cpu_rq(cpu)->cpu_power = power;
5652 sdg->sgp->power = power;
5655 void update_group_power(struct sched_domain *sd, int cpu)
5657 struct sched_domain *child = sd->child;
5658 struct sched_group *group, *sdg = sd->groups;
5659 unsigned long power, power_orig;
5660 unsigned long interval;
5662 interval = msecs_to_jiffies(sd->balance_interval);
5663 interval = clamp(interval, 1UL, max_load_balance_interval);
5664 sdg->sgp->next_update = jiffies + interval;
5667 update_cpu_power(sd, cpu);
5671 power_orig = power = 0;
5673 if (child->flags & SD_OVERLAP) {
5675 * SD_OVERLAP domains cannot assume that child groups
5676 * span the current group.
5679 for_each_cpu(cpu, sched_group_cpus(sdg)) {
5680 struct sched_group_power *sgp;
5681 struct rq *rq = cpu_rq(cpu);
5684 * build_sched_domains() -> init_sched_groups_power()
5685 * gets here before we've attached the domains to the
5688 * Use power_of(), which is set irrespective of domains
5689 * in update_cpu_power().
5691 * This avoids power/power_orig from being 0 and
5692 * causing divide-by-zero issues on boot.
5694 * Runtime updates will correct power_orig.
5696 if (unlikely(!rq->sd)) {
5697 power_orig += power_of(cpu);
5698 power += power_of(cpu);
5702 sgp = rq->sd->groups->sgp;
5703 power_orig += sgp->power_orig;
5704 power += sgp->power;
5708 * !SD_OVERLAP domains can assume that child groups
5709 * span the current group.
5712 group = child->groups;
5714 power_orig += group->sgp->power_orig;
5715 power += group->sgp->power;
5716 group = group->next;
5717 } while (group != child->groups);
5720 sdg->sgp->power_orig = power_orig;
5721 sdg->sgp->power = power;
5725 * Try and fix up capacity for tiny siblings, this is needed when
5726 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5727 * which on its own isn't powerful enough.
5729 * See update_sd_pick_busiest() and check_asym_packing().
5732 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5735 * Only siblings can have significantly less than SCHED_POWER_SCALE
5737 if (!(sd->flags & SD_SHARE_CPUPOWER))
5741 * If ~90% of the cpu_power is still there, we're good.
5743 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
5750 * Group imbalance indicates (and tries to solve) the problem where balancing
5751 * groups is inadequate due to tsk_cpus_allowed() constraints.
5753 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5754 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5757 * { 0 1 2 3 } { 4 5 6 7 }
5760 * If we were to balance group-wise we'd place two tasks in the first group and
5761 * two tasks in the second group. Clearly this is undesired as it will overload
5762 * cpu 3 and leave one of the cpus in the second group unused.
5764 * The current solution to this issue is detecting the skew in the first group
5765 * by noticing the lower domain failed to reach balance and had difficulty
5766 * moving tasks due to affinity constraints.
5768 * When this is so detected; this group becomes a candidate for busiest; see
5769 * update_sd_pick_busiest(). And calculate_imbalance() and
5770 * find_busiest_group() avoid some of the usual balance conditions to allow it
5771 * to create an effective group imbalance.
5773 * This is a somewhat tricky proposition since the next run might not find the
5774 * group imbalance and decide the groups need to be balanced again. A most
5775 * subtle and fragile situation.
5778 static inline int sg_imbalanced(struct sched_group *group)
5780 return group->sgp->imbalance;
5784 * Compute the group capacity.
5786 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
5787 * first dividing out the smt factor and computing the actual number of cores
5788 * and limit power unit capacity with that.
5790 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
5792 unsigned int capacity, smt, cpus;
5793 unsigned int power, power_orig;
5795 power = group->sgp->power;
5796 power_orig = group->sgp->power_orig;
5797 cpus = group->group_weight;
5799 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
5800 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
5801 capacity = cpus / smt; /* cores */
5803 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
5805 capacity = fix_small_capacity(env->sd, group);
5811 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5812 * @env: The load balancing environment.
5813 * @group: sched_group whose statistics are to be updated.
5814 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5815 * @local_group: Does group contain this_cpu.
5816 * @sgs: variable to hold the statistics for this group.
5818 static inline void update_sg_lb_stats(struct lb_env *env,
5819 struct sched_group *group, int load_idx,
5820 int local_group, struct sg_lb_stats *sgs)
5825 memset(sgs, 0, sizeof(*sgs));
5827 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5828 struct rq *rq = cpu_rq(i);
5830 /* Bias balancing toward cpus of our domain */
5832 load = target_load(i, load_idx);
5834 load = source_load(i, load_idx);
5836 sgs->group_load += load;
5837 sgs->sum_nr_running += rq->nr_running;
5838 #ifdef CONFIG_NUMA_BALANCING
5839 sgs->nr_numa_running += rq->nr_numa_running;
5840 sgs->nr_preferred_running += rq->nr_preferred_running;
5842 sgs->sum_weighted_load += weighted_cpuload(i);
5847 /* Adjust by relative CPU power of the group */
5848 sgs->group_power = group->sgp->power;
5849 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
5851 if (sgs->sum_nr_running)
5852 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
5854 sgs->group_weight = group->group_weight;
5856 sgs->group_imb = sg_imbalanced(group);
5857 sgs->group_capacity = sg_capacity(env, group);
5859 if (sgs->group_capacity > sgs->sum_nr_running)
5860 sgs->group_has_capacity = 1;
5864 * update_sd_pick_busiest - return 1 on busiest group
5865 * @env: The load balancing environment.
5866 * @sds: sched_domain statistics
5867 * @sg: sched_group candidate to be checked for being the busiest
5868 * @sgs: sched_group statistics
5870 * Determine if @sg is a busier group than the previously selected
5873 * Return: %true if @sg is a busier group than the previously selected
5874 * busiest group. %false otherwise.
5876 static bool update_sd_pick_busiest(struct lb_env *env,
5877 struct sd_lb_stats *sds,
5878 struct sched_group *sg,
5879 struct sg_lb_stats *sgs)
5881 if (sgs->avg_load <= sds->busiest_stat.avg_load)
5884 if (sgs->sum_nr_running > sgs->group_capacity)
5891 * ASYM_PACKING needs to move all the work to the lowest
5892 * numbered CPUs in the group, therefore mark all groups
5893 * higher than ourself as busy.
