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 capacity_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 compute_capacity;
1031 /* Approximate capacity in terms of runnable tasks on a node */
1032 unsigned long task_capacity;
1033 int has_free_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->compute_capacity += capacity_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_free_capacity, or we'll detect a huge
1060 * imbalance and bail there.
1065 ns->load = (ns->load * SCHED_CAPACITY_SCALE) / ns->compute_capacity;
1067 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE);
1068 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1071 struct task_numa_env {
1072 struct task_struct *p;
1074 int src_cpu, src_nid;
1075 int dst_cpu, dst_nid;
1077 struct numa_stats src_stats, dst_stats;
1081 struct task_struct *best_task;
1086 static void task_numa_assign(struct task_numa_env *env,
1087 struct task_struct *p, long imp)
1090 put_task_struct(env->best_task);
1095 env->best_imp = imp;
1096 env->best_cpu = env->dst_cpu;
1099 static bool load_too_imbalanced(long orig_src_load, long orig_dst_load,
1100 long src_load, long dst_load,
1101 struct task_numa_env *env)
1105 /* We care about the slope of the imbalance, not the direction. */
1106 if (dst_load < src_load)
1107 swap(dst_load, src_load);
1109 /* Is the difference below the threshold? */
1110 imb = dst_load * 100 - src_load * env->imbalance_pct;
1115 * The imbalance is above the allowed threshold.
1116 * Compare it with the old imbalance.
1118 if (orig_dst_load < orig_src_load)
1119 swap(orig_dst_load, orig_src_load);
1121 old_imb = orig_dst_load * 100 - orig_src_load * env->imbalance_pct;
1123 /* Would this change make things worse? */
1124 return (imb > old_imb);
1128 * This checks if the overall compute and NUMA accesses of the system would
1129 * be improved if the source tasks was migrated to the target dst_cpu taking
1130 * into account that it might be best if task running on the dst_cpu should
1131 * be exchanged with the source task
1133 static void task_numa_compare(struct task_numa_env *env,
1134 long taskimp, long groupimp)
1136 struct rq *src_rq = cpu_rq(env->src_cpu);
1137 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1138 struct task_struct *cur;
1139 long orig_src_load, src_load;
1140 long orig_dst_load, dst_load;
1142 long imp = (groupimp > 0) ? groupimp : taskimp;
1145 cur = ACCESS_ONCE(dst_rq->curr);
1146 if (cur->pid == 0) /* idle */
1150 * "imp" is the fault differential for the source task between the
1151 * source and destination node. Calculate the total differential for
1152 * the source task and potential destination task. The more negative
1153 * the value is, the more rmeote accesses that would be expected to
1154 * be incurred if the tasks were swapped.
1157 /* Skip this swap candidate if cannot move to the source cpu */
1158 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1162 * If dst and source tasks are in the same NUMA group, or not
1163 * in any group then look only at task weights.
1165 if (cur->numa_group == env->p->numa_group) {
1166 imp = taskimp + task_weight(cur, env->src_nid) -
1167 task_weight(cur, env->dst_nid);
1169 * Add some hysteresis to prevent swapping the
1170 * tasks within a group over tiny differences.
1172 if (cur->numa_group)
1176 * Compare the group weights. If a task is all by
1177 * itself (not part of a group), use the task weight
1180 if (env->p->numa_group)
1185 if (cur->numa_group)
1186 imp += group_weight(cur, env->src_nid) -
1187 group_weight(cur, env->dst_nid);
1189 imp += task_weight(cur, env->src_nid) -
1190 task_weight(cur, env->dst_nid);
1194 if (imp < env->best_imp)
1198 /* Is there capacity at our destination? */
1199 if (env->src_stats.has_free_capacity &&
1200 !env->dst_stats.has_free_capacity)
1206 /* Balance doesn't matter much if we're running a task per cpu */
1207 if (src_rq->nr_running == 1 && dst_rq->nr_running == 1)
1211 * In the overloaded case, try and keep the load balanced.
1214 orig_dst_load = env->dst_stats.load;
1215 orig_src_load = env->src_stats.load;
1217 /* XXX missing capacity terms */
1218 load = task_h_load(env->p);
1219 dst_load = orig_dst_load + load;
1220 src_load = orig_src_load - load;
1223 load = task_h_load(cur);
1228 if (load_too_imbalanced(orig_src_load, orig_dst_load,
1229 src_load, dst_load, env))
1233 task_numa_assign(env, cur, imp);
1238 static void task_numa_find_cpu(struct task_numa_env *env,
1239 long taskimp, long groupimp)
1243 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1244 /* Skip this CPU if the source task cannot migrate */
1245 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1249 task_numa_compare(env, taskimp, groupimp);
1253 static int task_numa_migrate(struct task_struct *p)
1255 struct task_numa_env env = {
1258 .src_cpu = task_cpu(p),
1259 .src_nid = task_node(p),
1261 .imbalance_pct = 112,
1267 struct sched_domain *sd;
1268 unsigned long taskweight, groupweight;
1270 long taskimp, groupimp;
1273 * Pick the lowest SD_NUMA domain, as that would have the smallest
1274 * imbalance and would be the first to start moving tasks about.
1276 * And we want to avoid any moving of tasks about, as that would create
1277 * random movement of tasks -- counter the numa conditions we're trying
1281 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1283 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1287 * Cpusets can break the scheduler domain tree into smaller
1288 * balance domains, some of which do not cross NUMA boundaries.
1289 * Tasks that are "trapped" in such domains cannot be migrated
1290 * elsewhere, so there is no point in (re)trying.
1292 if (unlikely(!sd)) {
1293 p->numa_preferred_nid = task_node(p);
1297 taskweight = task_weight(p, env.src_nid);
1298 groupweight = group_weight(p, env.src_nid);
1299 update_numa_stats(&env.src_stats, env.src_nid);
1300 env.dst_nid = p->numa_preferred_nid;
1301 taskimp = task_weight(p, env.dst_nid) - taskweight;
1302 groupimp = group_weight(p, env.dst_nid) - groupweight;
1303 update_numa_stats(&env.dst_stats, env.dst_nid);
1305 /* Try to find a spot on the preferred nid. */
1306 task_numa_find_cpu(&env, taskimp, groupimp);
1308 /* No space available on the preferred nid. Look elsewhere. */
1309 if (env.best_cpu == -1) {
1310 for_each_online_node(nid) {
1311 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1314 /* Only consider nodes where both task and groups benefit */
1315 taskimp = task_weight(p, nid) - taskweight;
1316 groupimp = group_weight(p, nid) - groupweight;
1317 if (taskimp < 0 && groupimp < 0)
1321 update_numa_stats(&env.dst_stats, env.dst_nid);
1322 task_numa_find_cpu(&env, taskimp, groupimp);
1326 /* No better CPU than the current one was found. */
1327 if (env.best_cpu == -1)
1331 * If the task is part of a workload that spans multiple NUMA nodes,
1332 * and is migrating into one of the workload's active nodes, remember
1333 * this node as the task's preferred numa node, so the workload can
1335 * A task that migrated to a second choice node will be better off
1336 * trying for a better one later. Do not set the preferred node here.
1338 if (p->numa_group && node_isset(env.dst_nid, p->numa_group->active_nodes))
1339 sched_setnuma(p, env.dst_nid);
1342 * Reset the scan period if the task is being rescheduled on an
1343 * alternative node to recheck if the tasks is now properly placed.
1345 p->numa_scan_period = task_scan_min(p);
1347 if (env.best_task == NULL) {
1348 ret = migrate_task_to(p, env.best_cpu);
1350 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1354 ret = migrate_swap(p, env.best_task);
1356 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1357 put_task_struct(env.best_task);
1361 /* Attempt to migrate a task to a CPU on the preferred node. */
1362 static void numa_migrate_preferred(struct task_struct *p)
1364 unsigned long interval = HZ;
1366 /* This task has no NUMA fault statistics yet */
1367 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults_memory))
1370 /* Periodically retry migrating the task to the preferred node */
1371 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1372 p->numa_migrate_retry = jiffies + interval;
1374 /* Success if task is already running on preferred CPU */
1375 if (task_node(p) == p->numa_preferred_nid)
1378 /* Otherwise, try migrate to a CPU on the preferred node */
1379 task_numa_migrate(p);
1383 * Find the nodes on which the workload is actively running. We do this by
1384 * tracking the nodes from which NUMA hinting faults are triggered. This can
1385 * be different from the set of nodes where the workload's memory is currently
1388 * The bitmask is used to make smarter decisions on when to do NUMA page
1389 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1390 * are added when they cause over 6/16 of the maximum number of faults, but
1391 * only removed when they drop below 3/16.
1393 static void update_numa_active_node_mask(struct numa_group *numa_group)
1395 unsigned long faults, max_faults = 0;
1398 for_each_online_node(nid) {
1399 faults = group_faults_cpu(numa_group, nid);
1400 if (faults > max_faults)
1401 max_faults = faults;
1404 for_each_online_node(nid) {
1405 faults = group_faults_cpu(numa_group, nid);
1406 if (!node_isset(nid, numa_group->active_nodes)) {
1407 if (faults > max_faults * 6 / 16)
1408 node_set(nid, numa_group->active_nodes);
1409 } else if (faults < max_faults * 3 / 16)
1410 node_clear(nid, numa_group->active_nodes);
1415 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1416 * increments. The more local the fault statistics are, the higher the scan
1417 * period will be for the next scan window. If local/remote ratio is below
1418 * NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) the
1419 * scan period will decrease
1421 #define NUMA_PERIOD_SLOTS 10
1422 #define NUMA_PERIOD_THRESHOLD 3
1425 * Increase the scan period (slow down scanning) if the majority of
1426 * our memory is already on our local node, or if the majority of
1427 * the page accesses are shared with other processes.
1428 * Otherwise, decrease the scan period.
1430 static void update_task_scan_period(struct task_struct *p,
1431 unsigned long shared, unsigned long private)
1433 unsigned int period_slot;
1437 unsigned long remote = p->numa_faults_locality[0];
1438 unsigned long local = p->numa_faults_locality[1];
1441 * If there were no record hinting faults then either the task is
1442 * completely idle or all activity is areas that are not of interest
1443 * to automatic numa balancing. Scan slower
1445 if (local + shared == 0) {
1446 p->numa_scan_period = min(p->numa_scan_period_max,
1447 p->numa_scan_period << 1);
1449 p->mm->numa_next_scan = jiffies +
1450 msecs_to_jiffies(p->numa_scan_period);
1456 * Prepare to scale scan period relative to the current period.
1457 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1458 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1459 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1461 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1462 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1463 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1464 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1467 diff = slot * period_slot;
1469 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1472 * Scale scan rate increases based on sharing. There is an
1473 * inverse relationship between the degree of sharing and
1474 * the adjustment made to the scanning period. Broadly
1475 * speaking the intent is that there is little point
1476 * scanning faster if shared accesses dominate as it may
1477 * simply bounce migrations uselessly
1479 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared));
1480 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1483 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1484 task_scan_min(p), task_scan_max(p));
1485 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1489 * Get the fraction of time the task has been running since the last
1490 * NUMA placement cycle. The scheduler keeps similar statistics, but
1491 * decays those on a 32ms period, which is orders of magnitude off
1492 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1493 * stats only if the task is so new there are no NUMA statistics yet.
1495 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1497 u64 runtime, delta, now;
1498 /* Use the start of this time slice to avoid calculations. */
1499 now = p->se.exec_start;
1500 runtime = p->se.sum_exec_runtime;
1502 if (p->last_task_numa_placement) {
1503 delta = runtime - p->last_sum_exec_runtime;
1504 *period = now - p->last_task_numa_placement;
1506 delta = p->se.avg.runnable_avg_sum;
1507 *period = p->se.avg.runnable_avg_period;
1510 p->last_sum_exec_runtime = runtime;
1511 p->last_task_numa_placement = now;
1516 static void task_numa_placement(struct task_struct *p)
1518 int seq, nid, max_nid = -1, max_group_nid = -1;
1519 unsigned long max_faults = 0, max_group_faults = 0;
1520 unsigned long fault_types[2] = { 0, 0 };
1521 unsigned long total_faults;
1522 u64 runtime, period;
1523 spinlock_t *group_lock = NULL;
1525 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1526 if (p->numa_scan_seq == seq)
1528 p->numa_scan_seq = seq;
1529 p->numa_scan_period_max = task_scan_max(p);
1531 total_faults = p->numa_faults_locality[0] +
1532 p->numa_faults_locality[1];
1533 runtime = numa_get_avg_runtime(p, &period);
1535 /* If the task is part of a group prevent parallel updates to group stats */
1536 if (p->numa_group) {
1537 group_lock = &p->numa_group->lock;
1538 spin_lock_irq(group_lock);
1541 /* Find the node with the highest number of faults */
1542 for_each_online_node(nid) {
1543 unsigned long faults = 0, group_faults = 0;
1546 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1547 long diff, f_diff, f_weight;
1549 i = task_faults_idx(nid, priv);
1551 /* Decay existing window, copy faults since last scan */
1552 diff = p->numa_faults_buffer_memory[i] - p->numa_faults_memory[i] / 2;
1553 fault_types[priv] += p->numa_faults_buffer_memory[i];
1554 p->numa_faults_buffer_memory[i] = 0;
1557 * Normalize the faults_from, so all tasks in a group
1558 * count according to CPU use, instead of by the raw
1559 * number of faults. Tasks with little runtime have
1560 * little over-all impact on throughput, and thus their
1561 * faults are less important.
1563 f_weight = div64_u64(runtime << 16, period + 1);
1564 f_weight = (f_weight * p->numa_faults_buffer_cpu[i]) /
1566 f_diff = f_weight - p->numa_faults_cpu[i] / 2;
1567 p->numa_faults_buffer_cpu[i] = 0;
1569 p->numa_faults_memory[i] += diff;
1570 p->numa_faults_cpu[i] += f_diff;
1571 faults += p->numa_faults_memory[i];
1572 p->total_numa_faults += diff;
1573 if (p->numa_group) {
1574 /* safe because we can only change our own group */
1575 p->numa_group->faults[i] += diff;
1576 p->numa_group->faults_cpu[i] += f_diff;
1577 p->numa_group->total_faults += diff;
1578 group_faults += p->numa_group->faults[i];
1582 if (faults > max_faults) {
1583 max_faults = faults;
1587 if (group_faults > max_group_faults) {
1588 max_group_faults = group_faults;
1589 max_group_nid = nid;
1593 update_task_scan_period(p, fault_types[0], fault_types[1]);
1595 if (p->numa_group) {
1596 update_numa_active_node_mask(p->numa_group);
1598 * If the preferred task and group nids are different,
1599 * iterate over the nodes again to find the best place.
1601 if (max_nid != max_group_nid) {
1602 unsigned long weight, max_weight = 0;
1604 for_each_online_node(nid) {
1605 weight = task_weight(p, nid) + group_weight(p, nid);
1606 if (weight > max_weight) {
1607 max_weight = weight;
1613 spin_unlock_irq(group_lock);
1617 /* Set the new preferred node */
1618 if (max_nid != p->numa_preferred_nid)
1619 sched_setnuma(p, max_nid);
1621 if (task_node(p) != p->numa_preferred_nid)
1622 numa_migrate_preferred(p);
1626 static inline int get_numa_group(struct numa_group *grp)
1628 return atomic_inc_not_zero(&grp->refcount);
1631 static inline void put_numa_group(struct numa_group *grp)
1633 if (atomic_dec_and_test(&grp->refcount))
1634 kfree_rcu(grp, rcu);
1637 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1640 struct numa_group *grp, *my_grp;
1641 struct task_struct *tsk;
1643 int cpu = cpupid_to_cpu(cpupid);
1646 if (unlikely(!p->numa_group)) {
1647 unsigned int size = sizeof(struct numa_group) +
1648 4*nr_node_ids*sizeof(unsigned long);
1650 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1654 atomic_set(&grp->refcount, 1);
1655 spin_lock_init(&grp->lock);
1656 INIT_LIST_HEAD(&grp->task_list);
1658 /* Second half of the array tracks nids where faults happen */
1659 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1662 node_set(task_node(current), grp->active_nodes);
1664 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1665 grp->faults[i] = p->numa_faults_memory[i];
1667 grp->total_faults = p->total_numa_faults;
1669 list_add(&p->numa_entry, &grp->task_list);
1671 rcu_assign_pointer(p->numa_group, grp);
1675 tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1677 if (!cpupid_match_pid(tsk, cpupid))
1680 grp = rcu_dereference(tsk->numa_group);
1684 my_grp = p->numa_group;
1689 * Only join the other group if its bigger; if we're the bigger group,
1690 * the other task will join us.
1692 if (my_grp->nr_tasks > grp->nr_tasks)
1696 * Tie-break on the grp address.
1698 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1701 /* Always join threads in the same process. */
1702 if (tsk->mm == current->mm)
1705 /* Simple filter to avoid false positives due to PID collisions */
1706 if (flags & TNF_SHARED)
1709 /* Update priv based on whether false sharing was detected */
1712 if (join && !get_numa_group(grp))
1720 BUG_ON(irqs_disabled());
1721 double_lock_irq(&my_grp->lock, &grp->lock);
1723 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
1724 my_grp->faults[i] -= p->numa_faults_memory[i];
1725 grp->faults[i] += p->numa_faults_memory[i];
1727 my_grp->total_faults -= p->total_numa_faults;
1728 grp->total_faults += p->total_numa_faults;
1730 list_move(&p->numa_entry, &grp->task_list);
1734 spin_unlock(&my_grp->lock);
1735 spin_unlock_irq(&grp->lock);
1737 rcu_assign_pointer(p->numa_group, grp);
1739 put_numa_group(my_grp);
1747 void task_numa_free(struct task_struct *p)
1749 struct numa_group *grp = p->numa_group;
1750 void *numa_faults = p->numa_faults_memory;
1751 unsigned long flags;
1755 spin_lock_irqsave(&grp->lock, flags);
1756 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1757 grp->faults[i] -= p->numa_faults_memory[i];
1758 grp->total_faults -= p->total_numa_faults;
1760 list_del(&p->numa_entry);
1762 spin_unlock_irqrestore(&grp->lock, flags);
1763 rcu_assign_pointer(p->numa_group, NULL);
1764 put_numa_group(grp);
1767 p->numa_faults_memory = NULL;
1768 p->numa_faults_buffer_memory = NULL;
1769 p->numa_faults_cpu= NULL;
1770 p->numa_faults_buffer_cpu = NULL;
1775 * Got a PROT_NONE fault for a page on @node.
1777 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
1779 struct task_struct *p = current;
1780 bool migrated = flags & TNF_MIGRATED;
1781 int cpu_node = task_node(current);
1782 int local = !!(flags & TNF_FAULT_LOCAL);
1785 if (!numabalancing_enabled)
1788 /* for example, ksmd faulting in a user's mm */
1792 /* Do not worry about placement if exiting */
1793 if (p->state == TASK_DEAD)
1796 /* Allocate buffer to track faults on a per-node basis */
1797 if (unlikely(!p->numa_faults_memory)) {
1798 int size = sizeof(*p->numa_faults_memory) *
1799 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
1801 p->numa_faults_memory = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
1802 if (!p->numa_faults_memory)
1805 BUG_ON(p->numa_faults_buffer_memory);
1807 * The averaged statistics, shared & private, memory & cpu,
1808 * occupy the first half of the array. The second half of the
1809 * array is for current counters, which are averaged into the
1810 * first set by task_numa_placement.
