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 void account_offnode_enqueue(struct rq *rq, struct task_struct *p)
802 p->numa_contrib = task_h_load(p);
803 rq->offnode_weight += p->numa_contrib;
804 rq->offnode_running++;
807 static void account_offnode_dequeue(struct rq *rq, struct task_struct *p)
809 rq->offnode_weight -= p->numa_contrib;
810 rq->offnode_running--;
814 * numa task sample period in ms: 5s
816 unsigned int sysctl_sched_numa_task_period = 5000;
819 * Wait for the 2-sample stuff to settle before migrating again
821 unsigned int sysctl_sched_numa_settle_count = 2;
824 * Got a PROT_NONE fault for a page on @node.
826 void __task_numa_fault(int node)
828 struct task_struct *p = current;
830 if (!p->numa_faults) {
831 p->numa_faults = kzalloc(sizeof(unsigned long) * nr_node_ids,
838 p->numa_faults[node]++;
841 void task_numa_placement(void)
843 unsigned long faults, max_faults = 0;
844 struct task_struct *p = current;
845 int node, max_node = -1;
846 int seq = ACCESS_ONCE(p->mm->numa_scan_seq);
848 if (p->numa_scan_seq == seq)
851 p->numa_scan_seq = seq;
853 if (unlikely(!p->numa_faults))
856 for (node = 0; node < nr_node_ids; node++) {
857 faults = p->numa_faults[node];
859 if (faults > max_faults) {
864 p->numa_faults[node] /= 2;
867 if (max_node != -1 && p->node != max_node) {
868 if (sched_feat(NUMA_SETTLE) &&
869 (seq - p->numa_migrate_seq) <= (int)sysctl_sched_numa_settle_count)
871 p->numa_migrate_seq = seq;
872 sched_setnode(p, max_node);
877 * The expensive part of numa migration is done from task_work context.
878 * Triggered from task_tick_numa().
880 void task_numa_work(struct callback_head *work)
882 unsigned long migrate, next_scan, now = jiffies;
883 struct task_struct *p = current;
884 struct mm_struct *mm = p->mm;
886 WARN_ON_ONCE(p != container_of(work, struct task_struct, rcu));
889 * Who cares about NUMA placement when they're dying.
891 * NOTE: make sure not to dereference p->mm before this check,
892 * exit_task_work() happens _after_ exit_mm() so we could be called
893 * without p->mm even though we still had it when we enqueued this
896 if (p->flags & PF_EXITING)
900 * Enforce maximal scan/migration frequency..
902 migrate = mm->numa_next_scan;
903 if (time_before(now, migrate))
906 next_scan = now + 2*msecs_to_jiffies(sysctl_sched_numa_task_period);
907 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
910 ACCESS_ONCE(mm->numa_scan_seq)++;
911 lazy_migrate_process(mm);
915 * Drive the periodic memory faults..
917 void task_tick_numa(struct rq *rq, struct task_struct *curr)
922 * We don't care about NUMA placement if we don't have memory.
928 * Using runtime rather than walltime has the dual advantage that
929 * we (mostly) drive the selection from busy threads and that the
930 * task needs to have done some actual work before we bother with
933 now = curr->se.sum_exec_runtime;
934 period = (u64)sysctl_sched_numa_task_period * NSEC_PER_MSEC;
936 if (now - curr->node_stamp > period) {
937 curr->node_stamp = now;
939 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
941 * We can re-use curr->rcu because we checked curr->mm
942 * != NULL so release_task()->call_rcu() was not called
943 * yet and exit_task_work() is called before
946 init_task_work(&curr->rcu, task_numa_work);
947 task_work_add(curr, &curr->rcu, true);
952 static void account_offnode_enqueue(struct rq *rq, struct task_struct *p)
956 static void account_offnode_dequeue(struct rq *rq, struct task_struct *p)
960 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
963 #endif /* CONFIG_SCHED_NUMA */
965 /**************************************************
966 * Scheduling class queueing methods:
970 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
972 update_load_add(&cfs_rq->load, se->load.weight);
973 if (!parent_entity(se))
974 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
976 if (entity_is_task(se)) {
977 struct rq *rq = rq_of(cfs_rq);
978 struct task_struct *p = task_of(se);
979 struct list_head *tasks = &rq->cfs_tasks;
981 if (offnode_task(p)) {
982 account_offnode_enqueue(rq, p);
983 tasks = offnode_tasks(rq);
986 list_add(&se->group_node, tasks);
988 #endif /* CONFIG_SMP */
989 cfs_rq->nr_running++;
993 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
995 update_load_sub(&cfs_rq->load, se->load.weight);
996 if (!parent_entity(se))
997 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
998 if (entity_is_task(se)) {
999 struct task_struct *p = task_of(se);
1001 list_del_init(&se->group_node);
1003 if (offnode_task(p))
1004 account_offnode_dequeue(rq_of(cfs_rq), p);
1006 cfs_rq->nr_running--;
1009 #ifdef CONFIG_FAIR_GROUP_SCHED
1010 /* we need this in update_cfs_load and load-balance functions below */
1011 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1013 static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq,
1016 struct task_group *tg = cfs_rq->tg;
1019 load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1);
1020 load_avg -= cfs_rq->load_contribution;
1022 if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) {
1023 atomic_add(load_avg, &tg->load_weight);
1024 cfs_rq->load_contribution += load_avg;
1028 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
1030 u64 period = sysctl_sched_shares_window;
1032 unsigned long load = cfs_rq->load.weight;
1034 if (cfs_rq->tg == &root_task_group || throttled_hierarchy(cfs_rq))
1037 now = rq_of(cfs_rq)->clock_task;
1038 delta = now - cfs_rq->load_stamp;
1040 /* truncate load history at 4 idle periods */
1041 if (cfs_rq->load_stamp > cfs_rq->load_last &&
1042 now - cfs_rq->load_last > 4 * period) {
1043 cfs_rq->load_period = 0;
1044 cfs_rq->load_avg = 0;
1048 cfs_rq->load_stamp = now;
1049 cfs_rq->load_unacc_exec_time = 0;
1050 cfs_rq->load_period += delta;
1052 cfs_rq->load_last = now;
1053 cfs_rq->load_avg += delta * load;
1056 /* consider updating load contribution on each fold or truncate */
1057 if (global_update || cfs_rq->load_period > period
1058 || !cfs_rq->load_period)
1059 update_cfs_rq_load_contribution(cfs_rq, global_update);
1061 while (cfs_rq->load_period > period) {
1063 * Inline assembly required to prevent the compiler
1064 * optimising this loop into a divmod call.
1065 * See __iter_div_u64_rem() for another example of this.
1067 asm("" : "+rm" (cfs_rq->load_period));
1068 cfs_rq->load_period /= 2;
1069 cfs_rq->load_avg /= 2;
1072 if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg)
1073 list_del_leaf_cfs_rq(cfs_rq);
1076 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1081 * Use this CPU's actual weight instead of the last load_contribution
1082 * to gain a more accurate current total weight. See
1083 * update_cfs_rq_load_contribution().
1085 tg_weight = atomic_read(&tg->load_weight);
1086 tg_weight -= cfs_rq->load_contribution;
1087 tg_weight += cfs_rq->load.weight;
1092 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1094 long tg_weight, load, shares;
1096 tg_weight = calc_tg_weight(tg, cfs_rq);
1097 load = cfs_rq->load.weight;
1099 shares = (tg->shares * load);
1101 shares /= tg_weight;
1103 if (shares < MIN_SHARES)
1104 shares = MIN_SHARES;
1105 if (shares > tg->shares)
1106 shares = tg->shares;
1111 static void update_entity_shares_tick(struct cfs_rq *cfs_rq)
1113 if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) {
1114 update_cfs_load(cfs_rq, 0);
1115 update_cfs_shares(cfs_rq);
1118 # else /* CONFIG_SMP */
1119 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
1123 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1128 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
1131 # endif /* CONFIG_SMP */
1132 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1133 unsigned long weight)
1136 /* commit outstanding execution time */
1137 if (cfs_rq->curr == se)
1138 update_curr(cfs_rq);
1139 account_entity_dequeue(cfs_rq, se);
1142 update_load_set(&se->load, weight);
1145 account_entity_enqueue(cfs_rq, se);
1148 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1150 struct task_group *tg;
1151 struct sched_entity *se;
1155 se = tg->se[cpu_of(rq_of(cfs_rq))];
1156 if (!se || throttled_hierarchy(cfs_rq))
1159 if (likely(se->load.weight == tg->shares))
1162 shares = calc_cfs_shares(cfs_rq, tg);
1164 reweight_entity(cfs_rq_of(se), se, shares);
1166 #else /* CONFIG_FAIR_GROUP_SCHED */
1167 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
1171 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1175 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
1178 #endif /* CONFIG_FAIR_GROUP_SCHED */
1180 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1182 #ifdef CONFIG_SCHEDSTATS
1183 struct task_struct *tsk = NULL;
1185 if (entity_is_task(se))
1188 if (se->statistics.sleep_start) {
1189 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
1194 if (unlikely(delta > se->statistics.sleep_max))
1195 se->statistics.sleep_max = delta;
1197 se->statistics.sleep_start = 0;
1198 se->statistics.sum_sleep_runtime += delta;
1201 account_scheduler_latency(tsk, delta >> 10, 1);
1202 trace_sched_stat_sleep(tsk, delta);
1205 if (se->statistics.block_start) {
1206 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
1211 if (unlikely(delta > se->statistics.block_max))
1212 se->statistics.block_max = delta;
1214 se->statistics.block_start = 0;
1215 se->statistics.sum_sleep_runtime += delta;
1218 if (tsk->in_iowait) {
1219 se->statistics.iowait_sum += delta;
1220 se->statistics.iowait_count++;
1221 trace_sched_stat_iowait(tsk, delta);
1224 trace_sched_stat_blocked(tsk, delta);
1227 * Blocking time is in units of nanosecs, so shift by
1228 * 20 to get a milliseconds-range estimation of the
1229 * amount of time that the task spent sleeping:
1231 if (unlikely(prof_on == SLEEP_PROFILING)) {
1232 profile_hits(SLEEP_PROFILING,
1233 (void *)get_wchan(tsk),
1236 account_scheduler_latency(tsk, delta >> 10, 0);
1242 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1244 #ifdef CONFIG_SCHED_DEBUG
1245 s64 d = se->vruntime - cfs_rq->min_vruntime;
1250 if (d > 3*sysctl_sched_latency)
1251 schedstat_inc(cfs_rq, nr_spread_over);
1256 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1258 u64 vruntime = cfs_rq->min_vruntime;
1261 * The 'current' period is already promised to the current tasks,
1262 * however the extra weight of the new task will slow them down a
1263 * little, place the new task so that it fits in the slot that
1264 * stays open at the end.
1266 if (initial && sched_feat(START_DEBIT))
1267 vruntime += sched_vslice(cfs_rq, se);
1269 /* sleeps up to a single latency don't count. */
1271 unsigned long thresh = sysctl_sched_latency;
1274 * Halve their sleep time's effect, to allow
1275 * for a gentler effect of sleepers:
1277 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1283 /* ensure we never gain time by being placed backwards. */
1284 vruntime = max_vruntime(se->vruntime, vruntime);
1286 se->vruntime = vruntime;
1289 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1292 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1295 * Update the normalized vruntime before updating min_vruntime
1296 * through callig update_curr().
1298 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1299 se->vruntime += cfs_rq->min_vruntime;
1302 * Update run-time statistics of the 'current'.
1304 update_curr(cfs_rq);
1305 update_cfs_load(cfs_rq, 0);
1306 account_entity_enqueue(cfs_rq, se);
1307 update_cfs_shares(cfs_rq);
1309 if (flags & ENQUEUE_WAKEUP) {
1310 place_entity(cfs_rq, se, 0);
1311 enqueue_sleeper(cfs_rq, se);
1314 update_stats_enqueue(cfs_rq, se);
1315 check_spread(cfs_rq, se);
1316 if (se != cfs_rq->curr)
1317 __enqueue_entity(cfs_rq, se);
1320 if (cfs_rq->nr_running == 1) {
1321 list_add_leaf_cfs_rq(cfs_rq);
1322 check_enqueue_throttle(cfs_rq);
1326 static void __clear_buddies_last(struct sched_entity *se)
1328 for_each_sched_entity(se) {
1329 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1330 if (cfs_rq->last == se)
1331 cfs_rq->last = NULL;
1337 static void __clear_buddies_next(struct sched_entity *se)
1339 for_each_sched_entity(se) {
1340 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1341 if (cfs_rq->next == se)
1342 cfs_rq->next = NULL;
1348 static void __clear_buddies_skip(struct sched_entity *se)
1350 for_each_sched_entity(se) {
1351 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1352 if (cfs_rq->skip == se)
1353 cfs_rq->skip = NULL;
1359 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1361 if (cfs_rq->last == se)
1362 __clear_buddies_last(se);
1364 if (cfs_rq->next == se)
1365 __clear_buddies_next(se);
1367 if (cfs_rq->skip == se)
1368 __clear_buddies_skip(se);
1371 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1374 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1377 * Update run-time statistics of the 'current'.
1379 update_curr(cfs_rq);
1381 update_stats_dequeue(cfs_rq, se);
1382 if (flags & DEQUEUE_SLEEP) {
1383 #ifdef CONFIG_SCHEDSTATS
1384 if (entity_is_task(se)) {
1385 struct task_struct *tsk = task_of(se);
1387 if (tsk->state & TASK_INTERRUPTIBLE)
1388 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1389 if (tsk->state & TASK_UNINTERRUPTIBLE)
1390 se->statistics.block_start = rq_of(cfs_rq)->clock;
1395 clear_buddies(cfs_rq, se);
1397 if (se != cfs_rq->curr)
1398 __dequeue_entity(cfs_rq, se);
1400 update_cfs_load(cfs_rq, 0);
1401 account_entity_dequeue(cfs_rq, se);
1404 * Normalize the entity after updating the min_vruntime because the
1405 * update can refer to the ->curr item and we need to reflect this
1406 * movement in our normalized position.
1408 if (!(flags & DEQUEUE_SLEEP))
1409 se->vruntime -= cfs_rq->min_vruntime;
1411 /* return excess runtime on last dequeue */
1412 return_cfs_rq_runtime(cfs_rq);
1414 update_min_vruntime(cfs_rq);
1415 update_cfs_shares(cfs_rq);
1419 * Preempt the current task with a newly woken task if needed:
1422 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1424 unsigned long ideal_runtime, delta_exec;
1425 struct sched_entity *se;
1428 ideal_runtime = sched_slice(cfs_rq, curr);
1429 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1430 if (delta_exec > ideal_runtime) {
1431 resched_task(rq_of(cfs_rq)->curr);
1433 * The current task ran long enough, ensure it doesn't get
1434 * re-elected due to buddy favours.
