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/random.h>
30 #include <linux/mempolicy.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;
117 * Increase the granularity value when there are more CPUs,
118 * because with more CPUs the 'effective latency' as visible
119 * to users decreases. But the relationship is not linear,
120 * so pick a second-best guess by going with the log2 of the
123 * This idea comes from the SD scheduler of Con Kolivas:
125 static int get_update_sysctl_factor(void)
127 unsigned int cpus = min_t(int, num_online_cpus(), 8);
130 switch (sysctl_sched_tunable_scaling) {
131 case SCHED_TUNABLESCALING_NONE:
134 case SCHED_TUNABLESCALING_LINEAR:
137 case SCHED_TUNABLESCALING_LOG:
139 factor = 1 + ilog2(cpus);
146 static void update_sysctl(void)
148 unsigned int factor = get_update_sysctl_factor();
150 #define SET_SYSCTL(name) \
151 (sysctl_##name = (factor) * normalized_sysctl_##name)
152 SET_SYSCTL(sched_min_granularity);
153 SET_SYSCTL(sched_latency);
154 SET_SYSCTL(sched_wakeup_granularity);
158 void sched_init_granularity(void)
163 #if BITS_PER_LONG == 32
164 # define WMULT_CONST (~0UL)
166 # define WMULT_CONST (1UL << 32)
169 #define WMULT_SHIFT 32
172 * Shift right and round:
174 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
177 * delta *= weight / lw
180 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
181 struct load_weight *lw)
186 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
187 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
188 * 2^SCHED_LOAD_RESOLUTION.
190 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
191 tmp = (u64)delta_exec * scale_load_down(weight);
193 tmp = (u64)delta_exec;
195 if (!lw->inv_weight) {
196 unsigned long w = scale_load_down(lw->weight);
198 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
200 else if (unlikely(!w))
201 lw->inv_weight = WMULT_CONST;
203 lw->inv_weight = WMULT_CONST / w;
207 * Check whether we'd overflow the 64-bit multiplication:
209 if (unlikely(tmp > WMULT_CONST))
210 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
213 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
215 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
219 const struct sched_class fair_sched_class;
221 /**************************************************************
222 * CFS operations on generic schedulable entities:
225 #ifdef CONFIG_FAIR_GROUP_SCHED
227 /* cpu runqueue to which this cfs_rq is attached */
228 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
233 /* An entity is a task if it doesn't "own" a runqueue */
234 #define entity_is_task(se) (!se->my_q)
236 static inline struct task_struct *task_of(struct sched_entity *se)
238 #ifdef CONFIG_SCHED_DEBUG
239 WARN_ON_ONCE(!entity_is_task(se));
241 return container_of(se, struct task_struct, se);
244 /* Walk up scheduling entities hierarchy */
245 #define for_each_sched_entity(se) \
246 for (; se; se = se->parent)
248 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
253 /* runqueue on which this entity is (to be) queued */
254 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
259 /* runqueue "owned" by this group */
260 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
265 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
267 if (!cfs_rq->on_list) {
269 * Ensure we either appear before our parent (if already
270 * enqueued) or force our parent to appear after us when it is
271 * enqueued. The fact that we always enqueue bottom-up
272 * reduces this to two cases.
274 if (cfs_rq->tg->parent &&
275 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
276 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
277 &rq_of(cfs_rq)->leaf_cfs_rq_list);
279 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
280 &rq_of(cfs_rq)->leaf_cfs_rq_list);
287 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
289 if (cfs_rq->on_list) {
290 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
295 /* Iterate thr' all leaf cfs_rq's on a runqueue */
296 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
297 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
299 /* Do the two (enqueued) entities belong to the same group ? */
301 is_same_group(struct sched_entity *se, struct sched_entity *pse)
303 if (se->cfs_rq == pse->cfs_rq)
309 static inline struct sched_entity *parent_entity(struct sched_entity *se)
314 /* return depth at which a sched entity is present in the hierarchy */
315 static inline int depth_se(struct sched_entity *se)
319 for_each_sched_entity(se)
326 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
328 int se_depth, pse_depth;
331 * preemption test can be made between sibling entities who are in the
332 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
333 * both tasks until we find their ancestors who are siblings of common
337 /* First walk up until both entities are at same depth */
338 se_depth = depth_se(*se);
339 pse_depth = depth_se(*pse);
341 while (se_depth > pse_depth) {
343 *se = parent_entity(*se);
346 while (pse_depth > se_depth) {
348 *pse = parent_entity(*pse);
351 while (!is_same_group(*se, *pse)) {
352 *se = parent_entity(*se);
353 *pse = parent_entity(*pse);
357 #else /* !CONFIG_FAIR_GROUP_SCHED */
359 static inline struct task_struct *task_of(struct sched_entity *se)
361 return container_of(se, struct task_struct, se);
364 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
366 return container_of(cfs_rq, struct rq, cfs);
369 #define entity_is_task(se) 1
371 #define for_each_sched_entity(se) \
372 for (; se; se = NULL)
374 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
376 return &task_rq(p)->cfs;
379 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
381 struct task_struct *p = task_of(se);
382 struct rq *rq = task_rq(p);
387 /* runqueue "owned" by this group */
388 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
393 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
397 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
401 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
402 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
405 is_same_group(struct sched_entity *se, struct sched_entity *pse)
410 static inline struct sched_entity *parent_entity(struct sched_entity *se)
416 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
420 #endif /* CONFIG_FAIR_GROUP_SCHED */
422 static __always_inline
423 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec);
425 /**************************************************************
426 * Scheduling class tree data structure manipulation methods:
429 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
431 s64 delta = (s64)(vruntime - min_vruntime);
433 min_vruntime = vruntime;
438 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
440 s64 delta = (s64)(vruntime - min_vruntime);
442 min_vruntime = vruntime;
447 static inline int entity_before(struct sched_entity *a,
448 struct sched_entity *b)
450 return (s64)(a->vruntime - b->vruntime) < 0;
453 static void update_min_vruntime(struct cfs_rq *cfs_rq)
455 u64 vruntime = cfs_rq->min_vruntime;
458 vruntime = cfs_rq->curr->vruntime;
460 if (cfs_rq->rb_leftmost) {
461 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
466 vruntime = se->vruntime;
468 vruntime = min_vruntime(vruntime, se->vruntime);
471 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
474 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
479 * Enqueue an entity into the rb-tree:
481 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
483 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
484 struct rb_node *parent = NULL;
485 struct sched_entity *entry;
489 * Find the right place in the rbtree:
493 entry = rb_entry(parent, struct sched_entity, run_node);
495 * We dont care about collisions. Nodes with
496 * the same key stay together.
498 if (entity_before(se, entry)) {
499 link = &parent->rb_left;
501 link = &parent->rb_right;
507 * Maintain a cache of leftmost tree entries (it is frequently
511 cfs_rq->rb_leftmost = &se->run_node;
513 rb_link_node(&se->run_node, parent, link);
514 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
517 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
519 if (cfs_rq->rb_leftmost == &se->run_node) {
520 struct rb_node *next_node;
522 next_node = rb_next(&se->run_node);
523 cfs_rq->rb_leftmost = next_node;
526 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
529 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
531 struct rb_node *left = cfs_rq->rb_leftmost;
536 return rb_entry(left, struct sched_entity, run_node);
539 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
541 struct rb_node *next = rb_next(&se->run_node);
546 return rb_entry(next, struct sched_entity, run_node);
549 #ifdef CONFIG_SCHED_DEBUG
550 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
552 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
557 return rb_entry(last, struct sched_entity, run_node);
560 /**************************************************************
561 * Scheduling class statistics methods:
564 int sched_proc_update_handler(struct ctl_table *table, int write,
565 void __user *buffer, size_t *lenp,
568 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
569 int factor = get_update_sysctl_factor();
574 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
575 sysctl_sched_min_granularity);
577 #define WRT_SYSCTL(name) \
578 (normalized_sysctl_##name = sysctl_##name / (factor))
579 WRT_SYSCTL(sched_min_granularity);
580 WRT_SYSCTL(sched_latency);
581 WRT_SYSCTL(sched_wakeup_granularity);
591 static inline unsigned long
592 calc_delta_fair(unsigned long delta, struct sched_entity *se)
594 if (unlikely(se->load.weight != NICE_0_LOAD))
595 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
601 * The idea is to set a period in which each task runs once.
603 * When there are too many tasks (sched_nr_latency) we have to stretch
604 * this period because otherwise the slices get too small.
606 * p = (nr <= nl) ? l : l*nr/nl
608 static u64 __sched_period(unsigned long nr_running)
610 u64 period = sysctl_sched_latency;
611 unsigned long nr_latency = sched_nr_latency;
613 if (unlikely(nr_running > nr_latency)) {
614 period = sysctl_sched_min_granularity;
615 period *= nr_running;
622 * We calculate the wall-time slice from the period by taking a part
623 * proportional to the weight.
627 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
629 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
631 for_each_sched_entity(se) {
632 struct load_weight *load;
633 struct load_weight lw;
635 cfs_rq = cfs_rq_of(se);
636 load = &cfs_rq->load;
638 if (unlikely(!se->on_rq)) {
641 update_load_add(&lw, se->load.weight);
644 slice = calc_delta_mine(slice, se->load.weight, load);
650 * We calculate the vruntime slice of a to be inserted task
654 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
656 return calc_delta_fair(sched_slice(cfs_rq, se), se);
659 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update);
660 static void update_cfs_shares(struct cfs_rq *cfs_rq);
663 * Update the current task's runtime statistics. Skip current tasks that
664 * are not in our scheduling class.
667 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
668 unsigned long delta_exec)
670 unsigned long delta_exec_weighted;
672 schedstat_set(curr->statistics.exec_max,
673 max((u64)delta_exec, curr->statistics.exec_max));
675 curr->sum_exec_runtime += delta_exec;
676 schedstat_add(cfs_rq, exec_clock, delta_exec);
677 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
679 curr->vruntime += delta_exec_weighted;
680 update_min_vruntime(cfs_rq);
682 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
683 cfs_rq->load_unacc_exec_time += delta_exec;
687 static void update_curr(struct cfs_rq *cfs_rq)
689 struct sched_entity *curr = cfs_rq->curr;
690 u64 now = rq_of(cfs_rq)->clock_task;
691 unsigned long delta_exec;
697 * Get the amount of time the current task was running
698 * since the last time we changed load (this cannot
699 * overflow on 32 bits):
701 delta_exec = (unsigned long)(now - curr->exec_start);
705 __update_curr(cfs_rq, curr, delta_exec);
706 curr->exec_start = now;
708 if (entity_is_task(curr)) {
709 struct task_struct *curtask = task_of(curr);
711 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
712 cpuacct_charge(curtask, delta_exec);
713 account_group_exec_runtime(curtask, delta_exec);
716 account_cfs_rq_runtime(cfs_rq, delta_exec);
720 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
722 schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock);
726 * Task is being enqueued - update stats:
728 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
731 * Are we enqueueing a waiting task? (for current tasks
732 * a dequeue/enqueue event is a NOP)
734 if (se != cfs_rq->curr)
735 update_stats_wait_start(cfs_rq, se);
739 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
741 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
742 rq_of(cfs_rq)->clock - se->statistics.wait_start));
743 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
744 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
745 rq_of(cfs_rq)->clock - se->statistics.wait_start);
746 #ifdef CONFIG_SCHEDSTATS
747 if (entity_is_task(se)) {
748 trace_sched_stat_wait(task_of(se),
749 rq_of(cfs_rq)->clock - se->statistics.wait_start);
752 schedstat_set(se->statistics.wait_start, 0);
756 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
759 * Mark the end of the wait period if dequeueing a
762 if (se != cfs_rq->curr)
763 update_stats_wait_end(cfs_rq, se);
767 * We are picking a new current task - update its stats:
770 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
773 * We are starting a new run period:
775 se->exec_start = rq_of(cfs_rq)->clock_task;
778 /**************************************************
779 * Scheduling class numa methods.
781 * The purpose of the NUMA bits are to maintain compute (task) and data
782 * (memory) locality. We try and achieve this by making tasks stick to
783 * a particular node (their home node) but if fairness mandates they run
784 * elsewhere for long enough, we let the memory follow them.
786 * Tasks start out with their home-node unset (-1) this effectively means
787 * they act !NUMA until we've established the task is busy enough to bother
790 * We keep a home-node per task and use periodic fault scans to try and
791 * estalish a task<->page relation. This assumes the task<->page relation is a
792 * compute<->data relation, this is false for things like virt. and n:m
793 * threading solutions but its the best we can do given the information we
797 static unsigned long task_h_load(struct task_struct *p);
799 #ifdef CONFIG_SCHED_NUMA
800 static struct list_head *account_numa_enqueue(struct rq *rq, struct task_struct *p)
802 struct list_head *tasks = &rq->cfs_tasks;
804 if (tsk_home_node(p) != cpu_to_node(task_cpu(p))) {
805 p->numa_contrib = task_h_load(p);
806 rq->offnode_weight += p->numa_contrib;
807 rq->offnode_running++;
808 tasks = &rq->offnode_tasks;
810 rq->onnode_running++;
815 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
817 if (tsk_home_node(p) != cpu_to_node(task_cpu(p))) {
818 rq->offnode_weight -= p->numa_contrib;
819 rq->offnode_running--;
821 rq->onnode_running--;
825 * numa task sample period in ms: 5s
827 unsigned int sysctl_sched_numa_task_period_min = 5000;
828 unsigned int sysctl_sched_numa_task_period_max = 5000*16;
831 * Wait for the 2-sample stuff to settle before migrating again
833 unsigned int sysctl_sched_numa_settle_count = 2;
836 * Got a PROT_NONE fault for a page on @node.
838 void task_numa_fault(int node)
840 struct task_struct *p = current;
842 if (unlikely(!p->numa_faults)) {
843 p->numa_faults = kzalloc(sizeof(unsigned long) * nr_node_ids,
849 p->numa_faults[node]++;
852 void task_numa_placement(void)
854 unsigned long faults, max_faults = 0;
855 struct task_struct *p = current;
856 int node, max_node = -1;
857 int seq = ACCESS_ONCE(p->mm->numa_scan_seq);
859 if (p->numa_scan_seq == seq)
862 p->numa_scan_seq = seq;
864 if (unlikely(!p->numa_faults))
867 for (node = 0; node < nr_node_ids; node++) {
868 faults = p->numa_faults[node];
870 if (faults > max_faults) {
875 p->numa_faults[node] /= 2;
881 if (p->node != max_node) {
882 p->numa_task_period = sysctl_sched_numa_task_period_min;
883 if (sched_feat(NUMA_SETTLE) &&
884 (seq - p->numa_migrate_seq) <= (int)sysctl_sched_numa_settle_count)
886 p->numa_migrate_seq = seq;
887 sched_setnode(p, max_node);
889 p->numa_task_period = min(sysctl_sched_numa_task_period_max,
890 p->numa_task_period * 2);
895 * The expensive part of numa migration is done from task_work context.
896 * Triggered from task_tick_numa().
898 void task_numa_work(struct callback_head *work)
900 unsigned long migrate, next_scan, now = jiffies;
901 struct task_struct *p = current;
902 struct mm_struct *mm = p->mm;
904 WARN_ON_ONCE(p != container_of(work, struct task_struct, rcu));
907 * Who cares about NUMA placement when they're dying.
909 * NOTE: make sure not to dereference p->mm before this check,
910 * exit_task_work() happens _after_ exit_mm() so we could be called
911 * without p->mm even though we still had it when we enqueued this
914 if (p->flags & PF_EXITING)
918 * Enforce maximal scan/migration frequency..
920 migrate = mm->numa_next_scan;
921 if (time_before(now, migrate))
924 next_scan = now + 2*msecs_to_jiffies(sysctl_sched_numa_task_period_min);
925 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
928 ACCESS_ONCE(mm->numa_scan_seq)++;
929 lazy_migrate_process(mm);
933 * Drive the periodic memory faults..
935 void task_tick_numa(struct rq *rq, struct task_struct *curr)
940 * We don't care about NUMA placement if we don't have memory.