1812 p->numa_faults_cpu = p->numa_faults_memory + (2 * nr_node_ids);
1813 p->numa_faults_buffer_memory = p->numa_faults_memory + (4 * nr_node_ids);
1814 p->numa_faults_buffer_cpu = p->numa_faults_memory + (6 * nr_node_ids);
1815 p->total_numa_faults = 0;
1816 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1820 * First accesses are treated as private, otherwise consider accesses
1821 * to be private if the accessing pid has not changed
1823 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
1826 priv = cpupid_match_pid(p, last_cpupid);
1827 if (!priv && !(flags & TNF_NO_GROUP))
1828 task_numa_group(p, last_cpupid, flags, &priv);
1832 * If a workload spans multiple NUMA nodes, a shared fault that
1833 * occurs wholly within the set of nodes that the workload is
1834 * actively using should be counted as local. This allows the
1835 * scan rate to slow down when a workload has settled down.
1837 if (!priv && !local && p->numa_group &&
1838 node_isset(cpu_node, p->numa_group->active_nodes) &&
1839 node_isset(mem_node, p->numa_group->active_nodes))
1842 task_numa_placement(p);
1845 * Retry task to preferred node migration periodically, in case it
1846 * case it previously failed, or the scheduler moved us.
1848 if (time_after(jiffies, p->numa_migrate_retry))
1849 numa_migrate_preferred(p);
1852 p->numa_pages_migrated += pages;
1854 p->numa_faults_buffer_memory[task_faults_idx(mem_node, priv)] += pages;
1855 p->numa_faults_buffer_cpu[task_faults_idx(cpu_node, priv)] += pages;
1856 p->numa_faults_locality[local] += pages;
1859 static void reset_ptenuma_scan(struct task_struct *p)
1861 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1862 p->mm->numa_scan_offset = 0;
1866 * The expensive part of numa migration is done from task_work context.
1867 * Triggered from task_tick_numa().
1869 void task_numa_work(struct callback_head *work)
1871 unsigned long migrate, next_scan, now = jiffies;
1872 struct task_struct *p = current;
1873 struct mm_struct *mm = p->mm;
1874 struct vm_area_struct *vma;
1875 unsigned long start, end;
1876 unsigned long nr_pte_updates = 0;
1879 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1881 work->next = work; /* protect against double add */
1883 * Who cares about NUMA placement when they're dying.
1885 * NOTE: make sure not to dereference p->mm before this check,
1886 * exit_task_work() happens _after_ exit_mm() so we could be called
1887 * without p->mm even though we still had it when we enqueued this
1890 if (p->flags & PF_EXITING)
1893 if (!mm->numa_next_scan) {
1894 mm->numa_next_scan = now +
1895 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1899 * Enforce maximal scan/migration frequency..
1901 migrate = mm->numa_next_scan;
1902 if (time_before(now, migrate))
1905 if (p->numa_scan_period == 0) {
1906 p->numa_scan_period_max = task_scan_max(p);
1907 p->numa_scan_period = task_scan_min(p);
1910 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1911 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1915 * Delay this task enough that another task of this mm will likely win
1916 * the next time around.
1918 p->node_stamp += 2 * TICK_NSEC;
1920 start = mm->numa_scan_offset;
1921 pages = sysctl_numa_balancing_scan_size;
1922 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1926 down_read(&mm->mmap_sem);
1927 vma = find_vma(mm, start);
1929 reset_ptenuma_scan(p);
1933 for (; vma; vma = vma->vm_next) {
1934 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1938 * Shared library pages mapped by multiple processes are not
1939 * migrated as it is expected they are cache replicated. Avoid
1940 * hinting faults in read-only file-backed mappings or the vdso
1941 * as migrating the pages will be of marginal benefit.
1944 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1948 * Skip inaccessible VMAs to avoid any confusion between
1949 * PROT_NONE and NUMA hinting ptes
1951 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
1955 start = max(start, vma->vm_start);
1956 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1957 end = min(end, vma->vm_end);
1958 nr_pte_updates += change_prot_numa(vma, start, end);
1961 * Scan sysctl_numa_balancing_scan_size but ensure that
1962 * at least one PTE is updated so that unused virtual
1963 * address space is quickly skipped.
1966 pages -= (end - start) >> PAGE_SHIFT;
1973 } while (end != vma->vm_end);
1978 * It is possible to reach the end of the VMA list but the last few
1979 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1980 * would find the !migratable VMA on the next scan but not reset the
1981 * scanner to the start so check it now.
1984 mm->numa_scan_offset = start;
1986 reset_ptenuma_scan(p);
1987 up_read(&mm->mmap_sem);
1991 * Drive the periodic memory faults..
1993 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1995 struct callback_head *work = &curr->numa_work;
1999 * We don't care about NUMA placement if we don't have memory.
2001 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2005 * Using runtime rather than walltime has the dual advantage that
2006 * we (mostly) drive the selection from busy threads and that the
2007 * task needs to have done some actual work before we bother with
2010 now = curr->se.sum_exec_runtime;
2011 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2013 if (now - curr->node_stamp > period) {
2014 if (!curr->node_stamp)
2015 curr->numa_scan_period = task_scan_min(curr);
2016 curr->node_stamp += period;
2018 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2019 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2020 task_work_add(curr, work, true);
2025 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2029 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2033 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2036 #endif /* CONFIG_NUMA_BALANCING */
2039 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2041 update_load_add(&cfs_rq->load, se->load.weight);
2042 if (!parent_entity(se))
2043 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2045 if (entity_is_task(se)) {
2046 struct rq *rq = rq_of(cfs_rq);
2048 account_numa_enqueue(rq, task_of(se));
2049 list_add(&se->group_node, &rq->cfs_tasks);
2052 cfs_rq->nr_running++;
2056 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2058 update_load_sub(&cfs_rq->load, se->load.weight);
2059 if (!parent_entity(se))
2060 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2061 if (entity_is_task(se)) {
2062 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2063 list_del_init(&se->group_node);
2065 cfs_rq->nr_running--;
2068 #ifdef CONFIG_FAIR_GROUP_SCHED
2070 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2075 * Use this CPU's actual weight instead of the last load_contribution
2076 * to gain a more accurate current total weight. See
2077 * update_cfs_rq_load_contribution().
2079 tg_weight = atomic_long_read(&tg->load_avg);
2080 tg_weight -= cfs_rq->tg_load_contrib;
2081 tg_weight += cfs_rq->load.weight;
2086 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2088 long tg_weight, load, shares;
2090 tg_weight = calc_tg_weight(tg, cfs_rq);
2091 load = cfs_rq->load.weight;
2093 shares = (tg->shares * load);
2095 shares /= tg_weight;
2097 if (shares < MIN_SHARES)
2098 shares = MIN_SHARES;
2099 if (shares > tg->shares)
2100 shares = tg->shares;
2104 # else /* CONFIG_SMP */
2105 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2109 # endif /* CONFIG_SMP */
2110 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2111 unsigned long weight)
2114 /* commit outstanding execution time */
2115 if (cfs_rq->curr == se)
2116 update_curr(cfs_rq);
2117 account_entity_dequeue(cfs_rq, se);
2120 update_load_set(&se->load, weight);
2123 account_entity_enqueue(cfs_rq, se);
2126 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2128 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2130 struct task_group *tg;
2131 struct sched_entity *se;
2135 se = tg->se[cpu_of(rq_of(cfs_rq))];
2136 if (!se || throttled_hierarchy(cfs_rq))
2139 if (likely(se->load.weight == tg->shares))
2142 shares = calc_cfs_shares(cfs_rq, tg);
2144 reweight_entity(cfs_rq_of(se), se, shares);
2146 #else /* CONFIG_FAIR_GROUP_SCHED */
2147 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2150 #endif /* CONFIG_FAIR_GROUP_SCHED */
2154 * We choose a half-life close to 1 scheduling period.
2155 * Note: The tables below are dependent on this value.
2157 #define LOAD_AVG_PERIOD 32
2158 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
2159 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
2161 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2162 static const u32 runnable_avg_yN_inv[] = {
2163 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2164 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2165 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2166 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2167 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2168 0x85aac367, 0x82cd8698,
2172 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2173 * over-estimates when re-combining.
2175 static const u32 runnable_avg_yN_sum[] = {
2176 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2177 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2178 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2183 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2185 static __always_inline u64 decay_load(u64 val, u64 n)
2187 unsigned int local_n;
2191 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2194 /* after bounds checking we can collapse to 32-bit */
2198 * As y^PERIOD = 1/2, we can combine
2199 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
2200 * With a look-up table which covers k^n (n<PERIOD)
2202 * To achieve constant time decay_load.
2204 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2205 val >>= local_n / LOAD_AVG_PERIOD;
2206 local_n %= LOAD_AVG_PERIOD;
2209 val *= runnable_avg_yN_inv[local_n];
2210 /* We don't use SRR here since we always want to round down. */
2215 * For updates fully spanning n periods, the contribution to runnable
2216 * average will be: \Sum 1024*y^n
2218 * We can compute this reasonably efficiently by combining:
2219 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2221 static u32 __compute_runnable_contrib(u64 n)
2225 if (likely(n <= LOAD_AVG_PERIOD))
2226 return runnable_avg_yN_sum[n];
2227 else if (unlikely(n >= LOAD_AVG_MAX_N))
2228 return LOAD_AVG_MAX;
2230 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2232 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2233 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2235 n -= LOAD_AVG_PERIOD;
2236 } while (n > LOAD_AVG_PERIOD);
2238 contrib = decay_load(contrib, n);
2239 return contrib + runnable_avg_yN_sum[n];
2243 * We can represent the historical contribution to runnable average as the
2244 * coefficients of a geometric series. To do this we sub-divide our runnable
2245 * history into segments of approximately 1ms (1024us); label the segment that
2246 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2248 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2250 * (now) (~1ms ago) (~2ms ago)
2252 * Let u_i denote the fraction of p_i that the entity was runnable.
2254 * We then designate the fractions u_i as our co-efficients, yielding the
2255 * following representation of historical load:
2256 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2258 * We choose y based on the with of a reasonably scheduling period, fixing:
2261 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2262 * approximately half as much as the contribution to load within the last ms
2265 * When a period "rolls over" and we have new u_0`, multiplying the previous
2266 * sum again by y is sufficient to update:
2267 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2268 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2270 static __always_inline int __update_entity_runnable_avg(u64 now,
2271 struct sched_avg *sa,
2275 u32 runnable_contrib;
2276 int delta_w, decayed = 0;
2278 delta = now - sa->last_runnable_update;
2280 * This should only happen when time goes backwards, which it
2281 * unfortunately does during sched clock init when we swap over to TSC.
2283 if ((s64)delta < 0) {
2284 sa->last_runnable_update = now;
2289 * Use 1024ns as the unit of measurement since it's a reasonable
2290 * approximation of 1us and fast to compute.
2295 sa->last_runnable_update = now;
2297 /* delta_w is the amount already accumulated against our next period */
2298 delta_w = sa->runnable_avg_period % 1024;
2299 if (delta + delta_w >= 1024) {
2300 /* period roll-over */
2304 * Now that we know we're crossing a period boundary, figure
2305 * out how much from delta we need to complete the current
2306 * period and accrue it.
2308 delta_w = 1024 - delta_w;
2310 sa->runnable_avg_sum += delta_w;
2311 sa->runnable_avg_period += delta_w;
2315 /* Figure out how many additional periods this update spans */
2316 periods = delta / 1024;
2319 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2321 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
2324 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2325 runnable_contrib = __compute_runnable_contrib(periods);
2327 sa->runnable_avg_sum += runnable_contrib;
2328 sa->runnable_avg_period += runnable_contrib;
2331 /* Remainder of delta accrued against u_0` */
2333 sa->runnable_avg_sum += delta;
2334 sa->runnable_avg_period += delta;
2339 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2340 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2342 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2343 u64 decays = atomic64_read(&cfs_rq->decay_counter);
2345 decays -= se->avg.decay_count;
2349 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2350 se->avg.decay_count = 0;
2355 #ifdef CONFIG_FAIR_GROUP_SCHED
2356 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2359 struct task_group *tg = cfs_rq->tg;
2362 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2363 tg_contrib -= cfs_rq->tg_load_contrib;
2365 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2366 atomic_long_add(tg_contrib, &tg->load_avg);
2367 cfs_rq->tg_load_contrib += tg_contrib;
2372 * Aggregate cfs_rq runnable averages into an equivalent task_group
2373 * representation for computing load contributions.
2375 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2376 struct cfs_rq *cfs_rq)
2378 struct task_group *tg = cfs_rq->tg;
2381 /* The fraction of a cpu used by this cfs_rq */
2382 contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT,
2383 sa->runnable_avg_period + 1);
2384 contrib -= cfs_rq->tg_runnable_contrib;
2386 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2387 atomic_add(contrib, &tg->runnable_avg);
2388 cfs_rq->tg_runnable_contrib += contrib;
2392 static inline void __update_group_entity_contrib(struct sched_entity *se)
2394 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2395 struct task_group *tg = cfs_rq->tg;
2400 contrib = cfs_rq->tg_load_contrib * tg->shares;
2401 se->avg.load_avg_contrib = div_u64(contrib,
2402 atomic_long_read(&tg->load_avg) + 1);
2405 * For group entities we need to compute a correction term in the case
2406 * that they are consuming <1 cpu so that we would contribute the same
2407 * load as a task of equal weight.
2409 * Explicitly co-ordinating this measurement would be expensive, but
2410 * fortunately the sum of each cpus contribution forms a usable
2411 * lower-bound on the true value.
2413 * Consider the aggregate of 2 contributions. Either they are disjoint
2414 * (and the sum represents true value) or they are disjoint and we are
2415 * understating by the aggregate of their overlap.
2417 * Extending this to N cpus, for a given overlap, the maximum amount we
2418 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2419 * cpus that overlap for this interval and w_i is the interval width.
2421 * On a small machine; the first term is well-bounded which bounds the
2422 * total error since w_i is a subset of the period. Whereas on a
2423 * larger machine, while this first term can be larger, if w_i is the
2424 * of consequential size guaranteed to see n_i*w_i quickly converge to
2425 * our upper bound of 1-cpu.
2427 runnable_avg = atomic_read(&tg->runnable_avg);
2428 if (runnable_avg < NICE_0_LOAD) {
2429 se->avg.load_avg_contrib *= runnable_avg;
2430 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2434 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2436 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
2437 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2439 #else /* CONFIG_FAIR_GROUP_SCHED */
2440 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2441 int force_update) {}
2442 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2443 struct cfs_rq *cfs_rq) {}
2444 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2445 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2446 #endif /* CONFIG_FAIR_GROUP_SCHED */
2448 static inline void __update_task_entity_contrib(struct sched_entity *se)
2452 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2453 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2454 contrib /= (se->avg.runnable_avg_period + 1);
2455 se->avg.load_avg_contrib = scale_load(contrib);
2458 /* Compute the current contribution to load_avg by se, return any delta */
2459 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2461 long old_contrib = se->avg.load_avg_contrib;
2463 if (entity_is_task(se)) {
2464 __update_task_entity_contrib(se);
2466 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2467 __update_group_entity_contrib(se);
2470 return se->avg.load_avg_contrib - old_contrib;
2473 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2476 if (likely(load_contrib < cfs_rq->blocked_load_avg))
2477 cfs_rq->blocked_load_avg -= load_contrib;
2479 cfs_rq->blocked_load_avg = 0;
2482 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2484 /* Update a sched_entity's runnable average */
2485 static inline void update_entity_load_avg(struct sched_entity *se,
2488 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2493 * For a group entity we need to use their owned cfs_rq_clock_task() in
2494 * case they are the parent of a throttled hierarchy.
2496 if (entity_is_task(se))
2497 now = cfs_rq_clock_task(cfs_rq);
2499 now = cfs_rq_clock_task(group_cfs_rq(se));
2501 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
2504 contrib_delta = __update_entity_load_avg_contrib(se);
2510 cfs_rq->runnable_load_avg += contrib_delta;
2512 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2516 * Decay the load contributed by all blocked children and account this so that
2517 * their contribution may appropriately discounted when they wake up.
2519 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2521 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2524 decays = now - cfs_rq->last_decay;
2525 if (!decays && !force_update)
2528 if (atomic_long_read(&cfs_rq->removed_load)) {
2529 unsigned long removed_load;
2530 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2531 subtract_blocked_load_contrib(cfs_rq, removed_load);
2535 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2537 atomic64_add(decays, &cfs_rq->decay_counter);
2538 cfs_rq->last_decay = now;
2541 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2544 /* Add the load generated by se into cfs_rq's child load-average */
2545 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2546 struct sched_entity *se,
2550 * We track migrations using entity decay_count <= 0, on a wake-up
2551 * migration we use a negative decay count to track the remote decays
2552 * accumulated while sleeping.
2554 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2555 * are seen by enqueue_entity_load_avg() as a migration with an already
2556 * constructed load_avg_contrib.
2558 if (unlikely(se->avg.decay_count <= 0)) {
2559 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2560 if (se->avg.decay_count) {
2562 * In a wake-up migration we have to approximate the
2563 * time sleeping. This is because we can't synchronize
2564 * clock_task between the two cpus, and it is not
2565 * guaranteed to be read-safe. Instead, we can
2566 * approximate this using our carried decays, which are
2567 * explicitly atomically readable.
2569 se->avg.last_runnable_update -= (-se->avg.decay_count)
2571 update_entity_load_avg(se, 0);
2572 /* Indicate that we're now synchronized and on-rq */
2573 se->avg.decay_count = 0;
2577 __synchronize_entity_decay(se);
2580 /* migrated tasks did not contribute to our blocked load */
2582 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2583 update_entity_load_avg(se, 0);
2586 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2587 /* we force update consideration on load-balancer moves */
2588 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2592 * Remove se's load from this cfs_rq child load-average, if the entity is
2593 * transitioning to a blocked state we track its projected decay using
2596 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2597 struct sched_entity *se,
2600 update_entity_load_avg(se, 1);
2601 /* we force update consideration on load-balancer moves */
2602 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2604 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2606 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2607 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2608 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2612 * Update the rq's load with the elapsed running time before entering
2613 * idle. if the last scheduled task is not a CFS task, idle_enter will
2614 * be the only way to update the runnable statistic.
2616 void idle_enter_fair(struct rq *this_rq)
2618 update_rq_runnable_avg(this_rq, 1);
2622 * Update the rq's load with the elapsed idle time before a task is
2623 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2624 * be the only way to update the runnable statistic.
2626 void idle_exit_fair(struct rq *this_rq)
2628 update_rq_runnable_avg(this_rq, 0);
2631 static int idle_balance(struct rq *this_rq);
2633 #else /* CONFIG_SMP */
2635 static inline void update_entity_load_avg(struct sched_entity *se,
2636 int update_cfs_rq) {}
2637 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2638 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2639 struct sched_entity *se,
2641 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2642 struct sched_entity *se,
2644 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2645 int force_update) {}
2647 static inline int idle_balance(struct rq *rq)
2652 #endif /* CONFIG_SMP */
2654 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2656 #ifdef CONFIG_SCHEDSTATS
2657 struct task_struct *tsk = NULL;
2659 if (entity_is_task(se))
2662 if (se->statistics.sleep_start) {
2663 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2668 if (unlikely(delta > se->statistics.sleep_max))
2669 se->statistics.sleep_max = delta;
2671 se->statistics.sleep_start = 0;
2672 se->statistics.sum_sleep_runtime += delta;
2675 account_scheduler_latency(tsk, delta >> 10, 1);
2676 trace_sched_stat_sleep(tsk, delta);
2679 if (se->statistics.block_start) {
2680 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2685 if (unlikely(delta > se->statistics.block_max))
2686 se->statistics.block_max = delta;
2688 se->statistics.block_start = 0;
2689 se->statistics.sum_sleep_runtime += delta;
2692 if (tsk->in_iowait) {
2693 se->statistics.iowait_sum += delta;
2694 se->statistics.iowait_count++;
2695 trace_sched_stat_iowait(tsk, delta);
2698 trace_sched_stat_blocked(tsk, delta);
2701 * Blocking time is in units of nanosecs, so shift by
2702 * 20 to get a milliseconds-range estimation of the
2703 * amount of time that the task spent sleeping:
2705 if (unlikely(prof_on == SLEEP_PROFILING)) {
2706 profile_hits(SLEEP_PROFILING,
2707 (void *)get_wchan(tsk),
2710 account_scheduler_latency(tsk, delta >> 10, 0);
2716 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2718 #ifdef CONFIG_SCHED_DEBUG
2719 s64 d = se->vruntime - cfs_rq->min_vruntime;
2724 if (d > 3*sysctl_sched_latency)
2725 schedstat_inc(cfs_rq, nr_spread_over);
2730 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2732 u64 vruntime = cfs_rq->min_vruntime;
2735 * The 'current' period is already promised to the current tasks,
2736 * however the extra weight of the new task will slow them down a
2737 * little, place the new task so that it fits in the slot that
2738 * stays open at the end.
