2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/slab.h>
27 #include <linux/profile.h>
28 #include <linux/interrupt.h>
29 #include <linux/mempolicy.h>
30 #include <linux/migrate.h>
31 #include <linux/task_work.h>
33 #include <trace/events/sched.h>
38 * Targeted preemption latency for CPU-bound tasks:
39 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
41 * NOTE: this latency value is not the same as the concept of
42 * 'timeslice length' - timeslices in CFS are of variable length
43 * and have no persistent notion like in traditional, time-slice
44 * based scheduling concepts.
46 * (to see the precise effective timeslice length of your workload,
47 * run vmstat and monitor the context-switches (cs) field)
49 unsigned int sysctl_sched_latency = 6000000ULL;
50 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
53 * The initial- and re-scaling of tunables is configurable
54 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
57 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
58 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
59 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
61 enum sched_tunable_scaling sysctl_sched_tunable_scaling
62 = SCHED_TUNABLESCALING_LOG;
65 * Minimal preemption granularity for CPU-bound tasks:
66 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 unsigned int sysctl_sched_min_granularity = 750000ULL;
69 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
72 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
74 static unsigned int sched_nr_latency = 8;
77 * After fork, child runs first. If set to 0 (default) then
78 * parent will (try to) run first.
80 unsigned int sysctl_sched_child_runs_first __read_mostly;
83 * SCHED_OTHER wake-up granularity.
84 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
86 * This option delays the preemption effects of decoupled workloads
87 * and reduces their over-scheduling. Synchronous workloads will still
88 * have immediate wakeup/sleep latencies.
90 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
91 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
93 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
96 * The exponential sliding window over which load is averaged for shares
100 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
102 #ifdef CONFIG_CFS_BANDWIDTH
104 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
105 * each time a cfs_rq requests quota.
107 * Note: in the case that the slice exceeds the runtime remaining (either due
108 * to consumption or the quota being specified to be smaller than the slice)
109 * we will always only issue the remaining available time.
111 * default: 5 msec, units: microseconds
113 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
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 queueing methods:
782 #ifdef CONFIG_NUMA_BALANCING
784 * numa task sample period in ms
786 unsigned int sysctl_numa_balancing_scan_period_min = 100;
787 unsigned int sysctl_numa_balancing_scan_period_max = 100*16;
789 /* Portion of address space to scan in MB */
790 unsigned int sysctl_numa_balancing_scan_size = 256;
792 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
793 unsigned int sysctl_numa_balancing_scan_delay = 1000;
795 static void task_numa_placement(struct task_struct *p)
797 int seq = ACCESS_ONCE(p->mm->numa_scan_seq);
799 if (p->numa_scan_seq == seq)
801 p->numa_scan_seq = seq;
803 /* FIXME: Scheduling placement policy hints go here */
807 * Got a PROT_NONE fault for a page on @node.
809 void task_numa_fault(int node, int pages)
811 struct task_struct *p = current;
813 /* FIXME: Allocate task-specific structure for placement policy here */
816 * Assume that as faults occur that pages are getting properly placed
817 * and fewer NUMA hints are required. Note that this is a big
818 * assumption, it assumes processes reach a steady steady with no
819 * further phase changes.
821 p->numa_scan_period = min(sysctl_numa_balancing_scan_period_max,
822 p->numa_scan_period + jiffies_to_msecs(2));
824 task_numa_placement(p);
827 static void reset_ptenuma_scan(struct task_struct *p)
829 ACCESS_ONCE(p->mm->numa_scan_seq)++;
830 p->mm->numa_scan_offset = 0;
834 * The expensive part of numa migration is done from task_work context.
835 * Triggered from task_tick_numa().
837 void task_numa_work(struct callback_head *work)
839 unsigned long migrate, next_scan, now = jiffies;
840 struct task_struct *p = current;
841 struct mm_struct *mm = p->mm;
842 struct vm_area_struct *vma;
843 unsigned long start, end;
846 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
848 work->next = work; /* protect against double add */
850 * Who cares about NUMA placement when they're dying.
852 * NOTE: make sure not to dereference p->mm before this check,
853 * exit_task_work() happens _after_ exit_mm() so we could be called
854 * without p->mm even though we still had it when we enqueued this
857 if (p->flags & PF_EXITING)
861 * Enforce maximal scan/migration frequency..
863 migrate = mm->numa_next_scan;
864 if (time_before(now, migrate))
867 if (p->numa_scan_period == 0)
868 p->numa_scan_period = sysctl_numa_balancing_scan_period_min;
870 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
871 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
875 * Do not set pte_numa if the current running node is rate-limited.
876 * This loses statistics on the fault but if we are unwilling to
877 * migrate to this node, it is less likely we can do useful work
879 if (migrate_ratelimited(numa_node_id()))
882 start = mm->numa_scan_offset;
883 pages = sysctl_numa_balancing_scan_size;
884 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
888 down_read(&mm->mmap_sem);
889 vma = find_vma(mm, start);
891 reset_ptenuma_scan(p);
895 for (; vma; vma = vma->vm_next) {
896 if (!vma_migratable(vma))
899 /* Skip small VMAs. They are not likely to be of relevance */
900 if (((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) < HPAGE_PMD_NR)
904 start = max(start, vma->vm_start);
905 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
906 end = min(end, vma->vm_end);
907 pages -= change_prot_numa(vma, start, end);
912 } while (end != vma->vm_end);
917 * It is possible to reach the end of the VMA list but the last few VMAs are
918 * not guaranteed to the vma_migratable. If they are not, we would find the
919 * !migratable VMA on the next scan but not reset the scanner to the start
923 mm->numa_scan_offset = start;
925 reset_ptenuma_scan(p);
926 up_read(&mm->mmap_sem);
930 * Drive the periodic memory faults..
932 void task_tick_numa(struct rq *rq, struct task_struct *curr)
934 struct callback_head *work = &curr->numa_work;
938 * We don't care about NUMA placement if we don't have memory.
940 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
944 * Using runtime rather than walltime has the dual advantage that
945 * we (mostly) drive the selection from busy threads and that the
946 * task needs to have done some actual work before we bother with
949 now = curr->se.sum_exec_runtime;
950 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
952 if (now - curr->node_stamp > period) {
953 if (!curr->node_stamp)
954 curr->numa_scan_period = sysctl_numa_balancing_scan_period_min;
955 curr->node_stamp = now;
957 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
958 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
959 task_work_add(curr, work, true);
964 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
967 #endif /* CONFIG_NUMA_BALANCING */
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 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
979 cfs_rq->nr_running++;
983 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
985 update_load_sub(&cfs_rq->load, se->load.weight);
986 if (!parent_entity(se))
987 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
988 if (entity_is_task(se))
989 list_del_init(&se->group_node);
990 cfs_rq->nr_running--;
993 #ifdef CONFIG_FAIR_GROUP_SCHED
994 /* we need this in update_cfs_load and load-balance functions below */
995 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
997 static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq,
1000 struct task_group *tg = cfs_rq->tg;
1003 load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1);
1004 load_avg -= cfs_rq->load_contribution;
1006 if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) {
1007 atomic_add(load_avg, &tg->load_weight);
1008 cfs_rq->load_contribution += load_avg;
1012 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
1014 u64 period = sysctl_sched_shares_window;
1016 unsigned long load = cfs_rq->load.weight;
1018 if (cfs_rq->tg == &root_task_group || throttled_hierarchy(cfs_rq))
1021 now = rq_of(cfs_rq)->clock_task;
1022 delta = now - cfs_rq->load_stamp;
1024 /* truncate load history at 4 idle periods */
1025 if (cfs_rq->load_stamp > cfs_rq->load_last &&
1026 now - cfs_rq->load_last > 4 * period) {
1027 cfs_rq->load_period = 0;
1028 cfs_rq->load_avg = 0;
1032 cfs_rq->load_stamp = now;
1033 cfs_rq->load_unacc_exec_time = 0;
1034 cfs_rq->load_period += delta;
1036 cfs_rq->load_last = now;
1037 cfs_rq->load_avg += delta * load;
1040 /* consider updating load contribution on each fold or truncate */
1041 if (global_update || cfs_rq->load_period > period
1042 || !cfs_rq->load_period)
1043 update_cfs_rq_load_contribution(cfs_rq, global_update);
1045 while (cfs_rq->load_period > period) {
1047 * Inline assembly required to prevent the compiler
1048 * optimising this loop into a divmod call.
1049 * See __iter_div_u64_rem() for another example of this.
1051 asm("" : "+rm" (cfs_rq->load_period));
1052 cfs_rq->load_period /= 2;
1053 cfs_rq->load_avg /= 2;
1056 if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg)
1057 list_del_leaf_cfs_rq(cfs_rq);
1060 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1065 * Use this CPU's actual weight instead of the last load_contribution
1066 * to gain a more accurate current total weight. See
1067 * update_cfs_rq_load_contribution().