946 * Using runtime rather than walltime has the dual advantage that
947 * we (mostly) drive the selection from busy threads and that the
948 * task needs to have done some actual work before we bother with
951 now = curr->se.sum_exec_runtime;
952 period = (u64)curr->numa_task_period * NSEC_PER_MSEC;
954 if (now - curr->node_stamp > period) {
955 curr->node_stamp = now;
957 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
959 * We can re-use curr->rcu because we checked curr->mm
960 * != NULL so release_task()->call_rcu() was not called
961 * yet and exit_task_work() is called before
964 init_task_work(&curr->rcu, task_numa_work);
965 task_work_add(curr, &curr->rcu, true);
970 static struct list_head *account_numa_enqueue(struct rq *rq, struct task_struct *p)
974 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
978 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
981 #endif /* CONFIG_SCHED_NUMA */
983 /**************************************************
984 * Scheduling class queueing methods:
988 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
990 update_load_add(&cfs_rq->load, se->load.weight);
991 if (!parent_entity(se))
992 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
994 if (entity_is_task(se)) {
995 struct rq *rq = rq_of(cfs_rq);
996 struct task_struct *p = task_of(se);
997 struct list_head *tasks = &rq->cfs_tasks;
999 if (tsk_home_node(p) != -1)
1000 tasks = account_numa_enqueue(rq, p);
1002 list_add(&se->group_node, tasks);
1004 #endif /* CONFIG_SMP */
1005 cfs_rq->nr_running++;
1009 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1011 update_load_sub(&cfs_rq->load, se->load.weight);
1012 if (!parent_entity(se))
1013 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1014 if (entity_is_task(se)) {
1015 struct task_struct *p = task_of(se);
1017 list_del_init(&se->group_node);
1019 if (tsk_home_node(p) != -1)
1020 account_numa_dequeue(rq_of(cfs_rq), p);
1022 cfs_rq->nr_running--;
1025 #ifdef CONFIG_FAIR_GROUP_SCHED
1026 /* we need this in update_cfs_load and load-balance functions below */
1027 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1029 static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq,
1032 struct task_group *tg = cfs_rq->tg;
1035 load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1);
1036 load_avg -= cfs_rq->load_contribution;
1038 if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) {
1039 atomic_add(load_avg, &tg->load_weight);
1040 cfs_rq->load_contribution += load_avg;
1044 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
1046 u64 period = sysctl_sched_shares_window;
1048 unsigned long load = cfs_rq->load.weight;
1050 if (cfs_rq->tg == &root_task_group || throttled_hierarchy(cfs_rq))
1053 now = rq_of(cfs_rq)->clock_task;
1054 delta = now - cfs_rq->load_stamp;
1056 /* truncate load history at 4 idle periods */
1057 if (cfs_rq->load_stamp > cfs_rq->load_last &&
1058 now - cfs_rq->load_last > 4 * period) {
1059 cfs_rq->load_period = 0;
1060 cfs_rq->load_avg = 0;
1064 cfs_rq->load_stamp = now;
1065 cfs_rq->load_unacc_exec_time = 0;
1066 cfs_rq->load_period += delta;
1068 cfs_rq->load_last = now;
1069 cfs_rq->load_avg += delta * load;
1072 /* consider updating load contribution on each fold or truncate */
1073 if (global_update || cfs_rq->load_period > period
1074 || !cfs_rq->load_period)
1075 update_cfs_rq_load_contribution(cfs_rq, global_update);
1077 while (cfs_rq->load_period > period) {
1079 * Inline assembly required to prevent the compiler
1080 * optimising this loop into a divmod call.
1081 * See __iter_div_u64_rem() for another example of this.
1083 asm("" : "+rm" (cfs_rq->load_period));
1084 cfs_rq->load_period /= 2;
1085 cfs_rq->load_avg /= 2;
1088 if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg)
1089 list_del_leaf_cfs_rq(cfs_rq);
1092 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1097 * Use this CPU's actual weight instead of the last load_contribution
1098 * to gain a more accurate current total weight. See
1099 * update_cfs_rq_load_contribution().
1101 tg_weight = atomic_read(&tg->load_weight);
1102 tg_weight -= cfs_rq->load_contribution;
1103 tg_weight += cfs_rq->load.weight;
1108 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1110 long tg_weight, load, shares;
1112 tg_weight = calc_tg_weight(tg, cfs_rq);
1113 load = cfs_rq->load.weight;
1115 shares = (tg->shares * load);
1117 shares /= tg_weight;
1119 if (shares < MIN_SHARES)
1120 shares = MIN_SHARES;
1121 if (shares > tg->shares)
1122 shares = tg->shares;
1127 static void update_entity_shares_tick(struct cfs_rq *cfs_rq)
1129 if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) {
1130 update_cfs_load(cfs_rq, 0);
1131 update_cfs_shares(cfs_rq);
1134 # else /* CONFIG_SMP */
1135 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
1139 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1144 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
1147 # endif /* CONFIG_SMP */
1148 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1149 unsigned long weight)
1152 /* commit outstanding execution time */
1153 if (cfs_rq->curr == se)
1154 update_curr(cfs_rq);
1155 account_entity_dequeue(cfs_rq, se);
1158 update_load_set(&se->load, weight);
1161 account_entity_enqueue(cfs_rq, se);
1164 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1166 struct task_group *tg;
1167 struct sched_entity *se;
1171 se = tg->se[cpu_of(rq_of(cfs_rq))];
1172 if (!se || throttled_hierarchy(cfs_rq))
1175 if (likely(se->load.weight == tg->shares))
1178 shares = calc_cfs_shares(cfs_rq, tg);
1180 reweight_entity(cfs_rq_of(se), se, shares);
1182 #else /* CONFIG_FAIR_GROUP_SCHED */
1183 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
1187 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1191 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
1194 #endif /* CONFIG_FAIR_GROUP_SCHED */
1196 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1198 #ifdef CONFIG_SCHEDSTATS
1199 struct task_struct *tsk = NULL;
1201 if (entity_is_task(se))
1204 if (se->statistics.sleep_start) {
1205 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
1210 if (unlikely(delta > se->statistics.sleep_max))
1211 se->statistics.sleep_max = delta;
1213 se->statistics.sleep_start = 0;
1214 se->statistics.sum_sleep_runtime += delta;
1217 account_scheduler_latency(tsk, delta >> 10, 1);
1218 trace_sched_stat_sleep(tsk, delta);
1221 if (se->statistics.block_start) {
1222 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
1227 if (unlikely(delta > se->statistics.block_max))
1228 se->statistics.block_max = delta;
1230 se->statistics.block_start = 0;
1231 se->statistics.sum_sleep_runtime += delta;
1234 if (tsk->in_iowait) {
1235 se->statistics.iowait_sum += delta;
1236 se->statistics.iowait_count++;
1237 trace_sched_stat_iowait(tsk, delta);
1240 trace_sched_stat_blocked(tsk, delta);
1243 * Blocking time is in units of nanosecs, so shift by
1244 * 20 to get a milliseconds-range estimation of the
1245 * amount of time that the task spent sleeping:
1247 if (unlikely(prof_on == SLEEP_PROFILING)) {
1248 profile_hits(SLEEP_PROFILING,
1249 (void *)get_wchan(tsk),
1252 account_scheduler_latency(tsk, delta >> 10, 0);
1258 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1260 #ifdef CONFIG_SCHED_DEBUG
1261 s64 d = se->vruntime - cfs_rq->min_vruntime;
1266 if (d > 3*sysctl_sched_latency)
1267 schedstat_inc(cfs_rq, nr_spread_over);
1272 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1274 u64 vruntime = cfs_rq->min_vruntime;
1277 * The 'current' period is already promised to the current tasks,
1278 * however the extra weight of the new task will slow them down a
1279 * little, place the new task so that it fits in the slot that
1280 * stays open at the end.
1282 if (initial && sched_feat(START_DEBIT))
1283 vruntime += sched_vslice(cfs_rq, se);
1285 /* sleeps up to a single latency don't count. */
1287 unsigned long thresh = sysctl_sched_latency;
1290 * Halve their sleep time's effect, to allow
1291 * for a gentler effect of sleepers:
1293 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1299 /* ensure we never gain time by being placed backwards. */
1300 vruntime = max_vruntime(se->vruntime, vruntime);
1302 se->vruntime = vruntime;
1305 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1308 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1311 * Update the normalized vruntime before updating min_vruntime
1312 * through callig update_curr().
1314 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1315 se->vruntime += cfs_rq->min_vruntime;
1318 * Update run-time statistics of the 'current'.
1320 update_curr(cfs_rq);
1321 update_cfs_load(cfs_rq, 0);
1322 account_entity_enqueue(cfs_rq, se);
1323 update_cfs_shares(cfs_rq);
1325 if (flags & ENQUEUE_WAKEUP) {
1326 place_entity(cfs_rq, se, 0);
1327 enqueue_sleeper(cfs_rq, se);
1330 update_stats_enqueue(cfs_rq, se);
1331 check_spread(cfs_rq, se);
1332 if (se != cfs_rq->curr)
1333 __enqueue_entity(cfs_rq, se);
1336 if (cfs_rq->nr_running == 1) {
1337 list_add_leaf_cfs_rq(cfs_rq);
1338 check_enqueue_throttle(cfs_rq);
1342 static void __clear_buddies_last(struct sched_entity *se)
1344 for_each_sched_entity(se) {
1345 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1346 if (cfs_rq->last == se)
1347 cfs_rq->last = NULL;
1353 static void __clear_buddies_next(struct sched_entity *se)
1355 for_each_sched_entity(se) {
1356 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1357 if (cfs_rq->next == se)
1358 cfs_rq->next = NULL;
1364 static void __clear_buddies_skip(struct sched_entity *se)
1366 for_each_sched_entity(se) {
1367 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1368 if (cfs_rq->skip == se)
1369 cfs_rq->skip = NULL;
1375 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1377 if (cfs_rq->last == se)
1378 __clear_buddies_last(se);
1380 if (cfs_rq->next == se)
1381 __clear_buddies_next(se);
1383 if (cfs_rq->skip == se)
1384 __clear_buddies_skip(se);
1387 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1390 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1393 * Update run-time statistics of the 'current'.
1395 update_curr(cfs_rq);
1397 update_stats_dequeue(cfs_rq, se);
1398 if (flags & DEQUEUE_SLEEP) {
1399 #ifdef CONFIG_SCHEDSTATS
1400 if (entity_is_task(se)) {
1401 struct task_struct *tsk = task_of(se);
1403 if (tsk->state & TASK_INTERRUPTIBLE)
1404 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1405 if (tsk->state & TASK_UNINTERRUPTIBLE)
1406 se->statistics.block_start = rq_of(cfs_rq)->clock;
1411 clear_buddies(cfs_rq, se);
1413 if (se != cfs_rq->curr)
1414 __dequeue_entity(cfs_rq, se);
1416 update_cfs_load(cfs_rq, 0);
1417 account_entity_dequeue(cfs_rq, se);
1420 * Normalize the entity after updating the min_vruntime because the
1421 * update can refer to the ->curr item and we need to reflect this
1422 * movement in our normalized position.
1424 if (!(flags & DEQUEUE_SLEEP))
1425 se->vruntime -= cfs_rq->min_vruntime;
1427 /* return excess runtime on last dequeue */
1428 return_cfs_rq_runtime(cfs_rq);
1430 update_min_vruntime(cfs_rq);
1431 update_cfs_shares(cfs_rq);
1435 * Preempt the current task with a newly woken task if needed:
1438 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1440 unsigned long ideal_runtime, delta_exec;
1441 struct sched_entity *se;
1444 ideal_runtime = sched_slice(cfs_rq, curr);
1445 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1446 if (delta_exec > ideal_runtime) {
1447 resched_task(rq_of(cfs_rq)->curr);
1449 * The current task ran long enough, ensure it doesn't get
1450 * re-elected due to buddy favours.
1452 clear_buddies(cfs_rq, curr);
1457 * Ensure that a task that missed wakeup preemption by a
1458 * narrow margin doesn't have to wait for a full slice.
1459 * This also mitigates buddy induced latencies under load.
1461 if (delta_exec < sysctl_sched_min_granularity)
1464 se = __pick_first_entity(cfs_rq);
1465 delta = curr->vruntime - se->vruntime;
1470 if (delta > ideal_runtime)
1471 resched_task(rq_of(cfs_rq)->curr);
1475 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1477 /* 'current' is not kept within the tree. */
1480 * Any task has to be enqueued before it get to execute on
1481 * a CPU. So account for the time it spent waiting on the
1484 update_stats_wait_end(cfs_rq, se);
1485 __dequeue_entity(cfs_rq, se);
1488 update_stats_curr_start(cfs_rq, se);
1490 #ifdef CONFIG_SCHEDSTATS
1492 * Track our maximum slice length, if the CPU's load is at
1493 * least twice that of our own weight (i.e. dont track it
1494 * when there are only lesser-weight tasks around):
1496 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1497 se->statistics.slice_max = max(se->statistics.slice_max,
1498 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1501 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1505 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1508 * Pick the next process, keeping these things in mind, in this order:
1509 * 1) keep things fair between processes/task groups
1510 * 2) pick the "next" process, since someone really wants that to run
1511 * 3) pick the "last" process, for cache locality
1512 * 4) do not run the "skip" process, if something else is available
1514 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1516 struct sched_entity *se = __pick_first_entity(cfs_rq);
1517 struct sched_entity *left = se;
1520 * Avoid running the skip buddy, if running something else can
1521 * be done without getting too unfair.
1523 if (cfs_rq->skip == se) {
1524 struct sched_entity *second = __pick_next_entity(se);
1525 if (second && wakeup_preempt_entity(second, left) < 1)
1530 * Prefer last buddy, try to return the CPU to a preempted task.
1532 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1536 * Someone really wants this to run. If it's not unfair, run it.
1538 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1541 clear_buddies(cfs_rq, se);
1546 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1548 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1551 * If still on the runqueue then deactivate_task()
1552 * was not called and update_curr() has to be done:
1555 update_curr(cfs_rq);
1557 /* throttle cfs_rqs exceeding runtime */
1558 check_cfs_rq_runtime(cfs_rq);
1560 check_spread(cfs_rq, prev);
1562 update_stats_wait_start(cfs_rq, prev);
1563 /* Put 'current' back into the tree. */
1564 __enqueue_entity(cfs_rq, prev);
1566 cfs_rq->curr = NULL;
1570 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1573 * Update run-time statistics of the 'current'.
1575 update_curr(cfs_rq);
1578 * Update share accounting for long-running entities.
1580 update_entity_shares_tick(cfs_rq);
1582 #ifdef CONFIG_SCHED_HRTICK
1584 * queued ticks are scheduled to match the slice, so don't bother
1585 * validating it and just reschedule.
1588 resched_task(rq_of(cfs_rq)->curr);
1592 * don't let the period tick interfere with the hrtick preemption
1594 if (!sched_feat(DOUBLE_TICK) &&
1595 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1599 if (cfs_rq->nr_running > 1)
1600 check_preempt_tick(cfs_rq, curr);
1604 /**************************************************
1605 * CFS bandwidth control machinery
1608 #ifdef CONFIG_CFS_BANDWIDTH
1610 #ifdef HAVE_JUMP_LABEL
1611 static struct static_key __cfs_bandwidth_used;
1613 static inline bool cfs_bandwidth_used(void)
1615 return static_key_false(&__cfs_bandwidth_used);
1618 void account_cfs_bandwidth_used(int enabled, int was_enabled)
1620 /* only need to count groups transitioning between enabled/!enabled */
1621 if (enabled && !was_enabled)
1622 static_key_slow_inc(&__cfs_bandwidth_used);
1623 else if (!enabled && was_enabled)
1624 static_key_slow_dec(&__cfs_bandwidth_used);
1626 #else /* HAVE_JUMP_LABEL */
1627 static bool cfs_bandwidth_used(void)
1632 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
1633 #endif /* HAVE_JUMP_LABEL */
1636 * default period for cfs group bandwidth.
1637 * default: 0.1s, units: nanoseconds
1639 static inline u64 default_cfs_period(void)
1641 return 100000000ULL;
1644 static inline u64 sched_cfs_bandwidth_slice(void)
1646 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
1650 * Replenish runtime according to assigned quota and update expiration time.
1651 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
1652 * additional synchronization around rq->lock.
1654 * requires cfs_b->lock
1656 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
1660 if (cfs_b->quota == RUNTIME_INF)
1663 now = sched_clock_cpu(smp_processor_id());
1664 cfs_b->runtime = cfs_b->quota;
1665 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
1668 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
1670 return &tg->cfs_bandwidth;
1673 /* returns 0 on failure to allocate runtime */
1674 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1676 struct task_group *tg = cfs_rq->tg;
1677 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
1678 u64 amount = 0, min_amount, expires;
1680 /* note: this is a positive sum as runtime_remaining <= 0 */
1681 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
1683 raw_spin_lock(&cfs_b->lock);
1684 if (cfs_b->quota == RUNTIME_INF)
1685 amount = min_amount;
1688 * If the bandwidth pool has become inactive, then at least one
1689 * period must have elapsed since the last consumption.