2740 if (initial && sched_feat(START_DEBIT))
2741 vruntime += sched_vslice(cfs_rq, se);
2743 /* sleeps up to a single latency don't count. */
2745 unsigned long thresh = sysctl_sched_latency;
2748 * Halve their sleep time's effect, to allow
2749 * for a gentler effect of sleepers:
2751 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2757 /* ensure we never gain time by being placed backwards. */
2758 se->vruntime = max_vruntime(se->vruntime, vruntime);
2761 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2764 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2767 * Update the normalized vruntime before updating min_vruntime
2768 * through calling update_curr().
2770 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2771 se->vruntime += cfs_rq->min_vruntime;
2774 * Update run-time statistics of the 'current'.
2776 update_curr(cfs_rq);
2777 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2778 account_entity_enqueue(cfs_rq, se);
2779 update_cfs_shares(cfs_rq);
2781 if (flags & ENQUEUE_WAKEUP) {
2782 place_entity(cfs_rq, se, 0);
2783 enqueue_sleeper(cfs_rq, se);
2786 update_stats_enqueue(cfs_rq, se);
2787 check_spread(cfs_rq, se);
2788 if (se != cfs_rq->curr)
2789 __enqueue_entity(cfs_rq, se);
2792 if (cfs_rq->nr_running == 1) {
2793 list_add_leaf_cfs_rq(cfs_rq);
2794 check_enqueue_throttle(cfs_rq);
2798 static void __clear_buddies_last(struct sched_entity *se)
2800 for_each_sched_entity(se) {
2801 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2802 if (cfs_rq->last != se)
2805 cfs_rq->last = NULL;
2809 static void __clear_buddies_next(struct sched_entity *se)
2811 for_each_sched_entity(se) {
2812 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2813 if (cfs_rq->next != se)
2816 cfs_rq->next = NULL;
2820 static void __clear_buddies_skip(struct sched_entity *se)
2822 for_each_sched_entity(se) {
2823 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2824 if (cfs_rq->skip != se)
2827 cfs_rq->skip = NULL;
2831 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2833 if (cfs_rq->last == se)
2834 __clear_buddies_last(se);
2836 if (cfs_rq->next == se)
2837 __clear_buddies_next(se);
2839 if (cfs_rq->skip == se)
2840 __clear_buddies_skip(se);
2843 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2846 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2849 * Update run-time statistics of the 'current'.
2851 update_curr(cfs_rq);
2852 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2854 update_stats_dequeue(cfs_rq, se);
2855 if (flags & DEQUEUE_SLEEP) {
2856 #ifdef CONFIG_SCHEDSTATS
2857 if (entity_is_task(se)) {
2858 struct task_struct *tsk = task_of(se);
2860 if (tsk->state & TASK_INTERRUPTIBLE)
2861 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2862 if (tsk->state & TASK_UNINTERRUPTIBLE)
2863 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2868 clear_buddies(cfs_rq, se);
2870 if (se != cfs_rq->curr)
2871 __dequeue_entity(cfs_rq, se);
2873 account_entity_dequeue(cfs_rq, se);
2876 * Normalize the entity after updating the min_vruntime because the
2877 * update can refer to the ->curr item and we need to reflect this
2878 * movement in our normalized position.
2880 if (!(flags & DEQUEUE_SLEEP))
2881 se->vruntime -= cfs_rq->min_vruntime;
2883 /* return excess runtime on last dequeue */
2884 return_cfs_rq_runtime(cfs_rq);
2886 update_min_vruntime(cfs_rq);
2887 update_cfs_shares(cfs_rq);
2891 * Preempt the current task with a newly woken task if needed:
2894 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2896 unsigned long ideal_runtime, delta_exec;
2897 struct sched_entity *se;
2900 ideal_runtime = sched_slice(cfs_rq, curr);
2901 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2902 if (delta_exec > ideal_runtime) {
2903 resched_task(rq_of(cfs_rq)->curr);
2905 * The current task ran long enough, ensure it doesn't get
2906 * re-elected due to buddy favours.
2908 clear_buddies(cfs_rq, curr);
2913 * Ensure that a task that missed wakeup preemption by a
2914 * narrow margin doesn't have to wait for a full slice.
2915 * This also mitigates buddy induced latencies under load.
2917 if (delta_exec < sysctl_sched_min_granularity)
2920 se = __pick_first_entity(cfs_rq);
2921 delta = curr->vruntime - se->vruntime;
2926 if (delta > ideal_runtime)
2927 resched_task(rq_of(cfs_rq)->curr);
2931 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2933 /* 'current' is not kept within the tree. */
2936 * Any task has to be enqueued before it get to execute on
2937 * a CPU. So account for the time it spent waiting on the
2940 update_stats_wait_end(cfs_rq, se);
2941 __dequeue_entity(cfs_rq, se);
2944 update_stats_curr_start(cfs_rq, se);
2946 #ifdef CONFIG_SCHEDSTATS
2948 * Track our maximum slice length, if the CPU's load is at
2949 * least twice that of our own weight (i.e. dont track it
2950 * when there are only lesser-weight tasks around):
2952 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2953 se->statistics.slice_max = max(se->statistics.slice_max,
2954 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2957 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2961 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2964 * Pick the next process, keeping these things in mind, in this order:
2965 * 1) keep things fair between processes/task groups
2966 * 2) pick the "next" process, since someone really wants that to run
2967 * 3) pick the "last" process, for cache locality
2968 * 4) do not run the "skip" process, if something else is available
2970 static struct sched_entity *
2971 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2973 struct sched_entity *left = __pick_first_entity(cfs_rq);
2974 struct sched_entity *se;
2977 * If curr is set we have to see if its left of the leftmost entity
2978 * still in the tree, provided there was anything in the tree at all.
2980 if (!left || (curr && entity_before(curr, left)))
2983 se = left; /* ideally we run the leftmost entity */
2986 * Avoid running the skip buddy, if running something else can
2987 * be done without getting too unfair.
2989 if (cfs_rq->skip == se) {
2990 struct sched_entity *second;
2993 second = __pick_first_entity(cfs_rq);
2995 second = __pick_next_entity(se);
2996 if (!second || (curr && entity_before(curr, second)))
3000 if (second && wakeup_preempt_entity(second, left) < 1)
3005 * Prefer last buddy, try to return the CPU to a preempted task.
3007 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3011 * Someone really wants this to run. If it's not unfair, run it.
3013 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3016 clear_buddies(cfs_rq, se);
3021 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3023 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3026 * If still on the runqueue then deactivate_task()
3027 * was not called and update_curr() has to be done:
3030 update_curr(cfs_rq);
3032 /* throttle cfs_rqs exceeding runtime */
3033 check_cfs_rq_runtime(cfs_rq);
3035 check_spread(cfs_rq, prev);
3037 update_stats_wait_start(cfs_rq, prev);
3038 /* Put 'current' back into the tree. */
3039 __enqueue_entity(cfs_rq, prev);
3040 /* in !on_rq case, update occurred at dequeue */
3041 update_entity_load_avg(prev, 1);
3043 cfs_rq->curr = NULL;
3047 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3050 * Update run-time statistics of the 'current'.
3052 update_curr(cfs_rq);
3055 * Ensure that runnable average is periodically updated.
3057 update_entity_load_avg(curr, 1);
3058 update_cfs_rq_blocked_load(cfs_rq, 1);
3059 update_cfs_shares(cfs_rq);
3061 #ifdef CONFIG_SCHED_HRTICK
3063 * queued ticks are scheduled to match the slice, so don't bother
3064 * validating it and just reschedule.
3067 resched_task(rq_of(cfs_rq)->curr);
3071 * don't let the period tick interfere with the hrtick preemption
3073 if (!sched_feat(DOUBLE_TICK) &&
3074 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3078 if (cfs_rq->nr_running > 1)
3079 check_preempt_tick(cfs_rq, curr);
3083 /**************************************************
3084 * CFS bandwidth control machinery
3087 #ifdef CONFIG_CFS_BANDWIDTH
3089 #ifdef HAVE_JUMP_LABEL
3090 static struct static_key __cfs_bandwidth_used;
3092 static inline bool cfs_bandwidth_used(void)
3094 return static_key_false(&__cfs_bandwidth_used);
3097 void cfs_bandwidth_usage_inc(void)
3099 static_key_slow_inc(&__cfs_bandwidth_used);
3102 void cfs_bandwidth_usage_dec(void)
3104 static_key_slow_dec(&__cfs_bandwidth_used);
3106 #else /* HAVE_JUMP_LABEL */
3107 static bool cfs_bandwidth_used(void)
3112 void cfs_bandwidth_usage_inc(void) {}
3113 void cfs_bandwidth_usage_dec(void) {}
3114 #endif /* HAVE_JUMP_LABEL */
3117 * default period for cfs group bandwidth.
3118 * default: 0.1s, units: nanoseconds
3120 static inline u64 default_cfs_period(void)
3122 return 100000000ULL;
3125 static inline u64 sched_cfs_bandwidth_slice(void)
3127 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3131 * Replenish runtime according to assigned quota and update expiration time.
3132 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3133 * additional synchronization around rq->lock.
3135 * requires cfs_b->lock
3137 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3141 if (cfs_b->quota == RUNTIME_INF)
3144 now = sched_clock_cpu(smp_processor_id());
3145 cfs_b->runtime = cfs_b->quota;
3146 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3149 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3151 return &tg->cfs_bandwidth;
3154 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3155 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3157 if (unlikely(cfs_rq->throttle_count))
3158 return cfs_rq->throttled_clock_task;
3160 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3163 /* returns 0 on failure to allocate runtime */
3164 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3166 struct task_group *tg = cfs_rq->tg;
3167 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3168 u64 amount = 0, min_amount, expires;
3170 /* note: this is a positive sum as runtime_remaining <= 0 */
3171 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3173 raw_spin_lock(&cfs_b->lock);
3174 if (cfs_b->quota == RUNTIME_INF)
3175 amount = min_amount;
3178 * If the bandwidth pool has become inactive, then at least one
3179 * period must have elapsed since the last consumption.
3180 * Refresh the global state and ensure bandwidth timer becomes
3183 if (!cfs_b->timer_active) {
3184 __refill_cfs_bandwidth_runtime(cfs_b);
3185 __start_cfs_bandwidth(cfs_b, false);
3188 if (cfs_b->runtime > 0) {
3189 amount = min(cfs_b->runtime, min_amount);
3190 cfs_b->runtime -= amount;
3194 expires = cfs_b->runtime_expires;
3195 raw_spin_unlock(&cfs_b->lock);
3197 cfs_rq->runtime_remaining += amount;
3199 * we may have advanced our local expiration to account for allowed
3200 * spread between our sched_clock and the one on which runtime was
3203 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3204 cfs_rq->runtime_expires = expires;
3206 return cfs_rq->runtime_remaining > 0;
3210 * Note: This depends on the synchronization provided by sched_clock and the
3211 * fact that rq->clock snapshots this value.
3213 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3215 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3217 /* if the deadline is ahead of our clock, nothing to do */
3218 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3221 if (cfs_rq->runtime_remaining < 0)
3225 * If the local deadline has passed we have to consider the
3226 * possibility that our sched_clock is 'fast' and the global deadline
3227 * has not truly expired.
3229 * Fortunately we can check determine whether this the case by checking
3230 * whether the global deadline has advanced. It is valid to compare
3231 * cfs_b->runtime_expires without any locks since we only care about
3232 * exact equality, so a partial write will still work.
3235 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3236 /* extend local deadline, drift is bounded above by 2 ticks */
3237 cfs_rq->runtime_expires += TICK_NSEC;
3239 /* global deadline is ahead, expiration has passed */
3240 cfs_rq->runtime_remaining = 0;
3244 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3246 /* dock delta_exec before expiring quota (as it could span periods) */
3247 cfs_rq->runtime_remaining -= delta_exec;
3248 expire_cfs_rq_runtime(cfs_rq);
3250 if (likely(cfs_rq->runtime_remaining > 0))
3254 * if we're unable to extend our runtime we resched so that the active
3255 * hierarchy can be throttled
3257 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3258 resched_task(rq_of(cfs_rq)->curr);
3261 static __always_inline
3262 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3264 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3267 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3270 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3272 return cfs_bandwidth_used() && cfs_rq->throttled;
3275 /* check whether cfs_rq, or any parent, is throttled */
3276 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3278 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3282 * Ensure that neither of the group entities corresponding to src_cpu or
3283 * dest_cpu are members of a throttled hierarchy when performing group
3284 * load-balance operations.
3286 static inline int throttled_lb_pair(struct task_group *tg,
3287 int src_cpu, int dest_cpu)
3289 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3291 src_cfs_rq = tg->cfs_rq[src_cpu];
3292 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3294 return throttled_hierarchy(src_cfs_rq) ||
3295 throttled_hierarchy(dest_cfs_rq);
3298 /* updated child weight may affect parent so we have to do this bottom up */
3299 static int tg_unthrottle_up(struct task_group *tg, void *data)
3301 struct rq *rq = data;
3302 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3304 cfs_rq->throttle_count--;
3306 if (!cfs_rq->throttle_count) {
3307 /* adjust cfs_rq_clock_task() */
3308 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3309 cfs_rq->throttled_clock_task;
3316 static int tg_throttle_down(struct task_group *tg, void *data)
3318 struct rq *rq = data;
3319 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3321 /* group is entering throttled state, stop time */
3322 if (!cfs_rq->throttle_count)
3323 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3324 cfs_rq->throttle_count++;
3329 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3331 struct rq *rq = rq_of(cfs_rq);
3332 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3333 struct sched_entity *se;
3334 long task_delta, dequeue = 1;
3336 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3338 /* freeze hierarchy runnable averages while throttled */
3340 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3343 task_delta = cfs_rq->h_nr_running;
3344 for_each_sched_entity(se) {
3345 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3346 /* throttled entity or throttle-on-deactivate */
3351 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3352 qcfs_rq->h_nr_running -= task_delta;
3354 if (qcfs_rq->load.weight)
3359 sub_nr_running(rq, task_delta);
3361 cfs_rq->throttled = 1;
3362 cfs_rq->throttled_clock = rq_clock(rq);
3363 raw_spin_lock(&cfs_b->lock);
3364 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3365 if (!cfs_b->timer_active)
3366 __start_cfs_bandwidth(cfs_b, false);
3367 raw_spin_unlock(&cfs_b->lock);
3370 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3372 struct rq *rq = rq_of(cfs_rq);
3373 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3374 struct sched_entity *se;
3378 se = cfs_rq->tg->se[cpu_of(rq)];
3380 cfs_rq->throttled = 0;
3382 update_rq_clock(rq);
3384 raw_spin_lock(&cfs_b->lock);
3385 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3386 list_del_rcu(&cfs_rq->throttled_list);
3387 raw_spin_unlock(&cfs_b->lock);
3389 /* update hierarchical throttle state */
3390 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3392 if (!cfs_rq->load.weight)
3395 task_delta = cfs_rq->h_nr_running;
3396 for_each_sched_entity(se) {
3400 cfs_rq = cfs_rq_of(se);
3402 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3403 cfs_rq->h_nr_running += task_delta;
3405 if (cfs_rq_throttled(cfs_rq))
3410 add_nr_running(rq, task_delta);
3412 /* determine whether we need to wake up potentially idle cpu */
3413 if (rq->curr == rq->idle && rq->cfs.nr_running)
3414 resched_task(rq->curr);
3417 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3418 u64 remaining, u64 expires)
3420 struct cfs_rq *cfs_rq;
3421 u64 runtime = remaining;
3424 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3426 struct rq *rq = rq_of(cfs_rq);
3428 raw_spin_lock(&rq->lock);
3429 if (!cfs_rq_throttled(cfs_rq))
3432 runtime = -cfs_rq->runtime_remaining + 1;
3433 if (runtime > remaining)
3434 runtime = remaining;
3435 remaining -= runtime;
3437 cfs_rq->runtime_remaining += runtime;
3438 cfs_rq->runtime_expires = expires;
3440 /* we check whether we're throttled above */
3441 if (cfs_rq->runtime_remaining > 0)
3442 unthrottle_cfs_rq(cfs_rq);
3445 raw_spin_unlock(&rq->lock);
3456 * Responsible for refilling a task_group's bandwidth and unthrottling its
3457 * cfs_rqs as appropriate. If there has been no activity within the last
3458 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3459 * used to track this state.
3461 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3463 u64 runtime, runtime_expires;
3466 /* no need to continue the timer with no bandwidth constraint */
3467 if (cfs_b->quota == RUNTIME_INF)
3468 goto out_deactivate;
3470 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3471 cfs_b->nr_periods += overrun;
3474 * idle depends on !throttled (for the case of a large deficit), and if
3475 * we're going inactive then everything else can be deferred
3477 if (cfs_b->idle && !throttled)
3478 goto out_deactivate;
3481 * if we have relooped after returning idle once, we need to update our
3482 * status as actually running, so that other cpus doing
3483 * __start_cfs_bandwidth will stop trying to cancel us.
3485 cfs_b->timer_active = 1;
3487 __refill_cfs_bandwidth_runtime(cfs_b);
3490 /* mark as potentially idle for the upcoming period */
3495 /* account preceding periods in which throttling occurred */
3496 cfs_b->nr_throttled += overrun;
3499 * There are throttled entities so we must first use the new bandwidth
3500 * to unthrottle them before making it generally available. This
3501 * ensures that all existing debts will be paid before a new cfs_rq is
3504 runtime = cfs_b->runtime;
3505 runtime_expires = cfs_b->runtime_expires;
3509 * This check is repeated as we are holding onto the new bandwidth
3510 * while we unthrottle. This can potentially race with an unthrottled
3511 * group trying to acquire new bandwidth from the global pool.
3513 while (throttled && runtime > 0) {
3514 raw_spin_unlock(&cfs_b->lock);
3515 /* we can't nest cfs_b->lock while distributing bandwidth */
3516 runtime = distribute_cfs_runtime(cfs_b, runtime,
3518 raw_spin_lock(&cfs_b->lock);
3520 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3523 /* return (any) remaining runtime */
3524 cfs_b->runtime = runtime;
3526 * While we are ensured activity in the period following an
3527 * unthrottle, this also covers the case in which the new bandwidth is
3528 * insufficient to cover the existing bandwidth deficit. (Forcing the
3529 * timer to remain active while there are any throttled entities.)
3536 cfs_b->timer_active = 0;
3540 /* a cfs_rq won't donate quota below this amount */
3541 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3542 /* minimum remaining period time to redistribute slack quota */
3543 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3544 /* how long we wait to gather additional slack before distributing */
3545 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3548 * Are we near the end of the current quota period?