1436 clear_buddies(cfs_rq, curr);
1441 * Ensure that a task that missed wakeup preemption by a
1442 * narrow margin doesn't have to wait for a full slice.
1443 * This also mitigates buddy induced latencies under load.
1445 if (delta_exec < sysctl_sched_min_granularity)
1448 se = __pick_first_entity(cfs_rq);
1449 delta = curr->vruntime - se->vruntime;
1454 if (delta > ideal_runtime)
1455 resched_task(rq_of(cfs_rq)->curr);
1459 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1461 /* 'current' is not kept within the tree. */
1464 * Any task has to be enqueued before it get to execute on
1465 * a CPU. So account for the time it spent waiting on the
1468 update_stats_wait_end(cfs_rq, se);
1469 __dequeue_entity(cfs_rq, se);
1472 update_stats_curr_start(cfs_rq, se);
1474 #ifdef CONFIG_SCHEDSTATS
1476 * Track our maximum slice length, if the CPU's load is at
1477 * least twice that of our own weight (i.e. dont track it
1478 * when there are only lesser-weight tasks around):
1480 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1481 se->statistics.slice_max = max(se->statistics.slice_max,
1482 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1485 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1489 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1492 * Pick the next process, keeping these things in mind, in this order:
1493 * 1) keep things fair between processes/task groups
1494 * 2) pick the "next" process, since someone really wants that to run
1495 * 3) pick the "last" process, for cache locality
1496 * 4) do not run the "skip" process, if something else is available
1498 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1500 struct sched_entity *se = __pick_first_entity(cfs_rq);
1501 struct sched_entity *left = se;
1504 * Avoid running the skip buddy, if running something else can
1505 * be done without getting too unfair.
1507 if (cfs_rq->skip == se) {
1508 struct sched_entity *second = __pick_next_entity(se);
1509 if (second && wakeup_preempt_entity(second, left) < 1)
1514 * Prefer last buddy, try to return the CPU to a preempted task.
1516 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1520 * Someone really wants this to run. If it's not unfair, run it.
1522 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1525 clear_buddies(cfs_rq, se);
1530 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1532 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1535 * If still on the runqueue then deactivate_task()
1536 * was not called and update_curr() has to be done:
1539 update_curr(cfs_rq);
1541 /* throttle cfs_rqs exceeding runtime */
1542 check_cfs_rq_runtime(cfs_rq);
1544 check_spread(cfs_rq, prev);
1546 update_stats_wait_start(cfs_rq, prev);
1547 /* Put 'current' back into the tree. */
1548 __enqueue_entity(cfs_rq, prev);
1550 cfs_rq->curr = NULL;
1554 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1557 * Update run-time statistics of the 'current'.
1559 update_curr(cfs_rq);
1562 * Update share accounting for long-running entities.
1564 update_entity_shares_tick(cfs_rq);
1566 #ifdef CONFIG_SCHED_HRTICK
1568 * queued ticks are scheduled to match the slice, so don't bother
1569 * validating it and just reschedule.
1572 resched_task(rq_of(cfs_rq)->curr);
1576 * don't let the period tick interfere with the hrtick preemption
1578 if (!sched_feat(DOUBLE_TICK) &&
1579 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1583 if (cfs_rq->nr_running > 1)
1584 check_preempt_tick(cfs_rq, curr);
1588 /**************************************************
1589 * CFS bandwidth control machinery
1592 #ifdef CONFIG_CFS_BANDWIDTH
1594 #ifdef HAVE_JUMP_LABEL
1595 static struct static_key __cfs_bandwidth_used;
1597 static inline bool cfs_bandwidth_used(void)
1599 return static_key_false(&__cfs_bandwidth_used);
1602 void account_cfs_bandwidth_used(int enabled, int was_enabled)
1604 /* only need to count groups transitioning between enabled/!enabled */
1605 if (enabled && !was_enabled)
1606 static_key_slow_inc(&__cfs_bandwidth_used);
1607 else if (!enabled && was_enabled)
1608 static_key_slow_dec(&__cfs_bandwidth_used);
1610 #else /* HAVE_JUMP_LABEL */
1611 static bool cfs_bandwidth_used(void)
1616 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
1617 #endif /* HAVE_JUMP_LABEL */
1620 * default period for cfs group bandwidth.
1621 * default: 0.1s, units: nanoseconds
1623 static inline u64 default_cfs_period(void)
1625 return 100000000ULL;
1628 static inline u64 sched_cfs_bandwidth_slice(void)
1630 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
1634 * Replenish runtime according to assigned quota and update expiration time.
1635 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
1636 * additional synchronization around rq->lock.
1638 * requires cfs_b->lock
1640 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
1644 if (cfs_b->quota == RUNTIME_INF)
1647 now = sched_clock_cpu(smp_processor_id());
1648 cfs_b->runtime = cfs_b->quota;
1649 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
1652 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
1654 return &tg->cfs_bandwidth;
1657 /* returns 0 on failure to allocate runtime */
1658 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1660 struct task_group *tg = cfs_rq->tg;
1661 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
1662 u64 amount = 0, min_amount, expires;
1664 /* note: this is a positive sum as runtime_remaining <= 0 */
1665 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
1667 raw_spin_lock(&cfs_b->lock);
1668 if (cfs_b->quota == RUNTIME_INF)
1669 amount = min_amount;
1672 * If the bandwidth pool has become inactive, then at least one
1673 * period must have elapsed since the last consumption.
1674 * Refresh the global state and ensure bandwidth timer becomes
1677 if (!cfs_b->timer_active) {
1678 __refill_cfs_bandwidth_runtime(cfs_b);
1679 __start_cfs_bandwidth(cfs_b);
1682 if (cfs_b->runtime > 0) {
1683 amount = min(cfs_b->runtime, min_amount);
1684 cfs_b->runtime -= amount;
1688 expires = cfs_b->runtime_expires;
1689 raw_spin_unlock(&cfs_b->lock);
1691 cfs_rq->runtime_remaining += amount;
1693 * we may have advanced our local expiration to account for allowed
1694 * spread between our sched_clock and the one on which runtime was
1697 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
1698 cfs_rq->runtime_expires = expires;
1700 return cfs_rq->runtime_remaining > 0;
1704 * Note: This depends on the synchronization provided by sched_clock and the
1705 * fact that rq->clock snapshots this value.
1707 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1709 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1710 struct rq *rq = rq_of(cfs_rq);
1712 /* if the deadline is ahead of our clock, nothing to do */
1713 if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0))
1716 if (cfs_rq->runtime_remaining < 0)
1720 * If the local deadline has passed we have to consider the
1721 * possibility that our sched_clock is 'fast' and the global deadline
1722 * has not truly expired.
1724 * Fortunately we can check determine whether this the case by checking
1725 * whether the global deadline has advanced.
1728 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
1729 /* extend local deadline, drift is bounded above by 2 ticks */
1730 cfs_rq->runtime_expires += TICK_NSEC;
1732 /* global deadline is ahead, expiration has passed */
1733 cfs_rq->runtime_remaining = 0;
1737 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1738 unsigned long delta_exec)
1740 /* dock delta_exec before expiring quota (as it could span periods) */
1741 cfs_rq->runtime_remaining -= delta_exec;
1742 expire_cfs_rq_runtime(cfs_rq);
1744 if (likely(cfs_rq->runtime_remaining > 0))
1748 * if we're unable to extend our runtime we resched so that the active
1749 * hierarchy can be throttled
1751 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
1752 resched_task(rq_of(cfs_rq)->curr);
1755 static __always_inline
1756 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
1758 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
1761 __account_cfs_rq_runtime(cfs_rq, delta_exec);
1764 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
1766 return cfs_bandwidth_used() && cfs_rq->throttled;
1769 /* check whether cfs_rq, or any parent, is throttled */
1770 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
1772 return cfs_bandwidth_used() && cfs_rq->throttle_count;
1776 * Ensure that neither of the group entities corresponding to src_cpu or
1777 * dest_cpu are members of a throttled hierarchy when performing group
1778 * load-balance operations.
1780 static inline int throttled_lb_pair(struct task_group *tg,
1781 int src_cpu, int dest_cpu)
1783 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
1785 src_cfs_rq = tg->cfs_rq[src_cpu];
1786 dest_cfs_rq = tg->cfs_rq[dest_cpu];
1788 return throttled_hierarchy(src_cfs_rq) ||
1789 throttled_hierarchy(dest_cfs_rq);
1792 /* updated child weight may affect parent so we have to do this bottom up */
1793 static int tg_unthrottle_up(struct task_group *tg, void *data)
1795 struct rq *rq = data;
1796 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1798 cfs_rq->throttle_count--;
1800 if (!cfs_rq->throttle_count) {
1801 u64 delta = rq->clock_task - cfs_rq->load_stamp;
1803 /* leaving throttled state, advance shares averaging windows */
1804 cfs_rq->load_stamp += delta;
1805 cfs_rq->load_last += delta;
1807 /* update entity weight now that we are on_rq again */
1808 update_cfs_shares(cfs_rq);
1815 static int tg_throttle_down(struct task_group *tg, void *data)
1817 struct rq *rq = data;
1818 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1820 /* group is entering throttled state, record last load */
1821 if (!cfs_rq->throttle_count)
1822 update_cfs_load(cfs_rq, 0);
1823 cfs_rq->throttle_count++;
1828 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
1830 struct rq *rq = rq_of(cfs_rq);
1831 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1832 struct sched_entity *se;
1833 long task_delta, dequeue = 1;
1835 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1837 /* account load preceding throttle */
1839 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
1842 task_delta = cfs_rq->h_nr_running;
1843 for_each_sched_entity(se) {
1844 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
1845 /* throttled entity or throttle-on-deactivate */
1850 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
1851 qcfs_rq->h_nr_running -= task_delta;
1853 if (qcfs_rq->load.weight)
1858 rq->nr_running -= task_delta;
1860 cfs_rq->throttled = 1;
1861 cfs_rq->throttled_timestamp = rq->clock;
1862 raw_spin_lock(&cfs_b->lock);
1863 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
1864 raw_spin_unlock(&cfs_b->lock);
1867 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
1869 struct rq *rq = rq_of(cfs_rq);
1870 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1871 struct sched_entity *se;
1875 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1877 cfs_rq->throttled = 0;
1878 raw_spin_lock(&cfs_b->lock);
1879 cfs_b->throttled_time += rq->clock - cfs_rq->throttled_timestamp;
1880 list_del_rcu(&cfs_rq->throttled_list);
1881 raw_spin_unlock(&cfs_b->lock);
1882 cfs_rq->throttled_timestamp = 0;
1884 update_rq_clock(rq);
1885 /* update hierarchical throttle state */
1886 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
1888 if (!cfs_rq->load.weight)
1891 task_delta = cfs_rq->h_nr_running;
1892 for_each_sched_entity(se) {
1896 cfs_rq = cfs_rq_of(se);
1898 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
1899 cfs_rq->h_nr_running += task_delta;
1901 if (cfs_rq_throttled(cfs_rq))
1906 rq->nr_running += task_delta;
1908 /* determine whether we need to wake up potentially idle cpu */
1909 if (rq->curr == rq->idle && rq->cfs.nr_running)
1910 resched_task(rq->curr);
1913 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
1914 u64 remaining, u64 expires)
1916 struct cfs_rq *cfs_rq;
1917 u64 runtime = remaining;
1920 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
1922 struct rq *rq = rq_of(cfs_rq);
1924 raw_spin_lock(&rq->lock);
1925 if (!cfs_rq_throttled(cfs_rq))
1928 runtime = -cfs_rq->runtime_remaining + 1;
1929 if (runtime > remaining)
1930 runtime = remaining;
1931 remaining -= runtime;
1933 cfs_rq->runtime_remaining += runtime;
1934 cfs_rq->runtime_expires = expires;
1936 /* we check whether we're throttled above */
1937 if (cfs_rq->runtime_remaining > 0)
1938 unthrottle_cfs_rq(cfs_rq);
1941 raw_spin_unlock(&rq->lock);
1952 * Responsible for refilling a task_group's bandwidth and unthrottling its
1953 * cfs_rqs as appropriate. If there has been no activity within the last
1954 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
1955 * used to track this state.
1957 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
1959 u64 runtime, runtime_expires;
1960 int idle = 1, throttled;
1962 raw_spin_lock(&cfs_b->lock);
1963 /* no need to continue the timer with no bandwidth constraint */
1964 if (cfs_b->quota == RUNTIME_INF)
1967 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
1968 /* idle depends on !throttled (for the case of a large deficit) */
1969 idle = cfs_b->idle && !throttled;
1970 cfs_b->nr_periods += overrun;
1972 /* if we're going inactive then everything else can be deferred */
1976 __refill_cfs_bandwidth_runtime(cfs_b);
1979 /* mark as potentially idle for the upcoming period */
1984 /* account preceding periods in which throttling occurred */
1985 cfs_b->nr_throttled += overrun;
1988 * There are throttled entities so we must first use the new bandwidth
1989 * to unthrottle them before making it generally available. This
1990 * ensures that all existing debts will be paid before a new cfs_rq is
1993 runtime = cfs_b->runtime;
1994 runtime_expires = cfs_b->runtime_expires;
1998 * This check is repeated as we are holding onto the new bandwidth
1999 * while we unthrottle. This can potentially race with an unthrottled
2000 * group trying to acquire new bandwidth from the global pool.
2002 while (throttled && runtime > 0) {
2003 raw_spin_unlock(&cfs_b->lock);
2004 /* we can't nest cfs_b->lock while distributing bandwidth */
2005 runtime = distribute_cfs_runtime(cfs_b, runtime,
2007 raw_spin_lock(&cfs_b->lock);
2009 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2012 /* return (any) remaining runtime */
2013 cfs_b->runtime = runtime;
2015 * While we are ensured activity in the period following an
2016 * unthrottle, this also covers the case in which the new bandwidth is
2017 * insufficient to cover the existing bandwidth deficit. (Forcing the
2018 * timer to remain active while there are any throttled entities.)