1690 * Refresh the global state and ensure bandwidth timer becomes
1693 if (!cfs_b->timer_active) {
1694 __refill_cfs_bandwidth_runtime(cfs_b);
1695 __start_cfs_bandwidth(cfs_b);
1698 if (cfs_b->runtime > 0) {
1699 amount = min(cfs_b->runtime, min_amount);
1700 cfs_b->runtime -= amount;
1704 expires = cfs_b->runtime_expires;
1705 raw_spin_unlock(&cfs_b->lock);
1707 cfs_rq->runtime_remaining += amount;
1709 * we may have advanced our local expiration to account for allowed
1710 * spread between our sched_clock and the one on which runtime was
1713 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
1714 cfs_rq->runtime_expires = expires;
1716 return cfs_rq->runtime_remaining > 0;
1720 * Note: This depends on the synchronization provided by sched_clock and the
1721 * fact that rq->clock snapshots this value.
1723 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1725 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1726 struct rq *rq = rq_of(cfs_rq);
1728 /* if the deadline is ahead of our clock, nothing to do */
1729 if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0))
1732 if (cfs_rq->runtime_remaining < 0)
1736 * If the local deadline has passed we have to consider the
1737 * possibility that our sched_clock is 'fast' and the global deadline
1738 * has not truly expired.
1740 * Fortunately we can check determine whether this the case by checking
1741 * whether the global deadline has advanced.
1744 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
1745 /* extend local deadline, drift is bounded above by 2 ticks */
1746 cfs_rq->runtime_expires += TICK_NSEC;
1748 /* global deadline is ahead, expiration has passed */
1749 cfs_rq->runtime_remaining = 0;
1753 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1754 unsigned long delta_exec)
1756 /* dock delta_exec before expiring quota (as it could span periods) */
1757 cfs_rq->runtime_remaining -= delta_exec;
1758 expire_cfs_rq_runtime(cfs_rq);
1760 if (likely(cfs_rq->runtime_remaining > 0))
1764 * if we're unable to extend our runtime we resched so that the active
1765 * hierarchy can be throttled
1767 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
1768 resched_task(rq_of(cfs_rq)->curr);
1771 static __always_inline
1772 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
1774 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
1777 __account_cfs_rq_runtime(cfs_rq, delta_exec);
1780 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
1782 return cfs_bandwidth_used() && cfs_rq->throttled;
1785 /* check whether cfs_rq, or any parent, is throttled */
1786 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
1788 return cfs_bandwidth_used() && cfs_rq->throttle_count;
1792 * Ensure that neither of the group entities corresponding to src_cpu or
1793 * dest_cpu are members of a throttled hierarchy when performing group
1794 * load-balance operations.
1796 static inline int throttled_lb_pair(struct task_group *tg,
1797 int src_cpu, int dest_cpu)
1799 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
1801 src_cfs_rq = tg->cfs_rq[src_cpu];
1802 dest_cfs_rq = tg->cfs_rq[dest_cpu];
1804 return throttled_hierarchy(src_cfs_rq) ||
1805 throttled_hierarchy(dest_cfs_rq);
1808 /* updated child weight may affect parent so we have to do this bottom up */
1809 static int tg_unthrottle_up(struct task_group *tg, void *data)
1811 struct rq *rq = data;
1812 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1814 cfs_rq->throttle_count--;
1816 if (!cfs_rq->throttle_count) {
1817 u64 delta = rq->clock_task - cfs_rq->load_stamp;
1819 /* leaving throttled state, advance shares averaging windows */
1820 cfs_rq->load_stamp += delta;
1821 cfs_rq->load_last += delta;
1823 /* update entity weight now that we are on_rq again */
1824 update_cfs_shares(cfs_rq);
1831 static int tg_throttle_down(struct task_group *tg, void *data)
1833 struct rq *rq = data;
1834 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1836 /* group is entering throttled state, record last load */
1837 if (!cfs_rq->throttle_count)
1838 update_cfs_load(cfs_rq, 0);
1839 cfs_rq->throttle_count++;
1844 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
1846 struct rq *rq = rq_of(cfs_rq);
1847 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1848 struct sched_entity *se;
1849 long task_delta, dequeue = 1;
1851 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1853 /* account load preceding throttle */
1855 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
1858 task_delta = cfs_rq->h_nr_running;
1859 for_each_sched_entity(se) {
1860 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
1861 /* throttled entity or throttle-on-deactivate */
1866 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
1867 qcfs_rq->h_nr_running -= task_delta;
1869 if (qcfs_rq->load.weight)
1874 rq->nr_running -= task_delta;
1876 cfs_rq->throttled = 1;
1877 cfs_rq->throttled_timestamp = rq->clock;
1878 raw_spin_lock(&cfs_b->lock);
1879 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
1880 raw_spin_unlock(&cfs_b->lock);
1883 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
1885 struct rq *rq = rq_of(cfs_rq);
1886 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1887 struct sched_entity *se;
1891 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1893 cfs_rq->throttled = 0;
1894 raw_spin_lock(&cfs_b->lock);
1895 cfs_b->throttled_time += rq->clock - cfs_rq->throttled_timestamp;
1896 list_del_rcu(&cfs_rq->throttled_list);
1897 raw_spin_unlock(&cfs_b->lock);
1898 cfs_rq->throttled_timestamp = 0;
1900 update_rq_clock(rq);
1901 /* update hierarchical throttle state */
1902 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
1904 if (!cfs_rq->load.weight)
1907 task_delta = cfs_rq->h_nr_running;
1908 for_each_sched_entity(se) {
1912 cfs_rq = cfs_rq_of(se);
1914 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
1915 cfs_rq->h_nr_running += task_delta;
1917 if (cfs_rq_throttled(cfs_rq))
1922 rq->nr_running += task_delta;
1924 /* determine whether we need to wake up potentially idle cpu */
1925 if (rq->curr == rq->idle && rq->cfs.nr_running)
1926 resched_task(rq->curr);
1929 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
1930 u64 remaining, u64 expires)
1932 struct cfs_rq *cfs_rq;
1933 u64 runtime = remaining;
1936 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
1938 struct rq *rq = rq_of(cfs_rq);
1940 raw_spin_lock(&rq->lock);
1941 if (!cfs_rq_throttled(cfs_rq))
1944 runtime = -cfs_rq->runtime_remaining + 1;
1945 if (runtime > remaining)
1946 runtime = remaining;
1947 remaining -= runtime;
1949 cfs_rq->runtime_remaining += runtime;
1950 cfs_rq->runtime_expires = expires;
1952 /* we check whether we're throttled above */
1953 if (cfs_rq->runtime_remaining > 0)
1954 unthrottle_cfs_rq(cfs_rq);
1957 raw_spin_unlock(&rq->lock);
1968 * Responsible for refilling a task_group's bandwidth and unthrottling its
1969 * cfs_rqs as appropriate. If there has been no activity within the last
1970 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
1971 * used to track this state.
1973 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
1975 u64 runtime, runtime_expires;
1976 int idle = 1, throttled;
1978 raw_spin_lock(&cfs_b->lock);
1979 /* no need to continue the timer with no bandwidth constraint */
1980 if (cfs_b->quota == RUNTIME_INF)
1983 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
1984 /* idle depends on !throttled (for the case of a large deficit) */
1985 idle = cfs_b->idle && !throttled;
1986 cfs_b->nr_periods += overrun;
1988 /* if we're going inactive then everything else can be deferred */
1992 __refill_cfs_bandwidth_runtime(cfs_b);
1995 /* mark as potentially idle for the upcoming period */
2000 /* account preceding periods in which throttling occurred */
2001 cfs_b->nr_throttled += overrun;
2004 * There are throttled entities so we must first use the new bandwidth
2005 * to unthrottle them before making it generally available. This
2006 * ensures that all existing debts will be paid before a new cfs_rq is
2009 runtime = cfs_b->runtime;
2010 runtime_expires = cfs_b->runtime_expires;
2014 * This check is repeated as we are holding onto the new bandwidth
2015 * while we unthrottle. This can potentially race with an unthrottled
2016 * group trying to acquire new bandwidth from the global pool.
2018 while (throttled && runtime > 0) {
2019 raw_spin_unlock(&cfs_b->lock);
2020 /* we can't nest cfs_b->lock while distributing bandwidth */
2021 runtime = distribute_cfs_runtime(cfs_b, runtime,
2023 raw_spin_lock(&cfs_b->lock);
2025 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2028 /* return (any) remaining runtime */
2029 cfs_b->runtime = runtime;
2031 * While we are ensured activity in the period following an
2032 * unthrottle, this also covers the case in which the new bandwidth is
2033 * insufficient to cover the existing bandwidth deficit. (Forcing the
2034 * timer to remain active while there are any throttled entities.)
2039 cfs_b->timer_active = 0;
2040 raw_spin_unlock(&cfs_b->lock);
2045 /* a cfs_rq won't donate quota below this amount */
2046 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2047 /* minimum remaining period time to redistribute slack quota */
2048 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2049 /* how long we wait to gather additional slack before distributing */
2050 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2052 /* are we near the end of the current quota period? */
2053 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2055 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2058 /* if the call-back is running a quota refresh is already occurring */
2059 if (hrtimer_callback_running(refresh_timer))
2062 /* is a quota refresh about to occur? */
2063 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2064 if (remaining < min_expire)
2070 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2072 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2074 /* if there's a quota refresh soon don't bother with slack */
2075 if (runtime_refresh_within(cfs_b, min_left))
2078 start_bandwidth_timer(&cfs_b->slack_timer,
2079 ns_to_ktime(cfs_bandwidth_slack_period));
2082 /* we know any runtime found here is valid as update_curr() precedes return */
2083 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2085 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2086 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2088 if (slack_runtime <= 0)
2091 raw_spin_lock(&cfs_b->lock);
2092 if (cfs_b->quota != RUNTIME_INF &&
2093 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2094 cfs_b->runtime += slack_runtime;
2096 /* we are under rq->lock, defer unthrottling using a timer */
2097 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2098 !list_empty(&cfs_b->throttled_cfs_rq))
2099 start_cfs_slack_bandwidth(cfs_b);
2101 raw_spin_unlock(&cfs_b->lock);
2103 /* even if it's not valid for return we don't want to try again */
2104 cfs_rq->runtime_remaining -= slack_runtime;
2107 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2109 if (!cfs_bandwidth_used())
2112 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2115 __return_cfs_rq_runtime(cfs_rq);
2119 * This is done with a timer (instead of inline with bandwidth return) since
2120 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2122 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2124 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2127 /* confirm we're still not at a refresh boundary */
2128 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2131 raw_spin_lock(&cfs_b->lock);
2132 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2133 runtime = cfs_b->runtime;
2136 expires = cfs_b->runtime_expires;
2137 raw_spin_unlock(&cfs_b->lock);
2142 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2144 raw_spin_lock(&cfs_b->lock);
2145 if (expires == cfs_b->runtime_expires)
2146 cfs_b->runtime = runtime;
2147 raw_spin_unlock(&cfs_b->lock);
2151 * When a group wakes up we want to make sure that its quota is not already
2152 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2153 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2155 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2157 if (!cfs_bandwidth_used())
2160 /* an active group must be handled by the update_curr()->put() path */
2161 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2164 /* ensure the group is not already throttled */
2165 if (cfs_rq_throttled(cfs_rq))
2168 /* update runtime allocation */
2169 account_cfs_rq_runtime(cfs_rq, 0);
2170 if (cfs_rq->runtime_remaining <= 0)
2171 throttle_cfs_rq(cfs_rq);
2174 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2175 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2177 if (!cfs_bandwidth_used())
2180 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2184 * it's possible for a throttled entity to be forced into a running
2185 * state (e.g. set_curr_task), in this case we're finished.
2187 if (cfs_rq_throttled(cfs_rq))
2190 throttle_cfs_rq(cfs_rq);
2193 static inline u64 default_cfs_period(void);
2194 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun);
2195 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b);
2197 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2199 struct cfs_bandwidth *cfs_b =
2200 container_of(timer, struct cfs_bandwidth, slack_timer);
2201 do_sched_cfs_slack_timer(cfs_b);
2203 return HRTIMER_NORESTART;
2206 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2208 struct cfs_bandwidth *cfs_b =
2209 container_of(timer, struct cfs_bandwidth, period_timer);
2215 now = hrtimer_cb_get_time(timer);
2216 overrun = hrtimer_forward(timer, now, cfs_b->period);
2221 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2224 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2227 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2229 raw_spin_lock_init(&cfs_b->lock);
2231 cfs_b->quota = RUNTIME_INF;
2232 cfs_b->period = ns_to_ktime(default_cfs_period());
2234 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2235 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2236 cfs_b->period_timer.function = sched_cfs_period_timer;
2237 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2238 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2241 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2243 cfs_rq->runtime_enabled = 0;
2244 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2247 /* requires cfs_b->lock, may release to reprogram timer */
2248 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2251 * The timer may be active because we're trying to set a new bandwidth
2252 * period or because we're racing with the tear-down path
2253 * (timer_active==0 becomes visible before the hrtimer call-back
2254 * terminates). In either case we ensure that it's re-programmed
2256 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2257 raw_spin_unlock(&cfs_b->lock);
2258 /* ensure cfs_b->lock is available while we wait */
2259 hrtimer_cancel(&cfs_b->period_timer);
2261 raw_spin_lock(&cfs_b->lock);
2262 /* if someone else restarted the timer then we're done */
2263 if (cfs_b->timer_active)
2267 cfs_b->timer_active = 1;
2268 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2271 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2273 hrtimer_cancel(&cfs_b->period_timer);
2274 hrtimer_cancel(&cfs_b->slack_timer);
2277 static void unthrottle_offline_cfs_rqs(struct rq *rq)
2279 struct cfs_rq *cfs_rq;
2281 for_each_leaf_cfs_rq(rq, cfs_rq) {
2282 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2284 if (!cfs_rq->runtime_enabled)
2288 * clock_task is not advancing so we just need to make sure
2289 * there's some valid quota amount
2291 cfs_rq->runtime_remaining = cfs_b->quota;
2292 if (cfs_rq_throttled(cfs_rq))
2293 unthrottle_cfs_rq(cfs_rq);
2297 #else /* CONFIG_CFS_BANDWIDTH */
2298 static __always_inline
2299 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec) {}
2300 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2301 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2302 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2304 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2309 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2314 static inline int throttled_lb_pair(struct task_group *tg,
2315 int src_cpu, int dest_cpu)
2320 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2322 #ifdef CONFIG_FAIR_GROUP_SCHED
2323 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2326 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2330 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2331 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2333 #endif /* CONFIG_CFS_BANDWIDTH */
2335 /**************************************************
2336 * CFS operations on tasks:
2339 #ifdef CONFIG_SCHED_HRTICK
2340 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2342 struct sched_entity *se = &p->se;
2343 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2345 WARN_ON(task_rq(p) != rq);
2347 if (cfs_rq->nr_running > 1) {
2348 u64 slice = sched_slice(cfs_rq, se);
2349 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2350 s64 delta = slice - ran;
2359 * Don't schedule slices shorter than 10000ns, that just
2360 * doesn't make sense. Rely on vruntime for fairness.
2363 delta = max_t(s64, 10000LL, delta);
2365 hrtick_start(rq, delta);
2370 * called from enqueue/dequeue and updates the hrtick when the
2371 * current task is from our class and nr_running is low enough
2374 static void hrtick_update(struct rq *rq)
2376 struct task_struct *curr = rq->curr;
2378 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2381 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2382 hrtick_start_fair(rq, curr);
2384 #else /* !CONFIG_SCHED_HRTICK */
2386 hrtick_start_fair(struct rq *rq, struct task_struct *p)
2390 static inline void hrtick_update(struct rq *rq)
2396 * The enqueue_task method is called before nr_running is
2397 * increased. Here we update the fair scheduling stats and
2398 * then put the task into the rbtree:
2401 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2403 struct cfs_rq *cfs_rq;
2404 struct sched_entity *se = &p->se;
2406 for_each_sched_entity(se) {
2409 cfs_rq = cfs_rq_of(se);
2410 enqueue_entity(cfs_rq, se, flags);
2413 * end evaluation on encountering a throttled cfs_rq
2415 * note: in the case of encountering a throttled cfs_rq we will
2416 * post the final h_nr_running increment below.