3550 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3551 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3552 * migrate_hrtimers, base is never cleared, so we are fine.
3554 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3556 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3559 /* if the call-back is running a quota refresh is already occurring */
3560 if (hrtimer_callback_running(refresh_timer))
3563 /* is a quota refresh about to occur? */
3564 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3565 if (remaining < min_expire)
3571 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3573 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3575 /* if there's a quota refresh soon don't bother with slack */
3576 if (runtime_refresh_within(cfs_b, min_left))
3579 start_bandwidth_timer(&cfs_b->slack_timer,
3580 ns_to_ktime(cfs_bandwidth_slack_period));
3583 /* we know any runtime found here is valid as update_curr() precedes return */
3584 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3586 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3587 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3589 if (slack_runtime <= 0)
3592 raw_spin_lock(&cfs_b->lock);
3593 if (cfs_b->quota != RUNTIME_INF &&
3594 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3595 cfs_b->runtime += slack_runtime;
3597 /* we are under rq->lock, defer unthrottling using a timer */
3598 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3599 !list_empty(&cfs_b->throttled_cfs_rq))
3600 start_cfs_slack_bandwidth(cfs_b);
3602 raw_spin_unlock(&cfs_b->lock);
3604 /* even if it's not valid for return we don't want to try again */
3605 cfs_rq->runtime_remaining -= slack_runtime;
3608 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3610 if (!cfs_bandwidth_used())
3613 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3616 __return_cfs_rq_runtime(cfs_rq);
3620 * This is done with a timer (instead of inline with bandwidth return) since
3621 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3623 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3625 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3628 /* confirm we're still not at a refresh boundary */
3629 raw_spin_lock(&cfs_b->lock);
3630 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3631 raw_spin_unlock(&cfs_b->lock);
3635 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
3636 runtime = cfs_b->runtime;
3639 expires = cfs_b->runtime_expires;
3640 raw_spin_unlock(&cfs_b->lock);
3645 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3647 raw_spin_lock(&cfs_b->lock);
3648 if (expires == cfs_b->runtime_expires)
3649 cfs_b->runtime = runtime;
3650 raw_spin_unlock(&cfs_b->lock);
3654 * When a group wakes up we want to make sure that its quota is not already
3655 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3656 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3658 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3660 if (!cfs_bandwidth_used())
3663 /* an active group must be handled by the update_curr()->put() path */
3664 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3667 /* ensure the group is not already throttled */
3668 if (cfs_rq_throttled(cfs_rq))
3671 /* update runtime allocation */
3672 account_cfs_rq_runtime(cfs_rq, 0);
3673 if (cfs_rq->runtime_remaining <= 0)
3674 throttle_cfs_rq(cfs_rq);
3677 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3678 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3680 if (!cfs_bandwidth_used())
3683 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3687 * it's possible for a throttled entity to be forced into a running
3688 * state (e.g. set_curr_task), in this case we're finished.
3690 if (cfs_rq_throttled(cfs_rq))
3693 throttle_cfs_rq(cfs_rq);
3697 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3699 struct cfs_bandwidth *cfs_b =
3700 container_of(timer, struct cfs_bandwidth, slack_timer);
3701 do_sched_cfs_slack_timer(cfs_b);
3703 return HRTIMER_NORESTART;
3706 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3708 struct cfs_bandwidth *cfs_b =
3709 container_of(timer, struct cfs_bandwidth, period_timer);
3714 raw_spin_lock(&cfs_b->lock);
3716 now = hrtimer_cb_get_time(timer);
3717 overrun = hrtimer_forward(timer, now, cfs_b->period);
3722 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3724 raw_spin_unlock(&cfs_b->lock);
3726 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3729 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3731 raw_spin_lock_init(&cfs_b->lock);
3733 cfs_b->quota = RUNTIME_INF;
3734 cfs_b->period = ns_to_ktime(default_cfs_period());
3736 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3737 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3738 cfs_b->period_timer.function = sched_cfs_period_timer;
3739 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3740 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3743 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3745 cfs_rq->runtime_enabled = 0;
3746 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3749 /* requires cfs_b->lock, may release to reprogram timer */
3750 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b, bool force)
3753 * The timer may be active because we're trying to set a new bandwidth
3754 * period or because we're racing with the tear-down path
3755 * (timer_active==0 becomes visible before the hrtimer call-back
3756 * terminates). In either case we ensure that it's re-programmed
3758 while (unlikely(hrtimer_active(&cfs_b->period_timer)) &&
3759 hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) {
3760 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3761 raw_spin_unlock(&cfs_b->lock);
3763 raw_spin_lock(&cfs_b->lock);
3764 /* if someone else restarted the timer then we're done */
3765 if (!force && cfs_b->timer_active)
3769 cfs_b->timer_active = 1;
3770 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3773 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3775 hrtimer_cancel(&cfs_b->period_timer);
3776 hrtimer_cancel(&cfs_b->slack_timer);
3779 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3781 struct cfs_rq *cfs_rq;
3783 for_each_leaf_cfs_rq(rq, cfs_rq) {
3784 if (!cfs_rq->runtime_enabled)
3788 * clock_task is not advancing so we just need to make sure
3789 * there's some valid quota amount
3791 cfs_rq->runtime_remaining = 1;
3792 if (cfs_rq_throttled(cfs_rq))
3793 unthrottle_cfs_rq(cfs_rq);
3797 #else /* CONFIG_CFS_BANDWIDTH */
3798 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3800 return rq_clock_task(rq_of(cfs_rq));
3803 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
3804 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
3805 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3806 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3808 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3813 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3818 static inline int throttled_lb_pair(struct task_group *tg,
3819 int src_cpu, int dest_cpu)
3824 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3826 #ifdef CONFIG_FAIR_GROUP_SCHED
3827 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3830 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3834 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3835 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3837 #endif /* CONFIG_CFS_BANDWIDTH */
3839 /**************************************************
3840 * CFS operations on tasks:
3843 #ifdef CONFIG_SCHED_HRTICK
3844 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3846 struct sched_entity *se = &p->se;
3847 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3849 WARN_ON(task_rq(p) != rq);
3851 if (cfs_rq->nr_running > 1) {
3852 u64 slice = sched_slice(cfs_rq, se);
3853 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3854 s64 delta = slice - ran;
3863 * Don't schedule slices shorter than 10000ns, that just
3864 * doesn't make sense. Rely on vruntime for fairness.
3867 delta = max_t(s64, 10000LL, delta);
3869 hrtick_start(rq, delta);
3874 * called from enqueue/dequeue and updates the hrtick when the
3875 * current task is from our class and nr_running is low enough
3878 static void hrtick_update(struct rq *rq)
3880 struct task_struct *curr = rq->curr;
3882 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3885 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3886 hrtick_start_fair(rq, curr);
3888 #else /* !CONFIG_SCHED_HRTICK */
3890 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3894 static inline void hrtick_update(struct rq *rq)
3900 * The enqueue_task method is called before nr_running is
3901 * increased. Here we update the fair scheduling stats and
3902 * then put the task into the rbtree:
3905 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3907 struct cfs_rq *cfs_rq;
3908 struct sched_entity *se = &p->se;
3910 for_each_sched_entity(se) {
3913 cfs_rq = cfs_rq_of(se);
3914 enqueue_entity(cfs_rq, se, flags);
3917 * end evaluation on encountering a throttled cfs_rq
3919 * note: in the case of encountering a throttled cfs_rq we will
3920 * post the final h_nr_running increment below.
3922 if (cfs_rq_throttled(cfs_rq))
3924 cfs_rq->h_nr_running++;
3926 flags = ENQUEUE_WAKEUP;
3929 for_each_sched_entity(se) {
3930 cfs_rq = cfs_rq_of(se);
3931 cfs_rq->h_nr_running++;
3933 if (cfs_rq_throttled(cfs_rq))
3936 update_cfs_shares(cfs_rq);
3937 update_entity_load_avg(se, 1);
3941 update_rq_runnable_avg(rq, rq->nr_running);
3942 add_nr_running(rq, 1);
3947 static void set_next_buddy(struct sched_entity *se);
3950 * The dequeue_task method is called before nr_running is
3951 * decreased. We remove the task from the rbtree and
3952 * update the fair scheduling stats:
3954 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3956 struct cfs_rq *cfs_rq;
3957 struct sched_entity *se = &p->se;
3958 int task_sleep = flags & DEQUEUE_SLEEP;
3960 for_each_sched_entity(se) {
3961 cfs_rq = cfs_rq_of(se);
3962 dequeue_entity(cfs_rq, se, flags);
3965 * end evaluation on encountering a throttled cfs_rq
3967 * note: in the case of encountering a throttled cfs_rq we will
3968 * post the final h_nr_running decrement below.
3970 if (cfs_rq_throttled(cfs_rq))
3972 cfs_rq->h_nr_running--;
3974 /* Don't dequeue parent if it has other entities besides us */
3975 if (cfs_rq->load.weight) {
3977 * Bias pick_next to pick a task from this cfs_rq, as
3978 * p is sleeping when it is within its sched_slice.
3980 if (task_sleep && parent_entity(se))
3981 set_next_buddy(parent_entity(se));
3983 /* avoid re-evaluating load for this entity */
3984 se = parent_entity(se);
3987 flags |= DEQUEUE_SLEEP;
3990 for_each_sched_entity(se) {
3991 cfs_rq = cfs_rq_of(se);
3992 cfs_rq->h_nr_running--;
3994 if (cfs_rq_throttled(cfs_rq))
3997 update_cfs_shares(cfs_rq);
3998 update_entity_load_avg(se, 1);
4002 sub_nr_running(rq, 1);
4003 update_rq_runnable_avg(rq, 1);
4009 /* Used instead of source_load when we know the type == 0 */
4010 static unsigned long weighted_cpuload(const int cpu)
4012 return cpu_rq(cpu)->cfs.runnable_load_avg;
4016 * Return a low guess at the load of a migration-source cpu weighted
4017 * according to the scheduling class and "nice" value.
4019 * We want to under-estimate the load of migration sources, to
4020 * balance conservatively.
4022 static unsigned long source_load(int cpu, int type)
4024 struct rq *rq = cpu_rq(cpu);
4025 unsigned long total = weighted_cpuload(cpu);
4027 if (type == 0 || !sched_feat(LB_BIAS))
4030 return min(rq->cpu_load[type-1], total);
4034 * Return a high guess at the load of a migration-target cpu weighted
4035 * according to the scheduling class and "nice" value.
4037 static unsigned long target_load(int cpu, int type)
4039 struct rq *rq = cpu_rq(cpu);
4040 unsigned long total = weighted_cpuload(cpu);
4042 if (type == 0 || !sched_feat(LB_BIAS))
4045 return max(rq->cpu_load[type-1], total);
4048 static unsigned long capacity_of(int cpu)
4050 return cpu_rq(cpu)->cpu_capacity;
4053 static unsigned long cpu_avg_load_per_task(int cpu)
4055 struct rq *rq = cpu_rq(cpu);
4056 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
4057 unsigned long load_avg = rq->cfs.runnable_load_avg;
4060 return load_avg / nr_running;
4065 static void record_wakee(struct task_struct *p)
4068 * Rough decay (wiping) for cost saving, don't worry
4069 * about the boundary, really active task won't care
4072 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4073 current->wakee_flips >>= 1;
4074 current->wakee_flip_decay_ts = jiffies;
4077 if (current->last_wakee != p) {
4078 current->last_wakee = p;
4079 current->wakee_flips++;
4083 static void task_waking_fair(struct task_struct *p)
4085 struct sched_entity *se = &p->se;
4086 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4089 #ifndef CONFIG_64BIT
4090 u64 min_vruntime_copy;
4093 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4095 min_vruntime = cfs_rq->min_vruntime;
4096 } while (min_vruntime != min_vruntime_copy);
4098 min_vruntime = cfs_rq->min_vruntime;
4101 se->vruntime -= min_vruntime;
4105 #ifdef CONFIG_FAIR_GROUP_SCHED
4107 * effective_load() calculates the load change as seen from the root_task_group
4109 * Adding load to a group doesn't make a group heavier, but can cause movement
4110 * of group shares between cpus. Assuming the shares were perfectly aligned one
4111 * can calculate the shift in shares.
4113 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4114 * on this @cpu and results in a total addition (subtraction) of @wg to the
4115 * total group weight.
4117 * Given a runqueue weight distribution (rw_i) we can compute a shares
4118 * distribution (s_i) using:
4120 * s_i = rw_i / \Sum rw_j (1)
4122 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4123 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4124 * shares distribution (s_i):
4126 * rw_i = { 2, 4, 1, 0 }
4127 * s_i = { 2/7, 4/7, 1/7, 0 }
4129 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4130 * task used to run on and the CPU the waker is running on), we need to
4131 * compute the effect of waking a task on either CPU and, in case of a sync
4132 * wakeup, compute the effect of the current task going to sleep.
4134 * So for a change of @wl to the local @cpu with an overall group weight change
4135 * of @wl we can compute the new shares distribution (s'_i) using:
4137 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4139 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4140 * differences in waking a task to CPU 0. The additional task changes the
4141 * weight and shares distributions like:
4143 * rw'_i = { 3, 4, 1, 0 }
4144 * s'_i = { 3/8, 4/8, 1/8, 0 }
4146 * We can then compute the difference in effective weight by using:
4148 * dw_i = S * (s'_i - s_i) (3)
4150 * Where 'S' is the group weight as seen by its parent.
4152 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4153 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4154 * 4/7) times the weight of the group.
4156 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4158 struct sched_entity *se = tg->se[cpu];
4160 if (!tg->parent) /* the trivial, non-cgroup case */
4163 for_each_sched_entity(se) {
4169 * W = @wg + \Sum rw_j
4171 W = wg + calc_tg_weight(tg, se->my_q);
4176 w = se->my_q->load.weight + wl;
4179 * wl = S * s'_i; see (2)
4182 wl = (w * tg->shares) / W;
4187 * Per the above, wl is the new se->load.weight value; since
4188 * those are clipped to [MIN_SHARES, ...) do so now. See
4189 * calc_cfs_shares().
4191 if (wl < MIN_SHARES)
4195 * wl = dw_i = S * (s'_i - s_i); see (3)
4197 wl -= se->load.weight;
4200 * Recursively apply this logic to all parent groups to compute
4201 * the final effective load change on the root group. Since
4202 * only the @tg group gets extra weight, all parent groups can
4203 * only redistribute existing shares. @wl is the shift in shares
4204 * resulting from this level per the above.
4213 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4220 static int wake_wide(struct task_struct *p)
4222 int factor = this_cpu_read(sd_llc_size);
4225 * Yeah, it's the switching-frequency, could means many wakee or
4226 * rapidly switch, use factor here will just help to automatically
4227 * adjust the loose-degree, so bigger node will lead to more pull.
4229 if (p->wakee_flips > factor) {
4231 * wakee is somewhat hot, it needs certain amount of cpu
4232 * resource, so if waker is far more hot, prefer to leave
4235 if (current->wakee_flips > (factor * p->wakee_flips))
4242 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4244 s64 this_load, load;
4245 int idx, this_cpu, prev_cpu;
4246 unsigned long tl_per_task;
4247 struct task_group *tg;
4248 unsigned long weight;
4252 * If we wake multiple tasks be careful to not bounce
4253 * ourselves around too much.
4259 this_cpu = smp_processor_id();
4260 prev_cpu = task_cpu(p);
4261 load = source_load(prev_cpu, idx);
4262 this_load = target_load(this_cpu, idx);
4265 * If sync wakeup then subtract the (maximum possible)
4266 * effect of the currently running task from the load
4267 * of the current CPU:
4270 tg = task_group(current);
4271 weight = current->se.load.weight;
4273 this_load += effective_load(tg, this_cpu, -weight, -weight);
4274 load += effective_load(tg, prev_cpu, 0, -weight);
4278 weight = p->se.load.weight;
4281 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4282 * due to the sync cause above having dropped this_load to 0, we'll
4283 * always have an imbalance, but there's really nothing you can do
4284 * about that, so that's good too.
4286 * Otherwise check if either cpus are near enough in load to allow this
4287 * task to be woken on this_cpu.
4289 if (this_load > 0) {
4290 s64 this_eff_load, prev_eff_load;
4292 this_eff_load = 100;
4293 this_eff_load *= capacity_of(prev_cpu);
4294 this_eff_load *= this_load +
4295 effective_load(tg, this_cpu, weight, weight);
4297 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4298 prev_eff_load *= capacity_of(this_cpu);
4299 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4301 balanced = this_eff_load <= prev_eff_load;
4306 * If the currently running task will sleep within
4307 * a reasonable amount of time then attract this newly
4310 if (sync && balanced)
4313 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4314 tl_per_task = cpu_avg_load_per_task(this_cpu);
4317 (this_load <= load &&
4318 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
4320 * This domain has SD_WAKE_AFFINE and
4321 * p is cache cold in this domain, and
4322 * there is no bad imbalance.
4324 schedstat_inc(sd, ttwu_move_affine);
4325 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4333 * find_idlest_group finds and returns the least busy CPU group within the
4336 static struct sched_group *
4337 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4338 int this_cpu, int sd_flag)
4340 struct sched_group *idlest = NULL, *group = sd->groups;
4341 unsigned long min_load = ULONG_MAX, this_load = 0;
4342 int load_idx = sd->forkexec_idx;
4343 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4345 if (sd_flag & SD_BALANCE_WAKE)
4346 load_idx = sd->wake_idx;
4349 unsigned long load, avg_load;
4353 /* Skip over this group if it has no CPUs allowed */
4354 if (!cpumask_intersects(sched_group_cpus(group),
4355 tsk_cpus_allowed(p)))
4358 local_group = cpumask_test_cpu(this_cpu,
4359 sched_group_cpus(group));
4361 /* Tally up the load of all CPUs in the group */
4364 for_each_cpu(i, sched_group_cpus(group)) {
4365 /* Bias balancing toward cpus of our domain */
4367 load = source_load(i, load_idx);
4369 load = target_load(i, load_idx);
4374 /* Adjust by relative CPU capacity of the group */
4375 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
4378 this_load = avg_load;
4379 } else if (avg_load < min_load) {
4380 min_load = avg_load;
4383 } while (group = group->next, group != sd->groups);
4385 if (!idlest || 100*this_load < imbalance*min_load)
4391 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4394 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4396 unsigned long load, min_load = ULONG_MAX;
4400 /* Traverse only the allowed CPUs */
4401 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4402 load = weighted_cpuload(i);
4404 if (load < min_load || (load == min_load && i == this_cpu)) {
4414 * Try and locate an idle CPU in the sched_domain.
4416 static int select_idle_sibling(struct task_struct *p, int target)
4418 struct sched_domain *sd;
4419 struct sched_group *sg;
4420 int i = task_cpu(p);
4422 if (idle_cpu(target))
4426 * If the prevous cpu is cache affine and idle, don't be stupid.
4428 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4432 * Otherwise, iterate the domains and find an elegible idle cpu.
4434 sd = rcu_dereference(per_cpu(sd_llc, target));
4435 for_each_lower_domain(sd) {
4438 if (!cpumask_intersects(sched_group_cpus(sg),
4439 tsk_cpus_allowed(p)))
4442 for_each_cpu(i, sched_group_cpus(sg)) {
4443 if (i == target || !idle_cpu(i))
4447 target = cpumask_first_and(sched_group_cpus(sg),
4448 tsk_cpus_allowed(p));
4452 } while (sg != sd->groups);
4459 * select_task_rq_fair: Select target runqueue for the waking task in domains
4460 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4461 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4463 * Balances load by selecting the idlest cpu in the idlest group, or under
4464 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4466 * Returns the target cpu number.