2023 cfs_b->timer_active = 0;
2024 raw_spin_unlock(&cfs_b->lock);
2029 /* a cfs_rq won't donate quota below this amount */
2030 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2031 /* minimum remaining period time to redistribute slack quota */
2032 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2033 /* how long we wait to gather additional slack before distributing */
2034 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2036 /* are we near the end of the current quota period? */
2037 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2039 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2042 /* if the call-back is running a quota refresh is already occurring */
2043 if (hrtimer_callback_running(refresh_timer))
2046 /* is a quota refresh about to occur? */
2047 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2048 if (remaining < min_expire)
2054 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2056 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2058 /* if there's a quota refresh soon don't bother with slack */
2059 if (runtime_refresh_within(cfs_b, min_left))
2062 start_bandwidth_timer(&cfs_b->slack_timer,
2063 ns_to_ktime(cfs_bandwidth_slack_period));
2066 /* we know any runtime found here is valid as update_curr() precedes return */
2067 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2069 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2070 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2072 if (slack_runtime <= 0)
2075 raw_spin_lock(&cfs_b->lock);
2076 if (cfs_b->quota != RUNTIME_INF &&
2077 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2078 cfs_b->runtime += slack_runtime;
2080 /* we are under rq->lock, defer unthrottling using a timer */
2081 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2082 !list_empty(&cfs_b->throttled_cfs_rq))
2083 start_cfs_slack_bandwidth(cfs_b);
2085 raw_spin_unlock(&cfs_b->lock);
2087 /* even if it's not valid for return we don't want to try again */
2088 cfs_rq->runtime_remaining -= slack_runtime;
2091 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2093 if (!cfs_bandwidth_used())
2096 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2099 __return_cfs_rq_runtime(cfs_rq);
2103 * This is done with a timer (instead of inline with bandwidth return) since
2104 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2106 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2108 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2111 /* confirm we're still not at a refresh boundary */
2112 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2115 raw_spin_lock(&cfs_b->lock);
2116 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2117 runtime = cfs_b->runtime;
2120 expires = cfs_b->runtime_expires;
2121 raw_spin_unlock(&cfs_b->lock);
2126 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2128 raw_spin_lock(&cfs_b->lock);
2129 if (expires == cfs_b->runtime_expires)
2130 cfs_b->runtime = runtime;
2131 raw_spin_unlock(&cfs_b->lock);
2135 * When a group wakes up we want to make sure that its quota is not already
2136 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2137 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2139 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2141 if (!cfs_bandwidth_used())
2144 /* an active group must be handled by the update_curr()->put() path */
2145 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2148 /* ensure the group is not already throttled */
2149 if (cfs_rq_throttled(cfs_rq))
2152 /* update runtime allocation */
2153 account_cfs_rq_runtime(cfs_rq, 0);
2154 if (cfs_rq->runtime_remaining <= 0)
2155 throttle_cfs_rq(cfs_rq);
2158 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2159 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2161 if (!cfs_bandwidth_used())
2164 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2168 * it's possible for a throttled entity to be forced into a running
2169 * state (e.g. set_curr_task), in this case we're finished.
2171 if (cfs_rq_throttled(cfs_rq))
2174 throttle_cfs_rq(cfs_rq);
2177 static inline u64 default_cfs_period(void);
2178 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun);
2179 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b);
2181 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2183 struct cfs_bandwidth *cfs_b =
2184 container_of(timer, struct cfs_bandwidth, slack_timer);
2185 do_sched_cfs_slack_timer(cfs_b);
2187 return HRTIMER_NORESTART;
2190 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2192 struct cfs_bandwidth *cfs_b =
2193 container_of(timer, struct cfs_bandwidth, period_timer);
2199 now = hrtimer_cb_get_time(timer);
2200 overrun = hrtimer_forward(timer, now, cfs_b->period);
2205 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2208 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2211 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2213 raw_spin_lock_init(&cfs_b->lock);
2215 cfs_b->quota = RUNTIME_INF;
2216 cfs_b->period = ns_to_ktime(default_cfs_period());
2218 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2219 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2220 cfs_b->period_timer.function = sched_cfs_period_timer;
2221 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2222 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2225 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2227 cfs_rq->runtime_enabled = 0;
2228 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2231 /* requires cfs_b->lock, may release to reprogram timer */
2232 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2235 * The timer may be active because we're trying to set a new bandwidth
2236 * period or because we're racing with the tear-down path
2237 * (timer_active==0 becomes visible before the hrtimer call-back
2238 * terminates). In either case we ensure that it's re-programmed
2240 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2241 raw_spin_unlock(&cfs_b->lock);
2242 /* ensure cfs_b->lock is available while we wait */
2243 hrtimer_cancel(&cfs_b->period_timer);
2245 raw_spin_lock(&cfs_b->lock);
2246 /* if someone else restarted the timer then we're done */
2247 if (cfs_b->timer_active)
2251 cfs_b->timer_active = 1;
2252 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2255 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2257 hrtimer_cancel(&cfs_b->period_timer);
2258 hrtimer_cancel(&cfs_b->slack_timer);
2261 static void unthrottle_offline_cfs_rqs(struct rq *rq)
2263 struct cfs_rq *cfs_rq;
2265 for_each_leaf_cfs_rq(rq, cfs_rq) {
2266 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2268 if (!cfs_rq->runtime_enabled)
2272 * clock_task is not advancing so we just need to make sure
2273 * there's some valid quota amount
2275 cfs_rq->runtime_remaining = cfs_b->quota;
2276 if (cfs_rq_throttled(cfs_rq))
2277 unthrottle_cfs_rq(cfs_rq);
2281 #else /* CONFIG_CFS_BANDWIDTH */
2282 static __always_inline
2283 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec) {}
2284 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2285 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2286 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2288 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2293 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2298 static inline int throttled_lb_pair(struct task_group *tg,
2299 int src_cpu, int dest_cpu)
2304 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2306 #ifdef CONFIG_FAIR_GROUP_SCHED
2307 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2310 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2314 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2315 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2317 #endif /* CONFIG_CFS_BANDWIDTH */
2319 /**************************************************
2320 * CFS operations on tasks:
2323 #ifdef CONFIG_SCHED_HRTICK
2324 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2326 struct sched_entity *se = &p->se;
2327 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2329 WARN_ON(task_rq(p) != rq);
2331 if (cfs_rq->nr_running > 1) {
2332 u64 slice = sched_slice(cfs_rq, se);
2333 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2334 s64 delta = slice - ran;
2343 * Don't schedule slices shorter than 10000ns, that just
2344 * doesn't make sense. Rely on vruntime for fairness.
2347 delta = max_t(s64, 10000LL, delta);
2349 hrtick_start(rq, delta);
2354 * called from enqueue/dequeue and updates the hrtick when the
2355 * current task is from our class and nr_running is low enough
2358 static void hrtick_update(struct rq *rq)
2360 struct task_struct *curr = rq->curr;
2362 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2365 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2366 hrtick_start_fair(rq, curr);
2368 #else /* !CONFIG_SCHED_HRTICK */
2370 hrtick_start_fair(struct rq *rq, struct task_struct *p)
2374 static inline void hrtick_update(struct rq *rq)
2380 * The enqueue_task method is called before nr_running is
2381 * increased. Here we update the fair scheduling stats and
2382 * then put the task into the rbtree:
2385 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2387 struct cfs_rq *cfs_rq;
2388 struct sched_entity *se = &p->se;
2390 for_each_sched_entity(se) {
2393 cfs_rq = cfs_rq_of(se);
2394 enqueue_entity(cfs_rq, se, flags);
2397 * end evaluation on encountering a throttled cfs_rq
2399 * note: in the case of encountering a throttled cfs_rq we will
2400 * post the final h_nr_running increment below.
2402 if (cfs_rq_throttled(cfs_rq))
2404 cfs_rq->h_nr_running++;
2406 flags = ENQUEUE_WAKEUP;
2409 for_each_sched_entity(se) {
2410 cfs_rq = cfs_rq_of(se);
2411 cfs_rq->h_nr_running++;
2413 if (cfs_rq_throttled(cfs_rq))
2416 update_cfs_load(cfs_rq, 0);
2417 update_cfs_shares(cfs_rq);
2425 static void set_next_buddy(struct sched_entity *se);
2428 * The dequeue_task method is called before nr_running is
2429 * decreased. We remove the task from the rbtree and
2430 * update the fair scheduling stats:
2432 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2434 struct cfs_rq *cfs_rq;
2435 struct sched_entity *se = &p->se;
2436 int task_sleep = flags & DEQUEUE_SLEEP;
2438 for_each_sched_entity(se) {
2439 cfs_rq = cfs_rq_of(se);
2440 dequeue_entity(cfs_rq, se, flags);
2443 * end evaluation on encountering a throttled cfs_rq
2445 * note: in the case of encountering a throttled cfs_rq we will
2446 * post the final h_nr_running decrement below.
2448 if (cfs_rq_throttled(cfs_rq))
2450 cfs_rq->h_nr_running--;
2452 /* Don't dequeue parent if it has other entities besides us */
2453 if (cfs_rq->load.weight) {
2455 * Bias pick_next to pick a task from this cfs_rq, as
2456 * p is sleeping when it is within its sched_slice.
2458 if (task_sleep && parent_entity(se))
2459 set_next_buddy(parent_entity(se));
2461 /* avoid re-evaluating load for this entity */
2462 se = parent_entity(se);
2465 flags |= DEQUEUE_SLEEP;
2468 for_each_sched_entity(se) {
2469 cfs_rq = cfs_rq_of(se);
2470 cfs_rq->h_nr_running--;
2472 if (cfs_rq_throttled(cfs_rq))
2475 update_cfs_load(cfs_rq, 0);
2476 update_cfs_shares(cfs_rq);
2485 /* Used instead of source_load when we know the type == 0 */
2486 static unsigned long weighted_cpuload(const int cpu)
2488 return cpu_rq(cpu)->load.weight;
2492 * Return a low guess at the load of a migration-source cpu weighted
2493 * according to the scheduling class and "nice" value.
2495 * We want to under-estimate the load of migration sources, to
2496 * balance conservatively.
2498 static unsigned long source_load(int cpu, int type)
2500 struct rq *rq = cpu_rq(cpu);
2501 unsigned long total = weighted_cpuload(cpu);
2503 if (type == 0 || !sched_feat(LB_BIAS))
2506 return min(rq->cpu_load[type-1], total);
2510 * Return a high guess at the load of a migration-target cpu weighted
2511 * according to the scheduling class and "nice" value.
2513 static unsigned long target_load(int cpu, int type)
2515 struct rq *rq = cpu_rq(cpu);
2516 unsigned long total = weighted_cpuload(cpu);
2518 if (type == 0 || !sched_feat(LB_BIAS))
2521 return max(rq->cpu_load[type-1], total);
2524 static unsigned long power_of(int cpu)
2526 return cpu_rq(cpu)->cpu_power;
2529 static unsigned long cpu_avg_load_per_task(int cpu)
2531 struct rq *rq = cpu_rq(cpu);
2532 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
2535 return rq->load.weight / nr_running;
2541 static void task_waking_fair(struct task_struct *p)
2543 struct sched_entity *se = &p->se;
2544 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2547 #ifndef CONFIG_64BIT
2548 u64 min_vruntime_copy;
2551 min_vruntime_copy = cfs_rq->min_vruntime_copy;
2553 min_vruntime = cfs_rq->min_vruntime;
2554 } while (min_vruntime != min_vruntime_copy);
2556 min_vruntime = cfs_rq->min_vruntime;
2559 se->vruntime -= min_vruntime;
2562 #ifdef CONFIG_FAIR_GROUP_SCHED
2564 * effective_load() calculates the load change as seen from the root_task_group
2566 * Adding load to a group doesn't make a group heavier, but can cause movement
2567 * of group shares between cpus. Assuming the shares were perfectly aligned one
2568 * can calculate the shift in shares.
2570 * Calculate the effective load difference if @wl is added (subtracted) to @tg
2571 * on this @cpu and results in a total addition (subtraction) of @wg to the
2572 * total group weight.
2574 * Given a runqueue weight distribution (rw_i) we can compute a shares
2575 * distribution (s_i) using:
2577 * s_i = rw_i / \Sum rw_j (1)
2579 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
2580 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
2581 * shares distribution (s_i):
2583 * rw_i = { 2, 4, 1, 0 }
2584 * s_i = { 2/7, 4/7, 1/7, 0 }
2586 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
2587 * task used to run on and the CPU the waker is running on), we need to
2588 * compute the effect of waking a task on either CPU and, in case of a sync
2589 * wakeup, compute the effect of the current task going to sleep.
2591 * So for a change of @wl to the local @cpu with an overall group weight change
2592 * of @wl we can compute the new shares distribution (s'_i) using:
2594 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
2596 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
2597 * differences in waking a task to CPU 0. The additional task changes the
2598 * weight and shares distributions like:
2600 * rw'_i = { 3, 4, 1, 0 }
2601 * s'_i = { 3/8, 4/8, 1/8, 0 }
2603 * We can then compute the difference in effective weight by using:
2605 * dw_i = S * (s'_i - s_i) (3)
2607 * Where 'S' is the group weight as seen by its parent.
2609 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
2610 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
2611 * 4/7) times the weight of the group.
2613 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
2615 struct sched_entity *se = tg->se[cpu];
2617 if (!tg->parent) /* the trivial, non-cgroup case */
2620 for_each_sched_entity(se) {
2626 * W = @wg + \Sum rw_j
2628 W = wg + calc_tg_weight(tg, se->my_q);
2633 w = se->my_q->load.weight + wl;
2636 * wl = S * s'_i; see (2)
2639 wl = (w * tg->shares) / W;
2644 * Per the above, wl is the new se->load.weight value; since
2645 * those are clipped to [MIN_SHARES, ...) do so now. See
2646 * calc_cfs_shares().
2648 if (wl < MIN_SHARES)
2652 * wl = dw_i = S * (s'_i - s_i); see (3)
2654 wl -= se->load.weight;
2657 * Recursively apply this logic to all parent groups to compute
2658 * the final effective load change on the root group. Since
2659 * only the @tg group gets extra weight, all parent groups can
2660 * only redistribute existing shares. @wl is the shift in shares
2661 * resulting from this level per the above.
2670 static inline unsigned long effective_load(struct task_group *tg, int cpu,
2671 unsigned long wl, unsigned long wg)
2678 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
2680 s64 this_load, load;
2681 int idx, this_cpu, prev_cpu;
2682 unsigned long tl_per_task;
2683 struct task_group *tg;
2684 unsigned long weight;
2688 this_cpu = smp_processor_id();
2689 prev_cpu = task_cpu(p);
2690 load = source_load(prev_cpu, idx);
2691 this_load = target_load(this_cpu, idx);
2694 * If sync wakeup then subtract the (maximum possible)
2695 * effect of the currently running task from the load
2696 * of the current CPU:
2699 tg = task_group(current);
2700 weight = current->se.load.weight;
2702 this_load += effective_load(tg, this_cpu, -weight, -weight);
2703 load += effective_load(tg, prev_cpu, 0, -weight);
2707 weight = p->se.load.weight;
2710 * In low-load situations, where prev_cpu is idle and this_cpu is idle
2711 * due to the sync cause above having dropped this_load to 0, we'll
2712 * always have an imbalance, but there's really nothing you can do
2713 * about that, so that's good too.
2715 * Otherwise check if either cpus are near enough in load to allow this
2716 * task to be woken on this_cpu.