2418 if (cfs_rq_throttled(cfs_rq))
2420 cfs_rq->h_nr_running++;
2422 flags = ENQUEUE_WAKEUP;
2425 for_each_sched_entity(se) {
2426 cfs_rq = cfs_rq_of(se);
2427 cfs_rq->h_nr_running++;
2429 if (cfs_rq_throttled(cfs_rq))
2432 update_cfs_load(cfs_rq, 0);
2433 update_cfs_shares(cfs_rq);
2441 static void set_next_buddy(struct sched_entity *se);
2444 * The dequeue_task method is called before nr_running is
2445 * decreased. We remove the task from the rbtree and
2446 * update the fair scheduling stats:
2448 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2450 struct cfs_rq *cfs_rq;
2451 struct sched_entity *se = &p->se;
2452 int task_sleep = flags & DEQUEUE_SLEEP;
2454 for_each_sched_entity(se) {
2455 cfs_rq = cfs_rq_of(se);
2456 dequeue_entity(cfs_rq, se, flags);
2459 * end evaluation on encountering a throttled cfs_rq
2461 * note: in the case of encountering a throttled cfs_rq we will
2462 * post the final h_nr_running decrement below.
2464 if (cfs_rq_throttled(cfs_rq))
2466 cfs_rq->h_nr_running--;
2468 /* Don't dequeue parent if it has other entities besides us */
2469 if (cfs_rq->load.weight) {
2471 * Bias pick_next to pick a task from this cfs_rq, as
2472 * p is sleeping when it is within its sched_slice.
2474 if (task_sleep && parent_entity(se))
2475 set_next_buddy(parent_entity(se));
2477 /* avoid re-evaluating load for this entity */
2478 se = parent_entity(se);
2481 flags |= DEQUEUE_SLEEP;
2484 for_each_sched_entity(se) {
2485 cfs_rq = cfs_rq_of(se);
2486 cfs_rq->h_nr_running--;
2488 if (cfs_rq_throttled(cfs_rq))
2491 update_cfs_load(cfs_rq, 0);
2492 update_cfs_shares(cfs_rq);
2501 /* Used instead of source_load when we know the type == 0 */
2502 static unsigned long weighted_cpuload(const int cpu)
2504 return cpu_rq(cpu)->load.weight;
2508 * Return a low guess at the load of a migration-source cpu weighted
2509 * according to the scheduling class and "nice" value.
2511 * We want to under-estimate the load of migration sources, to
2512 * balance conservatively.
2514 static unsigned long source_load(int cpu, int type)
2516 struct rq *rq = cpu_rq(cpu);
2517 unsigned long total = weighted_cpuload(cpu);
2519 if (type == 0 || !sched_feat(LB_BIAS))
2522 return min(rq->cpu_load[type-1], total);
2526 * Return a high guess at the load of a migration-target cpu weighted
2527 * according to the scheduling class and "nice" value.
2529 static unsigned long target_load(int cpu, int type)
2531 struct rq *rq = cpu_rq(cpu);
2532 unsigned long total = weighted_cpuload(cpu);
2534 if (type == 0 || !sched_feat(LB_BIAS))
2537 return max(rq->cpu_load[type-1], total);
2540 static unsigned long power_of(int cpu)
2542 return cpu_rq(cpu)->cpu_power;
2545 static unsigned long cpu_avg_load_per_task(int cpu)
2547 struct rq *rq = cpu_rq(cpu);
2548 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
2551 return rq->load.weight / nr_running;
2557 static void task_waking_fair(struct task_struct *p)
2559 struct sched_entity *se = &p->se;
2560 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2563 #ifndef CONFIG_64BIT
2564 u64 min_vruntime_copy;
2567 min_vruntime_copy = cfs_rq->min_vruntime_copy;
2569 min_vruntime = cfs_rq->min_vruntime;
2570 } while (min_vruntime != min_vruntime_copy);
2572 min_vruntime = cfs_rq->min_vruntime;
2575 se->vruntime -= min_vruntime;
2578 #ifdef CONFIG_FAIR_GROUP_SCHED
2580 * effective_load() calculates the load change as seen from the root_task_group
2582 * Adding load to a group doesn't make a group heavier, but can cause movement
2583 * of group shares between cpus. Assuming the shares were perfectly aligned one
2584 * can calculate the shift in shares.
2586 * Calculate the effective load difference if @wl is added (subtracted) to @tg
2587 * on this @cpu and results in a total addition (subtraction) of @wg to the
2588 * total group weight.
2590 * Given a runqueue weight distribution (rw_i) we can compute a shares
2591 * distribution (s_i) using:
2593 * s_i = rw_i / \Sum rw_j (1)
2595 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
2596 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
2597 * shares distribution (s_i):
2599 * rw_i = { 2, 4, 1, 0 }
2600 * s_i = { 2/7, 4/7, 1/7, 0 }
2602 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
2603 * task used to run on and the CPU the waker is running on), we need to
2604 * compute the effect of waking a task on either CPU and, in case of a sync
2605 * wakeup, compute the effect of the current task going to sleep.
2607 * So for a change of @wl to the local @cpu with an overall group weight change
2608 * of @wl we can compute the new shares distribution (s'_i) using:
2610 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
2612 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
2613 * differences in waking a task to CPU 0. The additional task changes the
2614 * weight and shares distributions like:
2616 * rw'_i = { 3, 4, 1, 0 }
2617 * s'_i = { 3/8, 4/8, 1/8, 0 }
2619 * We can then compute the difference in effective weight by using:
2621 * dw_i = S * (s'_i - s_i) (3)
2623 * Where 'S' is the group weight as seen by its parent.
2625 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
2626 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
2627 * 4/7) times the weight of the group.
2629 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
2631 struct sched_entity *se = tg->se[cpu];
2633 if (!tg->parent) /* the trivial, non-cgroup case */
2636 for_each_sched_entity(se) {
2642 * W = @wg + \Sum rw_j
2644 W = wg + calc_tg_weight(tg, se->my_q);
2649 w = se->my_q->load.weight + wl;
2652 * wl = S * s'_i; see (2)
2655 wl = (w * tg->shares) / W;
2660 * Per the above, wl is the new se->load.weight value; since
2661 * those are clipped to [MIN_SHARES, ...) do so now. See
2662 * calc_cfs_shares().
2664 if (wl < MIN_SHARES)
2668 * wl = dw_i = S * (s'_i - s_i); see (3)
2670 wl -= se->load.weight;
2673 * Recursively apply this logic to all parent groups to compute
2674 * the final effective load change on the root group. Since
2675 * only the @tg group gets extra weight, all parent groups can
2676 * only redistribute existing shares. @wl is the shift in shares
2677 * resulting from this level per the above.
2686 static inline unsigned long effective_load(struct task_group *tg, int cpu,
2687 unsigned long wl, unsigned long wg)
2694 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
2696 s64 this_load, load;
2697 int idx, this_cpu, prev_cpu;
2698 unsigned long tl_per_task;
2699 struct task_group *tg;
2700 unsigned long weight;
2704 this_cpu = smp_processor_id();
2705 prev_cpu = task_cpu(p);
2706 load = source_load(prev_cpu, idx);
2707 this_load = target_load(this_cpu, idx);
2710 * If sync wakeup then subtract the (maximum possible)
2711 * effect of the currently running task from the load
2712 * of the current CPU:
2715 tg = task_group(current);
2716 weight = current->se.load.weight;
2718 this_load += effective_load(tg, this_cpu, -weight, -weight);
2719 load += effective_load(tg, prev_cpu, 0, -weight);
2723 weight = p->se.load.weight;
2726 * In low-load situations, where prev_cpu is idle and this_cpu is idle
2727 * due to the sync cause above having dropped this_load to 0, we'll
2728 * always have an imbalance, but there's really nothing you can do
2729 * about that, so that's good too.
2731 * Otherwise check if either cpus are near enough in load to allow this
2732 * task to be woken on this_cpu.
2734 if (this_load > 0) {
2735 s64 this_eff_load, prev_eff_load;
2737 this_eff_load = 100;
2738 this_eff_load *= power_of(prev_cpu);
2739 this_eff_load *= this_load +
2740 effective_load(tg, this_cpu, weight, weight);
2742 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
2743 prev_eff_load *= power_of(this_cpu);
2744 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
2746 balanced = this_eff_load <= prev_eff_load;
2751 * If the currently running task will sleep within
2752 * a reasonable amount of time then attract this newly
2755 if (sync && balanced)
2758 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
2759 tl_per_task = cpu_avg_load_per_task(this_cpu);
2762 (this_load <= load &&
2763 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
2765 * This domain has SD_WAKE_AFFINE and
2766 * p is cache cold in this domain, and
2767 * there is no bad imbalance.
2769 schedstat_inc(sd, ttwu_move_affine);
2770 schedstat_inc(p, se.statistics.nr_wakeups_affine);
2778 * find_idlest_group finds and returns the least busy CPU group within the
2781 static struct sched_group *
2782 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
2783 int this_cpu, int load_idx)
2785 struct sched_group *idlest = NULL, *group = sd->groups;
2786 unsigned long min_load = ULONG_MAX, this_load = 0;
2787 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2790 unsigned long load, avg_load;
2794 /* Skip over this group if it has no CPUs allowed */
2795 if (!cpumask_intersects(sched_group_cpus(group),
2796 tsk_cpus_allowed(p)))
2799 local_group = cpumask_test_cpu(this_cpu,
2800 sched_group_cpus(group));
2802 /* Tally up the load of all CPUs in the group */
2805 for_each_cpu(i, sched_group_cpus(group)) {
2806 /* Bias balancing toward cpus of our domain */
2808 load = source_load(i, load_idx);
2810 load = target_load(i, load_idx);
2815 /* Adjust by relative CPU power of the group */
2816 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
2819 this_load = avg_load;
2820 } else if (avg_load < min_load) {
2821 min_load = avg_load;
2824 } while (group = group->next, group != sd->groups);
2826 if (!idlest || 100*this_load < imbalance*min_load)
2832 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2835 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2837 unsigned long load, min_load = ULONG_MAX;
2841 /* Traverse only the allowed CPUs */
2842 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
2843 load = weighted_cpuload(i);
2845 if (load < min_load || (load == min_load && i == this_cpu)) {
2855 * Try and locate an idle CPU in the sched_domain.
2857 static int select_idle_sibling(struct task_struct *p, int target)
2859 int cpu = smp_processor_id();
2860 int prev_cpu = task_cpu(p);
2861 struct sched_domain *sd;
2862 struct sched_group *sg;
2866 * If the task is going to be woken-up on this cpu and if it is
2867 * already idle, then it is the right target.
2869 if (target == cpu && idle_cpu(cpu))
2873 * If the task is going to be woken-up on the cpu where it previously
2874 * ran and if it is currently idle, then it the right target.
2876 if (target == prev_cpu && idle_cpu(prev_cpu))
2880 * Otherwise, iterate the domains and find an elegible idle cpu.
2882 sd = rcu_dereference(per_cpu(sd_llc, target));
2883 for_each_lower_domain(sd) {
2886 if (!cpumask_intersects(sched_group_cpus(sg),
2887 tsk_cpus_allowed(p)))
2890 for_each_cpu(i, sched_group_cpus(sg)) {
2895 target = cpumask_first_and(sched_group_cpus(sg),
2896 tsk_cpus_allowed(p));
2900 } while (sg != sd->groups);
2906 #ifdef CONFIG_SCHED_NUMA
2907 static inline bool pick_numa_rand(int n)
2909 return !(get_random_int() % n);
2913 * Pick a random elegible CPU in the target node, hopefully faster
2914 * than doing a least-loaded scan.
2916 static int numa_select_node_cpu(struct task_struct *p, int node)
2918 int weight = cpumask_weight(cpumask_of_node(node));
2921 for_each_cpu_and(i, cpumask_of_node(node), tsk_cpus_allowed(p)) {
2922 if (cpu < 0 || pick_numa_rand(weight))
2929 static int numa_select_node_cpu(struct task_struct *p, int node)
2933 #endif /* CONFIG_SCHED_NUMA */
2936 * sched_balance_self: balance the current task (running on cpu) in domains
2937 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2940 * Balance, ie. select the least loaded group.
2942 * Returns the target CPU number, or the same CPU if no balancing is needed.
2944 * preempt must be disabled.
2947 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
2949 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
2950 int cpu = smp_processor_id();
2951 int prev_cpu = task_cpu(p);
2953 int want_affine = 0;
2954 int sync = wake_flags & WF_SYNC;
2955 int node = tsk_home_node(p);
2957 if (p->nr_cpus_allowed == 1)
2960 if (sd_flag & SD_BALANCE_WAKE) {
2961 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
2967 if (sched_feat_numa(NUMA_TTWU_BIAS) && node != -1) {
2969 * For fork,exec find the idlest cpu in the home-node.
2971 if (sd_flag & (SD_BALANCE_FORK|SD_BALANCE_EXEC)) {
2972 int node_cpu = numa_select_node_cpu(p, node);
2976 new_cpu = cpu = node_cpu;
2977 sd = per_cpu(sd_node, cpu);
2982 * For wake, pretend we were running in the home-node.
2984 if (cpu_to_node(prev_cpu) != node) {
2985 int node_cpu = numa_select_node_cpu(p, node);
2989 if (sched_feat_numa(NUMA_TTWU_TO))
2992 prev_cpu = node_cpu;
2997 for_each_domain(cpu, tmp) {
2998 if (!(tmp->flags & SD_LOAD_BALANCE))
3002 * If both cpu and prev_cpu are part of this domain,
3003 * cpu is a valid SD_WAKE_AFFINE target.
3005 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3006 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3011 if (tmp->flags & sd_flag)
3016 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3019 new_cpu = select_idle_sibling(p, prev_cpu);
3025 int load_idx = sd->forkexec_idx;
3026 struct sched_group *group;
3029 if (!(sd->flags & sd_flag)) {
3034 if (sd_flag & SD_BALANCE_WAKE)
3035 load_idx = sd->wake_idx;
3037 group = find_idlest_group(sd, p, cpu, load_idx);
3043 new_cpu = find_idlest_cpu(group, p, cpu);
3044 if (new_cpu == -1 || new_cpu == cpu) {
3045 /* Now try balancing at a lower domain level of cpu */
3050 /* Now try balancing at a lower domain level of new_cpu */
3052 weight = sd->span_weight;
3054 for_each_domain(cpu, tmp) {
3055 if (weight <= tmp->span_weight)
3057 if (tmp->flags & sd_flag)
3060 /* while loop will break here if sd == NULL */
3067 #endif /* CONFIG_SMP */
3069 static unsigned long
3070 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3072 unsigned long gran = sysctl_sched_wakeup_granularity;
3075 * Since its curr running now, convert the gran from real-time
3076 * to virtual-time in his units.
3078 * By using 'se' instead of 'curr' we penalize light tasks, so
3079 * they get preempted easier. That is, if 'se' < 'curr' then
3080 * the resulting gran will be larger, therefore penalizing the
3081 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3082 * be smaller, again penalizing the lighter task.
3084 * This is especially important for buddies when the leftmost
3085 * task is higher priority than the buddy.
3087 return calc_delta_fair(gran, se);
3091 * Should 'se' preempt 'curr'.
3105 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3107 s64 gran, vdiff = curr->vruntime - se->vruntime;
3112 gran = wakeup_gran(curr, se);
3119 static void set_last_buddy(struct sched_entity *se)
3121 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3124 for_each_sched_entity(se)
3125 cfs_rq_of(se)->last = se;
3128 static void set_next_buddy(struct sched_entity *se)
3130 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3133 for_each_sched_entity(se)
3134 cfs_rq_of(se)->next = se;
3137 static void set_skip_buddy(struct sched_entity *se)
3139 for_each_sched_entity(se)
3140 cfs_rq_of(se)->skip = se;
3144 * Preempt the current task with a newly woken task if needed:
3146 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3148 struct task_struct *curr = rq->curr;
3149 struct sched_entity *se = &curr->se, *pse = &p->se;
3150 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3151 int scale = cfs_rq->nr_running >= sched_nr_latency;
3152 int next_buddy_marked = 0;
3154 if (unlikely(se == pse))
3158 * This is possible from callers such as move_task(), in which we
3159 * unconditionally check_prempt_curr() after an enqueue (which may have
3160 * lead to a throttle). This both saves work and prevents false
3161 * next-buddy nomination below.