4468 * preempt must be disabled.
4471 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4473 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4474 int cpu = smp_processor_id();
4476 int want_affine = 0;
4477 int sync = wake_flags & WF_SYNC;
4479 if (p->nr_cpus_allowed == 1)
4482 if (sd_flag & SD_BALANCE_WAKE) {
4483 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
4489 for_each_domain(cpu, tmp) {
4490 if (!(tmp->flags & SD_LOAD_BALANCE))
4494 * If both cpu and prev_cpu are part of this domain,
4495 * cpu is a valid SD_WAKE_AFFINE target.
4497 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4498 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4503 if (tmp->flags & sd_flag)
4507 if (affine_sd && cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4510 if (sd_flag & SD_BALANCE_WAKE) {
4511 new_cpu = select_idle_sibling(p, prev_cpu);
4516 struct sched_group *group;
4519 if (!(sd->flags & sd_flag)) {
4524 group = find_idlest_group(sd, p, cpu, sd_flag);
4530 new_cpu = find_idlest_cpu(group, p, cpu);
4531 if (new_cpu == -1 || new_cpu == cpu) {
4532 /* Now try balancing at a lower domain level of cpu */
4537 /* Now try balancing at a lower domain level of new_cpu */
4539 weight = sd->span_weight;
4541 for_each_domain(cpu, tmp) {
4542 if (weight <= tmp->span_weight)
4544 if (tmp->flags & sd_flag)
4547 /* while loop will break here if sd == NULL */
4556 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4557 * cfs_rq_of(p) references at time of call are still valid and identify the
4558 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4559 * other assumptions, including the state of rq->lock, should be made.
4562 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4564 struct sched_entity *se = &p->se;
4565 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4568 * Load tracking: accumulate removed load so that it can be processed
4569 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4570 * to blocked load iff they have a positive decay-count. It can never
4571 * be negative here since on-rq tasks have decay-count == 0.
4573 if (se->avg.decay_count) {
4574 se->avg.decay_count = -__synchronize_entity_decay(se);
4575 atomic_long_add(se->avg.load_avg_contrib,
4576 &cfs_rq->removed_load);
4579 /* We have migrated, no longer consider this task hot */
4582 #endif /* CONFIG_SMP */
4584 static unsigned long
4585 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4587 unsigned long gran = sysctl_sched_wakeup_granularity;
4590 * Since its curr running now, convert the gran from real-time
4591 * to virtual-time in his units.
4593 * By using 'se' instead of 'curr' we penalize light tasks, so
4594 * they get preempted easier. That is, if 'se' < 'curr' then
4595 * the resulting gran will be larger, therefore penalizing the
4596 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4597 * be smaller, again penalizing the lighter task.
4599 * This is especially important for buddies when the leftmost
4600 * task is higher priority than the buddy.
4602 return calc_delta_fair(gran, se);
4606 * Should 'se' preempt 'curr'.
4620 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4622 s64 gran, vdiff = curr->vruntime - se->vruntime;
4627 gran = wakeup_gran(curr, se);
4634 static void set_last_buddy(struct sched_entity *se)
4636 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4639 for_each_sched_entity(se)
4640 cfs_rq_of(se)->last = se;
4643 static void set_next_buddy(struct sched_entity *se)
4645 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4648 for_each_sched_entity(se)
4649 cfs_rq_of(se)->next = se;
4652 static void set_skip_buddy(struct sched_entity *se)
4654 for_each_sched_entity(se)
4655 cfs_rq_of(se)->skip = se;
4659 * Preempt the current task with a newly woken task if needed:
4661 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4663 struct task_struct *curr = rq->curr;
4664 struct sched_entity *se = &curr->se, *pse = &p->se;
4665 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4666 int scale = cfs_rq->nr_running >= sched_nr_latency;
4667 int next_buddy_marked = 0;
4669 if (unlikely(se == pse))
4673 * This is possible from callers such as move_task(), in which we
4674 * unconditionally check_prempt_curr() after an enqueue (which may have
4675 * lead to a throttle). This both saves work and prevents false
4676 * next-buddy nomination below.
4678 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4681 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4682 set_next_buddy(pse);
4683 next_buddy_marked = 1;
4687 * We can come here with TIF_NEED_RESCHED already set from new task
4690 * Note: this also catches the edge-case of curr being in a throttled
4691 * group (e.g. via set_curr_task), since update_curr() (in the
4692 * enqueue of curr) will have resulted in resched being set. This
4693 * prevents us from potentially nominating it as a false LAST_BUDDY
4696 if (test_tsk_need_resched(curr))
4699 /* Idle tasks are by definition preempted by non-idle tasks. */
4700 if (unlikely(curr->policy == SCHED_IDLE) &&
4701 likely(p->policy != SCHED_IDLE))
4705 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4706 * is driven by the tick):
4708 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4711 find_matching_se(&se, &pse);
4712 update_curr(cfs_rq_of(se));
4714 if (wakeup_preempt_entity(se, pse) == 1) {
4716 * Bias pick_next to pick the sched entity that is
4717 * triggering this preemption.
4719 if (!next_buddy_marked)
4720 set_next_buddy(pse);
4729 * Only set the backward buddy when the current task is still
4730 * on the rq. This can happen when a wakeup gets interleaved
4731 * with schedule on the ->pre_schedule() or idle_balance()
4732 * point, either of which can * drop the rq lock.
4734 * Also, during early boot the idle thread is in the fair class,
4735 * for obvious reasons its a bad idea to schedule back to it.
4737 if (unlikely(!se->on_rq || curr == rq->idle))
4740 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4744 static struct task_struct *
4745 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
4747 struct cfs_rq *cfs_rq = &rq->cfs;
4748 struct sched_entity *se;
4749 struct task_struct *p;
4753 #ifdef CONFIG_FAIR_GROUP_SCHED
4754 if (!cfs_rq->nr_running)
4757 if (prev->sched_class != &fair_sched_class)
4761 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
4762 * likely that a next task is from the same cgroup as the current.
4764 * Therefore attempt to avoid putting and setting the entire cgroup
4765 * hierarchy, only change the part that actually changes.
4769 struct sched_entity *curr = cfs_rq->curr;
4772 * Since we got here without doing put_prev_entity() we also
4773 * have to consider cfs_rq->curr. If it is still a runnable
4774 * entity, update_curr() will update its vruntime, otherwise
4775 * forget we've ever seen it.
4777 if (curr && curr->on_rq)
4778 update_curr(cfs_rq);
4783 * This call to check_cfs_rq_runtime() will do the throttle and
4784 * dequeue its entity in the parent(s). Therefore the 'simple'
4785 * nr_running test will indeed be correct.
4787 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
4790 se = pick_next_entity(cfs_rq, curr);
4791 cfs_rq = group_cfs_rq(se);
4797 * Since we haven't yet done put_prev_entity and if the selected task
4798 * is a different task than we started out with, try and touch the
4799 * least amount of cfs_rqs.
4802 struct sched_entity *pse = &prev->se;
4804 while (!(cfs_rq = is_same_group(se, pse))) {
4805 int se_depth = se->depth;
4806 int pse_depth = pse->depth;
4808 if (se_depth <= pse_depth) {
4809 put_prev_entity(cfs_rq_of(pse), pse);
4810 pse = parent_entity(pse);
4812 if (se_depth >= pse_depth) {
4813 set_next_entity(cfs_rq_of(se), se);
4814 se = parent_entity(se);
4818 put_prev_entity(cfs_rq, pse);
4819 set_next_entity(cfs_rq, se);
4822 if (hrtick_enabled(rq))
4823 hrtick_start_fair(rq, p);
4830 if (!cfs_rq->nr_running)
4833 put_prev_task(rq, prev);
4836 se = pick_next_entity(cfs_rq, NULL);
4837 set_next_entity(cfs_rq, se);
4838 cfs_rq = group_cfs_rq(se);
4843 if (hrtick_enabled(rq))
4844 hrtick_start_fair(rq, p);
4849 new_tasks = idle_balance(rq);
4851 * Because idle_balance() releases (and re-acquires) rq->lock, it is
4852 * possible for any higher priority task to appear. In that case we
4853 * must re-start the pick_next_entity() loop.
4865 * Account for a descheduled task:
4867 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4869 struct sched_entity *se = &prev->se;
4870 struct cfs_rq *cfs_rq;
4872 for_each_sched_entity(se) {
4873 cfs_rq = cfs_rq_of(se);
4874 put_prev_entity(cfs_rq, se);
4879 * sched_yield() is very simple
4881 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4883 static void yield_task_fair(struct rq *rq)
4885 struct task_struct *curr = rq->curr;
4886 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4887 struct sched_entity *se = &curr->se;
4890 * Are we the only task in the tree?
4892 if (unlikely(rq->nr_running == 1))
4895 clear_buddies(cfs_rq, se);
4897 if (curr->policy != SCHED_BATCH) {
4898 update_rq_clock(rq);
4900 * Update run-time statistics of the 'current'.
4902 update_curr(cfs_rq);
4904 * Tell update_rq_clock() that we've just updated,
4905 * so we don't do microscopic update in schedule()
4906 * and double the fastpath cost.
4908 rq->skip_clock_update = 1;
4914 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4916 struct sched_entity *se = &p->se;
4918 /* throttled hierarchies are not runnable */
4919 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4922 /* Tell the scheduler that we'd really like pse to run next. */
4925 yield_task_fair(rq);
4931 /**************************************************
4932 * Fair scheduling class load-balancing methods.
4936 * The purpose of load-balancing is to achieve the same basic fairness the
4937 * per-cpu scheduler provides, namely provide a proportional amount of compute
4938 * time to each task. This is expressed in the following equation:
4940 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4942 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4943 * W_i,0 is defined as:
4945 * W_i,0 = \Sum_j w_i,j (2)
4947 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4948 * is derived from the nice value as per prio_to_weight[].
4950 * The weight average is an exponential decay average of the instantaneous
4953 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4955 * C_i is the compute capacity of cpu i, typically it is the
4956 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4957 * can also include other factors [XXX].
4959 * To achieve this balance we define a measure of imbalance which follows
4960 * directly from (1):
4962 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
4964 * We them move tasks around to minimize the imbalance. In the continuous
4965 * function space it is obvious this converges, in the discrete case we get
4966 * a few fun cases generally called infeasible weight scenarios.
4969 * - infeasible weights;
4970 * - local vs global optima in the discrete case. ]
4975 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4976 * for all i,j solution, we create a tree of cpus that follows the hardware
4977 * topology where each level pairs two lower groups (or better). This results
4978 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4979 * tree to only the first of the previous level and we decrease the frequency
4980 * of load-balance at each level inv. proportional to the number of cpus in
4986 * \Sum { --- * --- * 2^i } = O(n) (5)
4988 * `- size of each group
4989 * | | `- number of cpus doing load-balance
4991 * `- sum over all levels
4993 * Coupled with a limit on how many tasks we can migrate every balance pass,
4994 * this makes (5) the runtime complexity of the balancer.
4996 * An important property here is that each CPU is still (indirectly) connected
4997 * to every other cpu in at most O(log n) steps:
4999 * The adjacency matrix of the resulting graph is given by:
5002 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5005 * And you'll find that:
5007 * A^(log_2 n)_i,j != 0 for all i,j (7)
5009 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5010 * The task movement gives a factor of O(m), giving a convergence complexity
5013 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5018 * In order to avoid CPUs going idle while there's still work to do, new idle
5019 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5020 * tree itself instead of relying on other CPUs to bring it work.
5022 * This adds some complexity to both (5) and (8) but it reduces the total idle
5030 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5033 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5038 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5040 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5042 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5045 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5046 * rewrite all of this once again.]
5049 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5051 enum fbq_type { regular, remote, all };
5053 #define LBF_ALL_PINNED 0x01
5054 #define LBF_NEED_BREAK 0x02
5055 #define LBF_DST_PINNED 0x04
5056 #define LBF_SOME_PINNED 0x08
5059 struct sched_domain *sd;
5067 struct cpumask *dst_grpmask;
5069 enum cpu_idle_type idle;
5071 /* The set of CPUs under consideration for load-balancing */
5072 struct cpumask *cpus;
5077 unsigned int loop_break;
5078 unsigned int loop_max;
5080 enum fbq_type fbq_type;
5084 * move_task - move a task from one runqueue to another runqueue.
5085 * Both runqueues must be locked.
5087 static void move_task(struct task_struct *p, struct lb_env *env)
5089 deactivate_task(env->src_rq, p, 0);
5090 set_task_cpu(p, env->dst_cpu);
5091 activate_task(env->dst_rq, p, 0);
5092 check_preempt_curr(env->dst_rq, p, 0);
5096 * Is this task likely cache-hot:
5099 task_hot(struct task_struct *p, u64 now)
5103 if (p->sched_class != &fair_sched_class)
5106 if (unlikely(p->policy == SCHED_IDLE))
5110 * Buddy candidates are cache hot:
5112 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
5113 (&p->se == cfs_rq_of(&p->se)->next ||
5114 &p->se == cfs_rq_of(&p->se)->last))
5117 if (sysctl_sched_migration_cost == -1)
5119 if (sysctl_sched_migration_cost == 0)
5122 delta = now - p->se.exec_start;
5124 return delta < (s64)sysctl_sched_migration_cost;
5127 #ifdef CONFIG_NUMA_BALANCING
5128 /* Returns true if the destination node has incurred more faults */
5129 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
5131 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5132 int src_nid, dst_nid;
5134 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults_memory ||
5135 !(env->sd->flags & SD_NUMA)) {
5139 src_nid = cpu_to_node(env->src_cpu);
5140 dst_nid = cpu_to_node(env->dst_cpu);
5142 if (src_nid == dst_nid)
5146 /* Task is already in the group's interleave set. */
5147 if (node_isset(src_nid, numa_group->active_nodes))
5150 /* Task is moving into the group's interleave set. */
5151 if (node_isset(dst_nid, numa_group->active_nodes))
5154 return group_faults(p, dst_nid) > group_faults(p, src_nid);
5157 /* Encourage migration to the preferred node. */
5158 if (dst_nid == p->numa_preferred_nid)
5161 return task_faults(p, dst_nid) > task_faults(p, src_nid);
5165 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5167 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5168 int src_nid, dst_nid;
5170 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
5173 if (!p->numa_faults_memory || !(env->sd->flags & SD_NUMA))
5176 src_nid = cpu_to_node(env->src_cpu);
5177 dst_nid = cpu_to_node(env->dst_cpu);
5179 if (src_nid == dst_nid)
5183 /* Task is moving within/into the group's interleave set. */
5184 if (node_isset(dst_nid, numa_group->active_nodes))
5187 /* Task is moving out of the group's interleave set. */
5188 if (node_isset(src_nid, numa_group->active_nodes))
5191 return group_faults(p, dst_nid) < group_faults(p, src_nid);
5194 /* Migrating away from the preferred node is always bad. */
5195 if (src_nid == p->numa_preferred_nid)
5198 return task_faults(p, dst_nid) < task_faults(p, src_nid);
5202 static inline bool migrate_improves_locality(struct task_struct *p,
5208 static inline bool migrate_degrades_locality(struct task_struct *p,
5216 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5219 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5221 int tsk_cache_hot = 0;
5223 * We do not migrate tasks that are:
5224 * 1) throttled_lb_pair, or
5225 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5226 * 3) running (obviously), or
5227 * 4) are cache-hot on their current CPU.
5229 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5232 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5235 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5237 env->flags |= LBF_SOME_PINNED;
5240 * Remember if this task can be migrated to any other cpu in
5241 * our sched_group. We may want to revisit it if we couldn't
5242 * meet load balance goals by pulling other tasks on src_cpu.
5244 * Also avoid computing new_dst_cpu if we have already computed
5245 * one in current iteration.
5247 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5250 /* Prevent to re-select dst_cpu via env's cpus */
5251 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5252 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5253 env->flags |= LBF_DST_PINNED;
5254 env->new_dst_cpu = cpu;
5262 /* Record that we found atleast one task that could run on dst_cpu */
5263 env->flags &= ~LBF_ALL_PINNED;
5265 if (task_running(env->src_rq, p)) {
5266 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5271 * Aggressive migration if:
5272 * 1) destination numa is preferred
5273 * 2) task is cache cold, or
5274 * 3) too many balance attempts have failed.
5276 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq));
5278 tsk_cache_hot = migrate_degrades_locality(p, env);
5280 if (migrate_improves_locality(p, env)) {
5281 #ifdef CONFIG_SCHEDSTATS
5282 if (tsk_cache_hot) {
5283 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5284 schedstat_inc(p, se.statistics.nr_forced_migrations);
5290 if (!tsk_cache_hot ||
5291 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5293 if (tsk_cache_hot) {
5294 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5295 schedstat_inc(p, se.statistics.nr_forced_migrations);
5301 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5306 * move_one_task tries to move exactly one task from busiest to this_rq, as
5307 * part of active balancing operations within "domain".
5308 * Returns 1 if successful and 0 otherwise.
5310 * Called with both runqueues locked.
5312 static int move_one_task(struct lb_env *env)
5314 struct task_struct *p, *n;
5316 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5317 if (!can_migrate_task(p, env))
5322 * Right now, this is only the second place move_task()
5323 * is called, so we can safely collect move_task()
5324 * stats here rather than inside move_task().
5326 schedstat_inc(env->sd, lb_gained[env->idle]);
5332 static const unsigned int sched_nr_migrate_break = 32;
5335 * move_tasks tries to move up to imbalance weighted load from busiest to
5336 * this_rq, as part of a balancing operation within domain "sd".
5337 * Returns 1 if successful and 0 otherwise.
5339 * Called with both runqueues locked.
5341 static int move_tasks(struct lb_env *env)
5343 struct list_head *tasks = &env->src_rq->cfs_tasks;
5344 struct task_struct *p;
5348 if (env->imbalance <= 0)
5351 while (!list_empty(tasks)) {
5352 p = list_first_entry(tasks, struct task_struct, se.group_node);
5355 /* We've more or less seen every task there is, call it quits */
5356 if (env->loop > env->loop_max)
5359 /* take a breather every nr_migrate tasks */
5360 if (env->loop > env->loop_break) {
5361 env->loop_break += sched_nr_migrate_break;
5362 env->flags |= LBF_NEED_BREAK;
5366 if (!can_migrate_task(p, env))
5369 load = task_h_load(p);
5371 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5374 if ((load / 2) > env->imbalance)
5379 env->imbalance -= load;
5381 #ifdef CONFIG_PREEMPT
5383 * NEWIDLE balancing is a source of latency, so preemptible
5384 * kernels will stop after the first task is pulled to minimize
5385 * the critical section.
5387 if (env->idle == CPU_NEWLY_IDLE)
5392 * We only want to steal up to the prescribed amount of
5395 if (env->imbalance <= 0)
5400 list_move_tail(&p->se.group_node, tasks);
5404 * Right now, this is one of only two places move_task() is called,
5405 * so we can safely collect move_task() stats here rather than
5406 * inside move_task().
5408 schedstat_add(env->sd, lb_gained[env->idle], pulled);
5413 #ifdef CONFIG_FAIR_GROUP_SCHED
5415 * update tg->load_weight by folding this cpu's load_avg
5417 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5419 struct sched_entity *se = tg->se[cpu];
5420 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5422 /* throttled entities do not contribute to load */
5423 if (throttled_hierarchy(cfs_rq))
5426 update_cfs_rq_blocked_load(cfs_rq, 1);
5429 update_entity_load_avg(se, 1);
5431 * We pivot on our runnable average having decayed to zero for
5432 * list removal. This generally implies that all our children
5433 * have also been removed (modulo rounding error or bandwidth
5434 * control); however, such cases are rare and we can fix these
5437 * TODO: fix up out-of-order children on enqueue.