2718 if (this_load > 0) {
2719 s64 this_eff_load, prev_eff_load;
2721 this_eff_load = 100;
2722 this_eff_load *= power_of(prev_cpu);
2723 this_eff_load *= this_load +
2724 effective_load(tg, this_cpu, weight, weight);
2726 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
2727 prev_eff_load *= power_of(this_cpu);
2728 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
2730 balanced = this_eff_load <= prev_eff_load;
2735 * If the currently running task will sleep within
2736 * a reasonable amount of time then attract this newly
2739 if (sync && balanced)
2742 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
2743 tl_per_task = cpu_avg_load_per_task(this_cpu);
2746 (this_load <= load &&
2747 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
2749 * This domain has SD_WAKE_AFFINE and
2750 * p is cache cold in this domain, and
2751 * there is no bad imbalance.
2753 schedstat_inc(sd, ttwu_move_affine);
2754 schedstat_inc(p, se.statistics.nr_wakeups_affine);
2762 * find_idlest_group finds and returns the least busy CPU group within the
2765 static struct sched_group *
2766 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
2767 int this_cpu, int load_idx)
2769 struct sched_group *idlest = NULL, *group = sd->groups;
2770 unsigned long min_load = ULONG_MAX, this_load = 0;
2771 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2774 unsigned long load, avg_load;
2778 /* Skip over this group if it has no CPUs allowed */
2779 if (!cpumask_intersects(sched_group_cpus(group),
2780 tsk_cpus_allowed(p)))
2783 local_group = cpumask_test_cpu(this_cpu,
2784 sched_group_cpus(group));
2786 /* Tally up the load of all CPUs in the group */
2789 for_each_cpu(i, sched_group_cpus(group)) {
2790 /* Bias balancing toward cpus of our domain */
2792 load = source_load(i, load_idx);
2794 load = target_load(i, load_idx);
2799 /* Adjust by relative CPU power of the group */
2800 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
2803 this_load = avg_load;
2804 } else if (avg_load < min_load) {
2805 min_load = avg_load;
2808 } while (group = group->next, group != sd->groups);
2810 if (!idlest || 100*this_load < imbalance*min_load)
2816 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2819 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2821 unsigned long load, min_load = ULONG_MAX;
2825 /* Traverse only the allowed CPUs */
2826 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
2827 load = weighted_cpuload(i);
2829 if (load < min_load || (load == min_load && i == this_cpu)) {
2839 * Try and locate an idle CPU in the sched_domain.
2841 static int select_idle_sibling(struct task_struct *p, int target)
2843 int cpu = smp_processor_id();
2844 int prev_cpu = task_cpu(p);
2845 struct sched_domain *sd;
2846 struct sched_group *sg;
2850 * If the task is going to be woken-up on this cpu and if it is
2851 * already idle, then it is the right target.
2853 if (target == cpu && idle_cpu(cpu))
2857 * If the task is going to be woken-up on the cpu where it previously
2858 * ran and if it is currently idle, then it the right target.
2860 if (target == prev_cpu && idle_cpu(prev_cpu))
2864 * Otherwise, iterate the domains and find an elegible idle cpu.
2866 sd = rcu_dereference(per_cpu(sd_llc, target));
2867 for_each_lower_domain(sd) {
2870 if (!cpumask_intersects(sched_group_cpus(sg),
2871 tsk_cpus_allowed(p)))
2874 for_each_cpu(i, sched_group_cpus(sg)) {
2879 target = cpumask_first_and(sched_group_cpus(sg),
2880 tsk_cpus_allowed(p));
2884 } while (sg != sd->groups);
2890 #ifdef CONFIG_SCHED_NUMA
2891 static inline bool pick_numa_rand(int n)
2893 return !(get_random_int() % n);
2897 * Pick a random elegible CPU in the target node, hopefully faster
2898 * than doing a least-loaded scan.
2900 static int numa_select_node_cpu(struct task_struct *p, int node)
2902 int weight = cpumask_weight(cpumask_of_node(node));
2905 for_each_cpu_and(i, cpumask_of_node(node), tsk_cpus_allowed(p)) {
2906 if (cpu < 0 || pick_numa_rand(weight))
2913 static int numa_select_node_cpu(struct task_struct *p, int node)
2917 #endif /* CONFIG_SCHED_NUMA */
2920 * sched_balance_self: balance the current task (running on cpu) in domains
2921 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2924 * Balance, ie. select the least loaded group.
2926 * Returns the target CPU number, or the same CPU if no balancing is needed.
2928 * preempt must be disabled.
2931 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
2933 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
2934 int cpu = smp_processor_id();
2935 int prev_cpu = task_cpu(p);
2937 int want_affine = 0;
2938 int sync = wake_flags & WF_SYNC;
2939 int node = tsk_home_node(p);
2941 if (p->nr_cpus_allowed == 1)
2944 if (sd_flag & SD_BALANCE_WAKE) {
2945 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
2951 if (sched_feat_numa(NUMA_TTWU_BIAS) && node != -1) {
2953 * For fork,exec find the idlest cpu in the home-node.
2955 if (sd_flag & (SD_BALANCE_FORK|SD_BALANCE_EXEC)) {
2956 int node_cpu = numa_select_node_cpu(p, node);
2960 new_cpu = cpu = node_cpu;
2961 sd = per_cpu(sd_node, cpu);
2966 * For wake, pretend we were running in the home-node.
2968 if (cpu_to_node(prev_cpu) != node) {
2969 int node_cpu = numa_select_node_cpu(p, node);
2973 if (sched_feat_numa(NUMA_TTWU_TO))
2976 prev_cpu = node_cpu;
2981 for_each_domain(cpu, tmp) {
2982 if (!(tmp->flags & SD_LOAD_BALANCE))
2986 * If both cpu and prev_cpu are part of this domain,
2987 * cpu is a valid SD_WAKE_AFFINE target.
2989 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
2990 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
2995 if (tmp->flags & sd_flag)
3000 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3003 new_cpu = select_idle_sibling(p, prev_cpu);
3009 int load_idx = sd->forkexec_idx;
3010 struct sched_group *group;
3013 if (!(sd->flags & sd_flag)) {
3018 if (sd_flag & SD_BALANCE_WAKE)
3019 load_idx = sd->wake_idx;
3021 group = find_idlest_group(sd, p, cpu, load_idx);
3027 new_cpu = find_idlest_cpu(group, p, cpu);
3028 if (new_cpu == -1 || new_cpu == cpu) {
3029 /* Now try balancing at a lower domain level of cpu */
3034 /* Now try balancing at a lower domain level of new_cpu */
3036 weight = sd->span_weight;
3038 for_each_domain(cpu, tmp) {
3039 if (weight <= tmp->span_weight)
3041 if (tmp->flags & sd_flag)
3044 /* while loop will break here if sd == NULL */
3051 #endif /* CONFIG_SMP */
3053 static unsigned long
3054 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3056 unsigned long gran = sysctl_sched_wakeup_granularity;
3059 * Since its curr running now, convert the gran from real-time
3060 * to virtual-time in his units.
3062 * By using 'se' instead of 'curr' we penalize light tasks, so
3063 * they get preempted easier. That is, if 'se' < 'curr' then
3064 * the resulting gran will be larger, therefore penalizing the
3065 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3066 * be smaller, again penalizing the lighter task.
3068 * This is especially important for buddies when the leftmost
3069 * task is higher priority than the buddy.
3071 return calc_delta_fair(gran, se);
3075 * Should 'se' preempt 'curr'.
3089 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3091 s64 gran, vdiff = curr->vruntime - se->vruntime;
3096 gran = wakeup_gran(curr, se);
3103 static void set_last_buddy(struct sched_entity *se)
3105 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3108 for_each_sched_entity(se)
3109 cfs_rq_of(se)->last = se;
3112 static void set_next_buddy(struct sched_entity *se)
3114 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3117 for_each_sched_entity(se)
3118 cfs_rq_of(se)->next = se;
3121 static void set_skip_buddy(struct sched_entity *se)
3123 for_each_sched_entity(se)
3124 cfs_rq_of(se)->skip = se;
3128 * Preempt the current task with a newly woken task if needed:
3130 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3132 struct task_struct *curr = rq->curr;
3133 struct sched_entity *se = &curr->se, *pse = &p->se;
3134 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3135 int scale = cfs_rq->nr_running >= sched_nr_latency;
3136 int next_buddy_marked = 0;
3138 if (unlikely(se == pse))
3142 * This is possible from callers such as move_task(), in which we
3143 * unconditionally check_prempt_curr() after an enqueue (which may have
3144 * lead to a throttle). This both saves work and prevents false
3145 * next-buddy nomination below.
3147 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3150 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3151 set_next_buddy(pse);
3152 next_buddy_marked = 1;
3156 * We can come here with TIF_NEED_RESCHED already set from new task
3159 * Note: this also catches the edge-case of curr being in a throttled
3160 * group (e.g. via set_curr_task), since update_curr() (in the
3161 * enqueue of curr) will have resulted in resched being set. This
3162 * prevents us from potentially nominating it as a false LAST_BUDDY
3165 if (test_tsk_need_resched(curr))
3168 /* Idle tasks are by definition preempted by non-idle tasks. */
3169 if (unlikely(curr->policy == SCHED_IDLE) &&
3170 likely(p->policy != SCHED_IDLE))
3174 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3175 * is driven by the tick):
3177 if (unlikely(p->policy != SCHED_NORMAL))
3180 find_matching_se(&se, &pse);
3181 update_curr(cfs_rq_of(se));
3183 if (wakeup_preempt_entity(se, pse) == 1) {
3185 * Bias pick_next to pick the sched entity that is
3186 * triggering this preemption.
3188 if (!next_buddy_marked)
3189 set_next_buddy(pse);
3198 * Only set the backward buddy when the current task is still
3199 * on the rq. This can happen when a wakeup gets interleaved
3200 * with schedule on the ->pre_schedule() or idle_balance()
3201 * point, either of which can * drop the rq lock.
3203 * Also, during early boot the idle thread is in the fair class,
3204 * for obvious reasons its a bad idea to schedule back to it.
3206 if (unlikely(!se->on_rq || curr == rq->idle))
3209 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3213 static struct task_struct *pick_next_task_fair(struct rq *rq)
3215 struct task_struct *p;
3216 struct cfs_rq *cfs_rq = &rq->cfs;
3217 struct sched_entity *se;
3219 if (!cfs_rq->nr_running)
3223 se = pick_next_entity(cfs_rq);
3224 set_next_entity(cfs_rq, se);
3225 cfs_rq = group_cfs_rq(se);
3229 if (hrtick_enabled(rq))
3230 hrtick_start_fair(rq, p);
3236 * Account for a descheduled task:
3238 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3240 struct sched_entity *se = &prev->se;
3241 struct cfs_rq *cfs_rq;
3243 for_each_sched_entity(se) {
3244 cfs_rq = cfs_rq_of(se);
3245 put_prev_entity(cfs_rq, se);
3250 * sched_yield() is very simple
3252 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3254 static void yield_task_fair(struct rq *rq)
3256 struct task_struct *curr = rq->curr;
3257 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3258 struct sched_entity *se = &curr->se;
3261 * Are we the only task in the tree?
3263 if (unlikely(rq->nr_running == 1))
3266 clear_buddies(cfs_rq, se);
3268 if (curr->policy != SCHED_BATCH) {
3269 update_rq_clock(rq);
3271 * Update run-time statistics of the 'current'.
3273 update_curr(cfs_rq);
3275 * Tell update_rq_clock() that we've just updated,
3276 * so we don't do microscopic update in schedule()
3277 * and double the fastpath cost.
3279 rq->skip_clock_update = 1;
3285 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3287 struct sched_entity *se = &p->se;
3289 /* throttled hierarchies are not runnable */
3290 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3293 /* Tell the scheduler that we'd really like pse to run next. */
3296 yield_task_fair(rq);
3302 /**************************************************
3303 * Fair scheduling class load-balancing methods:
3306 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
3308 #define LBF_ALL_PINNED 0x01
3309 #define LBF_NEED_BREAK 0x02
3310 #define LBF_SOME_PINNED 0x04
3313 struct sched_domain *sd;
3321 struct cpumask *dst_grpmask;
3323 enum cpu_idle_type idle;
3325 /* The set of CPUs under consideration for load-balancing */
3326 struct cpumask *cpus;
3330 struct list_head *tasks;
3333 unsigned int loop_break;
3334 unsigned int loop_max;
3336 struct rq * (*find_busiest_queue)(struct lb_env *,
3337 struct sched_group *);
3341 * move_task - move a task from one runqueue to another runqueue.
3342 * Both runqueues must be locked.
3344 static void move_task(struct task_struct *p, struct lb_env *env)
3346 deactivate_task(env->src_rq, p, 0);
3347 set_task_cpu(p, env->dst_cpu);
3348 activate_task(env->dst_rq, p, 0);
3349 check_preempt_curr(env->dst_rq, p, 0);
3352 static int task_numa_hot(struct task_struct *p, struct lb_env *env)
3354 int from_dist, to_dist;
3355 int node = tsk_home_node(p);
3357 if (!sched_feat_numa(NUMA_HOT) || node == -1)
3358 return 0; /* no node preference */
3360 from_dist = node_distance(cpu_to_node(env->src_cpu), node);
3361 to_dist = node_distance(cpu_to_node(env->dst_cpu), node);
3363 if (to_dist < from_dist)
3364 return 0; /* getting closer is ok */
3366 return 1; /* stick to where we are */
3370 * Is this task likely cache-hot:
3373 task_hot(struct task_struct *p, struct lb_env *env)
3377 if (p->sched_class != &fair_sched_class)
3380 if (unlikely(p->policy == SCHED_IDLE))
3384 * Buddy candidates are cache hot:
3386 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
3387 (&p->se == cfs_rq_of(&p->se)->next ||
3388 &p->se == cfs_rq_of(&p->se)->last))
3391 if (sysctl_sched_migration_cost == -1)
3393 if (sysctl_sched_migration_cost == 0)
3396 delta = env->src_rq->clock_task - p->se.exec_start;
3398 return delta < (s64)sysctl_sched_migration_cost;
3402 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3405 int can_migrate_task(struct task_struct *p, struct lb_env *env)
3407 int tsk_cache_hot = 0;
3409 * We do not migrate tasks that are:
3410 * 1) running (obviously), or
3411 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3412 * 3) are cache-hot on their current CPU.
3414 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
3417 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
3420 * Remember if this task can be migrated to any other cpu in
3421 * our sched_group. We may want to revisit it if we couldn't
3422 * meet load balance goals by pulling other tasks on src_cpu.
3424 * Also avoid computing new_dst_cpu if we have already computed
3425 * one in current iteration.