5895 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5896 env->dst_cpu < group_first_cpu(sg)) {
5900 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5907 #ifdef CONFIG_NUMA_BALANCING
5908 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5910 if (sgs->sum_nr_running > sgs->nr_numa_running)
5912 if (sgs->sum_nr_running > sgs->nr_preferred_running)
5917 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5919 if (rq->nr_running > rq->nr_numa_running)
5921 if (rq->nr_running > rq->nr_preferred_running)
5926 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5931 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5935 #endif /* CONFIG_NUMA_BALANCING */
5938 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5939 * @env: The load balancing environment.
5940 * @sds: variable to hold the statistics for this sched_domain.
5942 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
5944 struct sched_domain *child = env->sd->child;
5945 struct sched_group *sg = env->sd->groups;
5946 struct sg_lb_stats tmp_sgs;
5947 int load_idx, prefer_sibling = 0;
5949 if (child && child->flags & SD_PREFER_SIBLING)
5952 load_idx = get_sd_load_idx(env->sd, env->idle);
5955 struct sg_lb_stats *sgs = &tmp_sgs;
5958 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
5961 sgs = &sds->local_stat;
5963 if (env->idle != CPU_NEWLY_IDLE ||
5964 time_after_eq(jiffies, sg->sgp->next_update))
5965 update_group_power(env->sd, env->dst_cpu);
5968 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
5974 * In case the child domain prefers tasks go to siblings
5975 * first, lower the sg capacity to one so that we'll try
5976 * and move all the excess tasks away. We lower the capacity
5977 * of a group only if the local group has the capacity to fit
5978 * these excess tasks, i.e. nr_running < group_capacity. The
5979 * extra check prevents the case where you always pull from the
5980 * heaviest group when it is already under-utilized (possible
5981 * with a large weight task outweighs the tasks on the system).
5983 if (prefer_sibling && sds->local &&
5984 sds->local_stat.group_has_capacity)
5985 sgs->group_capacity = min(sgs->group_capacity, 1U);
5987 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
5989 sds->busiest_stat = *sgs;
5993 /* Now, start updating sd_lb_stats */
5994 sds->total_load += sgs->group_load;
5995 sds->total_pwr += sgs->group_power;
5998 } while (sg != env->sd->groups);
6000 if (env->sd->flags & SD_NUMA)
6001 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6005 * check_asym_packing - Check to see if the group is packed into the
6008 * This is primarily intended to used at the sibling level. Some
6009 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6010 * case of POWER7, it can move to lower SMT modes only when higher
6011 * threads are idle. When in lower SMT modes, the threads will
6012 * perform better since they share less core resources. Hence when we
6013 * have idle threads, we want them to be the higher ones.
6015 * This packing function is run on idle threads. It checks to see if
6016 * the busiest CPU in this domain (core in the P7 case) has a higher
6017 * CPU number than the packing function is being run on. Here we are
6018 * assuming lower CPU number will be equivalent to lower a SMT thread
6021 * Return: 1 when packing is required and a task should be moved to
6022 * this CPU. The amount of the imbalance is returned in *imbalance.
6024 * @env: The load balancing environment.
6025 * @sds: Statistics of the sched_domain which is to be packed
6027 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6031 if (!(env->sd->flags & SD_ASYM_PACKING))
6037 busiest_cpu = group_first_cpu(sds->busiest);
6038 if (env->dst_cpu > busiest_cpu)
6041 env->imbalance = DIV_ROUND_CLOSEST(
6042 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
6049 * fix_small_imbalance - Calculate the minor imbalance that exists
6050 * amongst the groups of a sched_domain, during
6052 * @env: The load balancing environment.
6053 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6056 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6058 unsigned long tmp, pwr_now = 0, pwr_move = 0;
6059 unsigned int imbn = 2;
6060 unsigned long scaled_busy_load_per_task;
6061 struct sg_lb_stats *local, *busiest;
6063 local = &sds->local_stat;
6064 busiest = &sds->busiest_stat;
6066 if (!local->sum_nr_running)
6067 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6068 else if (busiest->load_per_task > local->load_per_task)
6071 scaled_busy_load_per_task =
6072 (busiest->load_per_task * SCHED_POWER_SCALE) /
6073 busiest->group_power;
6075 if (busiest->avg_load + scaled_busy_load_per_task >=
6076 local->avg_load + (scaled_busy_load_per_task * imbn)) {
6077 env->imbalance = busiest->load_per_task;
6082 * OK, we don't have enough imbalance to justify moving tasks,
6083 * however we may be able to increase total CPU power used by
6087 pwr_now += busiest->group_power *
6088 min(busiest->load_per_task, busiest->avg_load);
6089 pwr_now += local->group_power *
6090 min(local->load_per_task, local->avg_load);
6091 pwr_now /= SCHED_POWER_SCALE;
6093 /* Amount of load we'd subtract */
6094 if (busiest->avg_load > scaled_busy_load_per_task) {
6095 pwr_move += busiest->group_power *
6096 min(busiest->load_per_task,
6097 busiest->avg_load - scaled_busy_load_per_task);
6100 /* Amount of load we'd add */
6101 if (busiest->avg_load * busiest->group_power <
6102 busiest->load_per_task * SCHED_POWER_SCALE) {
6103 tmp = (busiest->avg_load * busiest->group_power) /
6106 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
6109 pwr_move += local->group_power *
6110 min(local->load_per_task, local->avg_load + tmp);
6111 pwr_move /= SCHED_POWER_SCALE;
6113 /* Move if we gain throughput */
6114 if (pwr_move > pwr_now)
6115 env->imbalance = busiest->load_per_task;
6119 * calculate_imbalance - Calculate the amount of imbalance present within the
6120 * groups of a given sched_domain during load balance.
6121 * @env: load balance environment
6122 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6124 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6126 unsigned long max_pull, load_above_capacity = ~0UL;
6127 struct sg_lb_stats *local, *busiest;
6129 local = &sds->local_stat;
6130 busiest = &sds->busiest_stat;
6132 if (busiest->group_imb) {
6134 * In the group_imb case we cannot rely on group-wide averages
6135 * to ensure cpu-load equilibrium, look at wider averages. XXX
6137 busiest->load_per_task =
6138 min(busiest->load_per_task, sds->avg_load);
6142 * In the presence of smp nice balancing, certain scenarios can have
6143 * max load less than avg load(as we skip the groups at or below
6144 * its cpu_power, while calculating max_load..)