5439 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5440 list_del_leaf_cfs_rq(cfs_rq);
5442 struct rq *rq = rq_of(cfs_rq);
5443 update_rq_runnable_avg(rq, rq->nr_running);
5447 static void update_blocked_averages(int cpu)
5449 struct rq *rq = cpu_rq(cpu);
5450 struct cfs_rq *cfs_rq;
5451 unsigned long flags;
5453 raw_spin_lock_irqsave(&rq->lock, flags);
5454 update_rq_clock(rq);
5456 * Iterates the task_group tree in a bottom up fashion, see
5457 * list_add_leaf_cfs_rq() for details.
5459 for_each_leaf_cfs_rq(rq, cfs_rq) {
5461 * Note: We may want to consider periodically releasing
5462 * rq->lock about these updates so that creating many task
5463 * groups does not result in continually extending hold time.
5465 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5468 raw_spin_unlock_irqrestore(&rq->lock, flags);
5472 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5473 * This needs to be done in a top-down fashion because the load of a child
5474 * group is a fraction of its parents load.
5476 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5478 struct rq *rq = rq_of(cfs_rq);
5479 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5480 unsigned long now = jiffies;
5483 if (cfs_rq->last_h_load_update == now)
5486 cfs_rq->h_load_next = NULL;
5487 for_each_sched_entity(se) {
5488 cfs_rq = cfs_rq_of(se);
5489 cfs_rq->h_load_next = se;
5490 if (cfs_rq->last_h_load_update == now)
5495 cfs_rq->h_load = cfs_rq->runnable_load_avg;
5496 cfs_rq->last_h_load_update = now;
5499 while ((se = cfs_rq->h_load_next) != NULL) {
5500 load = cfs_rq->h_load;
5501 load = div64_ul(load * se->avg.load_avg_contrib,
5502 cfs_rq->runnable_load_avg + 1);
5503 cfs_rq = group_cfs_rq(se);
5504 cfs_rq->h_load = load;
5505 cfs_rq->last_h_load_update = now;
5509 static unsigned long task_h_load(struct task_struct *p)
5511 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5513 update_cfs_rq_h_load(cfs_rq);
5514 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
5515 cfs_rq->runnable_load_avg + 1);
5518 static inline void update_blocked_averages(int cpu)
5522 static unsigned long task_h_load(struct task_struct *p)
5524 return p->se.avg.load_avg_contrib;
5528 /********** Helpers for find_busiest_group ************************/
5530 * sg_lb_stats - stats of a sched_group required for load_balancing
5532 struct sg_lb_stats {
5533 unsigned long avg_load; /*Avg load across the CPUs of the group */
5534 unsigned long group_load; /* Total load over the CPUs of the group */
5535 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5536 unsigned long load_per_task;
5537 unsigned long group_capacity;
5538 unsigned int sum_nr_running; /* Nr tasks running in the group */
5539 unsigned int group_capacity_factor;
5540 unsigned int idle_cpus;
5541 unsigned int group_weight;
5542 int group_imb; /* Is there an imbalance in the group ? */
5543 int group_has_free_capacity;
5544 #ifdef CONFIG_NUMA_BALANCING
5545 unsigned int nr_numa_running;
5546 unsigned int nr_preferred_running;
5551 * sd_lb_stats - Structure to store the statistics of a sched_domain
5552 * during load balancing.
5554 struct sd_lb_stats {
5555 struct sched_group *busiest; /* Busiest group in this sd */
5556 struct sched_group *local; /* Local group in this sd */
5557 unsigned long total_load; /* Total load of all groups in sd */
5558 unsigned long total_capacity; /* Total capacity of all groups in sd */
5559 unsigned long avg_load; /* Average load across all groups in sd */
5561 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5562 struct sg_lb_stats local_stat; /* Statistics of the local group */
5565 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5568 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5569 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5570 * We must however clear busiest_stat::avg_load because
5571 * update_sd_pick_busiest() reads this before assignment.
5573 *sds = (struct sd_lb_stats){
5577 .total_capacity = 0UL,
5585 * get_sd_load_idx - Obtain the load index for a given sched domain.
5586 * @sd: The sched_domain whose load_idx is to be obtained.
5587 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5589 * Return: The load index.
5591 static inline int get_sd_load_idx(struct sched_domain *sd,
5592 enum cpu_idle_type idle)
5598 load_idx = sd->busy_idx;
5601 case CPU_NEWLY_IDLE:
5602 load_idx = sd->newidle_idx;
5605 load_idx = sd->idle_idx;
5612 static unsigned long default_scale_capacity(struct sched_domain *sd, int cpu)
5614 return SCHED_CAPACITY_SCALE;
5617 unsigned long __weak arch_scale_freq_capacity(struct sched_domain *sd, int cpu)
5619 return default_scale_capacity(sd, cpu);
5622 static unsigned long default_scale_smt_capacity(struct sched_domain *sd, int cpu)
5624 unsigned long weight = sd->span_weight;
5625 unsigned long smt_gain = sd->smt_gain;
5632 unsigned long __weak arch_scale_smt_capacity(struct sched_domain *sd, int cpu)
5634 return default_scale_smt_capacity(sd, cpu);
5637 static unsigned long scale_rt_capacity(int cpu)
5639 struct rq *rq = cpu_rq(cpu);
5640 u64 total, available, age_stamp, avg;
5644 * Since we're reading these variables without serialization make sure
5645 * we read them once before doing sanity checks on them.
5647 age_stamp = ACCESS_ONCE(rq->age_stamp);
5648 avg = ACCESS_ONCE(rq->rt_avg);
5650 delta = rq_clock(rq) - age_stamp;
5651 if (unlikely(delta < 0))
5654 total = sched_avg_period() + delta;
5656 if (unlikely(total < avg)) {
5657 /* Ensures that capacity won't end up being negative */
5660 available = total - avg;
5663 if (unlikely((s64)total < SCHED_CAPACITY_SCALE))
5664 total = SCHED_CAPACITY_SCALE;
5666 total >>= SCHED_CAPACITY_SHIFT;
5668 return div_u64(available, total);
5671 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
5673 unsigned long weight = sd->span_weight;
5674 unsigned long capacity = SCHED_CAPACITY_SCALE;
5675 struct sched_group *sdg = sd->groups;
5677 if ((sd->flags & SD_SHARE_CPUCAPACITY) && weight > 1) {
5678 if (sched_feat(ARCH_CAPACITY))
5679 capacity *= arch_scale_smt_capacity(sd, cpu);
5681 capacity *= default_scale_smt_capacity(sd, cpu);
5683 capacity >>= SCHED_CAPACITY_SHIFT;
5686 sdg->sgc->capacity_orig = capacity;
5688 if (sched_feat(ARCH_CAPACITY))
5689 capacity *= arch_scale_freq_capacity(sd, cpu);
5691 capacity *= default_scale_capacity(sd, cpu);
5693 capacity >>= SCHED_CAPACITY_SHIFT;
5695 capacity *= scale_rt_capacity(cpu);
5696 capacity >>= SCHED_CAPACITY_SHIFT;
5701 cpu_rq(cpu)->cpu_capacity = capacity;
5702 sdg->sgc->capacity = capacity;
5705 void update_group_capacity(struct sched_domain *sd, int cpu)
5707 struct sched_domain *child = sd->child;
5708 struct sched_group *group, *sdg = sd->groups;
5709 unsigned long capacity, capacity_orig;
5710 unsigned long interval;
5712 interval = msecs_to_jiffies(sd->balance_interval);
5713 interval = clamp(interval, 1UL, max_load_balance_interval);
5714 sdg->sgc->next_update = jiffies + interval;
5717 update_cpu_capacity(sd, cpu);
5721 capacity_orig = capacity = 0;
5723 if (child->flags & SD_OVERLAP) {
5725 * SD_OVERLAP domains cannot assume that child groups
5726 * span the current group.
5729 for_each_cpu(cpu, sched_group_cpus(sdg)) {
5730 struct sched_group_capacity *sgc;
5731 struct rq *rq = cpu_rq(cpu);
5734 * build_sched_domains() -> init_sched_groups_capacity()
5735 * gets here before we've attached the domains to the
5738 * Use capacity_of(), which is set irrespective of domains
5739 * in update_cpu_capacity().
5741 * This avoids capacity/capacity_orig from being 0 and
5742 * causing divide-by-zero issues on boot.
5744 * Runtime updates will correct capacity_orig.
5746 if (unlikely(!rq->sd)) {
5747 capacity_orig += capacity_of(cpu);
5748 capacity += capacity_of(cpu);
5752 sgc = rq->sd->groups->sgc;
5753 capacity_orig += sgc->capacity_orig;
5754 capacity += sgc->capacity;
5758 * !SD_OVERLAP domains can assume that child groups
5759 * span the current group.
5762 group = child->groups;
5764 capacity_orig += group->sgc->capacity_orig;
5765 capacity += group->sgc->capacity;
5766 group = group->next;
5767 } while (group != child->groups);
5770 sdg->sgc->capacity_orig = capacity_orig;
5771 sdg->sgc->capacity = capacity;
5775 * Try and fix up capacity for tiny siblings, this is needed when
5776 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5777 * which on its own isn't powerful enough.
5779 * See update_sd_pick_busiest() and check_asym_packing().
5782 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5785 * Only siblings can have significantly less than SCHED_CAPACITY_SCALE
5787 if (!(sd->flags & SD_SHARE_CPUCAPACITY))
5791 * If ~90% of the cpu_capacity is still there, we're good.
5793 if (group->sgc->capacity * 32 > group->sgc->capacity_orig * 29)
5800 * Group imbalance indicates (and tries to solve) the problem where balancing
5801 * groups is inadequate due to tsk_cpus_allowed() constraints.
5803 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5804 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5807 * { 0 1 2 3 } { 4 5 6 7 }
5810 * If we were to balance group-wise we'd place two tasks in the first group and
5811 * two tasks in the second group. Clearly this is undesired as it will overload
5812 * cpu 3 and leave one of the cpus in the second group unused.
5814 * The current solution to this issue is detecting the skew in the first group
5815 * by noticing the lower domain failed to reach balance and had difficulty
5816 * moving tasks due to affinity constraints.
5818 * When this is so detected; this group becomes a candidate for busiest; see
5819 * update_sd_pick_busiest(). And calculate_imbalance() and
5820 * find_busiest_group() avoid some of the usual balance conditions to allow it
5821 * to create an effective group imbalance.
5823 * This is a somewhat tricky proposition since the next run might not find the
5824 * group imbalance and decide the groups need to be balanced again. A most
5825 * subtle and fragile situation.
5828 static inline int sg_imbalanced(struct sched_group *group)
5830 return group->sgc->imbalance;
5834 * Compute the group capacity factor.
5836 * Avoid the issue where N*frac(smt_capacity) >= 1 creates 'phantom' cores by
5837 * first dividing out the smt factor and computing the actual number of cores
5838 * and limit unit capacity with that.
5840 static inline int sg_capacity_factor(struct lb_env *env, struct sched_group *group)
5842 unsigned int capacity_factor, smt, cpus;
5843 unsigned int capacity, capacity_orig;
5845 capacity = group->sgc->capacity;
5846 capacity_orig = group->sgc->capacity_orig;
5847 cpus = group->group_weight;
5849 /* smt := ceil(cpus / capacity), assumes: 1 < smt_capacity < 2 */
5850 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, capacity_orig);
5851 capacity_factor = cpus / smt; /* cores */
5853 capacity_factor = min_t(unsigned,
5854 capacity_factor, DIV_ROUND_CLOSEST(capacity, SCHED_CAPACITY_SCALE));
5855 if (!capacity_factor)
5856 capacity_factor = fix_small_capacity(env->sd, group);
5858 return capacity_factor;
5862 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5863 * @env: The load balancing environment.
5864 * @group: sched_group whose statistics are to be updated.
5865 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5866 * @local_group: Does group contain this_cpu.
5867 * @sgs: variable to hold the statistics for this group.
5869 static inline void update_sg_lb_stats(struct lb_env *env,
5870 struct sched_group *group, int load_idx,
5871 int local_group, struct sg_lb_stats *sgs)
5876 memset(sgs, 0, sizeof(*sgs));
5878 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5879 struct rq *rq = cpu_rq(i);
5881 /* Bias balancing toward cpus of our domain */
5883 load = target_load(i, load_idx);
5885 load = source_load(i, load_idx);
5887 sgs->group_load += load;
5888 sgs->sum_nr_running += rq->nr_running;
5889 #ifdef CONFIG_NUMA_BALANCING
5890 sgs->nr_numa_running += rq->nr_numa_running;
5891 sgs->nr_preferred_running += rq->nr_preferred_running;
5893 sgs->sum_weighted_load += weighted_cpuload(i);
5898 /* Adjust by relative CPU capacity of the group */
5899 sgs->group_capacity = group->sgc->capacity;
5900 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
5902 if (sgs->sum_nr_running)
5903 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
5905 sgs->group_weight = group->group_weight;
5907 sgs->group_imb = sg_imbalanced(group);
5908 sgs->group_capacity_factor = sg_capacity_factor(env, group);
5910 if (sgs->group_capacity_factor > sgs->sum_nr_running)
5911 sgs->group_has_free_capacity = 1;
5915 * update_sd_pick_busiest - return 1 on busiest group
5916 * @env: The load balancing environment.
5917 * @sds: sched_domain statistics
5918 * @sg: sched_group candidate to be checked for being the busiest
5919 * @sgs: sched_group statistics
5921 * Determine if @sg is a busier group than the previously selected
5924 * Return: %true if @sg is a busier group than the previously selected
5925 * busiest group. %false otherwise.
5927 static bool update_sd_pick_busiest(struct lb_env *env,
5928 struct sd_lb_stats *sds,
5929 struct sched_group *sg,
5930 struct sg_lb_stats *sgs)
5932 if (sgs->avg_load <= sds->busiest_stat.avg_load)
5935 if (sgs->sum_nr_running > sgs->group_capacity_factor)
5942 * ASYM_PACKING needs to move all the work to the lowest
5943 * numbered CPUs in the group, therefore mark all groups
5944 * higher than ourself as busy.
5946 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5947 env->dst_cpu < group_first_cpu(sg)) {
5951 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5958 #ifdef CONFIG_NUMA_BALANCING
5959 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5961 if (sgs->sum_nr_running > sgs->nr_numa_running)
5963 if (sgs->sum_nr_running > sgs->nr_preferred_running)
5968 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5970 if (rq->nr_running > rq->nr_numa_running)
5972 if (rq->nr_running > rq->nr_preferred_running)
5977 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5982 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5986 #endif /* CONFIG_NUMA_BALANCING */
5989 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5990 * @env: The load balancing environment.
5991 * @sds: variable to hold the statistics for this sched_domain.
5993 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
5995 struct sched_domain *child = env->sd->child;
5996 struct sched_group *sg = env->sd->groups;
5997 struct sg_lb_stats tmp_sgs;
5998 int load_idx, prefer_sibling = 0;
6000 if (child && child->flags & SD_PREFER_SIBLING)
6003 load_idx = get_sd_load_idx(env->sd, env->idle);
6006 struct sg_lb_stats *sgs = &tmp_sgs;
6009 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6012 sgs = &sds->local_stat;
6014 if (env->idle != CPU_NEWLY_IDLE ||
6015 time_after_eq(jiffies, sg->sgc->next_update))
6016 update_group_capacity(env->sd, env->dst_cpu);
6019 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
6025 * In case the child domain prefers tasks go to siblings
6026 * first, lower the sg capacity factor to one so that we'll try
6027 * and move all the excess tasks away. We lower the capacity
6028 * of a group only if the local group has the capacity to fit
6029 * these excess tasks, i.e. nr_running < group_capacity_factor. The
6030 * extra check prevents the case where you always pull from the
6031 * heaviest group when it is already under-utilized (possible
6032 * with a large weight task outweighs the tasks on the system).
6034 if (prefer_sibling && sds->local &&
6035 sds->local_stat.group_has_free_capacity)
6036 sgs->group_capacity_factor = min(sgs->group_capacity_factor, 1U);
6038 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6040 sds->busiest_stat = *sgs;
6044 /* Now, start updating sd_lb_stats */
6045 sds->total_load += sgs->group_load;
6046 sds->total_capacity += sgs->group_capacity;
6049 } while (sg != env->sd->groups);
6051 if (env->sd->flags & SD_NUMA)
6052 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6056 * check_asym_packing - Check to see if the group is packed into the
6059 * This is primarily intended to used at the sibling level. Some
6060 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6061 * case of POWER7, it can move to lower SMT modes only when higher
6062 * threads are idle. When in lower SMT modes, the threads will
6063 * perform better since they share less core resources. Hence when we
6064 * have idle threads, we want them to be the higher ones.
6066 * This packing function is run on idle threads. It checks to see if
6067 * the busiest CPU in this domain (core in the P7 case) has a higher
6068 * CPU number than the packing function is being run on. Here we are
6069 * assuming lower CPU number will be equivalent to lower a SMT thread
6072 * Return: 1 when packing is required and a task should be moved to
6073 * this CPU. The amount of the imbalance is returned in *imbalance.
6075 * @env: The load balancing environment.
6076 * @sds: Statistics of the sched_domain which is to be packed
6078 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6082 if (!(env->sd->flags & SD_ASYM_PACKING))
6088 busiest_cpu = group_first_cpu(sds->busiest);
6089 if (env->dst_cpu > busiest_cpu)
6092 env->imbalance = DIV_ROUND_CLOSEST(
6093 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6094 SCHED_CAPACITY_SCALE);
6100 * fix_small_imbalance - Calculate the minor imbalance that exists
6101 * amongst the groups of a sched_domain, during
6103 * @env: The load balancing environment.
6104 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6107 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6109 unsigned long tmp, capa_now = 0, capa_move = 0;
6110 unsigned int imbn = 2;
6111 unsigned long scaled_busy_load_per_task;
6112 struct sg_lb_stats *local, *busiest;
6114 local = &sds->local_stat;
6115 busiest = &sds->busiest_stat;
6117 if (!local->sum_nr_running)
6118 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6119 else if (busiest->load_per_task > local->load_per_task)
6122 scaled_busy_load_per_task =
6123 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6124 busiest->group_capacity;
6126 if (busiest->avg_load + scaled_busy_load_per_task >=
6127 local->avg_load + (scaled_busy_load_per_task * imbn)) {
6128 env->imbalance = busiest->load_per_task;
6133 * OK, we don't have enough imbalance to justify moving tasks,
6134 * however we may be able to increase total CPU capacity used by
6138 capa_now += busiest->group_capacity *
6139 min(busiest->load_per_task, busiest->avg_load);
6140 capa_now += local->group_capacity *
6141 min(local->load_per_task, local->avg_load);
6142 capa_now /= SCHED_CAPACITY_SCALE;
6144 /* Amount of load we'd subtract */
6145 if (busiest->avg_load > scaled_busy_load_per_task) {
6146 capa_move += busiest->group_capacity *
6147 min(busiest->load_per_task,
6148 busiest->avg_load - scaled_busy_load_per_task);
6151 /* Amount of load we'd add */
6152 if (busiest->avg_load * busiest->group_capacity <
6153 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
6154 tmp = (busiest->avg_load * busiest->group_capacity) /
6155 local->group_capacity;
6157 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6158 local->group_capacity;
6160 capa_move += local->group_capacity *
6161 min(local->load_per_task, local->avg_load + tmp);
6162 capa_move /= SCHED_CAPACITY_SCALE;
6164 /* Move if we gain throughput */
6165 if (capa_move > capa_now)
6166 env->imbalance = busiest->load_per_task;
6170 * calculate_imbalance - Calculate the amount of imbalance present within the
6171 * groups of a given sched_domain during load balance.