3427 if (!env->dst_grpmask || (env->flags & LBF_SOME_PINNED))
3430 new_dst_cpu = cpumask_first_and(env->dst_grpmask,
3431 tsk_cpus_allowed(p));
3432 if (new_dst_cpu < nr_cpu_ids) {
3433 env->flags |= LBF_SOME_PINNED;
3434 env->new_dst_cpu = new_dst_cpu;
3439 /* Record that we found atleast one task that could run on dst_cpu */
3440 env->flags &= ~LBF_ALL_PINNED;
3442 if (task_running(env->src_rq, p)) {
3443 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
3448 * Aggressive migration if:
3449 * 1) task is cache cold, or
3450 * 2) too many balance attempts have failed.
3453 tsk_cache_hot = task_hot(p, env);
3454 if (env->idle == CPU_NOT_IDLE)
3455 tsk_cache_hot |= task_numa_hot(p, env);
3456 if (!tsk_cache_hot ||
3457 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
3458 #ifdef CONFIG_SCHEDSTATS
3459 if (tsk_cache_hot) {
3460 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
3461 schedstat_inc(p, se.statistics.nr_forced_migrations);
3467 if (tsk_cache_hot) {
3468 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
3475 * move_one_task tries to move exactly one task from busiest to this_rq, as
3476 * part of active balancing operations within "domain".
3477 * Returns 1 if successful and 0 otherwise.
3479 * Called with both runqueues locked.
3481 static int __move_one_task(struct lb_env *env)
3483 struct task_struct *p, *n;
3485 list_for_each_entry_safe(p, n, env->tasks, se.group_node) {
3486 if (throttled_lb_pair(task_group(p), env->src_rq->cpu, env->dst_cpu))
3489 if (!can_migrate_task(p, env))
3494 * Right now, this is only the second place move_task()
3495 * is called, so we can safely collect move_task()
3496 * stats here rather than inside move_task().
3498 schedstat_inc(env->sd, lb_gained[env->idle]);
3504 static int move_one_task(struct lb_env *env)
3506 if (sched_feat_numa(NUMA_PULL)) {
3507 env->tasks = offnode_tasks(env->src_rq);
3508 if (__move_one_task(env))
3512 env->tasks = &env->src_rq->cfs_tasks;
3513 if (__move_one_task(env))
3519 static const unsigned int sched_nr_migrate_break = 32;
3522 * move_tasks tries to move up to imbalance weighted load from busiest to
3523 * this_rq, as part of a balancing operation within domain "sd".
3524 * Returns 1 if successful and 0 otherwise.
3526 * Called with both runqueues locked.
3528 static int move_tasks(struct lb_env *env)
3530 struct task_struct *p;
3534 if (env->imbalance <= 0)
3538 while (!list_empty(env->tasks)) {
3539 p = list_first_entry(env->tasks, struct task_struct, se.group_node);
3542 /* We've more or less seen every task there is, call it quits */
3543 if (env->loop > env->loop_max)
3546 /* take a breather every nr_migrate tasks */
3547 if (env->loop > env->loop_break) {
3548 env->loop_break += sched_nr_migrate_break;
3549 env->flags |= LBF_NEED_BREAK;
3553 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
3556 load = task_h_load(p);
3558 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
3561 if ((load / 2) > env->imbalance)
3564 if (!can_migrate_task(p, env))
3569 env->imbalance -= load;
3571 #ifdef CONFIG_PREEMPT
3573 * NEWIDLE balancing is a source of latency, so preemptible
3574 * kernels will stop after the first task is pulled to minimize
3575 * the critical section.
3577 if (env->idle == CPU_NEWLY_IDLE)
3582 * We only want to steal up to the prescribed amount of
3585 if (env->imbalance <= 0)
3590 list_move_tail(&p->se.group_node, env->tasks);
3593 if (env->tasks == offnode_tasks(env->src_rq)) {
3594 env->tasks = &env->src_rq->cfs_tasks;
3601 * Right now, this is one of only two places move_task() is called,
3602 * so we can safely collect move_task() stats here rather than
3603 * inside move_task().
3605 schedstat_add(env->sd, lb_gained[env->idle], pulled);
3610 #ifdef CONFIG_FAIR_GROUP_SCHED
3612 * update tg->load_weight by folding this cpu's load_avg
3614 static int update_shares_cpu(struct task_group *tg, int cpu)
3616 struct cfs_rq *cfs_rq;
3617 unsigned long flags;
3624 cfs_rq = tg->cfs_rq[cpu];
3626 raw_spin_lock_irqsave(&rq->lock, flags);
3628 update_rq_clock(rq);
3629 update_cfs_load(cfs_rq, 1);
3632 * We need to update shares after updating tg->load_weight in
3633 * order to adjust the weight of groups with long running tasks.
3635 update_cfs_shares(cfs_rq);
3637 raw_spin_unlock_irqrestore(&rq->lock, flags);
3642 static void update_shares(int cpu)
3644 struct cfs_rq *cfs_rq;
3645 struct rq *rq = cpu_rq(cpu);
3649 * Iterates the task_group tree in a bottom up fashion, see
3650 * list_add_leaf_cfs_rq() for details.
3652 for_each_leaf_cfs_rq(rq, cfs_rq) {
3653 /* throttled entities do not contribute to load */
3654 if (throttled_hierarchy(cfs_rq))
3657 update_shares_cpu(cfs_rq->tg, cpu);
3663 * Compute the cpu's hierarchical load factor for each task group.
3664 * This needs to be done in a top-down fashion because the load of a child
3665 * group is a fraction of its parents load.
3667 static int tg_load_down(struct task_group *tg, void *data)
3670 long cpu = (long)data;
3673 load = cpu_rq(cpu)->load.weight;
3675 load = tg->parent->cfs_rq[cpu]->h_load;
3676 load *= tg->se[cpu]->load.weight;
3677 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
3680 tg->cfs_rq[cpu]->h_load = load;
3685 static void update_h_load(long cpu)
3687 struct rq *rq = cpu_rq(cpu);
3688 unsigned long now = jiffies;
3690 if (rq->h_load_throttle == now)
3693 rq->h_load_throttle = now;
3696 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
3700 static unsigned long task_h_load(struct task_struct *p)
3702 struct cfs_rq *cfs_rq = task_cfs_rq(p);
3705 load = p->se.load.weight;
3706 load = div_u64(load * cfs_rq->h_load, cfs_rq->load.weight + 1);
3711 static inline void update_shares(int cpu)
3715 static inline void update_h_load(long cpu)
3719 static unsigned long task_h_load(struct task_struct *p)
3721 return p->se.load.weight;
3725 /********** Helpers for find_busiest_group ************************/
3727 * sd_lb_stats - Structure to store the statistics of a sched_domain
3728 * during load balancing.
3730 struct sd_lb_stats {
3731 struct sched_group *busiest; /* Busiest group in this sd */
3732 struct sched_group *this; /* Local group in this sd */
3733 unsigned long total_load; /* Total load of all groups in sd */
3734 unsigned long total_pwr; /* Total power of all groups in sd */
3735 unsigned long avg_load; /* Average load across all groups in sd */
3737 /** Statistics of this group */
3738 unsigned long this_load;
3739 unsigned long this_load_per_task;
3740 unsigned long this_nr_running;
3741 unsigned long this_has_capacity;
3742 unsigned int this_idle_cpus;
3744 /* Statistics of the busiest group */
3745 unsigned int busiest_idle_cpus;
3746 unsigned long max_load;
3747 unsigned long busiest_load_per_task;
3748 unsigned long busiest_nr_running;
3749 unsigned long busiest_group_capacity;
3750 unsigned long busiest_has_capacity;
3751 unsigned int busiest_group_weight;
3753 int group_imb; /* Is there imbalance in this sd */
3754 #ifdef CONFIG_SCHED_NUMA
3755 struct sched_group *numa_group; /* group which has offnode_tasks */
3756 unsigned long numa_group_weight;
3757 unsigned long numa_group_running;
3762 * sg_lb_stats - stats of a sched_group required for load_balancing
3764 struct sg_lb_stats {
3765 unsigned long avg_load; /*Avg load across the CPUs of the group */
3766 unsigned long group_load; /* Total load over the CPUs of the group */
3767 unsigned long sum_nr_running; /* Nr tasks running in the group */
3768 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3769 unsigned long group_capacity;
3770 unsigned long idle_cpus;
3771 unsigned long group_weight;
3772 int group_imb; /* Is there an imbalance in the group ? */
3773 int group_has_capacity; /* Is there extra capacity in the group? */
3774 #ifdef CONFIG_SCHED_NUMA
3775 unsigned long numa_weight;
3776 unsigned long numa_running;
3781 * get_sd_load_idx - Obtain the load index for a given sched domain.
3782 * @sd: The sched_domain whose load_idx is to be obtained.
3783 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3785 static inline int get_sd_load_idx(struct sched_domain *sd,
3786 enum cpu_idle_type idle)
3792 load_idx = sd->busy_idx;
3795 case CPU_NEWLY_IDLE:
3796 load_idx = sd->newidle_idx;
3799 load_idx = sd->idle_idx;
3806 #ifdef CONFIG_SCHED_NUMA
3807 static inline void update_sg_numa_stats(struct sg_lb_stats *sgs, struct rq *rq)
3809 sgs->numa_weight += rq->offnode_weight;
3810 sgs->numa_running += rq->offnode_running;
3814 * Since the offnode lists are indiscriminate (they contain tasks for all other
3815 * nodes) it is impossible to say if there's any task on there that wants to
3816 * move towards the pulling cpu. Therefore select a random offnode list to pull
3817 * from such that eventually we'll try them all.
3819 * Select a random group that has offnode tasks as sds->numa_group
3821 static inline void update_sd_numa_stats(struct sched_domain *sd,
3822 struct sched_group *group, struct sd_lb_stats *sds,
3823 int local_group, struct sg_lb_stats *sgs)
3825 if (!(sd->flags & SD_NUMA))
3831 if (!sgs->numa_running)
3834 if (!sds->numa_group || pick_numa_rand(sd->span_weight / group->group_weight)) {
3835 sds->numa_group = group;
3836 sds->numa_group_weight = sgs->numa_weight;
3837 sds->numa_group_running = sgs->numa_running;
3842 * Pick a random queue from the group that has offnode tasks.
3844 static struct rq *find_busiest_numa_queue(struct lb_env *env,
3845 struct sched_group *group)
3847 struct rq *busiest = NULL, *rq;
3850 for_each_cpu_and(cpu, sched_group_cpus(group), env->cpus) {
3852 if (!rq->offnode_running)
3854 if (!busiest || pick_numa_rand(group->group_weight))
3862 * Called in case of no other imbalance, if there is a queue running offnode
3863 * tasksk we'll say we're imbalanced anyway to nudge these tasks towards their
3866 static inline int check_numa_busiest_group(struct lb_env *env, struct sd_lb_stats *sds)
3868 if (!sched_feat(NUMA_PULL_BIAS))
3871 if (!sds->numa_group)
3874 env->imbalance = sds->numa_group_weight / sds->numa_group_running;
3875 sds->busiest = sds->numa_group;
3876 env->find_busiest_queue = find_busiest_numa_queue;
3880 static inline bool need_active_numa_balance(struct lb_env *env)
3882 return env->find_busiest_queue == find_busiest_numa_queue &&
3883 env->src_rq->offnode_running == 1 &&
3884 env->src_rq->nr_running == 1;
3887 #else /* CONFIG_SCHED_NUMA */
3889 static inline void update_sg_numa_stats(struct sg_lb_stats *sgs, struct rq *rq)
3893 static inline void update_sd_numa_stats(struct sched_domain *sd,
3894 struct sched_group *group, struct sd_lb_stats *sds,
3895 int local_group, struct sg_lb_stats *sgs)
3899 static inline int check_numa_busiest_group(struct lb_env *env, struct sd_lb_stats *sds)
3904 static inline bool need_active_numa_balance(struct lb_env *env)
3908 #endif /* CONFIG_SCHED_NUMA */
3910 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3912 return SCHED_POWER_SCALE;
3915 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3917 return default_scale_freq_power(sd, cpu);
3920 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3922 unsigned long weight = sd->span_weight;
3923 unsigned long smt_gain = sd->smt_gain;
3930 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3932 return default_scale_smt_power(sd, cpu);
3935 unsigned long scale_rt_power(int cpu)
3937 struct rq *rq = cpu_rq(cpu);
3938 u64 total, available, age_stamp, avg;
3941 * Since we're reading these variables without serialization make sure
3942 * we read them once before doing sanity checks on them.
3944 age_stamp = ACCESS_ONCE(rq->age_stamp);
3945 avg = ACCESS_ONCE(rq->rt_avg);
3947 total = sched_avg_period() + (rq->clock - age_stamp);
3949 if (unlikely(total < avg)) {
3950 /* Ensures that power won't end up being negative */
3953 available = total - avg;
3956 if (unlikely((s64)total < SCHED_POWER_SCALE))
3957 total = SCHED_POWER_SCALE;
3959 total >>= SCHED_POWER_SHIFT;
3961 return div_u64(available, total);
3964 static void update_cpu_power(struct sched_domain *sd, int cpu)
3966 unsigned long weight = sd->span_weight;
3967 unsigned long power = SCHED_POWER_SCALE;
3968 struct sched_group *sdg = sd->groups;
3970 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3971 if (sched_feat(ARCH_POWER))
3972 power *= arch_scale_smt_power(sd, cpu);
3974 power *= default_scale_smt_power(sd, cpu);
3976 power >>= SCHED_POWER_SHIFT;
3979 sdg->sgp->power_orig = power;
3981 if (sched_feat(ARCH_POWER))
3982 power *= arch_scale_freq_power(sd, cpu);
3984 power *= default_scale_freq_power(sd, cpu);
3986 power >>= SCHED_POWER_SHIFT;
3988 power *= scale_rt_power(cpu);
3989 power >>= SCHED_POWER_SHIFT;
3994 cpu_rq(cpu)->cpu_power = power;
3995 sdg->sgp->power = power;
3998 void update_group_power(struct sched_domain *sd, int cpu)
4000 struct sched_domain *child = sd->child;
4001 struct sched_group *group, *sdg = sd->groups;
4002 unsigned long power;
4003 unsigned long interval;
4005 interval = msecs_to_jiffies(sd->balance_interval);
4006 interval = clamp(interval, 1UL, max_load_balance_interval);
4007 sdg->sgp->next_update = jiffies + interval;
4010 update_cpu_power(sd, cpu);
4016 if (child->flags & SD_OVERLAP) {
4018 * SD_OVERLAP domains cannot assume that child groups
4019 * span the current group.
4022 for_each_cpu(cpu, sched_group_cpus(sdg))
4023 power += power_of(cpu);
4026 * !SD_OVERLAP domains can assume that child groups
4027 * span the current group.