6146 if (busiest->avg_load <= sds->avg_load ||
6147 local->avg_load >= sds->avg_load) {
6149 return fix_small_imbalance(env, sds);
6152 if (!busiest->group_imb) {
6154 * Don't want to pull so many tasks that a group would go idle.
6155 * Except of course for the group_imb case, since then we might
6156 * have to drop below capacity to reach cpu-load equilibrium.
6158 load_above_capacity =
6159 (busiest->sum_nr_running - busiest->group_capacity);
6161 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
6162 load_above_capacity /= busiest->group_power;
6166 * We're trying to get all the cpus to the average_load, so we don't
6167 * want to push ourselves above the average load, nor do we wish to
6168 * reduce the max loaded cpu below the average load. At the same time,
6169 * we also don't want to reduce the group load below the group capacity
6170 * (so that we can implement power-savings policies etc). Thus we look
6171 * for the minimum possible imbalance.
6173 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6175 /* How much load to actually move to equalise the imbalance */
6176 env->imbalance = min(
6177 max_pull * busiest->group_power,
6178 (sds->avg_load - local->avg_load) * local->group_power
6179 ) / SCHED_POWER_SCALE;
6182 * if *imbalance is less than the average load per runnable task
6183 * there is no guarantee that any tasks will be moved so we'll have
6184 * a think about bumping its value to force at least one task to be
6187 if (env->imbalance < busiest->load_per_task)
6188 return fix_small_imbalance(env, sds);
6191 /******* find_busiest_group() helpers end here *********************/
6194 * find_busiest_group - Returns the busiest group within the sched_domain
6195 * if there is an imbalance. If there isn't an imbalance, and
6196 * the user has opted for power-savings, it returns a group whose
6197 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6198 * such a group exists.
6200 * Also calculates the amount of weighted load which should be moved
6201 * to restore balance.
6203 * @env: The load balancing environment.
6205 * Return: - The busiest group if imbalance exists.
6206 * - If no imbalance and user has opted for power-savings balance,
6207 * return the least loaded group whose CPUs can be
6208 * put to idle by rebalancing its tasks onto our group.
6210 static struct sched_group *find_busiest_group(struct lb_env *env)
6212 struct sg_lb_stats *local, *busiest;
6213 struct sd_lb_stats sds;
6215 init_sd_lb_stats(&sds);
6218 * Compute the various statistics relavent for load balancing at
6221 update_sd_lb_stats(env, &sds);
6222 local = &sds.local_stat;
6223 busiest = &sds.busiest_stat;
6225 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6226 check_asym_packing(env, &sds))
6229 /* There is no busy sibling group to pull tasks from */
6230 if (!sds.busiest || busiest->sum_nr_running == 0)
6233 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
6236 * If the busiest group is imbalanced the below checks don't
6237 * work because they assume all things are equal, which typically
6238 * isn't true due to cpus_allowed constraints and the like.
6240 if (busiest->group_imb)
6243 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6244 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
6245 !busiest->group_has_capacity)
6249 * If the local group is more busy than the selected busiest group
6250 * don't try and pull any tasks.
6252 if (local->avg_load >= busiest->avg_load)
6256 * Don't pull any tasks if this group is already above the domain
6259 if (local->avg_load >= sds.avg_load)
6262 if (env->idle == CPU_IDLE) {
6264 * This cpu is idle. If the busiest group load doesn't
6265 * have more tasks than the number of available cpu's and
6266 * there is no imbalance between this and busiest group
6267 * wrt to idle cpu's, it is balanced.
6269 if ((local->idle_cpus < busiest->idle_cpus) &&
6270 busiest->sum_nr_running <= busiest->group_weight)
6274 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6275 * imbalance_pct to be conservative.
6277 if (100 * busiest->avg_load <=
6278 env->sd->imbalance_pct * local->avg_load)
6283 /* Looks like there is an imbalance. Compute it */
6284 calculate_imbalance(env, &sds);
6293 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6295 static struct rq *find_busiest_queue(struct lb_env *env,
6296 struct sched_group *group)
6298 struct rq *busiest = NULL, *rq;
6299 unsigned long busiest_load = 0, busiest_power = 1;
6302 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6303 unsigned long power, capacity, wl;
6307 rt = fbq_classify_rq(rq);
6310 * We classify groups/runqueues into three groups:
6311 * - regular: there are !numa tasks
6312 * - remote: there are numa tasks that run on the 'wrong' node
6313 * - all: there is no distinction
6315 * In order to avoid migrating ideally placed numa tasks,
6316 * ignore those when there's better options.
6318 * If we ignore the actual busiest queue to migrate another
6319 * task, the next balance pass can still reduce the busiest
6320 * queue by moving tasks around inside the node.
6322 * If we cannot move enough load due to this classification
6323 * the next pass will adjust the group classification and
6324 * allow migration of more tasks.
6326 * Both cases only affect the total convergence complexity.
6328 if (rt > env->fbq_type)
6331 power = power_of(i);
6332 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
6334 capacity = fix_small_capacity(env->sd, group);
6336 wl = weighted_cpuload(i);
6339 * When comparing with imbalance, use weighted_cpuload()
6340 * which is not scaled with the cpu power.
6342 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
6346 * For the load comparisons with the other cpu's, consider
6347 * the weighted_cpuload() scaled with the cpu power, so that
6348 * the load can be moved away from the cpu that is potentially
6349 * running at a lower capacity.
6351 * Thus we're looking for max(wl_i / power_i), crosswise
6352 * multiplication to rid ourselves of the division works out
6353 * to: wl_i * power_j > wl_j * power_i; where j is our
6356 if (wl * busiest_power > busiest_load * power) {
6358 busiest_power = power;
6367 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6368 * so long as it is large enough.
6370 #define MAX_PINNED_INTERVAL 512
6372 /* Working cpumask for load_balance and load_balance_newidle. */
6373 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6375 static int need_active_balance(struct lb_env *env)
6377 struct sched_domain *sd = env->sd;
6379 if (env->idle == CPU_NEWLY_IDLE) {
6382 * ASYM_PACKING needs to force migrate tasks from busy but
6383 * higher numbered CPUs in order to pack all tasks in the
6384 * lowest numbered CPUs.
6386 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6390 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6393 static int active_load_balance_cpu_stop(void *data);
6395 static int should_we_balance(struct lb_env *env)
6397 struct sched_group *sg = env->sd->groups;
6398 struct cpumask *sg_cpus, *sg_mask;
6399 int cpu, balance_cpu = -1;
6402 * In the newly idle case, we will allow all the cpu's
6403 * to do the newly idle load balance.