1069 tg_weight = atomic_read(&tg->load_weight);
1070 tg_weight -= cfs_rq->load_contribution;
1071 tg_weight += cfs_rq->load.weight;
1076 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1078 long tg_weight, load, shares;
1080 tg_weight = calc_tg_weight(tg, cfs_rq);
1081 load = cfs_rq->load.weight;
1083 shares = (tg->shares * load);
1085 shares /= tg_weight;
1087 if (shares < MIN_SHARES)
1088 shares = MIN_SHARES;
1089 if (shares > tg->shares)
1090 shares = tg->shares;
1095 static void update_entity_shares_tick(struct cfs_rq *cfs_rq)
1097 if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) {
1098 update_cfs_load(cfs_rq, 0);
1099 update_cfs_shares(cfs_rq);
1102 # else /* CONFIG_SMP */
1103 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
1107 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1112 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
1115 # endif /* CONFIG_SMP */
1116 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1117 unsigned long weight)
1120 /* commit outstanding execution time */
1121 if (cfs_rq->curr == se)
1122 update_curr(cfs_rq);
1123 account_entity_dequeue(cfs_rq, se);
1126 update_load_set(&se->load, weight);
1129 account_entity_enqueue(cfs_rq, se);
1132 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1134 struct task_group *tg;
1135 struct sched_entity *se;
1139 se = tg->se[cpu_of(rq_of(cfs_rq))];
1140 if (!se || throttled_hierarchy(cfs_rq))
1143 if (likely(se->load.weight == tg->shares))
1146 shares = calc_cfs_shares(cfs_rq, tg);
1148 reweight_entity(cfs_rq_of(se), se, shares);
1150 #else /* CONFIG_FAIR_GROUP_SCHED */
1151 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
1155 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1159 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
1162 #endif /* CONFIG_FAIR_GROUP_SCHED */
1164 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1166 #ifdef CONFIG_SCHEDSTATS
1167 struct task_struct *tsk = NULL;
1169 if (entity_is_task(se))
1172 if (se->statistics.sleep_start) {
1173 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
1178 if (unlikely(delta > se->statistics.sleep_max))
1179 se->statistics.sleep_max = delta;
1181 se->statistics.sleep_start = 0;
1182 se->statistics.sum_sleep_runtime += delta;
1185 account_scheduler_latency(tsk, delta >> 10, 1);
1186 trace_sched_stat_sleep(tsk, delta);
1189 if (se->statistics.block_start) {
1190 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
1195 if (unlikely(delta > se->statistics.block_max))
1196 se->statistics.block_max = delta;
1198 se->statistics.block_start = 0;
1199 se->statistics.sum_sleep_runtime += delta;
1202 if (tsk->in_iowait) {
1203 se->statistics.iowait_sum += delta;
1204 se->statistics.iowait_count++;
1205 trace_sched_stat_iowait(tsk, delta);
1208 trace_sched_stat_blocked(tsk, delta);
1211 * Blocking time is in units of nanosecs, so shift by
1212 * 20 to get a milliseconds-range estimation of the
1213 * amount of time that the task spent sleeping:
1215 if (unlikely(prof_on == SLEEP_PROFILING)) {
1216 profile_hits(SLEEP_PROFILING,
1217 (void *)get_wchan(tsk),
1220 account_scheduler_latency(tsk, delta >> 10, 0);
1226 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1228 #ifdef CONFIG_SCHED_DEBUG
1229 s64 d = se->vruntime - cfs_rq->min_vruntime;
1234 if (d > 3*sysctl_sched_latency)
1235 schedstat_inc(cfs_rq, nr_spread_over);
1240 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1242 u64 vruntime = cfs_rq->min_vruntime;
1245 * The 'current' period is already promised to the current tasks,
1246 * however the extra weight of the new task will slow them down a
1247 * little, place the new task so that it fits in the slot that
1248 * stays open at the end.
1250 if (initial && sched_feat(START_DEBIT))
1251 vruntime += sched_vslice(cfs_rq, se);
1253 /* sleeps up to a single latency don't count. */
1255 unsigned long thresh = sysctl_sched_latency;
1258 * Halve their sleep time's effect, to allow
1259 * for a gentler effect of sleepers:
1261 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1267 /* ensure we never gain time by being placed backwards. */
1268 vruntime = max_vruntime(se->vruntime, vruntime);
1270 se->vruntime = vruntime;
1273 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1276 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1279 * Update the normalized vruntime before updating min_vruntime
1280 * through callig update_curr().
1282 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1283 se->vruntime += cfs_rq->min_vruntime;
1286 * Update run-time statistics of the 'current'.
1288 update_curr(cfs_rq);
1289 update_cfs_load(cfs_rq, 0);
1290 account_entity_enqueue(cfs_rq, se);
1291 update_cfs_shares(cfs_rq);
1293 if (flags & ENQUEUE_WAKEUP) {
1294 place_entity(cfs_rq, se, 0);
1295 enqueue_sleeper(cfs_rq, se);
1298 update_stats_enqueue(cfs_rq, se);
1299 check_spread(cfs_rq, se);
1300 if (se != cfs_rq->curr)
1301 __enqueue_entity(cfs_rq, se);
1304 if (cfs_rq->nr_running == 1) {
1305 list_add_leaf_cfs_rq(cfs_rq);
1306 check_enqueue_throttle(cfs_rq);
1310 static void __clear_buddies_last(struct sched_entity *se)
1312 for_each_sched_entity(se) {
1313 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1314 if (cfs_rq->last == se)
1315 cfs_rq->last = NULL;
1321 static void __clear_buddies_next(struct sched_entity *se)
1323 for_each_sched_entity(se) {
1324 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1325 if (cfs_rq->next == se)
1326 cfs_rq->next = NULL;
1332 static void __clear_buddies_skip(struct sched_entity *se)
1334 for_each_sched_entity(se) {
1335 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1336 if (cfs_rq->skip == se)
1337 cfs_rq->skip = NULL;
1343 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1345 if (cfs_rq->last == se)
1346 __clear_buddies_last(se);
1348 if (cfs_rq->next == se)
1349 __clear_buddies_next(se);
1351 if (cfs_rq->skip == se)
1352 __clear_buddies_skip(se);
1355 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1358 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1361 * Update run-time statistics of the 'current'.
1363 update_curr(cfs_rq);
1365 update_stats_dequeue(cfs_rq, se);
1366 if (flags & DEQUEUE_SLEEP) {
1367 #ifdef CONFIG_SCHEDSTATS
1368 if (entity_is_task(se)) {
1369 struct task_struct *tsk = task_of(se);
1371 if (tsk->state & TASK_INTERRUPTIBLE)
1372 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1373 if (tsk->state & TASK_UNINTERRUPTIBLE)
1374 se->statistics.block_start = rq_of(cfs_rq)->clock;
1379 clear_buddies(cfs_rq, se);
1381 if (se != cfs_rq->curr)
1382 __dequeue_entity(cfs_rq, se);
1384 update_cfs_load(cfs_rq, 0);
1385 account_entity_dequeue(cfs_rq, se);
1388 * Normalize the entity after updating the min_vruntime because the
1389 * update can refer to the ->curr item and we need to reflect this
1390 * movement in our normalized position.
1392 if (!(flags & DEQUEUE_SLEEP))
1393 se->vruntime -= cfs_rq->min_vruntime;
1395 /* return excess runtime on last dequeue */
1396 return_cfs_rq_runtime(cfs_rq);
1398 update_min_vruntime(cfs_rq);
1399 update_cfs_shares(cfs_rq);
1403 * Preempt the current task with a newly woken task if needed:
1406 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1408 unsigned long ideal_runtime, delta_exec;
1409 struct sched_entity *se;
1412 ideal_runtime = sched_slice(cfs_rq, curr);
1413 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1414 if (delta_exec > ideal_runtime) {
1415 resched_task(rq_of(cfs_rq)->curr);
1417 * The current task ran long enough, ensure it doesn't get
1418 * re-elected due to buddy favours.
1420 clear_buddies(cfs_rq, curr);
1425 * Ensure that a task that missed wakeup preemption by a
1426 * narrow margin doesn't have to wait for a full slice.
1427 * This also mitigates buddy induced latencies under load.
1429 if (delta_exec < sysctl_sched_min_granularity)
1432 se = __pick_first_entity(cfs_rq);
1433 delta = curr->vruntime - se->vruntime;
1438 if (delta > ideal_runtime)
1439 resched_task(rq_of(cfs_rq)->curr);
1443 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1445 /* 'current' is not kept within the tree. */
1448 * Any task has to be enqueued before it get to execute on
1449 * a CPU. So account for the time it spent waiting on the
1452 update_stats_wait_end(cfs_rq, se);
1453 __dequeue_entity(cfs_rq, se);
1456 update_stats_curr_start(cfs_rq, se);
1458 #ifdef CONFIG_SCHEDSTATS
1460 * Track our maximum slice length, if the CPU's load is at
1461 * least twice that of our own weight (i.e. dont track it
1462 * when there are only lesser-weight tasks around):
1464 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1465 se->statistics.slice_max = max(se->statistics.slice_max,
1466 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1469 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1473 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1476 * Pick the next process, keeping these things in mind, in this order:
1477 * 1) keep things fair between processes/task groups
1478 * 2) pick the "next" process, since someone really wants that to run
1479 * 3) pick the "last" process, for cache locality
1480 * 4) do not run the "skip" process, if something else is available
1482 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1484 struct sched_entity *se = __pick_first_entity(cfs_rq);
1485 struct sched_entity *left = se;
1488 * Avoid running the skip buddy, if running something else can
1489 * be done without getting too unfair.
1491 if (cfs_rq->skip == se) {
1492 struct sched_entity *second = __pick_next_entity(se);
1493 if (second && wakeup_preempt_entity(second, left) < 1)
1498 * Prefer last buddy, try to return the CPU to a preempted task.
1500 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1504 * Someone really wants this to run. If it's not unfair, run it.
1506 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1509 clear_buddies(cfs_rq, se);
1514 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1516 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1519 * If still on the runqueue then deactivate_task()
1520 * was not called and update_curr() has to be done:
1523 update_curr(cfs_rq);
1525 /* throttle cfs_rqs exceeding runtime */
1526 check_cfs_rq_runtime(cfs_rq);
1528 check_spread(cfs_rq, prev);
1530 update_stats_wait_start(cfs_rq, prev);
1531 /* Put 'current' back into the tree. */
1532 __enqueue_entity(cfs_rq, prev);
1534 cfs_rq->curr = NULL;
1538 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1541 * Update run-time statistics of the 'current'.
1543 update_curr(cfs_rq);
1546 * Update share accounting for long-running entities.
1548 update_entity_shares_tick(cfs_rq);
1550 #ifdef CONFIG_SCHED_HRTICK
1552 * queued ticks are scheduled to match the slice, so don't bother
1553 * validating it and just reschedule.
1556 resched_task(rq_of(cfs_rq)->curr);
1560 * don't let the period tick interfere with the hrtick preemption
1562 if (!sched_feat(DOUBLE_TICK) &&
1563 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1567 if (cfs_rq->nr_running > 1)
1568 check_preempt_tick(cfs_rq, curr);
1572 /**************************************************
1573 * CFS bandwidth control machinery
1576 #ifdef CONFIG_CFS_BANDWIDTH
1578 #ifdef HAVE_JUMP_LABEL
1579 static struct static_key __cfs_bandwidth_used;
1581 static inline bool cfs_bandwidth_used(void)
1583 return static_key_false(&__cfs_bandwidth_used);
1586 void account_cfs_bandwidth_used(int enabled, int was_enabled)
1588 /* only need to count groups transitioning between enabled/!enabled */
1589 if (enabled && !was_enabled)
1590 static_key_slow_inc(&__cfs_bandwidth_used);
1591 else if (!enabled && was_enabled)
1592 static_key_slow_dec(&__cfs_bandwidth_used);
1594 #else /* HAVE_JUMP_LABEL */
1595 static bool cfs_bandwidth_used(void)
1600 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
1601 #endif /* HAVE_JUMP_LABEL */
1604 * default period for cfs group bandwidth.
1605 * default: 0.1s, units: nanoseconds
1607 static inline u64 default_cfs_period(void)
1609 return 100000000ULL;
1612 static inline u64 sched_cfs_bandwidth_slice(void)
1614 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
1618 * Replenish runtime according to assigned quota and update expiration time.