4030 group = child->groups;
4032 power += group->sgp->power;
4033 group = group->next;
4034 } while (group != child->groups);
4037 sdg->sgp->power_orig = sdg->sgp->power = power;
4041 * Try and fix up capacity for tiny siblings, this is needed when
4042 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4043 * which on its own isn't powerful enough.
4045 * See update_sd_pick_busiest() and check_asym_packing().
4048 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
4051 * Only siblings can have significantly less than SCHED_POWER_SCALE
4053 if (!(sd->flags & SD_SHARE_CPUPOWER))
4057 * If ~90% of the cpu_power is still there, we're good.
4059 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
4066 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4067 * @env: The load balancing environment.
4068 * @group: sched_group whose statistics are to be updated.
4069 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4070 * @local_group: Does group contain this_cpu.
4071 * @balance: Should we balance.
4072 * @sgs: variable to hold the statistics for this group.
4074 static inline void update_sg_lb_stats(struct lb_env *env,
4075 struct sched_group *group, int load_idx,
4076 int local_group, int *balance, struct sg_lb_stats *sgs)
4078 unsigned long nr_running, max_nr_running, min_nr_running;
4079 unsigned long load, max_cpu_load, min_cpu_load;
4080 unsigned int balance_cpu = -1, first_idle_cpu = 0;
4081 unsigned long avg_load_per_task = 0;
4085 balance_cpu = group_balance_cpu(group);
4087 /* Tally up the load of all CPUs in the group */
4089 min_cpu_load = ~0UL;
4091 min_nr_running = ~0UL;
4093 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
4094 struct rq *rq = cpu_rq(i);
4096 nr_running = rq->nr_running;
4098 /* Bias balancing toward cpus of our domain */
4100 if (idle_cpu(i) && !first_idle_cpu &&
4101 cpumask_test_cpu(i, sched_group_mask(group))) {
4106 load = target_load(i, load_idx);
4108 load = source_load(i, load_idx);
4109 if (load > max_cpu_load)
4110 max_cpu_load = load;
4111 if (min_cpu_load > load)
4112 min_cpu_load = load;
4114 if (nr_running > max_nr_running)
4115 max_nr_running = nr_running;
4116 if (min_nr_running > nr_running)
4117 min_nr_running = nr_running;
4120 sgs->group_load += load;
4121 sgs->sum_nr_running += nr_running;
4122 sgs->sum_weighted_load += weighted_cpuload(i);
4126 update_sg_numa_stats(sgs, rq);
4130 * First idle cpu or the first cpu(busiest) in this sched group
4131 * is eligible for doing load balancing at this and above
4132 * domains. In the newly idle case, we will allow all the cpu's
4133 * to do the newly idle load balance.
4136 if (env->idle != CPU_NEWLY_IDLE) {
4137 if (balance_cpu != env->dst_cpu) {
4141 update_group_power(env->sd, env->dst_cpu);
4142 } else if (time_after_eq(jiffies, group->sgp->next_update))
4143 update_group_power(env->sd, env->dst_cpu);
4146 /* Adjust by relative CPU power of the group */
4147 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
4150 * Consider the group unbalanced when the imbalance is larger
4151 * than the average weight of a task.
4153 * APZ: with cgroup the avg task weight can vary wildly and
4154 * might not be a suitable number - should we keep a
4155 * normalized nr_running number somewhere that negates
4158 if (sgs->sum_nr_running)
4159 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
4161 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task &&
4162 (max_nr_running - min_nr_running) > 1)
4165 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
4167 if (!sgs->group_capacity)
4168 sgs->group_capacity = fix_small_capacity(env->sd, group);
4169 sgs->group_weight = group->group_weight;
4171 if (sgs->group_capacity > sgs->sum_nr_running)
4172 sgs->group_has_capacity = 1;
4176 * update_sd_pick_busiest - return 1 on busiest group
4177 * @env: The load balancing environment.
4178 * @sds: sched_domain statistics
4179 * @sg: sched_group candidate to be checked for being the busiest
4180 * @sgs: sched_group statistics
4182 * Determine if @sg is a busier group than the previously selected
4185 static bool update_sd_pick_busiest(struct lb_env *env,
4186 struct sd_lb_stats *sds,
4187 struct sched_group *sg,
4188 struct sg_lb_stats *sgs)
4190 if (sgs->avg_load <= sds->max_load)
4193 if (sgs->sum_nr_running > sgs->group_capacity)
4200 * ASYM_PACKING needs to move all the work to the lowest
4201 * numbered CPUs in the group, therefore mark all groups
4202 * higher than ourself as busy.
4204 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
4205 env->dst_cpu < group_first_cpu(sg)) {
4209 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
4217 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
4218 * @env: The load balancing environment.
4219 * @balance: Should we balance.
4220 * @sds: variable to hold the statistics for this sched_domain.
4222 static inline void update_sd_lb_stats(struct lb_env *env,
4223 int *balance, struct sd_lb_stats *sds)
4225 struct sched_domain *child = env->sd->child;
4226 struct sched_group *sg = env->sd->groups;
4227 struct sg_lb_stats sgs;
4228 int load_idx, prefer_sibling = 0;
4230 if (child && child->flags & SD_PREFER_SIBLING)
4233 load_idx = get_sd_load_idx(env->sd, env->idle);
4238 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
4239 memset(&sgs, 0, sizeof(sgs));
4240 update_sg_lb_stats(env, sg, load_idx, local_group, balance, &sgs);
4242 if (local_group && !(*balance))
4245 sds->total_load += sgs.group_load;
4246 sds->total_pwr += sg->sgp->power;
4249 * In case the child domain prefers tasks go to siblings
4250 * first, lower the sg capacity to one so that we'll try
4251 * and move all the excess tasks away. We lower the capacity
4252 * of a group only if the local group has the capacity to fit
4253 * these excess tasks, i.e. nr_running < group_capacity. The
4254 * extra check prevents the case where you always pull from the
4255 * heaviest group when it is already under-utilized (possible
4256 * with a large weight task outweighs the tasks on the system).
4258 if (prefer_sibling && !local_group && sds->this_has_capacity)
4259 sgs.group_capacity = min(sgs.group_capacity, 1UL);
4262 sds->this_load = sgs.avg_load;
4264 sds->this_nr_running = sgs.sum_nr_running;
4265 sds->this_load_per_task = sgs.sum_weighted_load;
4266 sds->this_has_capacity = sgs.group_has_capacity;
4267 sds->this_idle_cpus = sgs.idle_cpus;
4268 } else if (update_sd_pick_busiest(env, sds, sg, &sgs)) {
4269 sds->max_load = sgs.avg_load;
4271 sds->busiest_nr_running = sgs.sum_nr_running;
4272 sds->busiest_idle_cpus = sgs.idle_cpus;
4273 sds->busiest_group_capacity = sgs.group_capacity;
4274 sds->busiest_load_per_task = sgs.sum_weighted_load;
4275 sds->busiest_has_capacity = sgs.group_has_capacity;
4276 sds->busiest_group_weight = sgs.group_weight;
4277 sds->group_imb = sgs.group_imb;
4280 update_sd_numa_stats(env->sd, sg, sds, local_group, &sgs);
4283 } while (sg != env->sd->groups);
4287 * check_asym_packing - Check to see if the group is packed into the
4290 * This is primarily intended to used at the sibling level. Some
4291 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4292 * case of POWER7, it can move to lower SMT modes only when higher
4293 * threads are idle. When in lower SMT modes, the threads will
4294 * perform better since they share less core resources. Hence when we
4295 * have idle threads, we want them to be the higher ones.
4297 * This packing function is run on idle threads. It checks to see if
4298 * the busiest CPU in this domain (core in the P7 case) has a higher
4299 * CPU number than the packing function is being run on. Here we are
4300 * assuming lower CPU number will be equivalent to lower a SMT thread
4303 * Returns 1 when packing is required and a task should be moved to
4304 * this CPU. The amount of the imbalance is returned in *imbalance.
4306 * @env: The load balancing environment.
4307 * @sds: Statistics of the sched_domain which is to be packed
4309 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
4313 if (!(env->sd->flags & SD_ASYM_PACKING))
4319 busiest_cpu = group_first_cpu(sds->busiest);
4320 if (env->dst_cpu > busiest_cpu)
4323 env->imbalance = DIV_ROUND_CLOSEST(
4324 sds->max_load * sds->busiest->sgp->power, SCHED_POWER_SCALE);
4330 * fix_small_imbalance - Calculate the minor imbalance that exists
4331 * amongst the groups of a sched_domain, during
4333 * @env: The load balancing environment.
4334 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4337 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4339 unsigned long tmp, pwr_now = 0, pwr_move = 0;
4340 unsigned int imbn = 2;
4341 unsigned long scaled_busy_load_per_task;
4343 if (sds->this_nr_running) {
4344 sds->this_load_per_task /= sds->this_nr_running;
4345 if (sds->busiest_load_per_task >
4346 sds->this_load_per_task)
4349 sds->this_load_per_task =
4350 cpu_avg_load_per_task(env->dst_cpu);
4353 scaled_busy_load_per_task = sds->busiest_load_per_task
4354 * SCHED_POWER_SCALE;
4355 scaled_busy_load_per_task /= sds->busiest->sgp->power;
4357 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
4358 (scaled_busy_load_per_task * imbn)) {
4359 env->imbalance = sds->busiest_load_per_task;
4364 * OK, we don't have enough imbalance to justify moving tasks,
4365 * however we may be able to increase total CPU power used by
4369 pwr_now += sds->busiest->sgp->power *
4370 min(sds->busiest_load_per_task, sds->max_load);
4371 pwr_now += sds->this->sgp->power *
4372 min(sds->this_load_per_task, sds->this_load);
4373 pwr_now /= SCHED_POWER_SCALE;
4375 /* Amount of load we'd subtract */
4376 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4377 sds->busiest->sgp->power;
4378 if (sds->max_load > tmp)
4379 pwr_move += sds->busiest->sgp->power *
4380 min(sds->busiest_load_per_task, sds->max_load - tmp);
4382 /* Amount of load we'd add */
4383 if (sds->max_load * sds->busiest->sgp->power <
4384 sds->busiest_load_per_task * SCHED_POWER_SCALE)
4385 tmp = (sds->max_load * sds->busiest->sgp->power) /
4386 sds->this->sgp->power;
4388 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4389 sds->this->sgp->power;
4390 pwr_move += sds->this->sgp->power *
4391 min(sds->this_load_per_task, sds->this_load + tmp);
4392 pwr_move /= SCHED_POWER_SCALE;
4394 /* Move if we gain throughput */
4395 if (pwr_move > pwr_now)
4396 env->imbalance = sds->busiest_load_per_task;
4400 * calculate_imbalance - Calculate the amount of imbalance present within the
4401 * groups of a given sched_domain during load balance.
4402 * @env: load balance environment
4403 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4405 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4407 unsigned long max_pull, load_above_capacity = ~0UL;
4409 sds->busiest_load_per_task /= sds->busiest_nr_running;
4410 if (sds->group_imb) {
4411 sds->busiest_load_per_task =
4412 min(sds->busiest_load_per_task, sds->avg_load);
4416 * In the presence of smp nice balancing, certain scenarios can have
4417 * max load less than avg load(as we skip the groups at or below
4418 * its cpu_power, while calculating max_load..)
4420 if (sds->max_load < sds->avg_load) {
4422 return fix_small_imbalance(env, sds);
4425 if (!sds->group_imb) {
4427 * Don't want to pull so many tasks that a group would go idle.
4429 load_above_capacity = (sds->busiest_nr_running -
4430 sds->busiest_group_capacity);
4432 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
4434 load_above_capacity /= sds->busiest->sgp->power;
4438 * We're trying to get all the cpus to the average_load, so we don't
4439 * want to push ourselves above the average load, nor do we wish to
4440 * reduce the max loaded cpu below the average load. At the same time,
4441 * we also don't want to reduce the group load below the group capacity
4442 * (so that we can implement power-savings policies etc). Thus we look
4443 * for the minimum possible imbalance.
4444 * Be careful of negative numbers as they'll appear as very large values
4445 * with unsigned longs.
4447 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
4449 /* How much load to actually move to equalise the imbalance */
4450 env->imbalance = min(max_pull * sds->busiest->sgp->power,
4451 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
4452 / SCHED_POWER_SCALE;
4455 * if *imbalance is less than the average load per runnable task
4456 * there is no guarantee that any tasks will be moved so we'll have
4457 * a think about bumping its value to force at least one task to be
4460 if (env->imbalance < sds->busiest_load_per_task)
4461 return fix_small_imbalance(env, sds);
4465 /******* find_busiest_group() helpers end here *********************/
4468 * find_busiest_group - Returns the busiest group within the sched_domain
4469 * if there is an imbalance. If there isn't an imbalance, and
4470 * the user has opted for power-savings, it returns a group whose
4471 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4472 * such a group exists.
4474 * Also calculates the amount of weighted load which should be moved
4475 * to restore balance.
4477 * @env: The load balancing environment.
4478 * @balance: Pointer to a variable indicating if this_cpu
4479 * is the appropriate cpu to perform load balancing at this_level.
4481 * Returns: - the busiest group if imbalance exists.
4482 * - If no imbalance and user has opted for power-savings balance,
4483 * return the least loaded group whose CPUs can be
4484 * put to idle by rebalancing its tasks onto our group.
4486 static struct sched_group *
4487 find_busiest_group(struct lb_env *env, int *balance)
4489 struct sd_lb_stats sds;
4491 memset(&sds, 0, sizeof(sds));
4494 * Compute the various statistics relavent for load balancing at
4497 update_sd_lb_stats(env, balance, &sds);
4500 * this_cpu is not the appropriate cpu to perform load balancing at
4506 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
4507 check_asym_packing(env, &sds))
4510 /* There is no busy sibling group to pull tasks from */
4511 if (!sds.busiest || sds.busiest_nr_running == 0)
4514 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
4517 * If the busiest group is imbalanced the below checks don't
4518 * work because they assumes all things are equal, which typically
4519 * isn't true due to cpus_allowed constraints and the like.
4524 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
4525 if (env->idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
4526 !sds.busiest_has_capacity)
4530 * If the local group is more busy than the selected busiest group
4531 * don't try and pull any tasks.
4533 if (sds.this_load >= sds.max_load)
4537 * Don't pull any tasks if this group is already above the domain
4540 if (sds.this_load >= sds.avg_load)
4543 if (env->idle == CPU_IDLE) {
4545 * This cpu is idle. If the busiest group load doesn't
4546 * have more tasks than the number of available cpu's and
4547 * there is no imbalance between this and busiest group
4548 * wrt to idle cpu's, it is balanced.
4550 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
4551 sds.busiest_nr_running <= sds.busiest_group_weight)
4555 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
4556 * imbalance_pct to be conservative.