6405 if (env->idle == CPU_NEWLY_IDLE)
6408 sg_cpus = sched_group_cpus(sg);
6409 sg_mask = sched_group_mask(sg);
6410 /* Try to find first idle cpu */
6411 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6412 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6419 if (balance_cpu == -1)
6420 balance_cpu = group_balance_cpu(sg);
6423 * First idle cpu or the first cpu(busiest) in this sched group
6424 * is eligible for doing load balancing at this and above domains.
6426 return balance_cpu == env->dst_cpu;
6430 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6431 * tasks if there is an imbalance.
6433 static int load_balance(int this_cpu, struct rq *this_rq,
6434 struct sched_domain *sd, enum cpu_idle_type idle,
6435 int *continue_balancing)
6437 int ld_moved, cur_ld_moved, active_balance = 0;
6438 struct sched_domain *sd_parent = sd->parent;
6439 struct sched_group *group;
6441 unsigned long flags;
6442 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
6444 struct lb_env env = {
6446 .dst_cpu = this_cpu,
6448 .dst_grpmask = sched_group_cpus(sd->groups),
6450 .loop_break = sched_nr_migrate_break,
6456 * For NEWLY_IDLE load_balancing, we don't need to consider
6457 * other cpus in our group
6459 if (idle == CPU_NEWLY_IDLE)
6460 env.dst_grpmask = NULL;
6462 cpumask_copy(cpus, cpu_active_mask);
6464 schedstat_inc(sd, lb_count[idle]);
6467 if (!should_we_balance(&env)) {
6468 *continue_balancing = 0;
6472 group = find_busiest_group(&env);
6474 schedstat_inc(sd, lb_nobusyg[idle]);
6478 busiest = find_busiest_queue(&env, group);
6480 schedstat_inc(sd, lb_nobusyq[idle]);
6484 BUG_ON(busiest == env.dst_rq);
6486 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6489 if (busiest->nr_running > 1) {
6491 * Attempt to move tasks. If find_busiest_group has found
6492 * an imbalance but busiest->nr_running <= 1, the group is
6493 * still unbalanced. ld_moved simply stays zero, so it is
6494 * correctly treated as an imbalance.
6496 env.flags |= LBF_ALL_PINNED;
6497 env.src_cpu = busiest->cpu;
6498 env.src_rq = busiest;
6499 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
6502 local_irq_save(flags);
6503 double_rq_lock(env.dst_rq, busiest);
6506 * cur_ld_moved - load moved in current iteration
6507 * ld_moved - cumulative load moved across iterations
6509 cur_ld_moved = move_tasks(&env);
6510 ld_moved += cur_ld_moved;
6511 double_rq_unlock(env.dst_rq, busiest);
6512 local_irq_restore(flags);
6515 * some other cpu did the load balance for us.
6517 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
6518 resched_cpu(env.dst_cpu);
6520 if (env.flags & LBF_NEED_BREAK) {
6521 env.flags &= ~LBF_NEED_BREAK;
6526 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6527 * us and move them to an alternate dst_cpu in our sched_group
6528 * where they can run. The upper limit on how many times we
6529 * iterate on same src_cpu is dependent on number of cpus in our
6532 * This changes load balance semantics a bit on who can move
6533 * load to a given_cpu. In addition to the given_cpu itself
6534 * (or a ilb_cpu acting on its behalf where given_cpu is
6535 * nohz-idle), we now have balance_cpu in a position to move
6536 * load to given_cpu. In rare situations, this may cause
6537 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6538 * _independently_ and at _same_ time to move some load to
6539 * given_cpu) causing exceess load to be moved to given_cpu.
6540 * This however should not happen so much in practice and
6541 * moreover subsequent load balance cycles should correct the
6542 * excess load moved.
6544 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
6546 /* Prevent to re-select dst_cpu via env's cpus */
6547 cpumask_clear_cpu(env.dst_cpu, env.cpus);
6549 env.dst_rq = cpu_rq(env.new_dst_cpu);
6550 env.dst_cpu = env.new_dst_cpu;
6551 env.flags &= ~LBF_DST_PINNED;
6553 env.loop_break = sched_nr_migrate_break;
6556 * Go back to "more_balance" rather than "redo" since we
6557 * need to continue with same src_cpu.
6563 * We failed to reach balance because of affinity.
6566 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
6568 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
6569 *group_imbalance = 1;
6570 } else if (*group_imbalance)
6571 *group_imbalance = 0;
6574 /* All tasks on this runqueue were pinned by CPU affinity */
6575 if (unlikely(env.flags & LBF_ALL_PINNED)) {
6576 cpumask_clear_cpu(cpu_of(busiest), cpus);
6577 if (!cpumask_empty(cpus)) {
6579 env.loop_break = sched_nr_migrate_break;
6587 schedstat_inc(sd, lb_failed[idle]);
6589 * Increment the failure counter only on periodic balance.
6590 * We do not want newidle balance, which can be very
6591 * frequent, pollute the failure counter causing
6592 * excessive cache_hot migrations and active balances.
6594 if (idle != CPU_NEWLY_IDLE)
6595 sd->nr_balance_failed++;
6597 if (need_active_balance(&env)) {
6598 raw_spin_lock_irqsave(&busiest->lock, flags);
6600 /* don't kick the active_load_balance_cpu_stop,
6601 * if the curr task on busiest cpu can't be
6604 if (!cpumask_test_cpu(this_cpu,
6605 tsk_cpus_allowed(busiest->curr))) {
6606 raw_spin_unlock_irqrestore(&busiest->lock,
6608 env.flags |= LBF_ALL_PINNED;
6609 goto out_one_pinned;
6613 * ->active_balance synchronizes accesses to
6614 * ->active_balance_work. Once set, it's cleared
6615 * only after active load balance is finished.