1619 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
1620 * additional synchronization around rq->lock.
1622 * requires cfs_b->lock
1624 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
1628 if (cfs_b->quota == RUNTIME_INF)
1631 now = sched_clock_cpu(smp_processor_id());
1632 cfs_b->runtime = cfs_b->quota;
1633 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
1636 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
1638 return &tg->cfs_bandwidth;
1641 /* returns 0 on failure to allocate runtime */
1642 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1644 struct task_group *tg = cfs_rq->tg;
1645 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
1646 u64 amount = 0, min_amount, expires;
1648 /* note: this is a positive sum as runtime_remaining <= 0 */
1649 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
1651 raw_spin_lock(&cfs_b->lock);
1652 if (cfs_b->quota == RUNTIME_INF)
1653 amount = min_amount;
1656 * If the bandwidth pool has become inactive, then at least one
1657 * period must have elapsed since the last consumption.
1658 * Refresh the global state and ensure bandwidth timer becomes
1661 if (!cfs_b->timer_active) {
1662 __refill_cfs_bandwidth_runtime(cfs_b);
1663 __start_cfs_bandwidth(cfs_b);
1666 if (cfs_b->runtime > 0) {
1667 amount = min(cfs_b->runtime, min_amount);
1668 cfs_b->runtime -= amount;
1672 expires = cfs_b->runtime_expires;
1673 raw_spin_unlock(&cfs_b->lock);
1675 cfs_rq->runtime_remaining += amount;
1677 * we may have advanced our local expiration to account for allowed
1678 * spread between our sched_clock and the one on which runtime was
1681 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
1682 cfs_rq->runtime_expires = expires;
1684 return cfs_rq->runtime_remaining > 0;
1688 * Note: This depends on the synchronization provided by sched_clock and the
1689 * fact that rq->clock snapshots this value.
1691 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1693 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1694 struct rq *rq = rq_of(cfs_rq);
1696 /* if the deadline is ahead of our clock, nothing to do */
1697 if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0))
1700 if (cfs_rq->runtime_remaining < 0)
1704 * If the local deadline has passed we have to consider the
1705 * possibility that our sched_clock is 'fast' and the global deadline
1706 * has not truly expired.
1708 * Fortunately we can check determine whether this the case by checking
1709 * whether the global deadline has advanced.
1712 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
1713 /* extend local deadline, drift is bounded above by 2 ticks */
1714 cfs_rq->runtime_expires += TICK_NSEC;
1716 /* global deadline is ahead, expiration has passed */
1717 cfs_rq->runtime_remaining = 0;
1721 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1722 unsigned long delta_exec)
1724 /* dock delta_exec before expiring quota (as it could span periods) */
1725 cfs_rq->runtime_remaining -= delta_exec;
1726 expire_cfs_rq_runtime(cfs_rq);
1728 if (likely(cfs_rq->runtime_remaining > 0))
1732 * if we're unable to extend our runtime we resched so that the active
1733 * hierarchy can be throttled
1735 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
1736 resched_task(rq_of(cfs_rq)->curr);
1739 static __always_inline
1740 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
1742 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
1745 __account_cfs_rq_runtime(cfs_rq, delta_exec);
1748 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
1750 return cfs_bandwidth_used() && cfs_rq->throttled;
1753 /* check whether cfs_rq, or any parent, is throttled */
1754 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
1756 return cfs_bandwidth_used() && cfs_rq->throttle_count;
1760 * Ensure that neither of the group entities corresponding to src_cpu or
1761 * dest_cpu are members of a throttled hierarchy when performing group
1762 * load-balance operations.
1764 static inline int throttled_lb_pair(struct task_group *tg,
1765 int src_cpu, int dest_cpu)
1767 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
1769 src_cfs_rq = tg->cfs_rq[src_cpu];
1770 dest_cfs_rq = tg->cfs_rq[dest_cpu];
1772 return throttled_hierarchy(src_cfs_rq) ||
1773 throttled_hierarchy(dest_cfs_rq);
1776 /* updated child weight may affect parent so we have to do this bottom up */
1777 static int tg_unthrottle_up(struct task_group *tg, void *data)
1779 struct rq *rq = data;
1780 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1782 cfs_rq->throttle_count--;
1784 if (!cfs_rq->throttle_count) {
1785 u64 delta = rq->clock_task - cfs_rq->load_stamp;
1787 /* leaving throttled state, advance shares averaging windows */
1788 cfs_rq->load_stamp += delta;
1789 cfs_rq->load_last += delta;
1791 /* update entity weight now that we are on_rq again */
1792 update_cfs_shares(cfs_rq);
1799 static int tg_throttle_down(struct task_group *tg, void *data)
1801 struct rq *rq = data;
1802 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1804 /* group is entering throttled state, record last load */
1805 if (!cfs_rq->throttle_count)
1806 update_cfs_load(cfs_rq, 0);
1807 cfs_rq->throttle_count++;
1812 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
1814 struct rq *rq = rq_of(cfs_rq);
1815 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1816 struct sched_entity *se;
1817 long task_delta, dequeue = 1;
1819 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1821 /* account load preceding throttle */
1823 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
1826 task_delta = cfs_rq->h_nr_running;
1827 for_each_sched_entity(se) {
1828 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
1829 /* throttled entity or throttle-on-deactivate */
1834 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
1835 qcfs_rq->h_nr_running -= task_delta;
1837 if (qcfs_rq->load.weight)
1842 rq->nr_running -= task_delta;
1844 cfs_rq->throttled = 1;
1845 cfs_rq->throttled_timestamp = rq->clock;
1846 raw_spin_lock(&cfs_b->lock);
1847 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
1848 raw_spin_unlock(&cfs_b->lock);
1851 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
1853 struct rq *rq = rq_of(cfs_rq);
1854 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1855 struct sched_entity *se;
1859 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1861 cfs_rq->throttled = 0;
1862 raw_spin_lock(&cfs_b->lock);
1863 cfs_b->throttled_time += rq->clock - cfs_rq->throttled_timestamp;
1864 list_del_rcu(&cfs_rq->throttled_list);
1865 raw_spin_unlock(&cfs_b->lock);
1866 cfs_rq->throttled_timestamp = 0;
1868 update_rq_clock(rq);
1869 /* update hierarchical throttle state */
1870 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
1872 if (!cfs_rq->load.weight)
1875 task_delta = cfs_rq->h_nr_running;
1876 for_each_sched_entity(se) {
1880 cfs_rq = cfs_rq_of(se);
1882 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
1883 cfs_rq->h_nr_running += task_delta;
1885 if (cfs_rq_throttled(cfs_rq))
1890 rq->nr_running += task_delta;
1892 /* determine whether we need to wake up potentially idle cpu */
1893 if (rq->curr == rq->idle && rq->cfs.nr_running)
1894 resched_task(rq->curr);
1897 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
1898 u64 remaining, u64 expires)
1900 struct cfs_rq *cfs_rq;
1901 u64 runtime = remaining;
1904 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
1906 struct rq *rq = rq_of(cfs_rq);
1908 raw_spin_lock(&rq->lock);
1909 if (!cfs_rq_throttled(cfs_rq))
1912 runtime = -cfs_rq->runtime_remaining + 1;
1913 if (runtime > remaining)
1914 runtime = remaining;
1915 remaining -= runtime;
1917 cfs_rq->runtime_remaining += runtime;
1918 cfs_rq->runtime_expires = expires;
1920 /* we check whether we're throttled above */
1921 if (cfs_rq->runtime_remaining > 0)
1922 unthrottle_cfs_rq(cfs_rq);
1925 raw_spin_unlock(&rq->lock);
1936 * Responsible for refilling a task_group's bandwidth and unthrottling its
1937 * cfs_rqs as appropriate. If there has been no activity within the last
1938 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
1939 * used to track this state.
1941 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
1943 u64 runtime, runtime_expires;
1944 int idle = 1, throttled;
1946 raw_spin_lock(&cfs_b->lock);
1947 /* no need to continue the timer with no bandwidth constraint */
1948 if (cfs_b->quota == RUNTIME_INF)
1951 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
1952 /* idle depends on !throttled (for the case of a large deficit) */
1953 idle = cfs_b->idle && !throttled;
1954 cfs_b->nr_periods += overrun;
1956 /* if we're going inactive then everything else can be deferred */
1960 __refill_cfs_bandwidth_runtime(cfs_b);
1963 /* mark as potentially idle for the upcoming period */
1968 /* account preceding periods in which throttling occurred */
1969 cfs_b->nr_throttled += overrun;
1972 * There are throttled entities so we must first use the new bandwidth
1973 * to unthrottle them before making it generally available. This
1974 * ensures that all existing debts will be paid before a new cfs_rq is
1977 runtime = cfs_b->runtime;
1978 runtime_expires = cfs_b->runtime_expires;
1982 * This check is repeated as we are holding onto the new bandwidth
1983 * while we unthrottle. This can potentially race with an unthrottled
1984 * group trying to acquire new bandwidth from the global pool.
1986 while (throttled && runtime > 0) {
1987 raw_spin_unlock(&cfs_b->lock);
1988 /* we can't nest cfs_b->lock while distributing bandwidth */
1989 runtime = distribute_cfs_runtime(cfs_b, runtime,
1991 raw_spin_lock(&cfs_b->lock);
1993 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
1996 /* return (any) remaining runtime */
1997 cfs_b->runtime = runtime;
1999 * While we are ensured activity in the period following an
2000 * unthrottle, this also covers the case in which the new bandwidth is
2001 * insufficient to cover the existing bandwidth deficit. (Forcing the
2002 * timer to remain active while there are any throttled entities.)