4558 if (100 * sds.max_load <= env->sd->imbalance_pct * sds.this_load)
4563 /* Looks like there is an imbalance. Compute it */
4564 calculate_imbalance(env, &sds);
4568 if (check_numa_busiest_group(env, &sds))
4577 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4579 static struct rq *find_busiest_queue(struct lb_env *env,
4580 struct sched_group *group)
4582 struct rq *busiest = NULL, *rq;
4583 unsigned long max_load = 0;
4586 for_each_cpu(i, sched_group_cpus(group)) {
4587 unsigned long power = power_of(i);
4588 unsigned long capacity = DIV_ROUND_CLOSEST(power,
4593 capacity = fix_small_capacity(env->sd, group);
4595 if (!cpumask_test_cpu(i, env->cpus))
4599 wl = weighted_cpuload(i);
4602 * When comparing with imbalance, use weighted_cpuload()
4603 * which is not scaled with the cpu power.
4605 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
4609 * For the load comparisons with the other cpu's, consider
4610 * the weighted_cpuload() scaled with the cpu power, so that
4611 * the load can be moved away from the cpu that is potentially
4612 * running at a lower capacity.
4614 wl = (wl * SCHED_POWER_SCALE) / power;
4616 if (wl > max_load) {
4626 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4627 * so long as it is large enough.
4629 #define MAX_PINNED_INTERVAL 512
4631 /* Working cpumask for load_balance and load_balance_newidle. */
4632 DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4634 static int need_active_balance(struct lb_env *env)
4636 struct sched_domain *sd = env->sd;
4638 if (env->idle == CPU_NEWLY_IDLE) {
4641 * ASYM_PACKING needs to force migrate tasks from busy but
4642 * higher numbered CPUs in order to pack all tasks in the
4643 * lowest numbered CPUs.
4645 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
4649 if (need_active_numa_balance(env))
4652 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
4655 static int active_load_balance_cpu_stop(void *data);
4658 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4659 * tasks if there is an imbalance.
4661 static int load_balance(int this_cpu, struct rq *this_rq,
4662 struct sched_domain *sd, enum cpu_idle_type idle,
4665 int ld_moved, cur_ld_moved, active_balance = 0;
4666 int lb_iterations, max_lb_iterations;
4667 struct sched_group *group;
4669 unsigned long flags;
4670 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4672 struct lb_env env = {
4674 .dst_cpu = this_cpu,
4676 .dst_grpmask = sched_group_cpus(sd->groups),
4678 .loop_break = sched_nr_migrate_break,
4680 .find_busiest_queue = find_busiest_queue,
4683 cpumask_copy(cpus, cpu_active_mask);
4684 max_lb_iterations = cpumask_weight(env.dst_grpmask);
4686 schedstat_inc(sd, lb_count[idle]);
4689 group = find_busiest_group(&env, balance);
4695 schedstat_inc(sd, lb_nobusyg[idle]);
4699 busiest = env.find_busiest_queue(&env, group);
4701 schedstat_inc(sd, lb_nobusyq[idle]);
4704 env.src_rq = busiest;
4705 env.src_cpu = busiest->cpu;
4707 BUG_ON(busiest == env.dst_rq);
4709 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
4713 if (busiest->nr_running > 1) {
4715 * Attempt to move tasks. If find_busiest_group has found
4716 * an imbalance but busiest->nr_running <= 1, the group is
4717 * still unbalanced. ld_moved simply stays zero, so it is
4718 * correctly treated as an imbalance.
4720 env.flags |= LBF_ALL_PINNED;
4721 env.src_cpu = busiest->cpu;
4722 env.src_rq = busiest;
4723 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
4724 if (sched_feat_numa(NUMA_PULL))
4725 env.tasks = offnode_tasks(busiest);
4727 env.tasks = &busiest->cfs_tasks;
4729 update_h_load(env.src_cpu);
4731 local_irq_save(flags);
4732 double_rq_lock(env.dst_rq, busiest);
4735 * cur_ld_moved - load moved in current iteration
4736 * ld_moved - cumulative load moved across iterations
4738 cur_ld_moved = move_tasks(&env);
4739 ld_moved += cur_ld_moved;
4740 double_rq_unlock(env.dst_rq, busiest);
4741 local_irq_restore(flags);
4743 if (env.flags & LBF_NEED_BREAK) {
4744 env.flags &= ~LBF_NEED_BREAK;
4749 * some other cpu did the load balance for us.
4751 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
4752 resched_cpu(env.dst_cpu);
4755 * Revisit (affine) tasks on src_cpu that couldn't be moved to
4756 * us and move them to an alternate dst_cpu in our sched_group
4757 * where they can run. The upper limit on how many times we
4758 * iterate on same src_cpu is dependent on number of cpus in our
4761 * This changes load balance semantics a bit on who can move
4762 * load to a given_cpu. In addition to the given_cpu itself
4763 * (or a ilb_cpu acting on its behalf where given_cpu is
4764 * nohz-idle), we now have balance_cpu in a position to move
4765 * load to given_cpu. In rare situations, this may cause
4766 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
4767 * _independently_ and at _same_ time to move some load to
4768 * given_cpu) causing exceess load to be moved to given_cpu.
4769 * This however should not happen so much in practice and
4770 * moreover subsequent load balance cycles should correct the
4771 * excess load moved.
4773 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0 &&
4774 lb_iterations++ < max_lb_iterations) {
4776 env.dst_rq = cpu_rq(env.new_dst_cpu);
4777 env.dst_cpu = env.new_dst_cpu;
4778 env.flags &= ~LBF_SOME_PINNED;
4780 env.loop_break = sched_nr_migrate_break;
4782 * Go back to "more_balance" rather than "redo" since we
4783 * need to continue with same src_cpu.
4788 /* All tasks on this runqueue were pinned by CPU affinity */
4789 if (unlikely(env.flags & LBF_ALL_PINNED)) {
4790 cpumask_clear_cpu(cpu_of(busiest), cpus);
4791 if (!cpumask_empty(cpus)) {
4793 env.loop_break = sched_nr_migrate_break;
4801 schedstat_inc(sd, lb_failed[idle]);
4803 * Increment the failure counter only on periodic balance.
4804 * We do not want newidle balance, which can be very
4805 * frequent, pollute the failure counter causing
4806 * excessive cache_hot migrations and active balances.
4808 if (idle != CPU_NEWLY_IDLE)
4809 sd->nr_balance_failed++;
4811 if (need_active_balance(&env)) {
4812 raw_spin_lock_irqsave(&busiest->lock, flags);
4814 /* don't kick the active_load_balance_cpu_stop,
4815 * if the curr task on busiest cpu can't be
4818 if (!cpumask_test_cpu(this_cpu,
4819 tsk_cpus_allowed(busiest->curr))) {
4820 raw_spin_unlock_irqrestore(&busiest->lock,
4822 env.flags |= LBF_ALL_PINNED;
4823 goto out_one_pinned;
4827 * ->active_balance synchronizes accesses to
4828 * ->active_balance_work. Once set, it's cleared
4829 * only after active load balance is finished.
4831 if (!busiest->active_balance) {
4832 busiest->active_balance = 1;
4833 busiest->push_cpu = this_cpu;
4836 raw_spin_unlock_irqrestore(&busiest->lock, flags);
4838 if (active_balance) {
4839 stop_one_cpu_nowait(cpu_of(busiest),
4840 active_load_balance_cpu_stop, busiest,
4841 &busiest->active_balance_work);
4845 * We've kicked active balancing, reset the failure
4848 sd->nr_balance_failed = sd->cache_nice_tries+1;
4851 sd->nr_balance_failed = 0;
4853 if (likely(!active_balance)) {
4854 /* We were unbalanced, so reset the balancing interval */
4855 sd->balance_interval = sd->min_interval;
4858 * If we've begun active balancing, start to back off. This
4859 * case may not be covered by the all_pinned logic if there
4860 * is only 1 task on the busy runqueue (because we don't call
4863 if (sd->balance_interval < sd->max_interval)
4864 sd->balance_interval *= 2;
4870 schedstat_inc(sd, lb_balanced[idle]);
4872 sd->nr_balance_failed = 0;
4875 /* tune up the balancing interval */
4876 if (((env.flags & LBF_ALL_PINNED) &&
4877 sd->balance_interval < MAX_PINNED_INTERVAL) ||
4878 (sd->balance_interval < sd->max_interval))
4879 sd->balance_interval *= 2;
4887 * idle_balance is called by schedule() if this_cpu is about to become
4888 * idle. Attempts to pull tasks from other CPUs.
4890 void idle_balance(int this_cpu, struct rq *this_rq)
4892 struct sched_domain *sd;
4893 int pulled_task = 0;
4894 unsigned long next_balance = jiffies + HZ;
4896 this_rq->idle_stamp = this_rq->clock;
4898 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4902 * Drop the rq->lock, but keep IRQ/preempt disabled.
4904 raw_spin_unlock(&this_rq->lock);
4906 update_shares(this_cpu);
4908 for_each_domain(this_cpu, sd) {
4909 unsigned long interval;
4912 if (!(sd->flags & SD_LOAD_BALANCE))
4915 if (sd->flags & SD_BALANCE_NEWIDLE) {
4916 /* If we've pulled tasks over stop searching: */
4917 pulled_task = load_balance(this_cpu, this_rq,
4918 sd, CPU_NEWLY_IDLE, &balance);
4921 interval = msecs_to_jiffies(sd->balance_interval);
4922 if (time_after(next_balance, sd->last_balance + interval))
4923 next_balance = sd->last_balance + interval;
4925 this_rq->idle_stamp = 0;
4931 raw_spin_lock(&this_rq->lock);
4933 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4935 * We are going idle. next_balance may be set based on
4936 * a busy processor. So reset next_balance.
4938 this_rq->next_balance = next_balance;
4943 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
4944 * running tasks off the busiest CPU onto idle CPUs. It requires at
4945 * least 1 task to be running on each physical CPU where possible, and
4946 * avoids physical / logical imbalances.
4948 static int active_load_balance_cpu_stop(void *data)
4950 struct rq *busiest_rq = data;
4951 int busiest_cpu = cpu_of(busiest_rq);
4952 int target_cpu = busiest_rq->push_cpu;
4953 struct rq *target_rq = cpu_rq(target_cpu);
4954 struct sched_domain *sd;
4956 raw_spin_lock_irq(&busiest_rq->lock);
4958 /* make sure the requested cpu hasn't gone down in the meantime */
4959 if (unlikely(busiest_cpu != smp_processor_id() ||
4960 !busiest_rq->active_balance))
4963 /* Is there any task to move? */
4964 if (busiest_rq->nr_running <= 1)
4968 * This condition is "impossible", if it occurs
4969 * we need to fix it. Originally reported by
4970 * Bjorn Helgaas on a 128-cpu setup.
4972 BUG_ON(busiest_rq == target_rq);
4974 /* move a task from busiest_rq to target_rq */
4975 double_lock_balance(busiest_rq, target_rq);
4977 /* Search for an sd spanning us and the target CPU. */
4979 for_each_domain(target_cpu, sd) {
4980 if ((sd->flags & SD_LOAD_BALANCE) &&
4981 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4986 struct lb_env env = {
4988 .dst_cpu = target_cpu,
4989 .dst_rq = target_rq,
4990 .src_cpu = busiest_rq->cpu,
4991 .src_rq = busiest_rq,
4995 schedstat_inc(sd, alb_count);
4997 if (move_one_task(&env))
4998 schedstat_inc(sd, alb_pushed);
5000 schedstat_inc(sd, alb_failed);
5003 double_unlock_balance(busiest_rq, target_rq);
5005 busiest_rq->active_balance = 0;
5006 raw_spin_unlock_irq(&busiest_rq->lock);
5012 * idle load balancing details
5013 * - When one of the busy CPUs notice that there may be an idle rebalancing
5014 * needed, they will kick the idle load balancer, which then does idle
5015 * load balancing for all the idle CPUs.
5018 cpumask_var_t idle_cpus_mask;
5020 unsigned long next_balance; /* in jiffy units */
5021 } nohz ____cacheline_aligned;
5023 static inline int find_new_ilb(int call_cpu)
5025 int ilb = cpumask_first(nohz.idle_cpus_mask);
5027 if (ilb < nr_cpu_ids && idle_cpu(ilb))
5034 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5035 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5036 * CPU (if there is one).
5038 static void nohz_balancer_kick(int cpu)
5042 nohz.next_balance++;
5044 ilb_cpu = find_new_ilb(cpu);
5046 if (ilb_cpu >= nr_cpu_ids)
5049 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
5052 * Use smp_send_reschedule() instead of resched_cpu().
5053 * This way we generate a sched IPI on the target cpu which
5054 * is idle. And the softirq performing nohz idle load balance
5055 * will be run before returning from the IPI.
5057 smp_send_reschedule(ilb_cpu);
5061 static inline void nohz_balance_exit_idle(int cpu)
5063 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5064 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5065 atomic_dec(&nohz.nr_cpus);
5066 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5070 static inline void set_cpu_sd_state_busy(void)
5072 struct sched_domain *sd;
5073 int cpu = smp_processor_id();
5075 if (!test_bit(NOHZ_IDLE, nohz_flags(cpu)))
5077 clear_bit(NOHZ_IDLE, nohz_flags(cpu));
5080 for_each_domain(cpu, sd)
5081 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
5085 void set_cpu_sd_state_idle(void)
5087 struct sched_domain *sd;
5088 int cpu = smp_processor_id();
5090 if (test_bit(NOHZ_IDLE, nohz_flags(cpu)))
5092 set_bit(NOHZ_IDLE, nohz_flags(cpu));
5095 for_each_domain(cpu, sd)
5096 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
5101 * This routine will record that the cpu is going idle with tick stopped.
5102 * This info will be used in performing idle load balancing in the future.
5104 void nohz_balance_enter_idle(int cpu)
5107 * If this cpu is going down, then nothing needs to be done.
5109 if (!cpu_active(cpu))
5112 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
5115 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
5116 atomic_inc(&nohz.nr_cpus);
5117 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5120 static int __cpuinit sched_ilb_notifier(struct notifier_block *nfb,
5121 unsigned long action, void *hcpu)
5123 switch (action & ~CPU_TASKS_FROZEN) {
5125 nohz_balance_exit_idle(smp_processor_id());
5133 static DEFINE_SPINLOCK(balancing);
5136 * Scale the max load_balance interval with the number of CPUs in the system.
5137 * This trades load-balance latency on larger machines for less cross talk.
5139 void update_max_interval(void)
5141 max_load_balance_interval = HZ*num_online_cpus()/10;
5145 * It checks each scheduling domain to see if it is due to be balanced,
5146 * and initiates a balancing operation if so.
5148 * Balancing parameters are set up in arch_init_sched_domains.