6617 if (!busiest->active_balance) {
6618 busiest->active_balance = 1;
6619 busiest->push_cpu = this_cpu;
6622 raw_spin_unlock_irqrestore(&busiest->lock, flags);
6624 if (active_balance) {
6625 stop_one_cpu_nowait(cpu_of(busiest),
6626 active_load_balance_cpu_stop, busiest,
6627 &busiest->active_balance_work);
6631 * We've kicked active balancing, reset the failure
6634 sd->nr_balance_failed = sd->cache_nice_tries+1;
6637 sd->nr_balance_failed = 0;
6639 if (likely(!active_balance)) {
6640 /* We were unbalanced, so reset the balancing interval */
6641 sd->balance_interval = sd->min_interval;
6644 * If we've begun active balancing, start to back off. This
6645 * case may not be covered by the all_pinned logic if there
6646 * is only 1 task on the busy runqueue (because we don't call
6649 if (sd->balance_interval < sd->max_interval)
6650 sd->balance_interval *= 2;
6656 schedstat_inc(sd, lb_balanced[idle]);
6658 sd->nr_balance_failed = 0;
6661 /* tune up the balancing interval */
6662 if (((env.flags & LBF_ALL_PINNED) &&
6663 sd->balance_interval < MAX_PINNED_INTERVAL) ||
6664 (sd->balance_interval < sd->max_interval))
6665 sd->balance_interval *= 2;
6673 * idle_balance is called by schedule() if this_cpu is about to become
6674 * idle. Attempts to pull tasks from other CPUs.
6676 static int idle_balance(struct rq *this_rq)
6678 struct sched_domain *sd;
6679 int pulled_task = 0;
6680 unsigned long next_balance = jiffies + HZ;
6682 int this_cpu = this_rq->cpu;
6684 idle_enter_fair(this_rq);
6687 * We must set idle_stamp _before_ calling idle_balance(), such that we
6688 * measure the duration of idle_balance() as idle time.
6690 this_rq->idle_stamp = rq_clock(this_rq);
6692 if (this_rq->avg_idle < sysctl_sched_migration_cost)
6696 * Drop the rq->lock, but keep IRQ/preempt disabled.
6698 raw_spin_unlock(&this_rq->lock);
6700 update_blocked_averages(this_cpu);
6702 for_each_domain(this_cpu, sd) {
6703 unsigned long interval;
6704 int continue_balancing = 1;
6705 u64 t0, domain_cost;
6707 if (!(sd->flags & SD_LOAD_BALANCE))
6710 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
6713 if (sd->flags & SD_BALANCE_NEWIDLE) {
6714 t0 = sched_clock_cpu(this_cpu);
6716 pulled_task = load_balance(this_cpu, this_rq,
6718 &continue_balancing);
6720 domain_cost = sched_clock_cpu(this_cpu) - t0;
6721 if (domain_cost > sd->max_newidle_lb_cost)
6722 sd->max_newidle_lb_cost = domain_cost;
6724 curr_cost += domain_cost;
6727 interval = msecs_to_jiffies(sd->balance_interval);
6728 if (time_after(next_balance, sd->last_balance + interval))
6729 next_balance = sd->last_balance + interval;
6732 * Stop searching for tasks to pull if there are
6733 * now runnable tasks on this rq.
6735 if (pulled_task || this_rq->nr_running > 0)
6740 raw_spin_lock(&this_rq->lock);
6742 if (curr_cost > this_rq->max_idle_balance_cost)
6743 this_rq->max_idle_balance_cost = curr_cost;
6746 * While browsing the domains, we released the rq lock, a task could
6747 * have been enqueued in the meantime. Since we're not going idle,
6748 * pretend we pulled a task.
6750 if (this_rq->cfs.h_nr_running && !pulled_task)
6753 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
6755 * We are going idle. next_balance may be set based on
6756 * a busy processor. So reset next_balance.
6758 this_rq->next_balance = next_balance;
6762 /* Is there a task of a high priority class? */
6763 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
6767 idle_exit_fair(this_rq);
6768 this_rq->idle_stamp = 0;
6775 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6776 * running tasks off the busiest CPU onto idle CPUs. It requires at
6777 * least 1 task to be running on each physical CPU where possible, and
6778 * avoids physical / logical imbalances.
6780 static int active_load_balance_cpu_stop(void *data)
6782 struct rq *busiest_rq = data;
6783 int busiest_cpu = cpu_of(busiest_rq);
6784 int target_cpu = busiest_rq->push_cpu;
6785 struct rq *target_rq = cpu_rq(target_cpu);
6786 struct sched_domain *sd;
6788 raw_spin_lock_irq(&busiest_rq->lock);
6790 /* make sure the requested cpu hasn't gone down in the meantime */
6791 if (unlikely(busiest_cpu != smp_processor_id() ||
6792 !busiest_rq->active_balance))
6795 /* Is there any task to move? */
6796 if (busiest_rq->nr_running <= 1)
6800 * This condition is "impossible", if it occurs
6801 * we need to fix it. Originally reported by
6802 * Bjorn Helgaas on a 128-cpu setup.
6804 BUG_ON(busiest_rq == target_rq);
6806 /* move a task from busiest_rq to target_rq */
6807 double_lock_balance(busiest_rq, target_rq);
6809 /* Search for an sd spanning us and the target CPU. */
6811 for_each_domain(target_cpu, sd) {
6812 if ((sd->flags & SD_LOAD_BALANCE) &&
6813 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6818 struct lb_env env = {
6820 .dst_cpu = target_cpu,
6821 .dst_rq = target_rq,
6822 .src_cpu = busiest_rq->cpu,
6823 .src_rq = busiest_rq,
6827 schedstat_inc(sd, alb_count);
6829 if (move_one_task(&env))
6830 schedstat_inc(sd, alb_pushed);
6832 schedstat_inc(sd, alb_failed);
6835 double_unlock_balance(busiest_rq, target_rq);
6837 busiest_rq->active_balance = 0;
6838 raw_spin_unlock_irq(&busiest_rq->lock);
6842 static inline int on_null_domain(struct rq *rq)
6844 return unlikely(!rcu_dereference_sched(rq->sd));
6847 #ifdef CONFIG_NO_HZ_COMMON
6849 * idle load balancing details
6850 * - When one of the busy CPUs notice that there may be an idle rebalancing
6851 * needed, they will kick the idle load balancer, which then does idle
6852 * load balancing for all the idle CPUs.
6855 cpumask_var_t idle_cpus_mask;
6857 unsigned long next_balance; /* in jiffy units */
6858 } nohz ____cacheline_aligned;
6860 static inline int find_new_ilb(void)
6862 int ilb = cpumask_first(nohz.idle_cpus_mask);
6864 if (ilb < nr_cpu_ids && idle_cpu(ilb))
6871 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6872 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6873 * CPU (if there is one).
6875 static void nohz_balancer_kick(void)
6879 nohz.next_balance++;
6881 ilb_cpu = find_new_ilb();
6883 if (ilb_cpu >= nr_cpu_ids)
6886 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
6889 * Use smp_send_reschedule() instead of resched_cpu().