2007 cfs_b->timer_active = 0;
2008 raw_spin_unlock(&cfs_b->lock);
2013 /* a cfs_rq won't donate quota below this amount */
2014 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2015 /* minimum remaining period time to redistribute slack quota */
2016 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2017 /* how long we wait to gather additional slack before distributing */
2018 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2020 /* are we near the end of the current quota period? */
2021 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2023 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2026 /* if the call-back is running a quota refresh is already occurring */
2027 if (hrtimer_callback_running(refresh_timer))
2030 /* is a quota refresh about to occur? */
2031 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2032 if (remaining < min_expire)
2038 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2040 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2042 /* if there's a quota refresh soon don't bother with slack */
2043 if (runtime_refresh_within(cfs_b, min_left))
2046 start_bandwidth_timer(&cfs_b->slack_timer,
2047 ns_to_ktime(cfs_bandwidth_slack_period));
2050 /* we know any runtime found here is valid as update_curr() precedes return */
2051 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2053 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2054 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2056 if (slack_runtime <= 0)
2059 raw_spin_lock(&cfs_b->lock);
2060 if (cfs_b->quota != RUNTIME_INF &&
2061 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2062 cfs_b->runtime += slack_runtime;
2064 /* we are under rq->lock, defer unthrottling using a timer */
2065 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2066 !list_empty(&cfs_b->throttled_cfs_rq))
2067 start_cfs_slack_bandwidth(cfs_b);
2069 raw_spin_unlock(&cfs_b->lock);
2071 /* even if it's not valid for return we don't want to try again */
2072 cfs_rq->runtime_remaining -= slack_runtime;
2075 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2077 if (!cfs_bandwidth_used())
2080 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2083 __return_cfs_rq_runtime(cfs_rq);
2087 * This is done with a timer (instead of inline with bandwidth return) since
2088 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2090 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2092 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2095 /* confirm we're still not at a refresh boundary */
2096 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2099 raw_spin_lock(&cfs_b->lock);
2100 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2101 runtime = cfs_b->runtime;
2104 expires = cfs_b->runtime_expires;
2105 raw_spin_unlock(&cfs_b->lock);
2110 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2112 raw_spin_lock(&cfs_b->lock);
2113 if (expires == cfs_b->runtime_expires)
2114 cfs_b->runtime = runtime;
2115 raw_spin_unlock(&cfs_b->lock);
2119 * When a group wakes up we want to make sure that its quota is not already
2120 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2121 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2123 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2125 if (!cfs_bandwidth_used())
2128 /* an active group must be handled by the update_curr()->put() path */
2129 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2132 /* ensure the group is not already throttled */
2133 if (cfs_rq_throttled(cfs_rq))
2136 /* update runtime allocation */
2137 account_cfs_rq_runtime(cfs_rq, 0);
2138 if (cfs_rq->runtime_remaining <= 0)
2139 throttle_cfs_rq(cfs_rq);
2142 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2143 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2145 if (!cfs_bandwidth_used())
2148 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2152 * it's possible for a throttled entity to be forced into a running
2153 * state (e.g. set_curr_task), in this case we're finished.
2155 if (cfs_rq_throttled(cfs_rq))
2158 throttle_cfs_rq(cfs_rq);
2161 static inline u64 default_cfs_period(void);
2162 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun);
2163 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b);
2165 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2167 struct cfs_bandwidth *cfs_b =
2168 container_of(timer, struct cfs_bandwidth, slack_timer);
2169 do_sched_cfs_slack_timer(cfs_b);
2171 return HRTIMER_NORESTART;
2174 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2176 struct cfs_bandwidth *cfs_b =
2177 container_of(timer, struct cfs_bandwidth, period_timer);
2183 now = hrtimer_cb_get_time(timer);
2184 overrun = hrtimer_forward(timer, now, cfs_b->period);
2189 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2192 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2195 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2197 raw_spin_lock_init(&cfs_b->lock);
2199 cfs_b->quota = RUNTIME_INF;
2200 cfs_b->period = ns_to_ktime(default_cfs_period());
2202 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2203 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2204 cfs_b->period_timer.function = sched_cfs_period_timer;
2205 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2206 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2209 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2211 cfs_rq->runtime_enabled = 0;
2212 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2215 /* requires cfs_b->lock, may release to reprogram timer */
2216 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2219 * The timer may be active because we're trying to set a new bandwidth
2220 * period or because we're racing with the tear-down path
2221 * (timer_active==0 becomes visible before the hrtimer call-back
2222 * terminates). In either case we ensure that it's re-programmed
2224 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2225 raw_spin_unlock(&cfs_b->lock);
2226 /* ensure cfs_b->lock is available while we wait */
2227 hrtimer_cancel(&cfs_b->period_timer);
2229 raw_spin_lock(&cfs_b->lock);
2230 /* if someone else restarted the timer then we're done */
2231 if (cfs_b->timer_active)
2235 cfs_b->timer_active = 1;
2236 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2239 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2241 hrtimer_cancel(&cfs_b->period_timer);
2242 hrtimer_cancel(&cfs_b->slack_timer);
2245 static void unthrottle_offline_cfs_rqs(struct rq *rq)
2247 struct cfs_rq *cfs_rq;
2249 for_each_leaf_cfs_rq(rq, cfs_rq) {
2250 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2252 if (!cfs_rq->runtime_enabled)
2256 * clock_task is not advancing so we just need to make sure
2257 * there's some valid quota amount
2259 cfs_rq->runtime_remaining = cfs_b->quota;
2260 if (cfs_rq_throttled(cfs_rq))
2261 unthrottle_cfs_rq(cfs_rq);
2265 #else /* CONFIG_CFS_BANDWIDTH */
2266 static __always_inline
2267 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec) {}
2268 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2269 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2270 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2272 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2277 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2282 static inline int throttled_lb_pair(struct task_group *tg,
2283 int src_cpu, int dest_cpu)
2288 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2290 #ifdef CONFIG_FAIR_GROUP_SCHED
2291 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2294 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2298 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2299 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2301 #endif /* CONFIG_CFS_BANDWIDTH */
2303 /**************************************************
2304 * CFS operations on tasks:
2307 #ifdef CONFIG_SCHED_HRTICK
2308 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2310 struct sched_entity *se = &p->se;
2311 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2313 WARN_ON(task_rq(p) != rq);
2315 if (cfs_rq->nr_running > 1) {
2316 u64 slice = sched_slice(cfs_rq, se);
2317 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2318 s64 delta = slice - ran;
2327 * Don't schedule slices shorter than 10000ns, that just
2328 * doesn't make sense. Rely on vruntime for fairness.
2331 delta = max_t(s64, 10000LL, delta);
2333 hrtick_start(rq, delta);
2338 * called from enqueue/dequeue and updates the hrtick when the
2339 * current task is from our class and nr_running is low enough
2342 static void hrtick_update(struct rq *rq)
2344 struct task_struct *curr = rq->curr;
2346 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2349 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2350 hrtick_start_fair(rq, curr);
2352 #else /* !CONFIG_SCHED_HRTICK */
2354 hrtick_start_fair(struct rq *rq, struct task_struct *p)
2358 static inline void hrtick_update(struct rq *rq)
2364 * The enqueue_task method is called before nr_running is
2365 * increased. Here we update the fair scheduling stats and
2366 * then put the task into the rbtree:
2369 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2371 struct cfs_rq *cfs_rq;
2372 struct sched_entity *se = &p->se;
2374 for_each_sched_entity(se) {
2377 cfs_rq = cfs_rq_of(se);
2378 enqueue_entity(cfs_rq, se, flags);
2381 * end evaluation on encountering a throttled cfs_rq
2383 * note: in the case of encountering a throttled cfs_rq we will
2384 * post the final h_nr_running increment below.
2386 if (cfs_rq_throttled(cfs_rq))
2388 cfs_rq->h_nr_running++;
2390 flags = ENQUEUE_WAKEUP;
2393 for_each_sched_entity(se) {
2394 cfs_rq = cfs_rq_of(se);
2395 cfs_rq->h_nr_running++;
2397 if (cfs_rq_throttled(cfs_rq))
2400 update_cfs_load(cfs_rq, 0);
2401 update_cfs_shares(cfs_rq);
2409 static void set_next_buddy(struct sched_entity *se);
2412 * The dequeue_task method is called before nr_running is
2413 * decreased. We remove the task from the rbtree and
2414 * update the fair scheduling stats:
2416 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2418 struct cfs_rq *cfs_rq;
2419 struct sched_entity *se = &p->se;
2420 int task_sleep = flags & DEQUEUE_SLEEP;
2422 for_each_sched_entity(se) {
2423 cfs_rq = cfs_rq_of(se);
2424 dequeue_entity(cfs_rq, se, flags);
2427 * end evaluation on encountering a throttled cfs_rq
2429 * note: in the case of encountering a throttled cfs_rq we will
2430 * post the final h_nr_running decrement below.
2432 if (cfs_rq_throttled(cfs_rq))
2434 cfs_rq->h_nr_running--;
2436 /* Don't dequeue parent if it has other entities besides us */
2437 if (cfs_rq->load.weight) {
2439 * Bias pick_next to pick a task from this cfs_rq, as
2440 * p is sleeping when it is within its sched_slice.
2442 if (task_sleep && parent_entity(se))
2443 set_next_buddy(parent_entity(se));
2445 /* avoid re-evaluating load for this entity */
2446 se = parent_entity(se);
2449 flags |= DEQUEUE_SLEEP;
2452 for_each_sched_entity(se) {
2453 cfs_rq = cfs_rq_of(se);
2454 cfs_rq->h_nr_running--;
2456 if (cfs_rq_throttled(cfs_rq))
2459 update_cfs_load(cfs_rq, 0);
2460 update_cfs_shares(cfs_rq);
2469 /* Used instead of source_load when we know the type == 0 */
2470 static unsigned long weighted_cpuload(const int cpu)
2472 return cpu_rq(cpu)->load.weight;
2476 * Return a low guess at the load of a migration-source cpu weighted
2477 * according to the scheduling class and "nice" value.
2479 * We want to under-estimate the load of migration sources, to
2480 * balance conservatively.
2482 static unsigned long source_load(int cpu, int type)
2484 struct rq *rq = cpu_rq(cpu);
2485 unsigned long total = weighted_cpuload(cpu);
2487 if (type == 0 || !sched_feat(LB_BIAS))
2490 return min(rq->cpu_load[type-1], total);
2494 * Return a high guess at the load of a migration-target cpu weighted
2495 * according to the scheduling class and "nice" value.
2497 static unsigned long target_load(int cpu, int type)
2499 struct rq *rq = cpu_rq(cpu);
2500 unsigned long total = weighted_cpuload(cpu);
2502 if (type == 0 || !sched_feat(LB_BIAS))
2505 return max(rq->cpu_load[type-1], total);
2508 static unsigned long power_of(int cpu)
2510 return cpu_rq(cpu)->cpu_power;
2513 static unsigned long cpu_avg_load_per_task(int cpu)
2515 struct rq *rq = cpu_rq(cpu);
2516 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
2519 return rq->load.weight / nr_running;
2525 static void task_waking_fair(struct task_struct *p)
2527 struct sched_entity *se = &p->se;
2528 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2531 #ifndef CONFIG_64BIT
2532 u64 min_vruntime_copy;
2535 min_vruntime_copy = cfs_rq->min_vruntime_copy;
2537 min_vruntime = cfs_rq->min_vruntime;
2538 } while (min_vruntime != min_vruntime_copy);
2540 min_vruntime = cfs_rq->min_vruntime;
2543 se->vruntime -= min_vruntime;
2546 #ifdef CONFIG_FAIR_GROUP_SCHED
2548 * effective_load() calculates the load change as seen from the root_task_group
2550 * Adding load to a group doesn't make a group heavier, but can cause movement
2551 * of group shares between cpus. Assuming the shares were perfectly aligned one
2552 * can calculate the shift in shares.