6172 * @env: load balance environment
6173 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6175 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6177 unsigned long max_pull, load_above_capacity = ~0UL;
6178 struct sg_lb_stats *local, *busiest;
6180 local = &sds->local_stat;
6181 busiest = &sds->busiest_stat;
6183 if (busiest->group_imb) {
6185 * In the group_imb case we cannot rely on group-wide averages
6186 * to ensure cpu-load equilibrium, look at wider averages. XXX
6188 busiest->load_per_task =
6189 min(busiest->load_per_task, sds->avg_load);
6193 * In the presence of smp nice balancing, certain scenarios can have
6194 * max load less than avg load(as we skip the groups at or below
6195 * its cpu_capacity, while calculating max_load..)
6197 if (busiest->avg_load <= sds->avg_load ||
6198 local->avg_load >= sds->avg_load) {
6200 return fix_small_imbalance(env, sds);
6203 if (!busiest->group_imb) {
6205 * Don't want to pull so many tasks that a group would go idle.
6206 * Except of course for the group_imb case, since then we might
6207 * have to drop below capacity to reach cpu-load equilibrium.
6209 load_above_capacity =
6210 (busiest->sum_nr_running - busiest->group_capacity_factor);
6212 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_CAPACITY_SCALE);
6213 load_above_capacity /= busiest->group_capacity;
6217 * We're trying to get all the cpus to the average_load, so we don't
6218 * want to push ourselves above the average load, nor do we wish to
6219 * reduce the max loaded cpu below the average load. At the same time,
6220 * we also don't want to reduce the group load below the group capacity
6221 * (so that we can implement power-savings policies etc). Thus we look
6222 * for the minimum possible imbalance.
6224 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6226 /* How much load to actually move to equalise the imbalance */
6227 env->imbalance = min(
6228 max_pull * busiest->group_capacity,
6229 (sds->avg_load - local->avg_load) * local->group_capacity
6230 ) / SCHED_CAPACITY_SCALE;
6233 * if *imbalance is less than the average load per runnable task
6234 * there is no guarantee that any tasks will be moved so we'll have
6235 * a think about bumping its value to force at least one task to be
6238 if (env->imbalance < busiest->load_per_task)
6239 return fix_small_imbalance(env, sds);
6242 /******* find_busiest_group() helpers end here *********************/
6245 * find_busiest_group - Returns the busiest group within the sched_domain
6246 * if there is an imbalance. If there isn't an imbalance, and
6247 * the user has opted for power-savings, it returns a group whose
6248 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6249 * such a group exists.
6251 * Also calculates the amount of weighted load which should be moved
6252 * to restore balance.
6254 * @env: The load balancing environment.
6256 * Return: - The busiest group if imbalance exists.
6257 * - If no imbalance and user has opted for power-savings balance,
6258 * return the least loaded group whose CPUs can be
6259 * put to idle by rebalancing its tasks onto our group.
6261 static struct sched_group *find_busiest_group(struct lb_env *env)
6263 struct sg_lb_stats *local, *busiest;
6264 struct sd_lb_stats sds;
6266 init_sd_lb_stats(&sds);
6269 * Compute the various statistics relavent for load balancing at
6272 update_sd_lb_stats(env, &sds);
6273 local = &sds.local_stat;
6274 busiest = &sds.busiest_stat;
6276 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6277 check_asym_packing(env, &sds))
6280 /* There is no busy sibling group to pull tasks from */
6281 if (!sds.busiest || busiest->sum_nr_running == 0)
6284 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
6285 / sds.total_capacity;
6288 * If the busiest group is imbalanced the below checks don't
6289 * work because they assume all things are equal, which typically
6290 * isn't true due to cpus_allowed constraints and the like.
6292 if (busiest->group_imb)
6295 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6296 if (env->idle == CPU_NEWLY_IDLE && local->group_has_free_capacity &&
6297 !busiest->group_has_free_capacity)
6301 * If the local group is more busy than the selected busiest group
6302 * don't try and pull any tasks.
6304 if (local->avg_load >= busiest->avg_load)
6308 * Don't pull any tasks if this group is already above the domain
6311 if (local->avg_load >= sds.avg_load)
6314 if (env->idle == CPU_IDLE) {
6316 * This cpu is idle. If the busiest group load doesn't
6317 * have more tasks than the number of available cpu's and
6318 * there is no imbalance between this and busiest group
6319 * wrt to idle cpu's, it is balanced.
6321 if ((local->idle_cpus < busiest->idle_cpus) &&
6322 busiest->sum_nr_running <= busiest->group_weight)
6326 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6327 * imbalance_pct to be conservative.
6329 if (100 * busiest->avg_load <=
6330 env->sd->imbalance_pct * local->avg_load)
6335 /* Looks like there is an imbalance. Compute it */
6336 calculate_imbalance(env, &sds);
6345 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6347 static struct rq *find_busiest_queue(struct lb_env *env,
6348 struct sched_group *group)
6350 struct rq *busiest = NULL, *rq;
6351 unsigned long busiest_load = 0, busiest_capacity = 1;
6354 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6355 unsigned long capacity, capacity_factor, wl;
6359 rt = fbq_classify_rq(rq);
6362 * We classify groups/runqueues into three groups:
6363 * - regular: there are !numa tasks
6364 * - remote: there are numa tasks that run on the 'wrong' node
6365 * - all: there is no distinction
6367 * In order to avoid migrating ideally placed numa tasks,
6368 * ignore those when there's better options.
6370 * If we ignore the actual busiest queue to migrate another
6371 * task, the next balance pass can still reduce the busiest
6372 * queue by moving tasks around inside the node.
6374 * If we cannot move enough load due to this classification
6375 * the next pass will adjust the group classification and
6376 * allow migration of more tasks.
6378 * Both cases only affect the total convergence complexity.
6380 if (rt > env->fbq_type)
6383 capacity = capacity_of(i);
6384 capacity_factor = DIV_ROUND_CLOSEST(capacity, SCHED_CAPACITY_SCALE);
6385 if (!capacity_factor)
6386 capacity_factor = fix_small_capacity(env->sd, group);
6388 wl = weighted_cpuload(i);
6391 * When comparing with imbalance, use weighted_cpuload()
6392 * which is not scaled with the cpu capacity.
6394 if (capacity_factor && rq->nr_running == 1 && wl > env->imbalance)
6398 * For the load comparisons with the other cpu's, consider
6399 * the weighted_cpuload() scaled with the cpu capacity, so
6400 * that the load can be moved away from the cpu that is
6401 * potentially running at a lower capacity.
6403 * Thus we're looking for max(wl_i / capacity_i), crosswise
6404 * multiplication to rid ourselves of the division works out
6405 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
6406 * our previous maximum.
6408 if (wl * busiest_capacity > busiest_load * capacity) {
6410 busiest_capacity = capacity;
6419 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6420 * so long as it is large enough.
6422 #define MAX_PINNED_INTERVAL 512
6424 /* Working cpumask for load_balance and load_balance_newidle. */
6425 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6427 static int need_active_balance(struct lb_env *env)
6429 struct sched_domain *sd = env->sd;
6431 if (env->idle == CPU_NEWLY_IDLE) {
6434 * ASYM_PACKING needs to force migrate tasks from busy but
6435 * higher numbered CPUs in order to pack all tasks in the
6436 * lowest numbered CPUs.
6438 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6442 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6445 static int active_load_balance_cpu_stop(void *data);
6447 static int should_we_balance(struct lb_env *env)
6449 struct sched_group *sg = env->sd->groups;
6450 struct cpumask *sg_cpus, *sg_mask;
6451 int cpu, balance_cpu = -1;
6454 * In the newly idle case, we will allow all the cpu's
6455 * to do the newly idle load balance.
6457 if (env->idle == CPU_NEWLY_IDLE)
6460 sg_cpus = sched_group_cpus(sg);
6461 sg_mask = sched_group_mask(sg);
6462 /* Try to find first idle cpu */
6463 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6464 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6471 if (balance_cpu == -1)
6472 balance_cpu = group_balance_cpu(sg);
6475 * First idle cpu or the first cpu(busiest) in this sched group
6476 * is eligible for doing load balancing at this and above domains.
6478 return balance_cpu == env->dst_cpu;
6482 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6483 * tasks if there is an imbalance.
6485 static int load_balance(int this_cpu, struct rq *this_rq,
6486 struct sched_domain *sd, enum cpu_idle_type idle,
6487 int *continue_balancing)
6489 int ld_moved, cur_ld_moved, active_balance = 0;
6490 struct sched_domain *sd_parent = sd->parent;
6491 struct sched_group *group;
6493 unsigned long flags;
6494 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
6496 struct lb_env env = {
6498 .dst_cpu = this_cpu,
6500 .dst_grpmask = sched_group_cpus(sd->groups),
6502 .loop_break = sched_nr_migrate_break,
6508 * For NEWLY_IDLE load_balancing, we don't need to consider
6509 * other cpus in our group
6511 if (idle == CPU_NEWLY_IDLE)
6512 env.dst_grpmask = NULL;
6514 cpumask_copy(cpus, cpu_active_mask);
6516 schedstat_inc(sd, lb_count[idle]);
6519 if (!should_we_balance(&env)) {
6520 *continue_balancing = 0;
6524 group = find_busiest_group(&env);
6526 schedstat_inc(sd, lb_nobusyg[idle]);
6530 busiest = find_busiest_queue(&env, group);
6532 schedstat_inc(sd, lb_nobusyq[idle]);
6536 BUG_ON(busiest == env.dst_rq);
6538 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6541 if (busiest->nr_running > 1) {
6543 * Attempt to move tasks. If find_busiest_group has found
6544 * an imbalance but busiest->nr_running <= 1, the group is
6545 * still unbalanced. ld_moved simply stays zero, so it is
6546 * correctly treated as an imbalance.
6548 env.flags |= LBF_ALL_PINNED;
6549 env.src_cpu = busiest->cpu;
6550 env.src_rq = busiest;
6551 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
6554 local_irq_save(flags);
6555 double_rq_lock(env.dst_rq, busiest);
6558 * cur_ld_moved - load moved in current iteration
6559 * ld_moved - cumulative load moved across iterations
6561 cur_ld_moved = move_tasks(&env);
6562 ld_moved += cur_ld_moved;
6563 double_rq_unlock(env.dst_rq, busiest);
6564 local_irq_restore(flags);
6567 * some other cpu did the load balance for us.
6569 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
6570 resched_cpu(env.dst_cpu);
6572 if (env.flags & LBF_NEED_BREAK) {
6573 env.flags &= ~LBF_NEED_BREAK;
6578 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6579 * us and move them to an alternate dst_cpu in our sched_group
6580 * where they can run. The upper limit on how many times we
6581 * iterate on same src_cpu is dependent on number of cpus in our
6584 * This changes load balance semantics a bit on who can move
6585 * load to a given_cpu. In addition to the given_cpu itself
6586 * (or a ilb_cpu acting on its behalf where given_cpu is
6587 * nohz-idle), we now have balance_cpu in a position to move
6588 * load to given_cpu. In rare situations, this may cause
6589 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6590 * _independently_ and at _same_ time to move some load to
6591 * given_cpu) causing exceess load to be moved to given_cpu.
6592 * This however should not happen so much in practice and
6593 * moreover subsequent load balance cycles should correct the
6594 * excess load moved.
6596 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
6598 /* Prevent to re-select dst_cpu via env's cpus */
6599 cpumask_clear_cpu(env.dst_cpu, env.cpus);
6601 env.dst_rq = cpu_rq(env.new_dst_cpu);
6602 env.dst_cpu = env.new_dst_cpu;
6603 env.flags &= ~LBF_DST_PINNED;
6605 env.loop_break = sched_nr_migrate_break;
6608 * Go back to "more_balance" rather than "redo" since we
6609 * need to continue with same src_cpu.
6615 * We failed to reach balance because of affinity.
6618 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
6620 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
6621 *group_imbalance = 1;
6622 } else if (*group_imbalance)
6623 *group_imbalance = 0;
6626 /* All tasks on this runqueue were pinned by CPU affinity */
6627 if (unlikely(env.flags & LBF_ALL_PINNED)) {
6628 cpumask_clear_cpu(cpu_of(busiest), cpus);
6629 if (!cpumask_empty(cpus)) {
6631 env.loop_break = sched_nr_migrate_break;
6639 schedstat_inc(sd, lb_failed[idle]);
6641 * Increment the failure counter only on periodic balance.
6642 * We do not want newidle balance, which can be very
6643 * frequent, pollute the failure counter causing
6644 * excessive cache_hot migrations and active balances.
6646 if (idle != CPU_NEWLY_IDLE)
6647 sd->nr_balance_failed++;
6649 if (need_active_balance(&env)) {
6650 raw_spin_lock_irqsave(&busiest->lock, flags);
6652 /* don't kick the active_load_balance_cpu_stop,
6653 * if the curr task on busiest cpu can't be
6656 if (!cpumask_test_cpu(this_cpu,
6657 tsk_cpus_allowed(busiest->curr))) {
6658 raw_spin_unlock_irqrestore(&busiest->lock,
6660 env.flags |= LBF_ALL_PINNED;
6661 goto out_one_pinned;
6665 * ->active_balance synchronizes accesses to
6666 * ->active_balance_work. Once set, it's cleared
6667 * only after active load balance is finished.
6669 if (!busiest->active_balance) {
6670 busiest->active_balance = 1;
6671 busiest->push_cpu = this_cpu;
6674 raw_spin_unlock_irqrestore(&busiest->lock, flags);
6676 if (active_balance) {
6677 stop_one_cpu_nowait(cpu_of(busiest),
6678 active_load_balance_cpu_stop, busiest,
6679 &busiest->active_balance_work);
6683 * We've kicked active balancing, reset the failure
6686 sd->nr_balance_failed = sd->cache_nice_tries+1;
6689 sd->nr_balance_failed = 0;
6691 if (likely(!active_balance)) {
6692 /* We were unbalanced, so reset the balancing interval */
6693 sd->balance_interval = sd->min_interval;
6696 * If we've begun active balancing, start to back off. This
6697 * case may not be covered by the all_pinned logic if there
6698 * is only 1 task on the busy runqueue (because we don't call
6701 if (sd->balance_interval < sd->max_interval)
6702 sd->balance_interval *= 2;
6708 schedstat_inc(sd, lb_balanced[idle]);
6710 sd->nr_balance_failed = 0;
6713 /* tune up the balancing interval */
6714 if (((env.flags & LBF_ALL_PINNED) &&
6715 sd->balance_interval < MAX_PINNED_INTERVAL) ||
6716 (sd->balance_interval < sd->max_interval))
6717 sd->balance_interval *= 2;
6724 static inline unsigned long
6725 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
6727 unsigned long interval = sd->balance_interval;
6730 interval *= sd->busy_factor;
6732 /* scale ms to jiffies */
6733 interval = msecs_to_jiffies(interval);
6734 interval = clamp(interval, 1UL, max_load_balance_interval);
6740 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
6742 unsigned long interval, next;
6744 interval = get_sd_balance_interval(sd, cpu_busy);
6745 next = sd->last_balance + interval;
6747 if (time_after(*next_balance, next))
6748 *next_balance = next;
6752 * idle_balance is called by schedule() if this_cpu is about to become
6753 * idle. Attempts to pull tasks from other CPUs.
6755 static int idle_balance(struct rq *this_rq)
6757 unsigned long next_balance = jiffies + HZ;
6758 int this_cpu = this_rq->cpu;
6759 struct sched_domain *sd;
6760 int pulled_task = 0;
6763 idle_enter_fair(this_rq);
6766 * We must set idle_stamp _before_ calling idle_balance(), such that we
6767 * measure the duration of idle_balance() as idle time.
6769 this_rq->idle_stamp = rq_clock(this_rq);
6771 if (this_rq->avg_idle < sysctl_sched_migration_cost) {
6773 sd = rcu_dereference_check_sched_domain(this_rq->sd);
6775 update_next_balance(sd, 0, &next_balance);
6782 * Drop the rq->lock, but keep IRQ/preempt disabled.
6784 raw_spin_unlock(&this_rq->lock);
6786 update_blocked_averages(this_cpu);
6788 for_each_domain(this_cpu, sd) {
6789 int continue_balancing = 1;
6790 u64 t0, domain_cost;
6792 if (!(sd->flags & SD_LOAD_BALANCE))
6795 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
6796 update_next_balance(sd, 0, &next_balance);
6800 if (sd->flags & SD_BALANCE_NEWIDLE) {
6801 t0 = sched_clock_cpu(this_cpu);
6803 pulled_task = load_balance(this_cpu, this_rq,
6805 &continue_balancing);
6807 domain_cost = sched_clock_cpu(this_cpu) - t0;
6808 if (domain_cost > sd->max_newidle_lb_cost)
6809 sd->max_newidle_lb_cost = domain_cost;
6811 curr_cost += domain_cost;
6814 update_next_balance(sd, 0, &next_balance);
6817 * Stop searching for tasks to pull if there are
6818 * now runnable tasks on this rq.
6820 if (pulled_task || this_rq->nr_running > 0)
6825 raw_spin_lock(&this_rq->lock);
6827 if (curr_cost > this_rq->max_idle_balance_cost)
6828 this_rq->max_idle_balance_cost = curr_cost;
6831 * While browsing the domains, we released the rq lock, a task could
6832 * have been enqueued in the meantime. Since we're not going idle,
6833 * pretend we pulled a task.
6835 if (this_rq->cfs.h_nr_running && !pulled_task)
6839 /* Move the next balance forward */
6840 if (time_after(this_rq->next_balance, next_balance))
6841 this_rq->next_balance = next_balance;
6843 /* Is there a task of a high priority class? */
6844 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
6848 idle_exit_fair(this_rq);
6849 this_rq->idle_stamp = 0;
6856 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6857 * running tasks off the busiest CPU onto idle CPUs. It requires at
6858 * least 1 task to be running on each physical CPU where possible, and
6859 * avoids physical / logical imbalances.
6861 static int active_load_balance_cpu_stop(void *data)
6863 struct rq *busiest_rq = data;
6864 int busiest_cpu = cpu_of(busiest_rq);
6865 int target_cpu = busiest_rq->push_cpu;
6866 struct rq *target_rq = cpu_rq(target_cpu);
6867 struct sched_domain *sd;
6869 raw_spin_lock_irq(&busiest_rq->lock);
6871 /* make sure the requested cpu hasn't gone down in the meantime */
6872 if (unlikely(busiest_cpu != smp_processor_id() ||
6873 !busiest_rq->active_balance))
6876 /* Is there any task to move? */
6877 if (busiest_rq->nr_running <= 1)
6881 * This condition is "impossible", if it occurs
6882 * we need to fix it. Originally reported by
6883 * Bjorn Helgaas on a 128-cpu setup.
6885 BUG_ON(busiest_rq == target_rq);
6887 /* move a task from busiest_rq to target_rq */
6888 double_lock_balance(busiest_rq, target_rq);
6890 /* Search for an sd spanning us and the target CPU. */
6892 for_each_domain(target_cpu, sd) {
6893 if ((sd->flags & SD_LOAD_BALANCE) &&
6894 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6899 struct lb_env env = {
6901 .dst_cpu = target_cpu,
6902 .dst_rq = target_rq,
6903 .src_cpu = busiest_rq->cpu,
6904 .src_rq = busiest_rq,
6908 schedstat_inc(sd, alb_count);
6910 if (move_one_task(&env))
6911 schedstat_inc(sd, alb_pushed);
6913 schedstat_inc(sd, alb_failed);
6916 double_unlock_balance(busiest_rq, target_rq);
6918 busiest_rq->active_balance = 0;
6919 raw_spin_unlock_irq(&busiest_rq->lock);
6923 static inline int on_null_domain(struct rq *rq)
6925 return unlikely(!rcu_dereference_sched(rq->sd));
6928 #ifdef CONFIG_NO_HZ_COMMON
6930 * idle load balancing details
6931 * - When one of the busy CPUs notice that there may be an idle rebalancing
6932 * needed, they will kick the idle load balancer, which then does idle
6933 * load balancing for all the idle CPUs.