6890 * This way we generate a sched IPI on the target cpu which
6891 * is idle. And the softirq performing nohz idle load balance
6892 * will be run before returning from the IPI.
6894 smp_send_reschedule(ilb_cpu);
6898 static inline void nohz_balance_exit_idle(int cpu)
6900 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
6902 * Completely isolated CPUs don't ever set, so we must test.
6904 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
6905 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
6906 atomic_dec(&nohz.nr_cpus);
6908 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6912 static inline void set_cpu_sd_state_busy(void)
6914 struct sched_domain *sd;
6915 int cpu = smp_processor_id();
6918 sd = rcu_dereference(per_cpu(sd_busy, cpu));
6920 if (!sd || !sd->nohz_idle)
6924 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
6929 void set_cpu_sd_state_idle(void)
6931 struct sched_domain *sd;
6932 int cpu = smp_processor_id();
6935 sd = rcu_dereference(per_cpu(sd_busy, cpu));
6937 if (!sd || sd->nohz_idle)
6941 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
6947 * This routine will record that the cpu is going idle with tick stopped.
6948 * This info will be used in performing idle load balancing in the future.
6950 void nohz_balance_enter_idle(int cpu)
6953 * If this cpu is going down, then nothing needs to be done.
6955 if (!cpu_active(cpu))
6958 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
6962 * If we're a completely isolated CPU, we don't play.
6964 if (on_null_domain(cpu_rq(cpu)))
6967 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
6968 atomic_inc(&nohz.nr_cpus);
6969 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6972 static int sched_ilb_notifier(struct notifier_block *nfb,
6973 unsigned long action, void *hcpu)
6975 switch (action & ~CPU_TASKS_FROZEN) {
6977 nohz_balance_exit_idle(smp_processor_id());
6985 static DEFINE_SPINLOCK(balancing);
6988 * Scale the max load_balance interval with the number of CPUs in the system.
6989 * This trades load-balance latency on larger machines for less cross talk.
6991 void update_max_interval(void)
6993 max_load_balance_interval = HZ*num_online_cpus()/10;
6997 * It checks each scheduling domain to see if it is due to be balanced,
6998 * and initiates a balancing operation if so.
7000 * Balancing parameters are set up in init_sched_domains.
7002 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7004 int continue_balancing = 1;
7006 unsigned long interval;
7007 struct sched_domain *sd;
7008 /* Earliest time when we have to do rebalance again */
7009 unsigned long next_balance = jiffies + 60*HZ;
7010 int update_next_balance = 0;
7011 int need_serialize, need_decay = 0;
7014 update_blocked_averages(cpu);
7017 for_each_domain(cpu, sd) {
7019 * Decay the newidle max times here because this is a regular
7020 * visit to all the domains. Decay ~1% per second.
7022 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7023 sd->max_newidle_lb_cost =
7024 (sd->max_newidle_lb_cost * 253) / 256;
7025 sd->next_decay_max_lb_cost = jiffies + HZ;
7028 max_cost += sd->max_newidle_lb_cost;
7030 if (!(sd->flags & SD_LOAD_BALANCE))
7034 * Stop the load balance at this level. There is another
7035 * CPU in our sched group which is doing load balancing more
7038 if (!continue_balancing) {
7044 interval = sd->balance_interval;
7045 if (idle != CPU_IDLE)
7046 interval *= sd->busy_factor;
7048 /* scale ms to jiffies */
7049 interval = msecs_to_jiffies(interval);
7050 interval = clamp(interval, 1UL, max_load_balance_interval);
7052 need_serialize = sd->flags & SD_SERIALIZE;
7054 if (need_serialize) {
7055 if (!spin_trylock(&balancing))
7059 if (time_after_eq(jiffies, sd->last_balance + interval)) {
7060 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7062 * The LBF_DST_PINNED logic could have changed
7063 * env->dst_cpu, so we can't know our idle
7064 * state even if we migrated tasks. Update it.
7066 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7068 sd->last_balance = jiffies;
7071 spin_unlock(&balancing);
7073 if (time_after(next_balance, sd->last_balance + interval)) {
7074 next_balance = sd->last_balance + interval;
7075 update_next_balance = 1;
7080 * Ensure the rq-wide value also decays but keep it at a
7081 * reasonable floor to avoid funnies with rq->avg_idle.
7083 rq->max_idle_balance_cost =
7084 max((u64)sysctl_sched_migration_cost, max_cost);
7089 * next_balance will be updated only when there is a need.
7090 * When the cpu is attached to null domain for ex, it will not be
7093 if (likely(update_next_balance))
7094 rq->next_balance = next_balance;
7097 #ifdef CONFIG_NO_HZ_COMMON
7099 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7100 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7102 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7104 int this_cpu = this_rq->cpu;
7108 if (idle != CPU_IDLE ||
7109 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7112 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7113 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7117 * If this cpu gets work to do, stop the load balancing
7118 * work being done for other cpus. Next load
7119 * balancing owner will pick it up.
7124 rq = cpu_rq(balance_cpu);
7126 raw_spin_lock_irq(&rq->lock);
7127 update_rq_clock(rq);
7128 update_idle_cpu_load(rq);
7129 raw_spin_unlock_irq(&rq->lock);
7131 rebalance_domains(rq, CPU_IDLE);
7133 if (time_after(this_rq->next_balance, rq->next_balance))
7134 this_rq->next_balance = rq->next_balance;
7136 nohz.next_balance = this_rq->next_balance;
7138 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7142 * Current heuristic for kicking the idle load balancer in the presence
7143 * of an idle cpu is the system.
7144 * - This rq has more than one task.
7145 * - At any scheduler domain level, this cpu's scheduler group has multiple
7146 * busy cpu's exceeding the group's power.
7147 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7148 * domain span are idle.
7150 static inline int nohz_kick_needed(struct rq *rq)
7152 unsigned long now = jiffies;
7153 struct sched_domain *sd;
7154 struct sched_group_power *sgp;
7155 int nr_busy, cpu = rq->cpu;
7157 if (unlikely(rq->idle_balance))
7161 * We may be recently in ticked or tickless idle mode. At the first
7162 * busy tick after returning from idle, we will update the busy stats.