2554 * Calculate the effective load difference if @wl is added (subtracted) to @tg
2555 * on this @cpu and results in a total addition (subtraction) of @wg to the
2556 * total group weight.
2558 * Given a runqueue weight distribution (rw_i) we can compute a shares
2559 * distribution (s_i) using:
2561 * s_i = rw_i / \Sum rw_j (1)
2563 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
2564 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
2565 * shares distribution (s_i):
2567 * rw_i = { 2, 4, 1, 0 }
2568 * s_i = { 2/7, 4/7, 1/7, 0 }
2570 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
2571 * task used to run on and the CPU the waker is running on), we need to
2572 * compute the effect of waking a task on either CPU and, in case of a sync
2573 * wakeup, compute the effect of the current task going to sleep.
2575 * So for a change of @wl to the local @cpu with an overall group weight change
2576 * of @wl we can compute the new shares distribution (s'_i) using:
2578 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
2580 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
2581 * differences in waking a task to CPU 0. The additional task changes the
2582 * weight and shares distributions like:
2584 * rw'_i = { 3, 4, 1, 0 }
2585 * s'_i = { 3/8, 4/8, 1/8, 0 }
2587 * We can then compute the difference in effective weight by using:
2589 * dw_i = S * (s'_i - s_i) (3)
2591 * Where 'S' is the group weight as seen by its parent.
2593 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
2594 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
2595 * 4/7) times the weight of the group.
2597 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
2599 struct sched_entity *se = tg->se[cpu];
2601 if (!tg->parent) /* the trivial, non-cgroup case */
2604 for_each_sched_entity(se) {
2610 * W = @wg + \Sum rw_j
2612 W = wg + calc_tg_weight(tg, se->my_q);
2617 w = se->my_q->load.weight + wl;
2620 * wl = S * s'_i; see (2)
2623 wl = (w * tg->shares) / W;
2628 * Per the above, wl is the new se->load.weight value; since
2629 * those are clipped to [MIN_SHARES, ...) do so now. See
2630 * calc_cfs_shares().
2632 if (wl < MIN_SHARES)
2636 * wl = dw_i = S * (s'_i - s_i); see (3)
2638 wl -= se->load.weight;
2641 * Recursively apply this logic to all parent groups to compute
2642 * the final effective load change on the root group. Since
2643 * only the @tg group gets extra weight, all parent groups can
2644 * only redistribute existing shares. @wl is the shift in shares
2645 * resulting from this level per the above.
2654 static inline unsigned long effective_load(struct task_group *tg, int cpu,
2655 unsigned long wl, unsigned long wg)
2662 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
2664 s64 this_load, load;
2665 int idx, this_cpu, prev_cpu;
2666 unsigned long tl_per_task;
2667 struct task_group *tg;
2668 unsigned long weight;
2672 this_cpu = smp_processor_id();
2673 prev_cpu = task_cpu(p);
2674 load = source_load(prev_cpu, idx);
2675 this_load = target_load(this_cpu, idx);
2678 * If sync wakeup then subtract the (maximum possible)
2679 * effect of the currently running task from the load
2680 * of the current CPU:
2683 tg = task_group(current);
2684 weight = current->se.load.weight;
2686 this_load += effective_load(tg, this_cpu, -weight, -weight);
2687 load += effective_load(tg, prev_cpu, 0, -weight);
2691 weight = p->se.load.weight;
2694 * In low-load situations, where prev_cpu is idle and this_cpu is idle
2695 * due to the sync cause above having dropped this_load to 0, we'll
2696 * always have an imbalance, but there's really nothing you can do
2697 * about that, so that's good too.
2699 * Otherwise check if either cpus are near enough in load to allow this
2700 * task to be woken on this_cpu.
2702 if (this_load > 0) {
2703 s64 this_eff_load, prev_eff_load;
2705 this_eff_load = 100;
2706 this_eff_load *= power_of(prev_cpu);
2707 this_eff_load *= this_load +
2708 effective_load(tg, this_cpu, weight, weight);
2710 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
2711 prev_eff_load *= power_of(this_cpu);
2712 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
2714 balanced = this_eff_load <= prev_eff_load;
2719 * If the currently running task will sleep within
2720 * a reasonable amount of time then attract this newly
2723 if (sync && balanced)
2726 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
2727 tl_per_task = cpu_avg_load_per_task(this_cpu);
2730 (this_load <= load &&
2731 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
2733 * This domain has SD_WAKE_AFFINE and
2734 * p is cache cold in this domain, and
2735 * there is no bad imbalance.
2737 schedstat_inc(sd, ttwu_move_affine);
2738 schedstat_inc(p, se.statistics.nr_wakeups_affine);
2746 * find_idlest_group finds and returns the least busy CPU group within the
2749 static struct sched_group *
2750 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
2751 int this_cpu, int load_idx)
2753 struct sched_group *idlest = NULL, *group = sd->groups;
2754 unsigned long min_load = ULONG_MAX, this_load = 0;
2755 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2758 unsigned long load, avg_load;
2762 /* Skip over this group if it has no CPUs allowed */
2763 if (!cpumask_intersects(sched_group_cpus(group),
2764 tsk_cpus_allowed(p)))
2767 local_group = cpumask_test_cpu(this_cpu,
2768 sched_group_cpus(group));
2770 /* Tally up the load of all CPUs in the group */
2773 for_each_cpu(i, sched_group_cpus(group)) {
2774 /* Bias balancing toward cpus of our domain */
2776 load = source_load(i, load_idx);
2778 load = target_load(i, load_idx);
2783 /* Adjust by relative CPU power of the group */
2784 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
2787 this_load = avg_load;
2788 } else if (avg_load < min_load) {
2789 min_load = avg_load;
2792 } while (group = group->next, group != sd->groups);
2794 if (!idlest || 100*this_load < imbalance*min_load)
2800 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2803 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2805 unsigned long load, min_load = ULONG_MAX;
2809 /* Traverse only the allowed CPUs */
2810 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
2811 load = weighted_cpuload(i);
2813 if (load < min_load || (load == min_load && i == this_cpu)) {
2823 * Try and locate an idle CPU in the sched_domain.
2825 static int select_idle_sibling(struct task_struct *p, int target)
2827 int cpu = smp_processor_id();
2828 int prev_cpu = task_cpu(p);
2829 struct sched_domain *sd;
2830 struct sched_group *sg;
2834 * If the task is going to be woken-up on this cpu and if it is
2835 * already idle, then it is the right target.
2837 if (target == cpu && idle_cpu(cpu))
2841 * If the task is going to be woken-up on the cpu where it previously
2842 * ran and if it is currently idle, then it the right target.
2844 if (target == prev_cpu && idle_cpu(prev_cpu))
2848 * Otherwise, iterate the domains and find an elegible idle cpu.
2850 sd = rcu_dereference(per_cpu(sd_llc, target));
2851 for_each_lower_domain(sd) {
2854 if (!cpumask_intersects(sched_group_cpus(sg),
2855 tsk_cpus_allowed(p)))
2858 for_each_cpu(i, sched_group_cpus(sg)) {
2863 target = cpumask_first_and(sched_group_cpus(sg),
2864 tsk_cpus_allowed(p));
2868 } while (sg != sd->groups);
2875 * sched_balance_self: balance the current task (running on cpu) in domains
2876 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2879 * Balance, ie. select the least loaded group.
2881 * Returns the target CPU number, or the same CPU if no balancing is needed.
2883 * preempt must be disabled.
2886 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
2888 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
2889 int cpu = smp_processor_id();
2890 int prev_cpu = task_cpu(p);
2892 int want_affine = 0;
2893 int sync = wake_flags & WF_SYNC;
2895 if (p->nr_cpus_allowed == 1)
2898 if (sd_flag & SD_BALANCE_WAKE) {
2899 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
2905 for_each_domain(cpu, tmp) {
2906 if (!(tmp->flags & SD_LOAD_BALANCE))
2910 * If both cpu and prev_cpu are part of this domain,
2911 * cpu is a valid SD_WAKE_AFFINE target.
2913 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
2914 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
2919 if (tmp->flags & sd_flag)
2924 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
2927 new_cpu = select_idle_sibling(p, prev_cpu);
2932 int load_idx = sd->forkexec_idx;
2933 struct sched_group *group;
2936 if (!(sd->flags & sd_flag)) {
2941 if (sd_flag & SD_BALANCE_WAKE)
2942 load_idx = sd->wake_idx;
2944 group = find_idlest_group(sd, p, cpu, load_idx);
2950 new_cpu = find_idlest_cpu(group, p, cpu);
2951 if (new_cpu == -1 || new_cpu == cpu) {
2952 /* Now try balancing at a lower domain level of cpu */
2957 /* Now try balancing at a lower domain level of new_cpu */
2959 weight = sd->span_weight;
2961 for_each_domain(cpu, tmp) {
2962 if (weight <= tmp->span_weight)
2964 if (tmp->flags & sd_flag)
2967 /* while loop will break here if sd == NULL */
2974 #endif /* CONFIG_SMP */
2976 static unsigned long
2977 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
2979 unsigned long gran = sysctl_sched_wakeup_granularity;
2982 * Since its curr running now, convert the gran from real-time
2983 * to virtual-time in his units.
2985 * By using 'se' instead of 'curr' we penalize light tasks, so
2986 * they get preempted easier. That is, if 'se' < 'curr' then
2987 * the resulting gran will be larger, therefore penalizing the
2988 * lighter, if otoh 'se' > 'curr' then the resulting gran will
2989 * be smaller, again penalizing the lighter task.
2991 * This is especially important for buddies when the leftmost
2992 * task is higher priority than the buddy.
2994 return calc_delta_fair(gran, se);
2998 * Should 'se' preempt 'curr'.