3163 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3166 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3167 set_next_buddy(pse);
3168 next_buddy_marked = 1;
3172 * We can come here with TIF_NEED_RESCHED already set from new task
3175 * Note: this also catches the edge-case of curr being in a throttled
3176 * group (e.g. via set_curr_task), since update_curr() (in the
3177 * enqueue of curr) will have resulted in resched being set. This
3178 * prevents us from potentially nominating it as a false LAST_BUDDY
3181 if (test_tsk_need_resched(curr))
3184 /* Idle tasks are by definition preempted by non-idle tasks. */
3185 if (unlikely(curr->policy == SCHED_IDLE) &&
3186 likely(p->policy != SCHED_IDLE))
3190 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3191 * is driven by the tick):
3193 if (unlikely(p->policy != SCHED_NORMAL))
3196 find_matching_se(&se, &pse);
3197 update_curr(cfs_rq_of(se));
3199 if (wakeup_preempt_entity(se, pse) == 1) {
3201 * Bias pick_next to pick the sched entity that is
3202 * triggering this preemption.
3204 if (!next_buddy_marked)
3205 set_next_buddy(pse);
3214 * Only set the backward buddy when the current task is still
3215 * on the rq. This can happen when a wakeup gets interleaved
3216 * with schedule on the ->pre_schedule() or idle_balance()
3217 * point, either of which can * drop the rq lock.
3219 * Also, during early boot the idle thread is in the fair class,
3220 * for obvious reasons its a bad idea to schedule back to it.
3222 if (unlikely(!se->on_rq || curr == rq->idle))
3225 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3229 static struct task_struct *pick_next_task_fair(struct rq *rq)
3231 struct task_struct *p;
3232 struct cfs_rq *cfs_rq = &rq->cfs;
3233 struct sched_entity *se;
3235 if (!cfs_rq->nr_running)
3239 se = pick_next_entity(cfs_rq);
3240 set_next_entity(cfs_rq, se);
3241 cfs_rq = group_cfs_rq(se);
3245 if (hrtick_enabled(rq))
3246 hrtick_start_fair(rq, p);
3252 * Account for a descheduled task:
3254 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3256 struct sched_entity *se = &prev->se;
3257 struct cfs_rq *cfs_rq;
3259 for_each_sched_entity(se) {
3260 cfs_rq = cfs_rq_of(se);
3261 put_prev_entity(cfs_rq, se);
3266 * sched_yield() is very simple
3268 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3270 static void yield_task_fair(struct rq *rq)
3272 struct task_struct *curr = rq->curr;
3273 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3274 struct sched_entity *se = &curr->se;
3277 * Are we the only task in the tree?
3279 if (unlikely(rq->nr_running == 1))
3282 clear_buddies(cfs_rq, se);
3284 if (curr->policy != SCHED_BATCH) {
3285 update_rq_clock(rq);
3287 * Update run-time statistics of the 'current'.
3289 update_curr(cfs_rq);
3291 * Tell update_rq_clock() that we've just updated,
3292 * so we don't do microscopic update in schedule()
3293 * and double the fastpath cost.
3295 rq->skip_clock_update = 1;
3301 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3303 struct sched_entity *se = &p->se;
3305 /* throttled hierarchies are not runnable */
3306 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3309 /* Tell the scheduler that we'd really like pse to run next. */
3312 yield_task_fair(rq);
3318 /**************************************************
3319 * Fair scheduling class load-balancing methods:
3322 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
3324 #define LBF_ALL_PINNED 0x01
3325 #define LBF_NEED_BREAK 0x02
3326 #define LBF_SOME_PINNED 0x04
3329 struct sched_domain *sd;
3337 struct cpumask *dst_grpmask;
3339 enum cpu_idle_type idle;
3341 /* The set of CPUs under consideration for load-balancing */
3342 struct cpumask *cpus;
3346 struct list_head *tasks;
3349 unsigned int loop_break;
3350 unsigned int loop_max;
3352 struct rq * (*find_busiest_queue)(struct lb_env *,
3353 struct sched_group *);
3357 * move_task - move a task from one runqueue to another runqueue.
3358 * Both runqueues must be locked.
3360 static void move_task(struct task_struct *p, struct lb_env *env)
3362 deactivate_task(env->src_rq, p, 0);
3363 set_task_cpu(p, env->dst_cpu);
3364 activate_task(env->dst_rq, p, 0);
3365 check_preempt_curr(env->dst_rq, p, 0);
3368 static int task_numa_hot(struct task_struct *p, struct lb_env *env)
3370 int from_dist, to_dist;
3371 int node = tsk_home_node(p);
3373 if (!sched_feat_numa(NUMA_HOT) || node == -1)
3374 return 0; /* no node preference */
3376 from_dist = node_distance(cpu_to_node(env->src_cpu), node);
3377 to_dist = node_distance(cpu_to_node(env->dst_cpu), node);
3379 if (to_dist < from_dist)
3380 return 0; /* getting closer is ok */
3382 return 1; /* stick to where we are */
3386 * Is this task likely cache-hot:
3389 task_hot(struct task_struct *p, struct lb_env *env)
3393 if (p->sched_class != &fair_sched_class)
3396 if (unlikely(p->policy == SCHED_IDLE))
3400 * Buddy candidates are cache hot:
3402 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
3403 (&p->se == cfs_rq_of(&p->se)->next ||
3404 &p->se == cfs_rq_of(&p->se)->last))
3407 if (sysctl_sched_migration_cost == -1)
3409 if (sysctl_sched_migration_cost == 0)
3412 delta = env->src_rq->clock_task - p->se.exec_start;
3414 return delta < (s64)sysctl_sched_migration_cost;
3418 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3421 int can_migrate_task(struct task_struct *p, struct lb_env *env)
3423 int tsk_cache_hot = 0;
3425 * We do not migrate tasks that are:
3426 * 1) running (obviously), or
3427 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3428 * 3) are cache-hot on their current CPU.
3430 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
3433 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
3436 * Remember if this task can be migrated to any other cpu in
3437 * our sched_group. We may want to revisit it if we couldn't
3438 * meet load balance goals by pulling other tasks on src_cpu.
3440 * Also avoid computing new_dst_cpu if we have already computed
3441 * one in current iteration.
3443 if (!env->dst_grpmask || (env->flags & LBF_SOME_PINNED))
3446 new_dst_cpu = cpumask_first_and(env->dst_grpmask,
3447 tsk_cpus_allowed(p));
3448 if (new_dst_cpu < nr_cpu_ids) {
3449 env->flags |= LBF_SOME_PINNED;
3450 env->new_dst_cpu = new_dst_cpu;
3455 /* Record that we found atleast one task that could run on dst_cpu */
3456 env->flags &= ~LBF_ALL_PINNED;
3458 if (task_running(env->src_rq, p)) {
3459 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
3464 * Aggressive migration if:
3465 * 1) task is cache cold, or
3466 * 2) too many balance attempts have failed.
3469 tsk_cache_hot = task_hot(p, env);
3470 if (env->idle == CPU_NOT_IDLE)
3471 tsk_cache_hot |= task_numa_hot(p, env);
3472 if (!tsk_cache_hot ||
3473 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
3474 #ifdef CONFIG_SCHEDSTATS
3475 if (tsk_cache_hot) {
3476 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
3477 schedstat_inc(p, se.statistics.nr_forced_migrations);
3483 if (tsk_cache_hot) {
3484 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
3491 * move_one_task tries to move exactly one task from busiest to this_rq, as
3492 * part of active balancing operations within "domain".
3493 * Returns 1 if successful and 0 otherwise.
3495 * Called with both runqueues locked.
3497 static int __move_one_task(struct lb_env *env)
3499 struct task_struct *p, *n;
3501 list_for_each_entry_safe(p, n, env->tasks, se.group_node) {
3502 if (throttled_lb_pair(task_group(p), env->src_rq->cpu, env->dst_cpu))
3505 if (!can_migrate_task(p, env))
3510 * Right now, this is only the second place move_task()
3511 * is called, so we can safely collect move_task()
3512 * stats here rather than inside move_task().
3514 schedstat_inc(env->sd, lb_gained[env->idle]);
3520 static int move_one_task(struct lb_env *env)
3522 if (sched_feat_numa(NUMA_PULL)) {
3523 env->tasks = offnode_tasks(env->src_rq);
3524 if (__move_one_task(env))
3528 env->tasks = &env->src_rq->cfs_tasks;
3529 if (__move_one_task(env))
3535 static const unsigned int sched_nr_migrate_break = 32;
3538 * move_tasks tries to move up to imbalance weighted load from busiest to
3539 * this_rq, as part of a balancing operation within domain "sd".
3540 * Returns 1 if successful and 0 otherwise.
3542 * Called with both runqueues locked.
3544 static int move_tasks(struct lb_env *env)
3546 struct task_struct *p;
3550 if (env->imbalance <= 0)
3554 while (!list_empty(env->tasks)) {
3555 p = list_first_entry(env->tasks, struct task_struct, se.group_node);
3558 /* We've more or less seen every task there is, call it quits */
3559 if (env->loop > env->loop_max)
3562 /* take a breather every nr_migrate tasks */
3563 if (env->loop > env->loop_break) {
3564 env->loop_break += sched_nr_migrate_break;
3565 env->flags |= LBF_NEED_BREAK;
3569 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
3572 load = task_h_load(p);
3574 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
3577 if ((load / 2) > env->imbalance)
3580 if (!can_migrate_task(p, env))
3585 env->imbalance -= load;
3587 #ifdef CONFIG_PREEMPT
3589 * NEWIDLE balancing is a source of latency, so preemptible
3590 * kernels will stop after the first task is pulled to minimize
3591 * the critical section.
3593 if (env->idle == CPU_NEWLY_IDLE)
3598 * We only want to steal up to the prescribed amount of
3601 if (env->imbalance <= 0)
3606 list_move_tail(&p->se.group_node, env->tasks);
3609 if (env->tasks == offnode_tasks(env->src_rq)) {
3610 env->tasks = &env->src_rq->cfs_tasks;
3617 * Right now, this is one of only two places move_task() is called,
3618 * so we can safely collect move_task() stats here rather than
3619 * inside move_task().
3621 schedstat_add(env->sd, lb_gained[env->idle], pulled);
3626 #ifdef CONFIG_FAIR_GROUP_SCHED
3628 * update tg->load_weight by folding this cpu's load_avg
3630 static int update_shares_cpu(struct task_group *tg, int cpu)
3632 struct cfs_rq *cfs_rq;
3633 unsigned long flags;
3640 cfs_rq = tg->cfs_rq[cpu];
3642 raw_spin_lock_irqsave(&rq->lock, flags);
3644 update_rq_clock(rq);
3645 update_cfs_load(cfs_rq, 1);
3648 * We need to update shares after updating tg->load_weight in
3649 * order to adjust the weight of groups with long running tasks.
3651 update_cfs_shares(cfs_rq);
3653 raw_spin_unlock_irqrestore(&rq->lock, flags);
3658 static void update_shares(int cpu)
3660 struct cfs_rq *cfs_rq;
3661 struct rq *rq = cpu_rq(cpu);
3665 * Iterates the task_group tree in a bottom up fashion, see
3666 * list_add_leaf_cfs_rq() for details.
3668 for_each_leaf_cfs_rq(rq, cfs_rq) {
3669 /* throttled entities do not contribute to load */
3670 if (throttled_hierarchy(cfs_rq))
3673 update_shares_cpu(cfs_rq->tg, cpu);
3679 * Compute the cpu's hierarchical load factor for each task group.
3680 * This needs to be done in a top-down fashion because the load of a child
3681 * group is a fraction of its parents load.
3683 static int tg_load_down(struct task_group *tg, void *data)
3686 long cpu = (long)data;
3689 load = cpu_rq(cpu)->load.weight;
3691 load = tg->parent->cfs_rq[cpu]->h_load;
3692 load *= tg->se[cpu]->load.weight;
3693 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
3696 tg->cfs_rq[cpu]->h_load = load;
3701 static void update_h_load(long cpu)
3703 struct rq *rq = cpu_rq(cpu);
3704 unsigned long now = jiffies;
3706 if (rq->h_load_throttle == now)
3709 rq->h_load_throttle = now;
3712 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
3716 static unsigned long task_h_load(struct task_struct *p)
3718 struct cfs_rq *cfs_rq = task_cfs_rq(p);
3721 load = p->se.load.weight;
3722 load = div_u64(load * cfs_rq->h_load, cfs_rq->load.weight + 1);
3727 static inline void update_shares(int cpu)
3731 static inline void update_h_load(long cpu)
3735 static unsigned long task_h_load(struct task_struct *p)
3737 return p->se.load.weight;
3741 /********** Helpers for find_busiest_group ************************/
3743 * sd_lb_stats - Structure to store the statistics of a sched_domain
3744 * during load balancing.
3746 struct sd_lb_stats {
3747 struct sched_group *busiest; /* Busiest group in this sd */
3748 struct sched_group *this; /* Local group in this sd */
3749 unsigned long total_load; /* Total load of all groups in sd */
3750 unsigned long total_pwr; /* Total power of all groups in sd */
3751 unsigned long avg_load; /* Average load across all groups in sd */
3753 /** Statistics of this group */
3754 unsigned long this_load;
3755 unsigned long this_load_per_task;
3756 unsigned long this_nr_running;
3757 unsigned long this_has_capacity;
3758 unsigned int this_idle_cpus;
3760 /* Statistics of the busiest group */
3761 unsigned int busiest_idle_cpus;
3762 unsigned long max_load;
3763 unsigned long busiest_load_per_task;
3764 unsigned long busiest_nr_running;
3765 unsigned long busiest_group_capacity;
3766 unsigned long busiest_has_capacity;
3767 unsigned int busiest_group_weight;
3769 int group_imb; /* Is there imbalance in this sd */
3770 #ifdef CONFIG_SCHED_NUMA
3771 struct sched_group *numa_group; /* group which has offnode_tasks */
3772 unsigned long numa_group_weight;
3773 unsigned long numa_group_running;
3775 unsigned long this_offnode_running;
3776 unsigned long this_onnode_running;
3781 * sg_lb_stats - stats of a sched_group required for load_balancing
3783 struct sg_lb_stats {
3784 unsigned long avg_load; /*Avg load across the CPUs of the group */
3785 unsigned long group_load; /* Total load over the CPUs of the group */
3786 unsigned long sum_nr_running; /* Nr tasks running in the group */
3787 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3788 unsigned long group_capacity;
3789 unsigned long idle_cpus;
3790 unsigned long group_weight;
3791 int group_imb; /* Is there an imbalance in the group ? */
3792 int group_has_capacity; /* Is there extra capacity in the group? */
3793 #ifdef CONFIG_SCHED_NUMA
3794 unsigned long numa_offnode_weight;
3795 unsigned long numa_offnode_running;
3796 unsigned long numa_onnode_running;
3801 * get_sd_load_idx - Obtain the load index for a given sched domain.
3802 * @sd: The sched_domain whose load_idx is to be obtained.
3803 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3805 static inline int get_sd_load_idx(struct sched_domain *sd,
3806 enum cpu_idle_type idle)
3812 load_idx = sd->busy_idx;
3815 case CPU_NEWLY_IDLE:
3816 load_idx = sd->newidle_idx;
3819 load_idx = sd->idle_idx;
3826 #ifdef CONFIG_SCHED_NUMA
3827 static inline void update_sg_numa_stats(struct sg_lb_stats *sgs, struct rq *rq)
3829 sgs->numa_offnode_weight += rq->offnode_weight;
3830 sgs->numa_offnode_running += rq->offnode_running;
3831 sgs->numa_onnode_running += rq->onnode_running;
3835 * Since the offnode lists are indiscriminate (they contain tasks for all other
3836 * nodes) it is impossible to say if there's any task on there that wants to
3837 * move towards the pulling cpu. Therefore select a random offnode list to pull
3838 * from such that eventually we'll try them all.
3840 * Select a random group that has offnode tasks as sds->numa_group
3842 static inline void update_sd_numa_stats(struct sched_domain *sd,
3843 struct sched_group *group, struct sd_lb_stats *sds,
3844 int local_group, struct sg_lb_stats *sgs)
3846 if (!(sd->flags & SD_NUMA))
3850 sds->this_offnode_running = sgs->numa_offnode_running;
3851 sds->this_onnode_running = sgs->numa_onnode_running;
3855 if (!sgs->numa_offnode_running)
3858 if (!sds->numa_group || pick_numa_rand(sd->span_weight / group->group_weight)) {
3859 sds->numa_group = group;
3860 sds->numa_group_weight = sgs->numa_offnode_weight;
3861 sds->numa_group_running = sgs->numa_offnode_running;
3866 * Pick a random queue from the group that has offnode tasks.