7164 set_cpu_sd_state_busy();
7165 nohz_balance_exit_idle(cpu);
7168 * None are in tickless mode and hence no need for NOHZ idle load
7171 if (likely(!atomic_read(&nohz.nr_cpus)))
7174 if (time_before(now, nohz.next_balance))
7177 if (rq->nr_running >= 2)
7181 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7184 sgp = sd->groups->sgp;
7185 nr_busy = atomic_read(&sgp->nr_busy_cpus);
7188 goto need_kick_unlock;
7191 sd = rcu_dereference(per_cpu(sd_asym, cpu));
7193 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7194 sched_domain_span(sd)) < cpu))
7195 goto need_kick_unlock;
7206 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7210 * run_rebalance_domains is triggered when needed from the scheduler tick.
7211 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7213 static void run_rebalance_domains(struct softirq_action *h)
7215 struct rq *this_rq = this_rq();
7216 enum cpu_idle_type idle = this_rq->idle_balance ?
7217 CPU_IDLE : CPU_NOT_IDLE;
7219 rebalance_domains(this_rq, idle);
7222 * If this cpu has a pending nohz_balance_kick, then do the
7223 * balancing on behalf of the other idle cpus whose ticks are
7226 nohz_idle_balance(this_rq, idle);
7230 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7232 void trigger_load_balance(struct rq *rq)
7234 /* Don't need to rebalance while attached to NULL domain */
7235 if (unlikely(on_null_domain(rq)))
7238 if (time_after_eq(jiffies, rq->next_balance))
7239 raise_softirq(SCHED_SOFTIRQ);
7240 #ifdef CONFIG_NO_HZ_COMMON
7241 if (nohz_kick_needed(rq))
7242 nohz_balancer_kick();
7246 static void rq_online_fair(struct rq *rq)
7251 static void rq_offline_fair(struct rq *rq)
7255 /* Ensure any throttled groups are reachable by pick_next_task */
7256 unthrottle_offline_cfs_rqs(rq);
7259 #endif /* CONFIG_SMP */
7262 * scheduler tick hitting a task of our scheduling class:
7264 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7266 struct cfs_rq *cfs_rq;
7267 struct sched_entity *se = &curr->se;
7269 for_each_sched_entity(se) {
7270 cfs_rq = cfs_rq_of(se);
7271 entity_tick(cfs_rq, se, queued);
7274 if (numabalancing_enabled)
7275 task_tick_numa(rq, curr);
7277 update_rq_runnable_avg(rq, 1);
7281 * called on fork with the child task as argument from the parent's context
7282 * - child not yet on the tasklist
7283 * - preemption disabled
7285 static void task_fork_fair(struct task_struct *p)
7287 struct cfs_rq *cfs_rq;
7288 struct sched_entity *se = &p->se, *curr;
7289 int this_cpu = smp_processor_id();
7290 struct rq *rq = this_rq();
7291 unsigned long flags;
7293 raw_spin_lock_irqsave(&rq->lock, flags);
7295 update_rq_clock(rq);
7297 cfs_rq = task_cfs_rq(current);
7298 curr = cfs_rq->curr;
7301 * Not only the cpu but also the task_group of the parent might have
7302 * been changed after parent->se.parent,cfs_rq were copied to
7303 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7304 * of child point to valid ones.
7307 __set_task_cpu(p, this_cpu);
7310 update_curr(cfs_rq);
7313 se->vruntime = curr->vruntime;
7314 place_entity(cfs_rq, se, 1);
7316 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
7318 * Upon rescheduling, sched_class::put_prev_task() will place
7319 * 'current' within the tree based on its new key value.
7321 swap(curr->vruntime, se->vruntime);
7322 resched_task(rq->curr);
7325 se->vruntime -= cfs_rq->min_vruntime;
7327 raw_spin_unlock_irqrestore(&rq->lock, flags);
7331 * Priority of the task has changed. Check to see if we preempt
7335 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
7341 * Reschedule if we are currently running on this runqueue and
7342 * our priority decreased, or if we are not currently running on
7343 * this runqueue and our priority is higher than the current's
7345 if (rq->curr == p) {
7346 if (p->prio > oldprio)
7347 resched_task(rq->curr);
7349 check_preempt_curr(rq, p, 0);
7352 static void switched_from_fair(struct rq *rq, struct task_struct *p)
7354 struct sched_entity *se = &p->se;
7355 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7358 * Ensure the task's vruntime is normalized, so that when it's
7359 * switched back to the fair class the enqueue_entity(.flags=0) will
7360 * do the right thing.
7362 * If it's on_rq, then the dequeue_entity(.flags=0) will already
7363 * have normalized the vruntime, if it's !on_rq, then only when
7364 * the task is sleeping will it still have non-normalized vruntime.
7366 if (!p->on_rq && p->state != TASK_RUNNING) {
7368 * Fix up our vruntime so that the current sleep doesn't
7369 * cause 'unlimited' sleep bonus.
7371 place_entity(cfs_rq, se, 0);
7372 se->vruntime -= cfs_rq->min_vruntime;
7377 * Remove our load from contribution when we leave sched_fair
7378 * and ensure we don't carry in an old decay_count if we
7381 if (se->avg.decay_count) {
7382 __synchronize_entity_decay(se);
7383 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
7389 * We switched to the sched_fair class.
7391 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7393 struct sched_entity *se = &p->se;
7394 #ifdef CONFIG_FAIR_GROUP_SCHED
7396 * Since the real-depth could have been changed (only FAIR
7397 * class maintain depth value), reset depth properly.
7399 se->depth = se->parent ? se->parent->depth + 1 : 0;
7405 * We were most likely switched from sched_rt, so
7406 * kick off the schedule if running, otherwise just see
7407 * if we can still preempt the current task.
7410 resched_task(rq->curr);
7412 check_preempt_curr(rq, p, 0);
7415 /* Account for a task changing its policy or group.
7417 * This routine is mostly called to set cfs_rq->curr field when a task
7418 * migrates between groups/classes.
7420 static void set_curr_task_fair(struct rq *rq)
7422 struct sched_entity *se = &rq->curr->se;
7424 for_each_sched_entity(se) {
7425 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7427 set_next_entity(cfs_rq, se);
7428 /* ensure bandwidth has been allocated on our new cfs_rq */
7429 account_cfs_rq_runtime(cfs_rq, 0);
7433 void init_cfs_rq(struct cfs_rq *cfs_rq)
7435 cfs_rq->tasks_timeline = RB_ROOT;
7436 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7437 #ifndef CONFIG_64BIT
7438 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7441 atomic64_set(&cfs_rq->decay_counter, 1);
7442 atomic_long_set(&cfs_rq->removed_load, 0);
7446 #ifdef CONFIG_FAIR_GROUP_SCHED
7447 static void task_move_group_fair(struct task_struct *p, int on_rq)
7449 struct sched_entity *se = &p->se;
7450 struct cfs_rq *cfs_rq;
7453 * If the task was not on the rq at the time of this cgroup movement
7454 * it must have been asleep, sleeping tasks keep their ->vruntime
7455 * absolute on their old rq until wakeup (needed for the fair sleeper
7456 * bonus in place_entity()).