3012 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3014 s64 gran, vdiff = curr->vruntime - se->vruntime;
3019 gran = wakeup_gran(curr, se);
3026 static void set_last_buddy(struct sched_entity *se)
3028 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3031 for_each_sched_entity(se)
3032 cfs_rq_of(se)->last = se;
3035 static void set_next_buddy(struct sched_entity *se)
3037 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3040 for_each_sched_entity(se)
3041 cfs_rq_of(se)->next = se;
3044 static void set_skip_buddy(struct sched_entity *se)
3046 for_each_sched_entity(se)
3047 cfs_rq_of(se)->skip = se;
3051 * Preempt the current task with a newly woken task if needed:
3053 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3055 struct task_struct *curr = rq->curr;
3056 struct sched_entity *se = &curr->se, *pse = &p->se;
3057 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3058 int scale = cfs_rq->nr_running >= sched_nr_latency;
3059 int next_buddy_marked = 0;
3061 if (unlikely(se == pse))
3065 * This is possible from callers such as move_task(), in which we
3066 * unconditionally check_prempt_curr() after an enqueue (which may have
3067 * lead to a throttle). This both saves work and prevents false
3068 * next-buddy nomination below.
3070 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3073 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3074 set_next_buddy(pse);
3075 next_buddy_marked = 1;
3079 * We can come here with TIF_NEED_RESCHED already set from new task
3082 * Note: this also catches the edge-case of curr being in a throttled
3083 * group (e.g. via set_curr_task), since update_curr() (in the
3084 * enqueue of curr) will have resulted in resched being set. This
3085 * prevents us from potentially nominating it as a false LAST_BUDDY
3088 if (test_tsk_need_resched(curr))
3091 /* Idle tasks are by definition preempted by non-idle tasks. */
3092 if (unlikely(curr->policy == SCHED_IDLE) &&
3093 likely(p->policy != SCHED_IDLE))
3097 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3098 * is driven by the tick):
3100 if (unlikely(p->policy != SCHED_NORMAL))
3103 find_matching_se(&se, &pse);
3104 update_curr(cfs_rq_of(se));
3106 if (wakeup_preempt_entity(se, pse) == 1) {
3108 * Bias pick_next to pick the sched entity that is
3109 * triggering this preemption.
3111 if (!next_buddy_marked)
3112 set_next_buddy(pse);
3121 * Only set the backward buddy when the current task is still
3122 * on the rq. This can happen when a wakeup gets interleaved
3123 * with schedule on the ->pre_schedule() or idle_balance()
3124 * point, either of which can * drop the rq lock.
3126 * Also, during early boot the idle thread is in the fair class,
3127 * for obvious reasons its a bad idea to schedule back to it.
3129 if (unlikely(!se->on_rq || curr == rq->idle))
3132 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3136 static struct task_struct *pick_next_task_fair(struct rq *rq)
3138 struct task_struct *p;
3139 struct cfs_rq *cfs_rq = &rq->cfs;
3140 struct sched_entity *se;
3142 if (!cfs_rq->nr_running)
3146 se = pick_next_entity(cfs_rq);
3147 set_next_entity(cfs_rq, se);
3148 cfs_rq = group_cfs_rq(se);
3152 if (hrtick_enabled(rq))
3153 hrtick_start_fair(rq, p);
3159 * Account for a descheduled task:
3161 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3163 struct sched_entity *se = &prev->se;
3164 struct cfs_rq *cfs_rq;
3166 for_each_sched_entity(se) {
3167 cfs_rq = cfs_rq_of(se);
3168 put_prev_entity(cfs_rq, se);
3173 * sched_yield() is very simple
3175 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3177 static void yield_task_fair(struct rq *rq)
3179 struct task_struct *curr = rq->curr;
3180 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3181 struct sched_entity *se = &curr->se;
3184 * Are we the only task in the tree?
3186 if (unlikely(rq->nr_running == 1))
3189 clear_buddies(cfs_rq, se);
3191 if (curr->policy != SCHED_BATCH) {
3192 update_rq_clock(rq);
3194 * Update run-time statistics of the 'current'.
3196 update_curr(cfs_rq);
3198 * Tell update_rq_clock() that we've just updated,
3199 * so we don't do microscopic update in schedule()
3200 * and double the fastpath cost.
3202 rq->skip_clock_update = 1;
3208 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3210 struct sched_entity *se = &p->se;
3212 /* throttled hierarchies are not runnable */
3213 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3216 /* Tell the scheduler that we'd really like pse to run next. */
3219 yield_task_fair(rq);
3225 /**************************************************
3226 * Fair scheduling class load-balancing methods:
3229 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
3231 #define LBF_ALL_PINNED 0x01
3232 #define LBF_NEED_BREAK 0x02
3233 #define LBF_SOME_PINNED 0x04
3236 struct sched_domain *sd;
3244 struct cpumask *dst_grpmask;
3246 enum cpu_idle_type idle;
3248 /* The set of CPUs under consideration for load-balancing */
3249 struct cpumask *cpus;
3254 unsigned int loop_break;
3255 unsigned int loop_max;
3259 * move_task - move a task from one runqueue to another runqueue.
3260 * Both runqueues must be locked.
3262 static void move_task(struct task_struct *p, struct lb_env *env)
3264 deactivate_task(env->src_rq, p, 0);
3265 set_task_cpu(p, env->dst_cpu);
3266 activate_task(env->dst_rq, p, 0);
3267 check_preempt_curr(env->dst_rq, p, 0);
3271 * Is this task likely cache-hot:
3274 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
3278 if (p->sched_class != &fair_sched_class)
3281 if (unlikely(p->policy == SCHED_IDLE))
3285 * Buddy candidates are cache hot:
3287 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
3288 (&p->se == cfs_rq_of(&p->se)->next ||
3289 &p->se == cfs_rq_of(&p->se)->last))
3292 if (sysctl_sched_migration_cost == -1)
3294 if (sysctl_sched_migration_cost == 0)
3297 delta = now - p->se.exec_start;
3299 return delta < (s64)sysctl_sched_migration_cost;
3303 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3306 int can_migrate_task(struct task_struct *p, struct lb_env *env)
3308 int tsk_cache_hot = 0;
3310 * We do not migrate tasks that are:
3311 * 1) running (obviously), or
3312 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3313 * 3) are cache-hot on their current CPU.
3315 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
3318 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
3321 * Remember if this task can be migrated to any other cpu in
3322 * our sched_group. We may want to revisit it if we couldn't
3323 * meet load balance goals by pulling other tasks on src_cpu.
3325 * Also avoid computing new_dst_cpu if we have already computed
3326 * one in current iteration.
3328 if (!env->dst_grpmask || (env->flags & LBF_SOME_PINNED))
3331 new_dst_cpu = cpumask_first_and(env->dst_grpmask,
3332 tsk_cpus_allowed(p));
3333 if (new_dst_cpu < nr_cpu_ids) {
3334 env->flags |= LBF_SOME_PINNED;
3335 env->new_dst_cpu = new_dst_cpu;
3340 /* Record that we found atleast one task that could run on dst_cpu */
3341 env->flags &= ~LBF_ALL_PINNED;
3343 if (task_running(env->src_rq, p)) {
3344 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
3349 * Aggressive migration if:
3350 * 1) task is cache cold, or
3351 * 2) too many balance attempts have failed.
3354 tsk_cache_hot = task_hot(p, env->src_rq->clock_task, env->sd);
3355 if (!tsk_cache_hot ||
3356 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
3357 #ifdef CONFIG_SCHEDSTATS
3358 if (tsk_cache_hot) {
3359 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
3360 schedstat_inc(p, se.statistics.nr_forced_migrations);
3366 if (tsk_cache_hot) {
3367 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
3374 * move_one_task tries to move exactly one task from busiest to this_rq, as
3375 * part of active balancing operations within "domain".
3376 * Returns 1 if successful and 0 otherwise.
3378 * Called with both runqueues locked.
3380 static int move_one_task(struct lb_env *env)
3382 struct task_struct *p, *n;
3384 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
3385 if (throttled_lb_pair(task_group(p), env->src_rq->cpu, env->dst_cpu))
3388 if (!can_migrate_task(p, env))
3393 * Right now, this is only the second place move_task()
3394 * is called, so we can safely collect move_task()
3395 * stats here rather than inside move_task().
3397 schedstat_inc(env->sd, lb_gained[env->idle]);
3403 static unsigned long task_h_load(struct task_struct *p);
3405 static const unsigned int sched_nr_migrate_break = 32;
3408 * move_tasks tries to move up to imbalance weighted load from busiest to
3409 * this_rq, as part of a balancing operation within domain "sd".
3410 * Returns 1 if successful and 0 otherwise.
3412 * Called with both runqueues locked.
3414 static int move_tasks(struct lb_env *env)
3416 struct list_head *tasks = &env->src_rq->cfs_tasks;
3417 struct task_struct *p;
3421 if (env->imbalance <= 0)
3424 while (!list_empty(tasks)) {
3425 p = list_first_entry(tasks, struct task_struct, se.group_node);
3428 /* We've more or less seen every task there is, call it quits */
3429 if (env->loop > env->loop_max)
3432 /* take a breather every nr_migrate tasks */
3433 if (env->loop > env->loop_break) {
3434 env->loop_break += sched_nr_migrate_break;
3435 env->flags |= LBF_NEED_BREAK;
3439 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
3442 load = task_h_load(p);
3444 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
3447 if ((load / 2) > env->imbalance)
3450 if (!can_migrate_task(p, env))
3455 env->imbalance -= load;
3457 #ifdef CONFIG_PREEMPT
3459 * NEWIDLE balancing is a source of latency, so preemptible
3460 * kernels will stop after the first task is pulled to minimize
3461 * the critical section.
3463 if (env->idle == CPU_NEWLY_IDLE)
3468 * We only want to steal up to the prescribed amount of
3471 if (env->imbalance <= 0)
3476 list_move_tail(&p->se.group_node, tasks);
3480 * Right now, this is one of only two places move_task() is called,
3481 * so we can safely collect move_task() stats here rather than
3482 * inside move_task().
3484 schedstat_add(env->sd, lb_gained[env->idle], pulled);
3489 #ifdef CONFIG_FAIR_GROUP_SCHED
3491 * update tg->load_weight by folding this cpu's load_avg
3493 static int update_shares_cpu(struct task_group *tg, int cpu)
3495 struct cfs_rq *cfs_rq;
3496 unsigned long flags;
3503 cfs_rq = tg->cfs_rq[cpu];
3505 raw_spin_lock_irqsave(&rq->lock, flags);
3507 update_rq_clock(rq);
3508 update_cfs_load(cfs_rq, 1);
3511 * We need to update shares after updating tg->load_weight in
3512 * order to adjust the weight of groups with long running tasks.
3514 update_cfs_shares(cfs_rq);
3516 raw_spin_unlock_irqrestore(&rq->lock, flags);
3521 static void update_shares(int cpu)
3523 struct cfs_rq *cfs_rq;
3524 struct rq *rq = cpu_rq(cpu);
3528 * Iterates the task_group tree in a bottom up fashion, see
3529 * list_add_leaf_cfs_rq() for details.