5150 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
5153 struct rq *rq = cpu_rq(cpu);
5154 unsigned long interval;
5155 struct sched_domain *sd;
5156 /* Earliest time when we have to do rebalance again */
5157 unsigned long next_balance = jiffies + 60*HZ;
5158 int update_next_balance = 0;
5164 for_each_domain(cpu, sd) {
5165 if (!(sd->flags & SD_LOAD_BALANCE))
5168 interval = sd->balance_interval;
5169 if (idle != CPU_IDLE)
5170 interval *= sd->busy_factor;
5172 /* scale ms to jiffies */
5173 interval = msecs_to_jiffies(interval);
5174 interval = clamp(interval, 1UL, max_load_balance_interval);
5176 need_serialize = sd->flags & SD_SERIALIZE;
5178 if (need_serialize) {
5179 if (!spin_trylock(&balancing))
5183 if (time_after_eq(jiffies, sd->last_balance + interval)) {
5184 if (load_balance(cpu, rq, sd, idle, &balance)) {
5186 * We've pulled tasks over so either we're no
5189 idle = CPU_NOT_IDLE;
5191 sd->last_balance = jiffies;
5194 spin_unlock(&balancing);
5196 if (time_after(next_balance, sd->last_balance + interval)) {
5197 next_balance = sd->last_balance + interval;
5198 update_next_balance = 1;
5202 * Stop the load balance at this level. There is another
5203 * CPU in our sched group which is doing load balancing more
5212 * next_balance will be updated only when there is a need.
5213 * When the cpu is attached to null domain for ex, it will not be
5216 if (likely(update_next_balance))
5217 rq->next_balance = next_balance;
5222 * In CONFIG_NO_HZ case, the idle balance kickee will do the
5223 * rebalancing for all the cpus for whom scheduler ticks are stopped.
5225 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
5227 struct rq *this_rq = cpu_rq(this_cpu);
5231 if (idle != CPU_IDLE ||
5232 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
5235 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
5236 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
5240 * If this cpu gets work to do, stop the load balancing
5241 * work being done for other cpus. Next load
5242 * balancing owner will pick it up.
5247 rq = cpu_rq(balance_cpu);
5249 raw_spin_lock_irq(&rq->lock);
5250 update_rq_clock(rq);
5251 update_idle_cpu_load(rq);
5252 raw_spin_unlock_irq(&rq->lock);
5254 rebalance_domains(balance_cpu, CPU_IDLE);
5256 if (time_after(this_rq->next_balance, rq->next_balance))
5257 this_rq->next_balance = rq->next_balance;
5259 nohz.next_balance = this_rq->next_balance;
5261 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
5265 * Current heuristic for kicking the idle load balancer in the presence
5266 * of an idle cpu is the system.
5267 * - This rq has more than one task.
5268 * - At any scheduler domain level, this cpu's scheduler group has multiple
5269 * busy cpu's exceeding the group's power.
5270 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
5271 * domain span are idle.
5273 static inline int nohz_kick_needed(struct rq *rq, int cpu)
5275 unsigned long now = jiffies;
5276 struct sched_domain *sd;
5278 if (unlikely(idle_cpu(cpu)))
5282 * We may be recently in ticked or tickless idle mode. At the first
5283 * busy tick after returning from idle, we will update the busy stats.
5285 set_cpu_sd_state_busy();
5286 nohz_balance_exit_idle(cpu);
5289 * None are in tickless mode and hence no need for NOHZ idle load
5292 if (likely(!atomic_read(&nohz.nr_cpus)))
5295 if (time_before(now, nohz.next_balance))
5298 if (rq->nr_running >= 2)
5302 for_each_domain(cpu, sd) {
5303 struct sched_group *sg = sd->groups;
5304 struct sched_group_power *sgp = sg->sgp;
5305 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
5307 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
5308 goto need_kick_unlock;
5310 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
5311 && (cpumask_first_and(nohz.idle_cpus_mask,
5312 sched_domain_span(sd)) < cpu))
5313 goto need_kick_unlock;
5315 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
5327 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
5331 * run_rebalance_domains is triggered when needed from the scheduler tick.
5332 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
5334 static void run_rebalance_domains(struct softirq_action *h)
5336 int this_cpu = smp_processor_id();
5337 struct rq *this_rq = cpu_rq(this_cpu);
5338 enum cpu_idle_type idle = this_rq->idle_balance ?
5339 CPU_IDLE : CPU_NOT_IDLE;
5341 rebalance_domains(this_cpu, idle);
5344 * If this cpu has a pending nohz_balance_kick, then do the
5345 * balancing on behalf of the other idle cpus whose ticks are
5348 nohz_idle_balance(this_cpu, idle);
5351 static inline int on_null_domain(int cpu)
5353 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
5357 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
5359 void trigger_load_balance(struct rq *rq, int cpu)
5361 /* Don't need to rebalance while attached to NULL domain */
5362 if (time_after_eq(jiffies, rq->next_balance) &&
5363 likely(!on_null_domain(cpu)))
5364 raise_softirq(SCHED_SOFTIRQ);
5366 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
5367 nohz_balancer_kick(cpu);
5371 static void rq_online_fair(struct rq *rq)
5376 static void rq_offline_fair(struct rq *rq)
5380 /* Ensure any throttled groups are reachable by pick_next_task */
5381 unthrottle_offline_cfs_rqs(rq);
5384 #endif /* CONFIG_SMP */
5387 * scheduler tick hitting a task of our scheduling class:
5389 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
5391 struct cfs_rq *cfs_rq;
5392 struct sched_entity *se = &curr->se;
5394 for_each_sched_entity(se) {
5395 cfs_rq = cfs_rq_of(se);
5396 entity_tick(cfs_rq, se, queued);
5399 if (sched_feat_numa(NUMA))
5400 task_tick_numa(rq, curr);
5404 * called on fork with the child task as argument from the parent's context
5405 * - child not yet on the tasklist
5406 * - preemption disabled
5408 static void task_fork_fair(struct task_struct *p)
5410 struct cfs_rq *cfs_rq;
5411 struct sched_entity *se = &p->se, *curr;
5412 int this_cpu = smp_processor_id();
5413 struct rq *rq = this_rq();
5414 unsigned long flags;
5416 raw_spin_lock_irqsave(&rq->lock, flags);
5418 update_rq_clock(rq);
5420 cfs_rq = task_cfs_rq(current);
5421 curr = cfs_rq->curr;
5423 if (unlikely(task_cpu(p) != this_cpu)) {
5425 __set_task_cpu(p, this_cpu);
5429 update_curr(cfs_rq);
5432 se->vruntime = curr->vruntime;
5433 place_entity(cfs_rq, se, 1);
5435 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
5437 * Upon rescheduling, sched_class::put_prev_task() will place
5438 * 'current' within the tree based on its new key value.
5440 swap(curr->vruntime, se->vruntime);
5441 resched_task(rq->curr);
5444 se->vruntime -= cfs_rq->min_vruntime;
5446 raw_spin_unlock_irqrestore(&rq->lock, flags);
5450 * Priority of the task has changed. Check to see if we preempt
5454 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
5460 * Reschedule if we are currently running on this runqueue and
5461 * our priority decreased, or if we are not currently running on
5462 * this runqueue and our priority is higher than the current's
5464 if (rq->curr == p) {
5465 if (p->prio > oldprio)
5466 resched_task(rq->curr);
5468 check_preempt_curr(rq, p, 0);
5471 static void switched_from_fair(struct rq *rq, struct task_struct *p)
5473 struct sched_entity *se = &p->se;
5474 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5477 * Ensure the task's vruntime is normalized, so that when its
5478 * switched back to the fair class the enqueue_entity(.flags=0) will
5479 * do the right thing.
5481 * If it was on_rq, then the dequeue_entity(.flags=0) will already
5482 * have normalized the vruntime, if it was !on_rq, then only when
5483 * the task is sleeping will it still have non-normalized vruntime.
5485 if (!se->on_rq && p->state != TASK_RUNNING) {
5487 * Fix up our vruntime so that the current sleep doesn't
5488 * cause 'unlimited' sleep bonus.
5490 place_entity(cfs_rq, se, 0);
5491 se->vruntime -= cfs_rq->min_vruntime;
5496 * We switched to the sched_fair class.
5498 static void switched_to_fair(struct rq *rq, struct task_struct *p)
5504 * We were most likely switched from sched_rt, so
5505 * kick off the schedule if running, otherwise just see
5506 * if we can still preempt the current task.
5509 resched_task(rq->curr);
5511 check_preempt_curr(rq, p, 0);
5514 /* Account for a task changing its policy or group.
5516 * This routine is mostly called to set cfs_rq->curr field when a task
5517 * migrates between groups/classes.
5519 static void set_curr_task_fair(struct rq *rq)
5521 struct sched_entity *se = &rq->curr->se;
5523 for_each_sched_entity(se) {
5524 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5526 set_next_entity(cfs_rq, se);
5527 /* ensure bandwidth has been allocated on our new cfs_rq */
5528 account_cfs_rq_runtime(cfs_rq, 0);
5532 void init_cfs_rq(struct cfs_rq *cfs_rq)
5534 cfs_rq->tasks_timeline = RB_ROOT;
5535 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
5536 #ifndef CONFIG_64BIT
5537 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
5541 #ifdef CONFIG_FAIR_GROUP_SCHED
5542 static void task_move_group_fair(struct task_struct *p, int on_rq)
5545 * If the task was not on the rq at the time of this cgroup movement
5546 * it must have been asleep, sleeping tasks keep their ->vruntime
5547 * absolute on their old rq until wakeup (needed for the fair sleeper
5548 * bonus in place_entity()).
5550 * If it was on the rq, we've just 'preempted' it, which does convert
5551 * ->vruntime to a relative base.
5553 * Make sure both cases convert their relative position when migrating
5554 * to another cgroup's rq. This does somewhat interfere with the
5555 * fair sleeper stuff for the first placement, but who cares.
5558 * When !on_rq, vruntime of the task has usually NOT been normalized.
5559 * But there are some cases where it has already been normalized:
5561 * - Moving a forked child which is waiting for being woken up by
5562 * wake_up_new_task().
5563 * - Moving a task which has been woken up by try_to_wake_up() and
5564 * waiting for actually being woken up by sched_ttwu_pending().
5566 * To prevent boost or penalty in the new cfs_rq caused by delta
5567 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
5569 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
5573 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
5574 set_task_rq(p, task_cpu(p));
5576 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
5579 void free_fair_sched_group(struct task_group *tg)
5583 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
5585 for_each_possible_cpu(i) {
5587 kfree(tg->cfs_rq[i]);
5596 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
5598 struct cfs_rq *cfs_rq;
5599 struct sched_entity *se;
5602 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
5605 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
5609 tg->shares = NICE_0_LOAD;
5611 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
5613 for_each_possible_cpu(i) {
5614 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
5615 GFP_KERNEL, cpu_to_node(i));
5619 se = kzalloc_node(sizeof(struct sched_entity),
5620 GFP_KERNEL, cpu_to_node(i));
5624 init_cfs_rq(cfs_rq);
5625 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
5636 void unregister_fair_sched_group(struct task_group *tg, int cpu)
5638 struct rq *rq = cpu_rq(cpu);
5639 unsigned long flags;
5642 * Only empty task groups can be destroyed; so we can speculatively
5643 * check on_list without danger of it being re-added.
5645 if (!tg->cfs_rq[cpu]->on_list)
5648 raw_spin_lock_irqsave(&rq->lock, flags);
5649 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
5650 raw_spin_unlock_irqrestore(&rq->lock, flags);
5653 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
5654 struct sched_entity *se, int cpu,
5655 struct sched_entity *parent)
5657 struct rq *rq = cpu_rq(cpu);
5662 /* allow initial update_cfs_load() to truncate */
5663 cfs_rq->load_stamp = 1;
5665 init_cfs_rq_runtime(cfs_rq);
5667 tg->cfs_rq[cpu] = cfs_rq;
5670 /* se could be NULL for root_task_group */
5675 se->cfs_rq = &rq->cfs;
5677 se->cfs_rq = parent->my_q;
5680 update_load_set(&se->load, 0);
5681 se->parent = parent;
5684 static DEFINE_MUTEX(shares_mutex);
5686 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
5689 unsigned long flags;
5692 * We can't change the weight of the root cgroup.
5697 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
5699 mutex_lock(&shares_mutex);
5700 if (tg->shares == shares)
5703 tg->shares = shares;
5704 for_each_possible_cpu(i) {
5705 struct rq *rq = cpu_rq(i);
5706 struct sched_entity *se;
5709 /* Propagate contribution to hierarchy */
5710 raw_spin_lock_irqsave(&rq->lock, flags);
5711 for_each_sched_entity(se)
5712 update_cfs_shares(group_cfs_rq(se));
5713 raw_spin_unlock_irqrestore(&rq->lock, flags);
5717 mutex_unlock(&shares_mutex);
5720 #else /* CONFIG_FAIR_GROUP_SCHED */
5722 void free_fair_sched_group(struct task_group *tg) { }
5724 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
5729 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
5731 #endif /* CONFIG_FAIR_GROUP_SCHED */
5734 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
5736 struct sched_entity *se = &task->se;
5737 unsigned int rr_interval = 0;
5740 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
5743 if (rq->cfs.load.weight)
5744 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5750 * All the scheduling class methods:
5752 const struct sched_class fair_sched_class = {
5753 .next = &idle_sched_class,
5754 .enqueue_task = enqueue_task_fair,
5755 .dequeue_task = dequeue_task_fair,
5756 .yield_task = yield_task_fair,
5757 .yield_to_task = yield_to_task_fair,
5759 .check_preempt_curr = check_preempt_wakeup,
5761 .pick_next_task = pick_next_task_fair,
5762 .put_prev_task = put_prev_task_fair,
5765 .select_task_rq = select_task_rq_fair,
5767 .rq_online = rq_online_fair,
5768 .rq_offline = rq_offline_fair,
5770 .task_waking = task_waking_fair,
5773 .set_curr_task = set_curr_task_fair,
5774 .task_tick = task_tick_fair,
5775 .task_fork = task_fork_fair,
5777 .prio_changed = prio_changed_fair,
5778 .switched_from = switched_from_fair,
5779 .switched_to = switched_to_fair,
5781 .get_rr_interval = get_rr_interval_fair,
5783 #ifdef CONFIG_FAIR_GROUP_SCHED
5784 .task_move_group = task_move_group_fair,
5788 #ifdef CONFIG_SCHED_DEBUG
5789 void print_cfs_stats(struct seq_file *m, int cpu)
5791 struct cfs_rq *cfs_rq;
5794 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
5795 print_cfs_rq(m, cpu, cfs_rq);
5800 __init void init_sched_fair_class(void)
5803 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
5806 nohz.next_balance = jiffies;
5807 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
5808 cpu_notifier(sched_ilb_notifier, 0);