6936 cpumask_var_t idle_cpus_mask;
6938 unsigned long next_balance; /* in jiffy units */
6939 } nohz ____cacheline_aligned;
6941 static inline int find_new_ilb(void)
6943 int ilb = cpumask_first(nohz.idle_cpus_mask);
6945 if (ilb < nr_cpu_ids && idle_cpu(ilb))
6952 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6953 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6954 * CPU (if there is one).
6956 static void nohz_balancer_kick(void)
6960 nohz.next_balance++;
6962 ilb_cpu = find_new_ilb();
6964 if (ilb_cpu >= nr_cpu_ids)
6967 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
6970 * Use smp_send_reschedule() instead of resched_cpu().
6971 * This way we generate a sched IPI on the target cpu which
6972 * is idle. And the softirq performing nohz idle load balance
6973 * will be run before returning from the IPI.
6975 smp_send_reschedule(ilb_cpu);
6979 static inline void nohz_balance_exit_idle(int cpu)
6981 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
6983 * Completely isolated CPUs don't ever set, so we must test.
6985 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
6986 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
6987 atomic_dec(&nohz.nr_cpus);
6989 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6993 static inline void set_cpu_sd_state_busy(void)
6995 struct sched_domain *sd;
6996 int cpu = smp_processor_id();
6999 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7001 if (!sd || !sd->nohz_idle)
7005 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7010 void set_cpu_sd_state_idle(void)
7012 struct sched_domain *sd;
7013 int cpu = smp_processor_id();
7016 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7018 if (!sd || sd->nohz_idle)
7022 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7028 * This routine will record that the cpu is going idle with tick stopped.
7029 * This info will be used in performing idle load balancing in the future.
7031 void nohz_balance_enter_idle(int cpu)
7034 * If this cpu is going down, then nothing needs to be done.
7036 if (!cpu_active(cpu))
7039 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7043 * If we're a completely isolated CPU, we don't play.
7045 if (on_null_domain(cpu_rq(cpu)))
7048 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7049 atomic_inc(&nohz.nr_cpus);
7050 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7053 static int sched_ilb_notifier(struct notifier_block *nfb,
7054 unsigned long action, void *hcpu)
7056 switch (action & ~CPU_TASKS_FROZEN) {
7058 nohz_balance_exit_idle(smp_processor_id());
7066 static DEFINE_SPINLOCK(balancing);
7069 * Scale the max load_balance interval with the number of CPUs in the system.
7070 * This trades load-balance latency on larger machines for less cross talk.
7072 void update_max_interval(void)
7074 max_load_balance_interval = HZ*num_online_cpus()/10;
7078 * It checks each scheduling domain to see if it is due to be balanced,
7079 * and initiates a balancing operation if so.
7081 * Balancing parameters are set up in init_sched_domains.
7083 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7085 int continue_balancing = 1;
7087 unsigned long interval;
7088 struct sched_domain *sd;
7089 /* Earliest time when we have to do rebalance again */
7090 unsigned long next_balance = jiffies + 60*HZ;
7091 int update_next_balance = 0;
7092 int need_serialize, need_decay = 0;
7095 update_blocked_averages(cpu);
7098 for_each_domain(cpu, sd) {
7100 * Decay the newidle max times here because this is a regular
7101 * visit to all the domains. Decay ~1% per second.
7103 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7104 sd->max_newidle_lb_cost =
7105 (sd->max_newidle_lb_cost * 253) / 256;
7106 sd->next_decay_max_lb_cost = jiffies + HZ;
7109 max_cost += sd->max_newidle_lb_cost;
7111 if (!(sd->flags & SD_LOAD_BALANCE))
7115 * Stop the load balance at this level. There is another
7116 * CPU in our sched group which is doing load balancing more
7119 if (!continue_balancing) {
7125 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7127 need_serialize = sd->flags & SD_SERIALIZE;
7128 if (need_serialize) {
7129 if (!spin_trylock(&balancing))
7133 if (time_after_eq(jiffies, sd->last_balance + interval)) {
7134 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7136 * The LBF_DST_PINNED logic could have changed
7137 * env->dst_cpu, so we can't know our idle
7138 * state even if we migrated tasks. Update it.
7140 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7142 sd->last_balance = jiffies;
7143 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7146 spin_unlock(&balancing);
7148 if (time_after(next_balance, sd->last_balance + interval)) {
7149 next_balance = sd->last_balance + interval;
7150 update_next_balance = 1;
7155 * Ensure the rq-wide value also decays but keep it at a
7156 * reasonable floor to avoid funnies with rq->avg_idle.
7158 rq->max_idle_balance_cost =
7159 max((u64)sysctl_sched_migration_cost, max_cost);
7164 * next_balance will be updated only when there is a need.
7165 * When the cpu is attached to null domain for ex, it will not be
7168 if (likely(update_next_balance))
7169 rq->next_balance = next_balance;
7172 #ifdef CONFIG_NO_HZ_COMMON
7174 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7175 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7177 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7179 int this_cpu = this_rq->cpu;
7183 if (idle != CPU_IDLE ||
7184 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7187 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7188 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7192 * If this cpu gets work to do, stop the load balancing
7193 * work being done for other cpus. Next load
7194 * balancing owner will pick it up.
7199 rq = cpu_rq(balance_cpu);
7202 * If time for next balance is due,
7205 if (time_after_eq(jiffies, rq->next_balance)) {
7206 raw_spin_lock_irq(&rq->lock);
7207 update_rq_clock(rq);
7208 update_idle_cpu_load(rq);
7209 raw_spin_unlock_irq(&rq->lock);
7210 rebalance_domains(rq, CPU_IDLE);
7213 if (time_after(this_rq->next_balance, rq->next_balance))
7214 this_rq->next_balance = rq->next_balance;
7216 nohz.next_balance = this_rq->next_balance;
7218 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7222 * Current heuristic for kicking the idle load balancer in the presence
7223 * of an idle cpu is the system.
7224 * - This rq has more than one task.
7225 * - At any scheduler domain level, this cpu's scheduler group has multiple
7226 * busy cpu's exceeding the group's capacity.
7227 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7228 * domain span are idle.
7230 static inline int nohz_kick_needed(struct rq *rq)
7232 unsigned long now = jiffies;
7233 struct sched_domain *sd;
7234 struct sched_group_capacity *sgc;
7235 int nr_busy, cpu = rq->cpu;
7237 if (unlikely(rq->idle_balance))
7241 * We may be recently in ticked or tickless idle mode. At the first
7242 * busy tick after returning from idle, we will update the busy stats.
7244 set_cpu_sd_state_busy();
7245 nohz_balance_exit_idle(cpu);
7248 * None are in tickless mode and hence no need for NOHZ idle load
7251 if (likely(!atomic_read(&nohz.nr_cpus)))
7254 if (time_before(now, nohz.next_balance))
7257 if (rq->nr_running >= 2)
7261 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7264 sgc = sd->groups->sgc;
7265 nr_busy = atomic_read(&sgc->nr_busy_cpus);
7268 goto need_kick_unlock;
7271 sd = rcu_dereference(per_cpu(sd_asym, cpu));
7273 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7274 sched_domain_span(sd)) < cpu))
7275 goto need_kick_unlock;
7286 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7290 * run_rebalance_domains is triggered when needed from the scheduler tick.
7291 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7293 static void run_rebalance_domains(struct softirq_action *h)
7295 struct rq *this_rq = this_rq();
7296 enum cpu_idle_type idle = this_rq->idle_balance ?
7297 CPU_IDLE : CPU_NOT_IDLE;
7299 rebalance_domains(this_rq, idle);
7302 * If this cpu has a pending nohz_balance_kick, then do the
7303 * balancing on behalf of the other idle cpus whose ticks are
7306 nohz_idle_balance(this_rq, idle);
7310 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7312 void trigger_load_balance(struct rq *rq)
7314 /* Don't need to rebalance while attached to NULL domain */
7315 if (unlikely(on_null_domain(rq)))
7318 if (time_after_eq(jiffies, rq->next_balance))
7319 raise_softirq(SCHED_SOFTIRQ);
7320 #ifdef CONFIG_NO_HZ_COMMON
7321 if (nohz_kick_needed(rq))
7322 nohz_balancer_kick();
7326 static void rq_online_fair(struct rq *rq)
7331 static void rq_offline_fair(struct rq *rq)
7335 /* Ensure any throttled groups are reachable by pick_next_task */
7336 unthrottle_offline_cfs_rqs(rq);
7339 #endif /* CONFIG_SMP */
7342 * scheduler tick hitting a task of our scheduling class:
7344 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7346 struct cfs_rq *cfs_rq;
7347 struct sched_entity *se = &curr->se;
7349 for_each_sched_entity(se) {
7350 cfs_rq = cfs_rq_of(se);
7351 entity_tick(cfs_rq, se, queued);
7354 if (numabalancing_enabled)
7355 task_tick_numa(rq, curr);
7357 update_rq_runnable_avg(rq, 1);
7361 * called on fork with the child task as argument from the parent's context
7362 * - child not yet on the tasklist
7363 * - preemption disabled
7365 static void task_fork_fair(struct task_struct *p)
7367 struct cfs_rq *cfs_rq;
7368 struct sched_entity *se = &p->se, *curr;
7369 int this_cpu = smp_processor_id();
7370 struct rq *rq = this_rq();
7371 unsigned long flags;
7373 raw_spin_lock_irqsave(&rq->lock, flags);
7375 update_rq_clock(rq);
7377 cfs_rq = task_cfs_rq(current);
7378 curr = cfs_rq->curr;
7381 * Not only the cpu but also the task_group of the parent might have
7382 * been changed after parent->se.parent,cfs_rq were copied to
7383 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7384 * of child point to valid ones.
7387 __set_task_cpu(p, this_cpu);
7390 update_curr(cfs_rq);
7393 se->vruntime = curr->vruntime;
7394 place_entity(cfs_rq, se, 1);
7396 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
7398 * Upon rescheduling, sched_class::put_prev_task() will place
7399 * 'current' within the tree based on its new key value.
7401 swap(curr->vruntime, se->vruntime);
7402 resched_task(rq->curr);
7405 se->vruntime -= cfs_rq->min_vruntime;
7407 raw_spin_unlock_irqrestore(&rq->lock, flags);
7411 * Priority of the task has changed. Check to see if we preempt
7415 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
7421 * Reschedule if we are currently running on this runqueue and
7422 * our priority decreased, or if we are not currently running on
7423 * this runqueue and our priority is higher than the current's
7425 if (rq->curr == p) {
7426 if (p->prio > oldprio)
7427 resched_task(rq->curr);
7429 check_preempt_curr(rq, p, 0);
7432 static void switched_from_fair(struct rq *rq, struct task_struct *p)
7434 struct sched_entity *se = &p->se;
7435 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7438 * Ensure the task's vruntime is normalized, so that when it's
7439 * switched back to the fair class the enqueue_entity(.flags=0) will
7440 * do the right thing.
7442 * If it's on_rq, then the dequeue_entity(.flags=0) will already
7443 * have normalized the vruntime, if it's !on_rq, then only when
7444 * the task is sleeping will it still have non-normalized vruntime.
7446 if (!p->on_rq && p->state != TASK_RUNNING) {
7448 * Fix up our vruntime so that the current sleep doesn't
7449 * cause 'unlimited' sleep bonus.
7451 place_entity(cfs_rq, se, 0);
7452 se->vruntime -= cfs_rq->min_vruntime;
7457 * Remove our load from contribution when we leave sched_fair
7458 * and ensure we don't carry in an old decay_count if we
7461 if (se->avg.decay_count) {
7462 __synchronize_entity_decay(se);
7463 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
7469 * We switched to the sched_fair class.
7471 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7473 struct sched_entity *se = &p->se;
7474 #ifdef CONFIG_FAIR_GROUP_SCHED
7476 * Since the real-depth could have been changed (only FAIR
7477 * class maintain depth value), reset depth properly.
7479 se->depth = se->parent ? se->parent->depth + 1 : 0;
7485 * We were most likely switched from sched_rt, so
7486 * kick off the schedule if running, otherwise just see
7487 * if we can still preempt the current task.
7490 resched_task(rq->curr);
7492 check_preempt_curr(rq, p, 0);
7495 /* Account for a task changing its policy or group.
7497 * This routine is mostly called to set cfs_rq->curr field when a task
7498 * migrates between groups/classes.
7500 static void set_curr_task_fair(struct rq *rq)
7502 struct sched_entity *se = &rq->curr->se;
7504 for_each_sched_entity(se) {
7505 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7507 set_next_entity(cfs_rq, se);
7508 /* ensure bandwidth has been allocated on our new cfs_rq */
7509 account_cfs_rq_runtime(cfs_rq, 0);
7513 void init_cfs_rq(struct cfs_rq *cfs_rq)
7515 cfs_rq->tasks_timeline = RB_ROOT;
7516 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7517 #ifndef CONFIG_64BIT
7518 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7521 atomic64_set(&cfs_rq->decay_counter, 1);
7522 atomic_long_set(&cfs_rq->removed_load, 0);
7526 #ifdef CONFIG_FAIR_GROUP_SCHED
7527 static void task_move_group_fair(struct task_struct *p, int on_rq)
7529 struct sched_entity *se = &p->se;
7530 struct cfs_rq *cfs_rq;
7533 * If the task was not on the rq at the time of this cgroup movement
7534 * it must have been asleep, sleeping tasks keep their ->vruntime
7535 * absolute on their old rq until wakeup (needed for the fair sleeper
7536 * bonus in place_entity()).
7538 * If it was on the rq, we've just 'preempted' it, which does convert
7539 * ->vruntime to a relative base.
7541 * Make sure both cases convert their relative position when migrating
7542 * to another cgroup's rq. This does somewhat interfere with the
7543 * fair sleeper stuff for the first placement, but who cares.
7546 * When !on_rq, vruntime of the task has usually NOT been normalized.
7547 * But there are some cases where it has already been normalized:
7549 * - Moving a forked child which is waiting for being woken up by
7550 * wake_up_new_task().
7551 * - Moving a task which has been woken up by try_to_wake_up() and
7552 * waiting for actually being woken up by sched_ttwu_pending().
7554 * To prevent boost or penalty in the new cfs_rq caused by delta
7555 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7557 if (!on_rq && (!se->sum_exec_runtime || p->state == TASK_WAKING))
7561 se->vruntime -= cfs_rq_of(se)->min_vruntime;
7562 set_task_rq(p, task_cpu(p));
7563 se->depth = se->parent ? se->parent->depth + 1 : 0;
7565 cfs_rq = cfs_rq_of(se);
7566 se->vruntime += cfs_rq->min_vruntime;
7569 * migrate_task_rq_fair() will have removed our previous
7570 * contribution, but we must synchronize for ongoing future
7573 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7574 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
7579 void free_fair_sched_group(struct task_group *tg)
7583 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7585 for_each_possible_cpu(i) {
7587 kfree(tg->cfs_rq[i]);
7596 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7598 struct cfs_rq *cfs_rq;
7599 struct sched_entity *se;
7602 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7605 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7609 tg->shares = NICE_0_LOAD;
7611 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7613 for_each_possible_cpu(i) {
7614 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7615 GFP_KERNEL, cpu_to_node(i));
7619 se = kzalloc_node(sizeof(struct sched_entity),
7620 GFP_KERNEL, cpu_to_node(i));
7624 init_cfs_rq(cfs_rq);
7625 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7636 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7638 struct rq *rq = cpu_rq(cpu);
7639 unsigned long flags;
7642 * Only empty task groups can be destroyed; so we can speculatively
7643 * check on_list without danger of it being re-added.
7645 if (!tg->cfs_rq[cpu]->on_list)
7648 raw_spin_lock_irqsave(&rq->lock, flags);
7649 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7650 raw_spin_unlock_irqrestore(&rq->lock, flags);
7653 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7654 struct sched_entity *se, int cpu,
7655 struct sched_entity *parent)
7657 struct rq *rq = cpu_rq(cpu);
7661 init_cfs_rq_runtime(cfs_rq);
7663 tg->cfs_rq[cpu] = cfs_rq;
7666 /* se could be NULL for root_task_group */
7671 se->cfs_rq = &rq->cfs;
7674 se->cfs_rq = parent->my_q;
7675 se->depth = parent->depth + 1;
7679 /* guarantee group entities always have weight */
7680 update_load_set(&se->load, NICE_0_LOAD);
7681 se->parent = parent;
7684 static DEFINE_MUTEX(shares_mutex);
7686 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7689 unsigned long flags;
7692 * We can't change the weight of the root cgroup.
7697 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7699 mutex_lock(&shares_mutex);
7700 if (tg->shares == shares)
7703 tg->shares = shares;
7704 for_each_possible_cpu(i) {
7705 struct rq *rq = cpu_rq(i);
7706 struct sched_entity *se;
7709 /* Propagate contribution to hierarchy */
7710 raw_spin_lock_irqsave(&rq->lock, flags);
7712 /* Possible calls to update_curr() need rq clock */
7713 update_rq_clock(rq);
7714 for_each_sched_entity(se)
7715 update_cfs_shares(group_cfs_rq(se));
7716 raw_spin_unlock_irqrestore(&rq->lock, flags);
7720 mutex_unlock(&shares_mutex);
7723 #else /* CONFIG_FAIR_GROUP_SCHED */
7725 void free_fair_sched_group(struct task_group *tg) { }
7727 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7732 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7734 #endif /* CONFIG_FAIR_GROUP_SCHED */
7737 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
7739 struct sched_entity *se = &task->se;
7740 unsigned int rr_interval = 0;
7743 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7746 if (rq->cfs.load.weight)
7747 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
7753 * All the scheduling class methods:
7755 const struct sched_class fair_sched_class = {
7756 .next = &idle_sched_class,
7757 .enqueue_task = enqueue_task_fair,
7758 .dequeue_task = dequeue_task_fair,
7759 .yield_task = yield_task_fair,
7760 .yield_to_task = yield_to_task_fair,
7762 .check_preempt_curr = check_preempt_wakeup,
7764 .pick_next_task = pick_next_task_fair,
7765 .put_prev_task = put_prev_task_fair,
7768 .select_task_rq = select_task_rq_fair,
7769 .migrate_task_rq = migrate_task_rq_fair,
7771 .rq_online = rq_online_fair,
7772 .rq_offline = rq_offline_fair,
7774 .task_waking = task_waking_fair,
7777 .set_curr_task = set_curr_task_fair,
7778 .task_tick = task_tick_fair,
7779 .task_fork = task_fork_fair,
7781 .prio_changed = prio_changed_fair,
7782 .switched_from = switched_from_fair,
7783 .switched_to = switched_to_fair,
7785 .get_rr_interval = get_rr_interval_fair,
7787 #ifdef CONFIG_FAIR_GROUP_SCHED
7788 .task_move_group = task_move_group_fair,
7792 #ifdef CONFIG_SCHED_DEBUG
7793 void print_cfs_stats(struct seq_file *m, int cpu)
7795 struct cfs_rq *cfs_rq;
7798 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7799 print_cfs_rq(m, cpu, cfs_rq);
7804 __init void init_sched_fair_class(void)
7807 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7809 #ifdef CONFIG_NO_HZ_COMMON
7810 nohz.next_balance = jiffies;
7811 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7812 cpu_notifier(sched_ilb_notifier, 0);