3531 for_each_leaf_cfs_rq(rq, cfs_rq) {
3532 /* throttled entities do not contribute to load */
3533 if (throttled_hierarchy(cfs_rq))
3536 update_shares_cpu(cfs_rq->tg, cpu);
3542 * Compute the cpu's hierarchical load factor for each task group.
3543 * This needs to be done in a top-down fashion because the load of a child
3544 * group is a fraction of its parents load.
3546 static int tg_load_down(struct task_group *tg, void *data)
3549 long cpu = (long)data;
3552 load = cpu_rq(cpu)->load.weight;
3554 load = tg->parent->cfs_rq[cpu]->h_load;
3555 load *= tg->se[cpu]->load.weight;
3556 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
3559 tg->cfs_rq[cpu]->h_load = load;
3564 static void update_h_load(long cpu)
3566 struct rq *rq = cpu_rq(cpu);
3567 unsigned long now = jiffies;
3569 if (rq->h_load_throttle == now)
3572 rq->h_load_throttle = now;
3575 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
3579 static unsigned long task_h_load(struct task_struct *p)
3581 struct cfs_rq *cfs_rq = task_cfs_rq(p);
3584 load = p->se.load.weight;
3585 load = div_u64(load * cfs_rq->h_load, cfs_rq->load.weight + 1);
3590 static inline void update_shares(int cpu)
3594 static inline void update_h_load(long cpu)
3598 static unsigned long task_h_load(struct task_struct *p)
3600 return p->se.load.weight;
3604 /********** Helpers for find_busiest_group ************************/
3606 * sd_lb_stats - Structure to store the statistics of a sched_domain
3607 * during load balancing.
3609 struct sd_lb_stats {
3610 struct sched_group *busiest; /* Busiest group in this sd */
3611 struct sched_group *this; /* Local group in this sd */
3612 unsigned long total_load; /* Total load of all groups in sd */
3613 unsigned long total_pwr; /* Total power of all groups in sd */
3614 unsigned long avg_load; /* Average load across all groups in sd */
3616 /** Statistics of this group */
3617 unsigned long this_load;
3618 unsigned long this_load_per_task;
3619 unsigned long this_nr_running;
3620 unsigned long this_has_capacity;
3621 unsigned int this_idle_cpus;
3623 /* Statistics of the busiest group */
3624 unsigned int busiest_idle_cpus;
3625 unsigned long max_load;
3626 unsigned long busiest_load_per_task;
3627 unsigned long busiest_nr_running;
3628 unsigned long busiest_group_capacity;
3629 unsigned long busiest_has_capacity;
3630 unsigned int busiest_group_weight;
3632 int group_imb; /* Is there imbalance in this sd */
3636 * sg_lb_stats - stats of a sched_group required for load_balancing
3638 struct sg_lb_stats {
3639 unsigned long avg_load; /*Avg load across the CPUs of the group */
3640 unsigned long group_load; /* Total load over the CPUs of the group */
3641 unsigned long sum_nr_running; /* Nr tasks running in the group */
3642 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3643 unsigned long group_capacity;
3644 unsigned long idle_cpus;
3645 unsigned long group_weight;
3646 int group_imb; /* Is there an imbalance in the group ? */
3647 int group_has_capacity; /* Is there extra capacity in the group? */
3651 * get_sd_load_idx - Obtain the load index for a given sched domain.
3652 * @sd: The sched_domain whose load_idx is to be obtained.
3653 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3655 static inline int get_sd_load_idx(struct sched_domain *sd,
3656 enum cpu_idle_type idle)
3662 load_idx = sd->busy_idx;
3665 case CPU_NEWLY_IDLE:
3666 load_idx = sd->newidle_idx;
3669 load_idx = sd->idle_idx;
3676 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3678 return SCHED_POWER_SCALE;
3681 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3683 return default_scale_freq_power(sd, cpu);
3686 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3688 unsigned long weight = sd->span_weight;
3689 unsigned long smt_gain = sd->smt_gain;
3696 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3698 return default_scale_smt_power(sd, cpu);
3701 unsigned long scale_rt_power(int cpu)
3703 struct rq *rq = cpu_rq(cpu);
3704 u64 total, available, age_stamp, avg;
3707 * Since we're reading these variables without serialization make sure
3708 * we read them once before doing sanity checks on them.
3710 age_stamp = ACCESS_ONCE(rq->age_stamp);
3711 avg = ACCESS_ONCE(rq->rt_avg);
3713 total = sched_avg_period() + (rq->clock - age_stamp);
3715 if (unlikely(total < avg)) {
3716 /* Ensures that power won't end up being negative */
3719 available = total - avg;
3722 if (unlikely((s64)total < SCHED_POWER_SCALE))
3723 total = SCHED_POWER_SCALE;
3725 total >>= SCHED_POWER_SHIFT;
3727 return div_u64(available, total);
3730 static void update_cpu_power(struct sched_domain *sd, int cpu)
3732 unsigned long weight = sd->span_weight;
3733 unsigned long power = SCHED_POWER_SCALE;
3734 struct sched_group *sdg = sd->groups;
3736 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3737 if (sched_feat(ARCH_POWER))
3738 power *= arch_scale_smt_power(sd, cpu);
3740 power *= default_scale_smt_power(sd, cpu);
3742 power >>= SCHED_POWER_SHIFT;
3745 sdg->sgp->power_orig = power;
3747 if (sched_feat(ARCH_POWER))
3748 power *= arch_scale_freq_power(sd, cpu);
3750 power *= default_scale_freq_power(sd, cpu);
3752 power >>= SCHED_POWER_SHIFT;
3754 power *= scale_rt_power(cpu);
3755 power >>= SCHED_POWER_SHIFT;
3760 cpu_rq(cpu)->cpu_power = power;
3761 sdg->sgp->power = power;
3764 void update_group_power(struct sched_domain *sd, int cpu)
3766 struct sched_domain *child = sd->child;
3767 struct sched_group *group, *sdg = sd->groups;
3768 unsigned long power;
3769 unsigned long interval;
3771 interval = msecs_to_jiffies(sd->balance_interval);
3772 interval = clamp(interval, 1UL, max_load_balance_interval);
3773 sdg->sgp->next_update = jiffies + interval;
3776 update_cpu_power(sd, cpu);
3782 if (child->flags & SD_OVERLAP) {
3784 * SD_OVERLAP domains cannot assume that child groups
3785 * span the current group.
3788 for_each_cpu(cpu, sched_group_cpus(sdg))
3789 power += power_of(cpu);
3792 * !SD_OVERLAP domains can assume that child groups
3793 * span the current group.
3796 group = child->groups;
3798 power += group->sgp->power;
3799 group = group->next;
3800 } while (group != child->groups);
3803 sdg->sgp->power_orig = sdg->sgp->power = power;
3807 * Try and fix up capacity for tiny siblings, this is needed when
3808 * things like SD_ASYM_PACKING need f_b_g to select another sibling
3809 * which on its own isn't powerful enough.
3811 * See update_sd_pick_busiest() and check_asym_packing().
3814 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
3817 * Only siblings can have significantly less than SCHED_POWER_SCALE
3819 if (!(sd->flags & SD_SHARE_CPUPOWER))
3823 * If ~90% of the cpu_power is still there, we're good.
3825 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
3832 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3833 * @env: The load balancing environment.
3834 * @group: sched_group whose statistics are to be updated.
3835 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3836 * @local_group: Does group contain this_cpu.
3837 * @balance: Should we balance.
3838 * @sgs: variable to hold the statistics for this group.
3840 static inline void update_sg_lb_stats(struct lb_env *env,
3841 struct sched_group *group, int load_idx,
3842 int local_group, int *balance, struct sg_lb_stats *sgs)
3844 unsigned long nr_running, max_nr_running, min_nr_running;
3845 unsigned long load, max_cpu_load, min_cpu_load;
3846 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3847 unsigned long avg_load_per_task = 0;
3851 balance_cpu = group_balance_cpu(group);
3853 /* Tally up the load of all CPUs in the group */
3855 min_cpu_load = ~0UL;
3857 min_nr_running = ~0UL;
3859 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
3860 struct rq *rq = cpu_rq(i);
3862 nr_running = rq->nr_running;
3864 /* Bias balancing toward cpus of our domain */
3866 if (idle_cpu(i) && !first_idle_cpu &&
3867 cpumask_test_cpu(i, sched_group_mask(group))) {
3872 load = target_load(i, load_idx);
3874 load = source_load(i, load_idx);
3875 if (load > max_cpu_load)
3876 max_cpu_load = load;
3877 if (min_cpu_load > load)
3878 min_cpu_load = load;
3880 if (nr_running > max_nr_running)
3881 max_nr_running = nr_running;
3882 if (min_nr_running > nr_running)
3883 min_nr_running = nr_running;
3886 sgs->group_load += load;
3887 sgs->sum_nr_running += nr_running;
3888 sgs->sum_weighted_load += weighted_cpuload(i);
3894 * First idle cpu or the first cpu(busiest) in this sched group
3895 * is eligible for doing load balancing at this and above
3896 * domains. In the newly idle case, we will allow all the cpu's
3897 * to do the newly idle load balance.
3900 if (env->idle != CPU_NEWLY_IDLE) {
3901 if (balance_cpu != env->dst_cpu) {
3905 update_group_power(env->sd, env->dst_cpu);
3906 } else if (time_after_eq(jiffies, group->sgp->next_update))
3907 update_group_power(env->sd, env->dst_cpu);
3910 /* Adjust by relative CPU power of the group */
3911 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
3914 * Consider the group unbalanced when the imbalance is larger
3915 * than the average weight of a task.
3917 * APZ: with cgroup the avg task weight can vary wildly and
3918 * might not be a suitable number - should we keep a
3919 * normalized nr_running number somewhere that negates
3922 if (sgs->sum_nr_running)
3923 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
3925 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task &&
3926 (max_nr_running - min_nr_running) > 1)
3929 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
3931 if (!sgs->group_capacity)
3932 sgs->group_capacity = fix_small_capacity(env->sd, group);
3933 sgs->group_weight = group->group_weight;
3935 if (sgs->group_capacity > sgs->sum_nr_running)
3936 sgs->group_has_capacity = 1;
3940 * update_sd_pick_busiest - return 1 on busiest group
3941 * @env: The load balancing environment.