3868 static struct rq *find_busiest_numa_queue(struct lb_env *env,
3869 struct sched_group *group)
3871 struct rq *busiest = NULL, *rq;
3874 for_each_cpu_and(cpu, sched_group_cpus(group), env->cpus) {
3876 if (!rq->offnode_running)
3878 if (!busiest || pick_numa_rand(group->group_weight))
3886 * Called in case of no other imbalance, if there is a queue running offnode
3887 * tasksk we'll say we're imbalanced anyway to nudge these tasks towards their
3890 static inline int check_numa_busiest_group(struct lb_env *env, struct sd_lb_stats *sds)
3892 if (!sched_feat(NUMA_PULL_BIAS))
3895 if (!sds->numa_group)
3899 * Only pull an offnode task home if we've got offnode or !numa tasks to trade for it.
3901 if (!sds->this_offnode_running &&
3902 !(sds->this_nr_running - sds->this_onnode_running - sds->this_offnode_running))
3905 env->imbalance = sds->numa_group_weight / sds->numa_group_running;
3906 sds->busiest = sds->numa_group;
3907 env->find_busiest_queue = find_busiest_numa_queue;
3911 static inline bool need_active_numa_balance(struct lb_env *env)
3913 return env->find_busiest_queue == find_busiest_numa_queue &&
3914 env->src_rq->offnode_running == 1 &&
3915 env->src_rq->nr_running == 1;
3918 #else /* CONFIG_SCHED_NUMA */
3920 static inline void update_sg_numa_stats(struct sg_lb_stats *sgs, struct rq *rq)
3924 static inline void update_sd_numa_stats(struct sched_domain *sd,
3925 struct sched_group *group, struct sd_lb_stats *sds,
3926 int local_group, struct sg_lb_stats *sgs)
3930 static inline int check_numa_busiest_group(struct lb_env *env, struct sd_lb_stats *sds)
3935 static inline bool need_active_numa_balance(struct lb_env *env)
3939 #endif /* CONFIG_SCHED_NUMA */
3941 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3943 return SCHED_POWER_SCALE;
3946 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3948 return default_scale_freq_power(sd, cpu);
3951 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3953 unsigned long weight = sd->span_weight;
3954 unsigned long smt_gain = sd->smt_gain;
3961 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3963 return default_scale_smt_power(sd, cpu);
3966 unsigned long scale_rt_power(int cpu)
3968 struct rq *rq = cpu_rq(cpu);
3969 u64 total, available, age_stamp, avg;
3972 * Since we're reading these variables without serialization make sure
3973 * we read them once before doing sanity checks on them.
3975 age_stamp = ACCESS_ONCE(rq->age_stamp);
3976 avg = ACCESS_ONCE(rq->rt_avg);
3978 total = sched_avg_period() + (rq->clock - age_stamp);
3980 if (unlikely(total < avg)) {
3981 /* Ensures that power won't end up being negative */
3984 available = total - avg;
3987 if (unlikely((s64)total < SCHED_POWER_SCALE))
3988 total = SCHED_POWER_SCALE;
3990 total >>= SCHED_POWER_SHIFT;
3992 return div_u64(available, total);
3995 static void update_cpu_power(struct sched_domain *sd, int cpu)
3997 unsigned long weight = sd->span_weight;
3998 unsigned long power = SCHED_POWER_SCALE;
3999 struct sched_group *sdg = sd->groups;
4001 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
4002 if (sched_feat(ARCH_POWER))
4003 power *= arch_scale_smt_power(sd, cpu);
4005 power *= default_scale_smt_power(sd, cpu);
4007 power >>= SCHED_POWER_SHIFT;
4010 sdg->sgp->power_orig = power;
4012 if (sched_feat(ARCH_POWER))
4013 power *= arch_scale_freq_power(sd, cpu);
4015 power *= default_scale_freq_power(sd, cpu);
4017 power >>= SCHED_POWER_SHIFT;
4019 power *= scale_rt_power(cpu);
4020 power >>= SCHED_POWER_SHIFT;
4025 cpu_rq(cpu)->cpu_power = power;
4026 sdg->sgp->power = power;
4029 void update_group_power(struct sched_domain *sd, int cpu)
4031 struct sched_domain *child = sd->child;
4032 struct sched_group *group, *sdg = sd->groups;
4033 unsigned long power;
4034 unsigned long interval;
4036 interval = msecs_to_jiffies(sd->balance_interval);
4037 interval = clamp(interval, 1UL, max_load_balance_interval);
4038 sdg->sgp->next_update = jiffies + interval;
4041 update_cpu_power(sd, cpu);
4047 if (child->flags & SD_OVERLAP) {
4049 * SD_OVERLAP domains cannot assume that child groups
4050 * span the current group.
4053 for_each_cpu(cpu, sched_group_cpus(sdg))
4054 power += power_of(cpu);
4057 * !SD_OVERLAP domains can assume that child groups
4058 * span the current group.
4061 group = child->groups;
4063 power += group->sgp->power;
4064 group = group->next;
4065 } while (group != child->groups);
4068 sdg->sgp->power_orig = sdg->sgp->power = power;
4072 * Try and fix up capacity for tiny siblings, this is needed when
4073 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4074 * which on its own isn't powerful enough.
4076 * See update_sd_pick_busiest() and check_asym_packing().
4079 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
4082 * Only siblings can have significantly less than SCHED_POWER_SCALE
4084 if (!(sd->flags & SD_SHARE_CPUPOWER))
4088 * If ~90% of the cpu_power is still there, we're good.
4090 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
4097 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4098 * @env: The load balancing environment.
4099 * @group: sched_group whose statistics are to be updated.
4100 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4101 * @local_group: Does group contain this_cpu.
4102 * @balance: Should we balance.
4103 * @sgs: variable to hold the statistics for this group.
4105 static inline void update_sg_lb_stats(struct lb_env *env,
4106 struct sched_group *group, int load_idx,
4107 int local_group, int *balance, struct sg_lb_stats *sgs)
4109 unsigned long nr_running, max_nr_running, min_nr_running;
4110 unsigned long load, max_cpu_load, min_cpu_load;
4111 unsigned int balance_cpu = -1, first_idle_cpu = 0;
4112 unsigned long avg_load_per_task = 0;
4116 balance_cpu = group_balance_cpu(group);
4118 /* Tally up the load of all CPUs in the group */
4120 min_cpu_load = ~0UL;
4122 min_nr_running = ~0UL;
4124 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
4125 struct rq *rq = cpu_rq(i);
4127 nr_running = rq->nr_running;
4129 /* Bias balancing toward cpus of our domain */
4131 if (idle_cpu(i) && !first_idle_cpu &&
4132 cpumask_test_cpu(i, sched_group_mask(group))) {
4137 load = target_load(i, load_idx);
4139 load = source_load(i, load_idx);
4140 if (load > max_cpu_load)
4141 max_cpu_load = load;
4142 if (min_cpu_load > load)
4143 min_cpu_load = load;
4145 if (nr_running > max_nr_running)
4146 max_nr_running = nr_running;
4147 if (min_nr_running > nr_running)
4148 min_nr_running = nr_running;
4151 sgs->group_load += load;
4152 sgs->sum_nr_running += nr_running;
4153 sgs->sum_weighted_load += weighted_cpuload(i);
4157 update_sg_numa_stats(sgs, rq);
4161 * First idle cpu or the first cpu(busiest) in this sched group
4162 * is eligible for doing load balancing at this and above
4163 * domains. In the newly idle case, we will allow all the cpu's
4164 * to do the newly idle load balance.
4167 if (env->idle != CPU_NEWLY_IDLE) {
4168 if (balance_cpu != env->dst_cpu) {
4172 update_group_power(env->sd, env->dst_cpu);
4173 } else if (time_after_eq(jiffies, group->sgp->next_update))
4174 update_group_power(env->sd, env->dst_cpu);
4177 /* Adjust by relative CPU power of the group */
4178 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
4181 * Consider the group unbalanced when the imbalance is larger
4182 * than the average weight of a task.
4184 * APZ: with cgroup the avg task weight can vary wildly and
4185 * might not be a suitable number - should we keep a
4186 * normalized nr_running number somewhere that negates
4189 if (sgs->sum_nr_running)
4190 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
4192 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task &&
4193 (max_nr_running - min_nr_running) > 1)
4196 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
4198 if (!sgs->group_capacity)
4199 sgs->group_capacity = fix_small_capacity(env->sd, group);
4200 sgs->group_weight = group->group_weight;
4202 if (sgs->group_capacity > sgs->sum_nr_running)
4203 sgs->group_has_capacity = 1;
4207 * update_sd_pick_busiest - return 1 on busiest group
4208 * @env: The load balancing environment.
4209 * @sds: sched_domain statistics
4210 * @sg: sched_group candidate to be checked for being the busiest
4211 * @sgs: sched_group statistics
4213 * Determine if @sg is a busier group than the previously selected
4216 static bool update_sd_pick_busiest(struct lb_env *env,
4217 struct sd_lb_stats *sds,
4218 struct sched_group *sg,
4219 struct sg_lb_stats *sgs)
4221 if (sgs->avg_load <= sds->max_load)
4224 if (sgs->sum_nr_running > sgs->group_capacity)
4231 * ASYM_PACKING needs to move all the work to the lowest
4232 * numbered CPUs in the group, therefore mark all groups
4233 * higher than ourself as busy.
4235 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
4236 env->dst_cpu < group_first_cpu(sg)) {
4240 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
4248 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
4249 * @env: The load balancing environment.
4250 * @balance: Should we balance.
4251 * @sds: variable to hold the statistics for this sched_domain.
4253 static inline void update_sd_lb_stats(struct lb_env *env,
4254 int *balance, struct sd_lb_stats *sds)
4256 struct sched_domain *child = env->sd->child;
4257 struct sched_group *sg = env->sd->groups;
4258 struct sg_lb_stats sgs;
4259 int load_idx, prefer_sibling = 0;
4261 if (child && child->flags & SD_PREFER_SIBLING)
4264 load_idx = get_sd_load_idx(env->sd, env->idle);
4269 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
4270 memset(&sgs, 0, sizeof(sgs));
4271 update_sg_lb_stats(env, sg, load_idx, local_group, balance, &sgs);
4273 if (local_group && !(*balance))
4276 sds->total_load += sgs.group_load;
4277 sds->total_pwr += sg->sgp->power;
4280 * In case the child domain prefers tasks go to siblings
4281 * first, lower the sg capacity to one so that we'll try
4282 * and move all the excess tasks away. We lower the capacity
4283 * of a group only if the local group has the capacity to fit
4284 * these excess tasks, i.e. nr_running < group_capacity. The
4285 * extra check prevents the case where you always pull from the
4286 * heaviest group when it is already under-utilized (possible
4287 * with a large weight task outweighs the tasks on the system).
4289 if (prefer_sibling && !local_group && sds->this_has_capacity)
4290 sgs.group_capacity = min(sgs.group_capacity, 1UL);
4293 sds->this_load = sgs.avg_load;
4295 sds->this_nr_running = sgs.sum_nr_running;
4296 sds->this_load_per_task = sgs.sum_weighted_load;
4297 sds->this_has_capacity = sgs.group_has_capacity;
4298 sds->this_idle_cpus = sgs.idle_cpus;
4299 } else if (update_sd_pick_busiest(env, sds, sg, &sgs)) {
4300 sds->max_load = sgs.avg_load;
4302 sds->busiest_nr_running = sgs.sum_nr_running;
4303 sds->busiest_idle_cpus = sgs.idle_cpus;
4304 sds->busiest_group_capacity = sgs.group_capacity;
4305 sds->busiest_load_per_task = sgs.sum_weighted_load;
4306 sds->busiest_has_capacity = sgs.group_has_capacity;
4307 sds->busiest_group_weight = sgs.group_weight;
4308 sds->group_imb = sgs.group_imb;
4311 update_sd_numa_stats(env->sd, sg, sds, local_group, &sgs);
4314 } while (sg != env->sd->groups);
4318 * check_asym_packing - Check to see if the group is packed into the
4321 * This is primarily intended to used at the sibling level. Some
4322 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4323 * case of POWER7, it can move to lower SMT modes only when higher
4324 * threads are idle. When in lower SMT modes, the threads will
4325 * perform better since they share less core resources. Hence when we
4326 * have idle threads, we want them to be the higher ones.
4328 * This packing function is run on idle threads. It checks to see if
4329 * the busiest CPU in this domain (core in the P7 case) has a higher
4330 * CPU number than the packing function is being run on. Here we are
4331 * assuming lower CPU number will be equivalent to lower a SMT thread
4334 * Returns 1 when packing is required and a task should be moved to
4335 * this CPU. The amount of the imbalance is returned in *imbalance.
4337 * @env: The load balancing environment.
4338 * @sds: Statistics of the sched_domain which is to be packed
4340 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
4344 if (!(env->sd->flags & SD_ASYM_PACKING))
4350 busiest_cpu = group_first_cpu(sds->busiest);
4351 if (env->dst_cpu > busiest_cpu)
4354 env->imbalance = DIV_ROUND_CLOSEST(
4355 sds->max_load * sds->busiest->sgp->power, SCHED_POWER_SCALE);
4361 * fix_small_imbalance - Calculate the minor imbalance that exists
4362 * amongst the groups of a sched_domain, during
4364 * @env: The load balancing environment.
4365 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4368 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4370 unsigned long tmp, pwr_now = 0, pwr_move = 0;
4371 unsigned int imbn = 2;
4372 unsigned long scaled_busy_load_per_task;
4374 if (sds->this_nr_running) {
4375 sds->this_load_per_task /= sds->this_nr_running;
4376 if (sds->busiest_load_per_task >
4377 sds->this_load_per_task)
4380 sds->this_load_per_task =
4381 cpu_avg_load_per_task(env->dst_cpu);
4384 scaled_busy_load_per_task = sds->busiest_load_per_task
4385 * SCHED_POWER_SCALE;
4386 scaled_busy_load_per_task /= sds->busiest->sgp->power;
4388 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
4389 (scaled_busy_load_per_task * imbn)) {
4390 env->imbalance = sds->busiest_load_per_task;
4395 * OK, we don't have enough imbalance to justify moving tasks,
4396 * however we may be able to increase total CPU power used by
4400 pwr_now += sds->busiest->sgp->power *
4401 min(sds->busiest_load_per_task, sds->max_load);
4402 pwr_now += sds->this->sgp->power *
4403 min(sds->this_load_per_task, sds->this_load);
4404 pwr_now /= SCHED_POWER_SCALE;
4406 /* Amount of load we'd subtract */
4407 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4408 sds->busiest->sgp->power;
4409 if (sds->max_load > tmp)
4410 pwr_move += sds->busiest->sgp->power *
4411 min(sds->busiest_load_per_task, sds->max_load - tmp);
4413 /* Amount of load we'd add */
4414 if (sds->max_load * sds->busiest->sgp->power <
4415 sds->busiest_load_per_task * SCHED_POWER_SCALE)
4416 tmp = (sds->max_load * sds->busiest->sgp->power) /
4417 sds->this->sgp->power;
4419 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4420 sds->this->sgp->power;
4421 pwr_move += sds->this->sgp->power *
4422 min(sds->this_load_per_task, sds->this_load + tmp);
4423 pwr_move /= SCHED_POWER_SCALE;
4425 /* Move if we gain throughput */
4426 if (pwr_move > pwr_now)
4427 env->imbalance = sds->busiest_load_per_task;
4431 * calculate_imbalance - Calculate the amount of imbalance present within the
4432 * groups of a given sched_domain during load balance.
4433 * @env: load balance environment
4434 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4436 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4438 unsigned long max_pull, load_above_capacity = ~0UL;
4440 sds->busiest_load_per_task /= sds->busiest_nr_running;
4441 if (sds->group_imb) {
4442 sds->busiest_load_per_task =
4443 min(sds->busiest_load_per_task, sds->avg_load);
4447 * In the presence of smp nice balancing, certain scenarios can have
4448 * max load less than avg load(as we skip the groups at or below
4449 * its cpu_power, while calculating max_load..)
4451 if (sds->max_load < sds->avg_load) {
4453 return fix_small_imbalance(env, sds);
4456 if (!sds->group_imb) {
4458 * Don't want to pull so many tasks that a group would go idle.