7458 * If it was on the rq, we've just 'preempted' it, which does convert
7459 * ->vruntime to a relative base.
7461 * Make sure both cases convert their relative position when migrating
7462 * to another cgroup's rq. This does somewhat interfere with the
7463 * fair sleeper stuff for the first placement, but who cares.
7466 * When !on_rq, vruntime of the task has usually NOT been normalized.
7467 * But there are some cases where it has already been normalized:
7469 * - Moving a forked child which is waiting for being woken up by
7470 * wake_up_new_task().
7471 * - Moving a task which has been woken up by try_to_wake_up() and
7472 * waiting for actually being woken up by sched_ttwu_pending().
7474 * To prevent boost or penalty in the new cfs_rq caused by delta
7475 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7477 if (!on_rq && (!se->sum_exec_runtime || p->state == TASK_WAKING))
7481 se->vruntime -= cfs_rq_of(se)->min_vruntime;
7482 set_task_rq(p, task_cpu(p));
7483 se->depth = se->parent ? se->parent->depth + 1 : 0;
7485 cfs_rq = cfs_rq_of(se);
7486 se->vruntime += cfs_rq->min_vruntime;
7489 * migrate_task_rq_fair() will have removed our previous
7490 * contribution, but we must synchronize for ongoing future
7493 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7494 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
7499 void free_fair_sched_group(struct task_group *tg)
7503 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7505 for_each_possible_cpu(i) {
7507 kfree(tg->cfs_rq[i]);
7516 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7518 struct cfs_rq *cfs_rq;
7519 struct sched_entity *se;
7522 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7525 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7529 tg->shares = NICE_0_LOAD;
7531 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7533 for_each_possible_cpu(i) {
7534 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7535 GFP_KERNEL, cpu_to_node(i));
7539 se = kzalloc_node(sizeof(struct sched_entity),
7540 GFP_KERNEL, cpu_to_node(i));
7544 init_cfs_rq(cfs_rq);
7545 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7556 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7558 struct rq *rq = cpu_rq(cpu);
7559 unsigned long flags;
7562 * Only empty task groups can be destroyed; so we can speculatively
7563 * check on_list without danger of it being re-added.
7565 if (!tg->cfs_rq[cpu]->on_list)
7568 raw_spin_lock_irqsave(&rq->lock, flags);
7569 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7570 raw_spin_unlock_irqrestore(&rq->lock, flags);
7573 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7574 struct sched_entity *se, int cpu,
7575 struct sched_entity *parent)
7577 struct rq *rq = cpu_rq(cpu);
7581 init_cfs_rq_runtime(cfs_rq);
7583 tg->cfs_rq[cpu] = cfs_rq;
7586 /* se could be NULL for root_task_group */
7591 se->cfs_rq = &rq->cfs;
7594 se->cfs_rq = parent->my_q;
7595 se->depth = parent->depth + 1;
7599 /* guarantee group entities always have weight */
7600 update_load_set(&se->load, NICE_0_LOAD);
7601 se->parent = parent;
7604 static DEFINE_MUTEX(shares_mutex);
7606 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7609 unsigned long flags;
7612 * We can't change the weight of the root cgroup.
7617 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7619 mutex_lock(&shares_mutex);
7620 if (tg->shares == shares)
7623 tg->shares = shares;
7624 for_each_possible_cpu(i) {
7625 struct rq *rq = cpu_rq(i);
7626 struct sched_entity *se;
7629 /* Propagate contribution to hierarchy */
7630 raw_spin_lock_irqsave(&rq->lock, flags);
7632 /* Possible calls to update_curr() need rq clock */
7633 update_rq_clock(rq);
7634 for_each_sched_entity(se)
7635 update_cfs_shares(group_cfs_rq(se));
7636 raw_spin_unlock_irqrestore(&rq->lock, flags);
7640 mutex_unlock(&shares_mutex);
7643 #else /* CONFIG_FAIR_GROUP_SCHED */
7645 void free_fair_sched_group(struct task_group *tg) { }
7647 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7652 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7654 #endif /* CONFIG_FAIR_GROUP_SCHED */
7657 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
7659 struct sched_entity *se = &task->se;
7660 unsigned int rr_interval = 0;
7663 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7666 if (rq->cfs.load.weight)
7667 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
7673 * All the scheduling class methods:
7675 const struct sched_class fair_sched_class = {
7676 .next = &idle_sched_class,
7677 .enqueue_task = enqueue_task_fair,
7678 .dequeue_task = dequeue_task_fair,
7679 .yield_task = yield_task_fair,
7680 .yield_to_task = yield_to_task_fair,
7682 .check_preempt_curr = check_preempt_wakeup,
7684 .pick_next_task = pick_next_task_fair,
7685 .put_prev_task = put_prev_task_fair,
7688 .select_task_rq = select_task_rq_fair,
7689 .migrate_task_rq = migrate_task_rq_fair,
7691 .rq_online = rq_online_fair,
7692 .rq_offline = rq_offline_fair,
7694 .task_waking = task_waking_fair,
7697 .set_curr_task = set_curr_task_fair,
7698 .task_tick = task_tick_fair,
7699 .task_fork = task_fork_fair,
7701 .prio_changed = prio_changed_fair,
7702 .switched_from = switched_from_fair,
7703 .switched_to = switched_to_fair,
7705 .get_rr_interval = get_rr_interval_fair,
7707 #ifdef CONFIG_FAIR_GROUP_SCHED
7708 .task_move_group = task_move_group_fair,
7712 #ifdef CONFIG_SCHED_DEBUG
7713 void print_cfs_stats(struct seq_file *m, int cpu)
7715 struct cfs_rq *cfs_rq;
7718 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7719 print_cfs_rq(m, cpu, cfs_rq);
7724 __init void init_sched_fair_class(void)
7727 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7729 #ifdef CONFIG_NO_HZ_COMMON
7730 nohz.next_balance = jiffies;
7731 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7732 cpu_notifier(sched_ilb_notifier, 0);