4460 load_above_capacity = (sds->busiest_nr_running -
4461 sds->busiest_group_capacity);
4463 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
4465 load_above_capacity /= sds->busiest->sgp->power;
4469 * We're trying to get all the cpus to the average_load, so we don't
4470 * want to push ourselves above the average load, nor do we wish to
4471 * reduce the max loaded cpu below the average load. At the same time,
4472 * we also don't want to reduce the group load below the group capacity
4473 * (so that we can implement power-savings policies etc). Thus we look
4474 * for the minimum possible imbalance.
4475 * Be careful of negative numbers as they'll appear as very large values
4476 * with unsigned longs.
4478 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
4480 /* How much load to actually move to equalise the imbalance */
4481 env->imbalance = min(max_pull * sds->busiest->sgp->power,
4482 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
4483 / SCHED_POWER_SCALE;
4486 * if *imbalance is less than the average load per runnable task
4487 * there is no guarantee that any tasks will be moved so we'll have
4488 * a think about bumping its value to force at least one task to be
4491 if (env->imbalance < sds->busiest_load_per_task)
4492 return fix_small_imbalance(env, sds);
4496 /******* find_busiest_group() helpers end here *********************/
4499 * find_busiest_group - Returns the busiest group within the sched_domain
4500 * if there is an imbalance. If there isn't an imbalance, and
4501 * the user has opted for power-savings, it returns a group whose
4502 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4503 * such a group exists.
4505 * Also calculates the amount of weighted load which should be moved
4506 * to restore balance.
4508 * @env: The load balancing environment.
4509 * @balance: Pointer to a variable indicating if this_cpu
4510 * is the appropriate cpu to perform load balancing at this_level.
4512 * Returns: - the busiest group if imbalance exists.
4513 * - If no imbalance and user has opted for power-savings balance,
4514 * return the least loaded group whose CPUs can be
4515 * put to idle by rebalancing its tasks onto our group.
4517 static struct sched_group *
4518 find_busiest_group(struct lb_env *env, int *balance)
4520 struct sd_lb_stats sds;
4522 memset(&sds, 0, sizeof(sds));
4525 * Compute the various statistics relavent for load balancing at
4528 update_sd_lb_stats(env, balance, &sds);
4531 * this_cpu is not the appropriate cpu to perform load balancing at
4537 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
4538 check_asym_packing(env, &sds))
4541 /* There is no busy sibling group to pull tasks from */
4542 if (!sds.busiest || sds.busiest_nr_running == 0)
4545 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
4548 * If the busiest group is imbalanced the below checks don't
4549 * work because they assumes all things are equal, which typically
4550 * isn't true due to cpus_allowed constraints and the like.
4555 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
4556 if (env->idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
4557 !sds.busiest_has_capacity)
4561 * If the local group is more busy than the selected busiest group
4562 * don't try and pull any tasks.
4564 if (sds.this_load >= sds.max_load)
4568 * Don't pull any tasks if this group is already above the domain
4571 if (sds.this_load >= sds.avg_load)
4574 if (env->idle == CPU_IDLE) {
4576 * This cpu is idle. If the busiest group load doesn't
4577 * have more tasks than the number of available cpu's and
4578 * there is no imbalance between this and busiest group
4579 * wrt to idle cpu's, it is balanced.
4581 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
4582 sds.busiest_nr_running <= sds.busiest_group_weight)
4586 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
4587 * imbalance_pct to be conservative.
4589 if (100 * sds.max_load <= env->sd->imbalance_pct * sds.this_load)
4594 /* Looks like there is an imbalance. Compute it */
4595 calculate_imbalance(env, &sds);
4599 if (check_numa_busiest_group(env, &sds))
4608 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4610 static struct rq *find_busiest_queue(struct lb_env *env,
4611 struct sched_group *group)
4613 struct rq *busiest = NULL, *rq;
4614 unsigned long max_load = 0;
4617 for_each_cpu(i, sched_group_cpus(group)) {
4618 unsigned long power = power_of(i);
4619 unsigned long capacity = DIV_ROUND_CLOSEST(power,
4624 capacity = fix_small_capacity(env->sd, group);
4626 if (!cpumask_test_cpu(i, env->cpus))
4630 wl = weighted_cpuload(i);
4633 * When comparing with imbalance, use weighted_cpuload()
4634 * which is not scaled with the cpu power.
4636 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
4640 * For the load comparisons with the other cpu's, consider
4641 * the weighted_cpuload() scaled with the cpu power, so that
4642 * the load can be moved away from the cpu that is potentially
4643 * running at a lower capacity.
4645 wl = (wl * SCHED_POWER_SCALE) / power;
4647 if (wl > max_load) {
4657 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4658 * so long as it is large enough.
4660 #define MAX_PINNED_INTERVAL 512
4662 /* Working cpumask for load_balance and load_balance_newidle. */
4663 DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4665 static int need_active_balance(struct lb_env *env)
4667 struct sched_domain *sd = env->sd;
4669 if (env->idle == CPU_NEWLY_IDLE) {
4672 * ASYM_PACKING needs to force migrate tasks from busy but
4673 * higher numbered CPUs in order to pack all tasks in the
4674 * lowest numbered CPUs.
4676 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
4680 if (need_active_numa_balance(env))
4683 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
4686 static int active_load_balance_cpu_stop(void *data);
4689 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4690 * tasks if there is an imbalance.
4692 static int load_balance(int this_cpu, struct rq *this_rq,
4693 struct sched_domain *sd, enum cpu_idle_type idle,
4696 int ld_moved, cur_ld_moved, active_balance = 0;
4697 int lb_iterations, max_lb_iterations;
4698 struct sched_group *group;
4700 unsigned long flags;
4701 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4703 struct lb_env env = {
4705 .dst_cpu = this_cpu,
4707 .dst_grpmask = sched_group_cpus(sd->groups),
4709 .loop_break = sched_nr_migrate_break,
4711 .find_busiest_queue = find_busiest_queue,
4714 cpumask_copy(cpus, cpu_active_mask);
4715 max_lb_iterations = cpumask_weight(env.dst_grpmask);
4717 schedstat_inc(sd, lb_count[idle]);
4720 group = find_busiest_group(&env, balance);
4726 schedstat_inc(sd, lb_nobusyg[idle]);
4730 busiest = env.find_busiest_queue(&env, group);
4732 schedstat_inc(sd, lb_nobusyq[idle]);
4735 env.src_rq = busiest;
4736 env.src_cpu = busiest->cpu;
4738 BUG_ON(busiest == env.dst_rq);
4740 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
4744 if (busiest->nr_running > 1) {
4746 * Attempt to move tasks. If find_busiest_group has found
4747 * an imbalance but busiest->nr_running <= 1, the group is
4748 * still unbalanced. ld_moved simply stays zero, so it is
4749 * correctly treated as an imbalance.
4751 env.flags |= LBF_ALL_PINNED;
4752 env.src_cpu = busiest->cpu;
4753 env.src_rq = busiest;
4754 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
4755 if (sched_feat_numa(NUMA_PULL))
4756 env.tasks = offnode_tasks(busiest);
4758 env.tasks = &busiest->cfs_tasks;
4760 update_h_load(env.src_cpu);
4762 local_irq_save(flags);
4763 double_rq_lock(env.dst_rq, busiest);
4766 * cur_ld_moved - load moved in current iteration
4767 * ld_moved - cumulative load moved across iterations
4769 cur_ld_moved = move_tasks(&env);
4770 ld_moved += cur_ld_moved;
4771 double_rq_unlock(env.dst_rq, busiest);
4772 local_irq_restore(flags);
4774 if (env.flags & LBF_NEED_BREAK) {
4775 env.flags &= ~LBF_NEED_BREAK;
4780 * some other cpu did the load balance for us.
4782 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
4783 resched_cpu(env.dst_cpu);
4786 * Revisit (affine) tasks on src_cpu that couldn't be moved to
4787 * us and move them to an alternate dst_cpu in our sched_group
4788 * where they can run. The upper limit on how many times we
4789 * iterate on same src_cpu is dependent on number of cpus in our
4792 * This changes load balance semantics a bit on who can move
4793 * load to a given_cpu. In addition to the given_cpu itself
4794 * (or a ilb_cpu acting on its behalf where given_cpu is
4795 * nohz-idle), we now have balance_cpu in a position to move
4796 * load to given_cpu. In rare situations, this may cause
4797 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
4798 * _independently_ and at _same_ time to move some load to
4799 * given_cpu) causing exceess load to be moved to given_cpu.
4800 * This however should not happen so much in practice and
4801 * moreover subsequent load balance cycles should correct the
4802 * excess load moved.
4804 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0 &&
4805 lb_iterations++ < max_lb_iterations) {
4807 env.dst_rq = cpu_rq(env.new_dst_cpu);
4808 env.dst_cpu = env.new_dst_cpu;
4809 env.flags &= ~LBF_SOME_PINNED;
4811 env.loop_break = sched_nr_migrate_break;
4813 * Go back to "more_balance" rather than "redo" since we
4814 * need to continue with same src_cpu.
4819 /* All tasks on this runqueue were pinned by CPU affinity */
4820 if (unlikely(env.flags & LBF_ALL_PINNED)) {
4821 cpumask_clear_cpu(cpu_of(busiest), cpus);
4822 if (!cpumask_empty(cpus)) {
4824 env.loop_break = sched_nr_migrate_break;
4832 schedstat_inc(sd, lb_failed[idle]);
4834 * Increment the failure counter only on periodic balance.
4835 * We do not want newidle balance, which can be very
4836 * frequent, pollute the failure counter causing
4837 * excessive cache_hot migrations and active balances.
4839 if (idle != CPU_NEWLY_IDLE)
4840 sd->nr_balance_failed++;
4842 if (need_active_balance(&env)) {
4843 raw_spin_lock_irqsave(&busiest->lock, flags);
4845 /* don't kick the active_load_balance_cpu_stop,
4846 * if the curr task on busiest cpu can't be
4849 if (!cpumask_test_cpu(this_cpu,
4850 tsk_cpus_allowed(busiest->curr))) {
4851 raw_spin_unlock_irqrestore(&busiest->lock,
4853 env.flags |= LBF_ALL_PINNED;
4854 goto out_one_pinned;
4858 * ->active_balance synchronizes accesses to
4859 * ->active_balance_work. Once set, it's cleared
4860 * only after active load balance is finished.
4862 if (!busiest->active_balance) {
4863 busiest->active_balance = 1;
4864 busiest->push_cpu = this_cpu;
4867 raw_spin_unlock_irqrestore(&busiest->lock, flags);
4869 if (active_balance) {
4870 stop_one_cpu_nowait(cpu_of(busiest),
4871 active_load_balance_cpu_stop, busiest,
4872 &busiest->active_balance_work);
4876 * We've kicked active balancing, reset the failure
4879 sd->nr_balance_failed = sd->cache_nice_tries+1;
4882 sd->nr_balance_failed = 0;
4884 if (likely(!active_balance)) {
4885 /* We were unbalanced, so reset the balancing interval */
4886 sd->balance_interval = sd->min_interval;
4889 * If we've begun active balancing, start to back off. This
4890 * case may not be covered by the all_pinned logic if there
4891 * is only 1 task on the busy runqueue (because we don't call
4894 if (sd->balance_interval < sd->max_interval)
4895 sd->balance_interval *= 2;
4901 schedstat_inc(sd, lb_balanced[idle]);
4903 sd->nr_balance_failed = 0;
4906 /* tune up the balancing interval */
4907 if (((env.flags & LBF_ALL_PINNED) &&
4908 sd->balance_interval < MAX_PINNED_INTERVAL) ||
4909 (sd->balance_interval < sd->max_interval))
4910 sd->balance_interval *= 2;
4918 * idle_balance is called by schedule() if this_cpu is about to become
4919 * idle. Attempts to pull tasks from other CPUs.
4921 void idle_balance(int this_cpu, struct rq *this_rq)
4923 struct sched_domain *sd;
4924 int pulled_task = 0;
4925 unsigned long next_balance = jiffies + HZ;
4927 this_rq->idle_stamp = this_rq->clock;
4929 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4933 * Drop the rq->lock, but keep IRQ/preempt disabled.
4935 raw_spin_unlock(&this_rq->lock);
4937 update_shares(this_cpu);
4939 for_each_domain(this_cpu, sd) {
4940 unsigned long interval;
4943 if (!(sd->flags & SD_LOAD_BALANCE))
4946 if (sd->flags & SD_BALANCE_NEWIDLE) {
4947 /* If we've pulled tasks over stop searching: */
4948 pulled_task = load_balance(this_cpu, this_rq,
4949 sd, CPU_NEWLY_IDLE, &balance);
4952 interval = msecs_to_jiffies(sd->balance_interval);
4953 if (time_after(next_balance, sd->last_balance + interval))
4954 next_balance = sd->last_balance + interval;
4956 this_rq->idle_stamp = 0;
4962 raw_spin_lock(&this_rq->lock);
4964 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4966 * We are going idle. next_balance may be set based on
4967 * a busy processor. So reset next_balance.
4969 this_rq->next_balance = next_balance;
4974 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
4975 * running tasks off the busiest CPU onto idle CPUs. It requires at
4976 * least 1 task to be running on each physical CPU where possible, and
4977 * avoids physical / logical imbalances.
4979 static int active_load_balance_cpu_stop(void *data)
4981 struct rq *busiest_rq = data;
4982 int busiest_cpu = cpu_of(busiest_rq);
4983 int target_cpu = busiest_rq->push_cpu;
4984 struct rq *target_rq = cpu_rq(target_cpu);
4985 struct sched_domain *sd;
4987 raw_spin_lock_irq(&busiest_rq->lock);
4989 /* make sure the requested cpu hasn't gone down in the meantime */
4990 if (unlikely(busiest_cpu != smp_processor_id() ||
4991 !busiest_rq->active_balance))
4994 /* Is there any task to move? */
4995 if (busiest_rq->nr_running <= 1)
4999 * This condition is "impossible", if it occurs
5000 * we need to fix it. Originally reported by
5001 * Bjorn Helgaas on a 128-cpu setup.
5003 BUG_ON(busiest_rq == target_rq);
5005 /* move a task from busiest_rq to target_rq */
5006 double_lock_balance(busiest_rq, target_rq);
5008 /* Search for an sd spanning us and the target CPU. */
5010 for_each_domain(target_cpu, sd) {
5011 if ((sd->flags & SD_LOAD_BALANCE) &&
5012 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
5017 struct lb_env env = {
5019 .dst_cpu = target_cpu,
5020 .dst_rq = target_rq,
5021 .src_cpu = busiest_rq->cpu,
5022 .src_rq = busiest_rq,
5026 schedstat_inc(sd, alb_count);
5028 if (move_one_task(&env))
5029 schedstat_inc(sd, alb_pushed);
5031 schedstat_inc(sd, alb_failed);
5034 double_unlock_balance(busiest_rq, target_rq);
5036 busiest_rq->active_balance = 0;
5037 raw_spin_unlock_irq(&busiest_rq->lock);
5043 * idle load balancing details
5044 * - When one of the busy CPUs notice that there may be an idle rebalancing
5045 * needed, they will kick the idle load balancer, which then does idle
5046 * load balancing for all the idle CPUs.
5049 cpumask_var_t idle_cpus_mask;
5051 unsigned long next_balance; /* in jiffy units */
5052 } nohz ____cacheline_aligned;
5054 static inline int find_new_ilb(int call_cpu)
5056 int ilb = cpumask_first(nohz.idle_cpus_mask);
5058 if (ilb < nr_cpu_ids && idle_cpu(ilb))
5065 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5066 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5067 * CPU (if there is one).
5069 static void nohz_balancer_kick(int cpu)
5073 nohz.next_balance++;
5075 ilb_cpu = find_new_ilb(cpu);
5077 if (ilb_cpu >= nr_cpu_ids)
5080 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
5083 * Use smp_send_reschedule() instead of resched_cpu().
5084 * This way we generate a sched IPI on the target cpu which
5085 * is idle. And the softirq performing nohz idle load balance
5086 * will be run before returning from the IPI.
5088 smp_send_reschedule(ilb_cpu);
5092 static inline void nohz_balance_exit_idle(int cpu)
5094 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5095 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5096 atomic_dec(&nohz.nr_cpus);
5097 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5101 static inline void set_cpu_sd_state_busy(void)
5103 struct sched_domain *sd;
5104 int cpu = smp_processor_id();
5106 if (!test_bit(NOHZ_IDLE, nohz_flags(cpu)))
5108 clear_bit(NOHZ_IDLE, nohz_flags(cpu));
5111 for_each_domain(cpu, sd)
5112 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
5116 void set_cpu_sd_state_idle(void)
5118 struct sched_domain *sd;
5119 int cpu = smp_processor_id();
5121 if (test_bit(NOHZ_IDLE, nohz_flags(cpu)))
5123 set_bit(NOHZ_IDLE, nohz_flags(cpu));
5126 for_each_domain(cpu, sd)
5127 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
5132 * This routine will record that the cpu is going idle with tick stopped.
5133 * This info will be used in performing idle load balancing in the future.
5135 void nohz_balance_enter_idle(int cpu)
5138 * If this cpu is going down, then nothing needs to be done.
5140 if (!cpu_active(cpu))
5143 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
5146 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
5147 atomic_inc(&nohz.nr_cpus);
5148 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5151 static int __cpuinit sched_ilb_notifier(struct notifier_block *nfb,
5152 unsigned long action, void *hcpu)
5154 switch (action & ~CPU_TASKS_FROZEN) {
5156 nohz_balance_exit_idle(smp_processor_id());
5164 static DEFINE_SPINLOCK(balancing);
5167 * Scale the max load_balance interval with the number of CPUs in the system.
5168 * This trades load-balance latency on larger machines for less cross talk.
5170 void update_max_interval(void)
5172 max_load_balance_interval = HZ*num_online_cpus()/10;
5176 * It checks each scheduling domain to see if it is due to be balanced,
5177 * and initiates a balancing operation if so.
5179 * Balancing parameters are set up in arch_init_sched_domains.
5181 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
5184 struct rq *rq = cpu_rq(cpu);
5185 unsigned long interval;
5186 struct sched_domain *sd;
5187 /* Earliest time when we have to do rebalance again */
5188 unsigned long next_balance = jiffies + 60*HZ;
5189 int update_next_balance = 0;
5195 for_each_domain(cpu, sd) {
5196 if (!(sd->flags & SD_LOAD_BALANCE))
5199 interval = sd->balance_interval;
5200 if (idle != CPU_IDLE)
5201 interval *= sd->busy_factor;
5203 /* scale ms to jiffies */
5204 interval = msecs_to_jiffies(interval);
5205 interval = clamp(interval, 1UL, max_load_balance_interval);
5207 need_serialize = sd->flags & SD_SERIALIZE;
5209 if (need_serialize) {
5210 if (!spin_trylock(&balancing))
5214 if (time_after_eq(jiffies, sd->last_balance + interval)) {
5215 if (load_balance(cpu, rq, sd, idle, &balance)) {
5217 * We've pulled tasks over so either we're no
5220 idle = CPU_NOT_IDLE;
5222 sd->last_balance = jiffies;
5225 spin_unlock(&balancing);
5227 if (time_after(next_balance, sd->last_balance + interval)) {
5228 next_balance = sd->last_balance + interval;
5229 update_next_balance = 1;
5233 * Stop the load balance at this level. There is another
5234 * CPU in our sched group which is doing load balancing more
5243 * next_balance will be updated only when there is a need.
5244 * When the cpu is attached to null domain for ex, it will not be
5247 if (likely(update_next_balance))
5248 rq->next_balance = next_balance;
5253 * In CONFIG_NO_HZ case, the idle balance kickee will do the
5254 * rebalancing for all the cpus for whom scheduler ticks are stopped.
5256 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
5258 struct rq *this_rq = cpu_rq(this_cpu);
5262 if (idle != CPU_IDLE ||
5263 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
5266 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
5267 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
5271 * If this cpu gets work to do, stop the load balancing
5272 * work being done for other cpus. Next load
5273 * balancing owner will pick it up.
5278 rq = cpu_rq(balance_cpu);
5280 raw_spin_lock_irq(&rq->lock);
5281 update_rq_clock(rq);
5282 update_idle_cpu_load(rq);
5283 raw_spin_unlock_irq(&rq->lock);
5285 rebalance_domains(balance_cpu, CPU_IDLE);
5287 if (time_after(this_rq->next_balance, rq->next_balance))
5288 this_rq->next_balance = rq->next_balance;
5290 nohz.next_balance = this_rq->next_balance;
5292 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
5296 * Current heuristic for kicking the idle load balancer in the presence
5297 * of an idle cpu is the system.
5298 * - This rq has more than one task.
5299 * - At any scheduler domain level, this cpu's scheduler group has multiple
5300 * busy cpu's exceeding the group's power.
5301 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
5302 * domain span are idle.
5304 static inline int nohz_kick_needed(struct rq *rq, int cpu)
5306 unsigned long now = jiffies;
5307 struct sched_domain *sd;
5309 if (unlikely(idle_cpu(cpu)))
5313 * We may be recently in ticked or tickless idle mode. At the first
5314 * busy tick after returning from idle, we will update the busy stats.
5316 set_cpu_sd_state_busy();
5317 nohz_balance_exit_idle(cpu);
5320 * None are in tickless mode and hence no need for NOHZ idle load
5323 if (likely(!atomic_read(&nohz.nr_cpus)))
5326 if (time_before(now, nohz.next_balance))
5329 if (rq->nr_running >= 2)
5333 for_each_domain(cpu, sd) {
5334 struct sched_group *sg = sd->groups;
5335 struct sched_group_power *sgp = sg->sgp;
5336 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
5338 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
5339 goto need_kick_unlock;
5341 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
5342 && (cpumask_first_and(nohz.idle_cpus_mask,
5343 sched_domain_span(sd)) < cpu))
5344 goto need_kick_unlock;
5346 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
5358 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
5362 * run_rebalance_domains is triggered when needed from the scheduler tick.
5363 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
5365 static void run_rebalance_domains(struct softirq_action *h)
5367 int this_cpu = smp_processor_id();
5368 struct rq *this_rq = cpu_rq(this_cpu);
5369 enum cpu_idle_type idle = this_rq->idle_balance ?
5370 CPU_IDLE : CPU_NOT_IDLE;
5372 rebalance_domains(this_cpu, idle);
5375 * If this cpu has a pending nohz_balance_kick, then do the
5376 * balancing on behalf of the other idle cpus whose ticks are
5379 nohz_idle_balance(this_cpu, idle);
5382 static inline int on_null_domain(int cpu)
5384 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
5388 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
5390 void trigger_load_balance(struct rq *rq, int cpu)
5392 /* Don't need to rebalance while attached to NULL domain */
5393 if (time_after_eq(jiffies, rq->next_balance) &&
5394 likely(!on_null_domain(cpu)))
5395 raise_softirq(SCHED_SOFTIRQ);
5397 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
5398 nohz_balancer_kick(cpu);
5402 static void rq_online_fair(struct rq *rq)
5407 static void rq_offline_fair(struct rq *rq)
5411 /* Ensure any throttled groups are reachable by pick_next_task */
5412 unthrottle_offline_cfs_rqs(rq);
5415 #endif /* CONFIG_SMP */
5418 * scheduler tick hitting a task of our scheduling class:
5420 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
5422 struct cfs_rq *cfs_rq;
5423 struct sched_entity *se = &curr->se;
5425 for_each_sched_entity(se) {
5426 cfs_rq = cfs_rq_of(se);
5427 entity_tick(cfs_rq, se, queued);
5430 if (sched_feat_numa(NUMA))
5431 task_tick_numa(rq, curr);
5435 * called on fork with the child task as argument from the parent's context
5436 * - child not yet on the tasklist
5437 * - preemption disabled
5439 static void task_fork_fair(struct task_struct *p)
5441 struct cfs_rq *cfs_rq;
5442 struct sched_entity *se = &p->se, *curr;
5443 int this_cpu = smp_processor_id();
5444 struct rq *rq = this_rq();
5445 unsigned long flags;
5447 raw_spin_lock_irqsave(&rq->lock, flags);
5449 update_rq_clock(rq);
5451 cfs_rq = task_cfs_rq(current);
5452 curr = cfs_rq->curr;
5454 if (unlikely(task_cpu(p) != this_cpu)) {
5456 __set_task_cpu(p, this_cpu);
5460 update_curr(cfs_rq);
5463 se->vruntime = curr->vruntime;
5464 place_entity(cfs_rq, se, 1);
5466 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
5468 * Upon rescheduling, sched_class::put_prev_task() will place
5469 * 'current' within the tree based on its new key value.
5471 swap(curr->vruntime, se->vruntime);
5472 resched_task(rq->curr);
5475 se->vruntime -= cfs_rq->min_vruntime;
5477 raw_spin_unlock_irqrestore(&rq->lock, flags);
5481 * Priority of the task has changed. Check to see if we preempt
5485 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
5491 * Reschedule if we are currently running on this runqueue and
5492 * our priority decreased, or if we are not currently running on
5493 * this runqueue and our priority is higher than the current's
5495 if (rq->curr == p) {
5496 if (p->prio > oldprio)
5497 resched_task(rq->curr);
5499 check_preempt_curr(rq, p, 0);
5502 static void switched_from_fair(struct rq *rq, struct task_struct *p)
5504 struct sched_entity *se = &p->se;
5505 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5508 * Ensure the task's vruntime is normalized, so that when its
5509 * switched back to the fair class the enqueue_entity(.flags=0) will
5510 * do the right thing.
5512 * If it was on_rq, then the dequeue_entity(.flags=0) will already
5513 * have normalized the vruntime, if it was !on_rq, then only when
5514 * the task is sleeping will it still have non-normalized vruntime.
5516 if (!se->on_rq && p->state != TASK_RUNNING) {
5518 * Fix up our vruntime so that the current sleep doesn't
5519 * cause 'unlimited' sleep bonus.
5521 place_entity(cfs_rq, se, 0);
5522 se->vruntime -= cfs_rq->min_vruntime;
5527 * We switched to the sched_fair class.
5529 static void switched_to_fair(struct rq *rq, struct task_struct *p)
5535 * We were most likely switched from sched_rt, so
5536 * kick off the schedule if running, otherwise just see
5537 * if we can still preempt the current task.
5540 resched_task(rq->curr);
5542 check_preempt_curr(rq, p, 0);
5545 /* Account for a task changing its policy or group.
5547 * This routine is mostly called to set cfs_rq->curr field when a task
5548 * migrates between groups/classes.
5550 static void set_curr_task_fair(struct rq *rq)
5552 struct sched_entity *se = &rq->curr->se;
5554 for_each_sched_entity(se) {
5555 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5557 set_next_entity(cfs_rq, se);
5558 /* ensure bandwidth has been allocated on our new cfs_rq */
5559 account_cfs_rq_runtime(cfs_rq, 0);
5563 void init_cfs_rq(struct cfs_rq *cfs_rq)
5565 cfs_rq->tasks_timeline = RB_ROOT;
5566 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
5567 #ifndef CONFIG_64BIT
5568 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
5572 #ifdef CONFIG_FAIR_GROUP_SCHED
5573 static void task_move_group_fair(struct task_struct *p, int on_rq)
5576 * If the task was not on the rq at the time of this cgroup movement
5577 * it must have been asleep, sleeping tasks keep their ->vruntime
5578 * absolute on their old rq until wakeup (needed for the fair sleeper
5579 * bonus in place_entity()).
5581 * If it was on the rq, we've just 'preempted' it, which does convert
5582 * ->vruntime to a relative base.
5584 * Make sure both cases convert their relative position when migrating
5585 * to another cgroup's rq. This does somewhat interfere with the
5586 * fair sleeper stuff for the first placement, but who cares.
5589 * When !on_rq, vruntime of the task has usually NOT been normalized.
5590 * But there are some cases where it has already been normalized:
5592 * - Moving a forked child which is waiting for being woken up by
5593 * wake_up_new_task().
5594 * - Moving a task which has been woken up by try_to_wake_up() and
5595 * waiting for actually being woken up by sched_ttwu_pending().
5597 * To prevent boost or penalty in the new cfs_rq caused by delta
5598 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
5600 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
5604 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
5605 set_task_rq(p, task_cpu(p));
5607 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
5610 void free_fair_sched_group(struct task_group *tg)
5614 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
5616 for_each_possible_cpu(i) {
5618 kfree(tg->cfs_rq[i]);
5627 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
5629 struct cfs_rq *cfs_rq;
5630 struct sched_entity *se;
5633 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
5636 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
5640 tg->shares = NICE_0_LOAD;
5642 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
5644 for_each_possible_cpu(i) {
5645 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
5646 GFP_KERNEL, cpu_to_node(i));
5650 se = kzalloc_node(sizeof(struct sched_entity),
5651 GFP_KERNEL, cpu_to_node(i));
5655 init_cfs_rq(cfs_rq);
5656 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
5667 void unregister_fair_sched_group(struct task_group *tg, int cpu)
5669 struct rq *rq = cpu_rq(cpu);
5670 unsigned long flags;
5673 * Only empty task groups can be destroyed; so we can speculatively
5674 * check on_list without danger of it being re-added.
5676 if (!tg->cfs_rq[cpu]->on_list)
5679 raw_spin_lock_irqsave(&rq->lock, flags);
5680 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
5681 raw_spin_unlock_irqrestore(&rq->lock, flags);
5684 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
5685 struct sched_entity *se, int cpu,
5686 struct sched_entity *parent)
5688 struct rq *rq = cpu_rq(cpu);
5693 /* allow initial update_cfs_load() to truncate */
5694 cfs_rq->load_stamp = 1;
5696 init_cfs_rq_runtime(cfs_rq);
5698 tg->cfs_rq[cpu] = cfs_rq;
5701 /* se could be NULL for root_task_group */
5706 se->cfs_rq = &rq->cfs;
5708 se->cfs_rq = parent->my_q;
5711 update_load_set(&se->load, 0);
5712 se->parent = parent;
5715 static DEFINE_MUTEX(shares_mutex);
5717 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
5720 unsigned long flags;
5723 * We can't change the weight of the root cgroup.
5728 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
5730 mutex_lock(&shares_mutex);
5731 if (tg->shares == shares)
5734 tg->shares = shares;
5735 for_each_possible_cpu(i) {
5736 struct rq *rq = cpu_rq(i);
5737 struct sched_entity *se;
5740 /* Propagate contribution to hierarchy */
5741 raw_spin_lock_irqsave(&rq->lock, flags);
5742 for_each_sched_entity(se)
5743 update_cfs_shares(group_cfs_rq(se));
5744 raw_spin_unlock_irqrestore(&rq->lock, flags);
5748 mutex_unlock(&shares_mutex);
5751 #else /* CONFIG_FAIR_GROUP_SCHED */
5753 void free_fair_sched_group(struct task_group *tg) { }
5755 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
5760 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
5762 #endif /* CONFIG_FAIR_GROUP_SCHED */
5765 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
5767 struct sched_entity *se = &task->se;
5768 unsigned int rr_interval = 0;
5771 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
5774 if (rq->cfs.load.weight)
5775 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5781 * All the scheduling class methods:
5783 const struct sched_class fair_sched_class = {
5784 .next = &idle_sched_class,
5785 .enqueue_task = enqueue_task_fair,
5786 .dequeue_task = dequeue_task_fair,
5787 .yield_task = yield_task_fair,
5788 .yield_to_task = yield_to_task_fair,
5790 .check_preempt_curr = check_preempt_wakeup,
5792 .pick_next_task = pick_next_task_fair,
5793 .put_prev_task = put_prev_task_fair,
5796 .select_task_rq = select_task_rq_fair,
5798 .rq_online = rq_online_fair,
5799 .rq_offline = rq_offline_fair,
5801 .task_waking = task_waking_fair,
5804 .set_curr_task = set_curr_task_fair,
5805 .task_tick = task_tick_fair,
5806 .task_fork = task_fork_fair,
5808 .prio_changed = prio_changed_fair,
5809 .switched_from = switched_from_fair,
5810 .switched_to = switched_to_fair,
5812 .get_rr_interval = get_rr_interval_fair,
5814 #ifdef CONFIG_FAIR_GROUP_SCHED
5815 .task_move_group = task_move_group_fair,
5819 #ifdef CONFIG_SCHED_DEBUG
5820 void print_cfs_stats(struct seq_file *m, int cpu)
5822 struct cfs_rq *cfs_rq;
5825 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
5826 print_cfs_rq(m, cpu, cfs_rq);
5831 __init void init_sched_fair_class(void)
5834 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
5837 nohz.next_balance = jiffies;
5838 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
5839 cpu_notifier(sched_ilb_notifier, 0);