3942 * @sds: sched_domain statistics
3943 * @sg: sched_group candidate to be checked for being the busiest
3944 * @sgs: sched_group statistics
3946 * Determine if @sg is a busier group than the previously selected
3949 static bool update_sd_pick_busiest(struct lb_env *env,
3950 struct sd_lb_stats *sds,
3951 struct sched_group *sg,
3952 struct sg_lb_stats *sgs)
3954 if (sgs->avg_load <= sds->max_load)
3957 if (sgs->sum_nr_running > sgs->group_capacity)
3964 * ASYM_PACKING needs to move all the work to the lowest
3965 * numbered CPUs in the group, therefore mark all groups
3966 * higher than ourself as busy.
3968 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
3969 env->dst_cpu < group_first_cpu(sg)) {
3973 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
3981 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
3982 * @env: The load balancing environment.
3983 * @balance: Should we balance.
3984 * @sds: variable to hold the statistics for this sched_domain.
3986 static inline void update_sd_lb_stats(struct lb_env *env,
3987 int *balance, struct sd_lb_stats *sds)
3989 struct sched_domain *child = env->sd->child;
3990 struct sched_group *sg = env->sd->groups;
3991 struct sg_lb_stats sgs;
3992 int load_idx, prefer_sibling = 0;
3994 if (child && child->flags & SD_PREFER_SIBLING)
3997 load_idx = get_sd_load_idx(env->sd, env->idle);
4002 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
4003 memset(&sgs, 0, sizeof(sgs));
4004 update_sg_lb_stats(env, sg, load_idx, local_group, balance, &sgs);
4006 if (local_group && !(*balance))
4009 sds->total_load += sgs.group_load;
4010 sds->total_pwr += sg->sgp->power;
4013 * In case the child domain prefers tasks go to siblings
4014 * first, lower the sg capacity to one so that we'll try
4015 * and move all the excess tasks away. We lower the capacity
4016 * of a group only if the local group has the capacity to fit
4017 * these excess tasks, i.e. nr_running < group_capacity. The
4018 * extra check prevents the case where you always pull from the
4019 * heaviest group when it is already under-utilized (possible
4020 * with a large weight task outweighs the tasks on the system).
4022 if (prefer_sibling && !local_group && sds->this_has_capacity)
4023 sgs.group_capacity = min(sgs.group_capacity, 1UL);
4026 sds->this_load = sgs.avg_load;
4028 sds->this_nr_running = sgs.sum_nr_running;
4029 sds->this_load_per_task = sgs.sum_weighted_load;
4030 sds->this_has_capacity = sgs.group_has_capacity;
4031 sds->this_idle_cpus = sgs.idle_cpus;
4032 } else if (update_sd_pick_busiest(env, sds, sg, &sgs)) {
4033 sds->max_load = sgs.avg_load;
4035 sds->busiest_nr_running = sgs.sum_nr_running;
4036 sds->busiest_idle_cpus = sgs.idle_cpus;
4037 sds->busiest_group_capacity = sgs.group_capacity;
4038 sds->busiest_load_per_task = sgs.sum_weighted_load;
4039 sds->busiest_has_capacity = sgs.group_has_capacity;
4040 sds->busiest_group_weight = sgs.group_weight;
4041 sds->group_imb = sgs.group_imb;
4045 } while (sg != env->sd->groups);
4049 * check_asym_packing - Check to see if the group is packed into the
4052 * This is primarily intended to used at the sibling level. Some
4053 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4054 * case of POWER7, it can move to lower SMT modes only when higher
4055 * threads are idle. When in lower SMT modes, the threads will
4056 * perform better since they share less core resources. Hence when we
4057 * have idle threads, we want them to be the higher ones.
4059 * This packing function is run on idle threads. It checks to see if
4060 * the busiest CPU in this domain (core in the P7 case) has a higher
4061 * CPU number than the packing function is being run on. Here we are
4062 * assuming lower CPU number will be equivalent to lower a SMT thread
4065 * Returns 1 when packing is required and a task should be moved to
4066 * this CPU. The amount of the imbalance is returned in *imbalance.
4068 * @env: The load balancing environment.
4069 * @sds: Statistics of the sched_domain which is to be packed
4071 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
4075 if (!(env->sd->flags & SD_ASYM_PACKING))
4081 busiest_cpu = group_first_cpu(sds->busiest);
4082 if (env->dst_cpu > busiest_cpu)
4085 env->imbalance = DIV_ROUND_CLOSEST(
4086 sds->max_load * sds->busiest->sgp->power, SCHED_POWER_SCALE);
4092 * fix_small_imbalance - Calculate the minor imbalance that exists
4093 * amongst the groups of a sched_domain, during
4095 * @env: The load balancing environment.
4096 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4099 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4101 unsigned long tmp, pwr_now = 0, pwr_move = 0;
4102 unsigned int imbn = 2;
4103 unsigned long scaled_busy_load_per_task;
4105 if (sds->this_nr_running) {
4106 sds->this_load_per_task /= sds->this_nr_running;
4107 if (sds->busiest_load_per_task >
4108 sds->this_load_per_task)
4111 sds->this_load_per_task =
4112 cpu_avg_load_per_task(env->dst_cpu);
4115 scaled_busy_load_per_task = sds->busiest_load_per_task
4116 * SCHED_POWER_SCALE;
4117 scaled_busy_load_per_task /= sds->busiest->sgp->power;
4119 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
4120 (scaled_busy_load_per_task * imbn)) {
4121 env->imbalance = sds->busiest_load_per_task;
4126 * OK, we don't have enough imbalance to justify moving tasks,
4127 * however we may be able to increase total CPU power used by
4131 pwr_now += sds->busiest->sgp->power *
4132 min(sds->busiest_load_per_task, sds->max_load);
4133 pwr_now += sds->this->sgp->power *
4134 min(sds->this_load_per_task, sds->this_load);
4135 pwr_now /= SCHED_POWER_SCALE;
4137 /* Amount of load we'd subtract */
4138 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4139 sds->busiest->sgp->power;
4140 if (sds->max_load > tmp)
4141 pwr_move += sds->busiest->sgp->power *
4142 min(sds->busiest_load_per_task, sds->max_load - tmp);
4144 /* Amount of load we'd add */
4145 if (sds->max_load * sds->busiest->sgp->power <
4146 sds->busiest_load_per_task * SCHED_POWER_SCALE)
4147 tmp = (sds->max_load * sds->busiest->sgp->power) /
4148 sds->this->sgp->power;
4150 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4151 sds->this->sgp->power;
4152 pwr_move += sds->this->sgp->power *
4153 min(sds->this_load_per_task, sds->this_load + tmp);
4154 pwr_move /= SCHED_POWER_SCALE;
4156 /* Move if we gain throughput */
4157 if (pwr_move > pwr_now)
4158 env->imbalance = sds->busiest_load_per_task;
4162 * calculate_imbalance - Calculate the amount of imbalance present within the
4163 * groups of a given sched_domain during load balance.
4164 * @env: load balance environment
4165 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4167 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4169 unsigned long max_pull, load_above_capacity = ~0UL;
4171 sds->busiest_load_per_task /= sds->busiest_nr_running;
4172 if (sds->group_imb) {
4173 sds->busiest_load_per_task =
4174 min(sds->busiest_load_per_task, sds->avg_load);
4178 * In the presence of smp nice balancing, certain scenarios can have
4179 * max load less than avg load(as we skip the groups at or below
4180 * its cpu_power, while calculating max_load..)
4182 if (sds->max_load < sds->avg_load) {
4184 return fix_small_imbalance(env, sds);
4187 if (!sds->group_imb) {
4189 * Don't want to pull so many tasks that a group would go idle.
4191 load_above_capacity = (sds->busiest_nr_running -
4192 sds->busiest_group_capacity);
4194 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
4196 load_above_capacity /= sds->busiest->sgp->power;
4200 * We're trying to get all the cpus to the average_load, so we don't
4201 * want to push ourselves above the average load, nor do we wish to
4202 * reduce the max loaded cpu below the average load. At the same time,
4203 * we also don't want to reduce the group load below the group capacity
4204 * (so that we can implement power-savings policies etc). Thus we look
4205 * for the minimum possible imbalance.
4206 * Be careful of negative numbers as they'll appear as very large values
4207 * with unsigned longs.
4209 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
4211 /* How much load to actually move to equalise the imbalance */
4212 env->imbalance = min(max_pull * sds->busiest->sgp->power,
4213 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
4214 / SCHED_POWER_SCALE;
4217 * if *imbalance is less than the average load per runnable task
4218 * there is no guarantee that any tasks will be moved so we'll have
4219 * a think about bumping its value to force at least one task to be
4222 if (env->imbalance < sds->busiest_load_per_task)
4223 return fix_small_imbalance(env, sds);
4227 /******* find_busiest_group() helpers end here *********************/
4230 * find_busiest_group - Returns the busiest group within the sched_domain
4231 * if there is an imbalance. If there isn't an imbalance, and
4232 * the user has opted for power-savings, it returns a group whose
4233 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4234 * such a group exists.
4236 * Also calculates the amount of weighted load which should be moved
4237 * to restore balance.
4239 * @env: The load balancing environment.
4240 * @balance: Pointer to a variable indicating if this_cpu
4241 * is the appropriate cpu to perform load balancing at this_level.
4243 * Returns: - the busiest group if imbalance exists.
4244 * - If no imbalance and user has opted for power-savings balance,
4245 * return the least loaded group whose CPUs can be
4246 * put to idle by rebalancing its tasks onto our group.
4248 static struct sched_group *
4249 find_busiest_group(struct lb_env *env, int *balance)
4251 struct sd_lb_stats sds;
4253 memset(&sds, 0, sizeof(sds));
4256 * Compute the various statistics relavent for load balancing at
4259 update_sd_lb_stats(env, balance, &sds);
4262 * this_cpu is not the appropriate cpu to perform load balancing at
4268 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
4269 check_asym_packing(env, &sds))
4272 /* There is no busy sibling group to pull tasks from */
4273 if (!sds.busiest || sds.busiest_nr_running == 0)
4276 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
4279 * If the busiest group is imbalanced the below checks don't
4280 * work because they assumes all things are equal, which typically
4281 * isn't true due to cpus_allowed constraints and the like.
4286 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
4287 if (env->idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
4288 !sds.busiest_has_capacity)
4292 * If the local group is more busy than the selected busiest group
4293 * don't try and pull any tasks.
4295 if (sds.this_load >= sds.max_load)
4299 * Don't pull any tasks if this group is already above the domain
4302 if (sds.this_load >= sds.avg_load)
4305 if (env->idle == CPU_IDLE) {
4307 * This cpu is idle. If the busiest group load doesn't
4308 * have more tasks than the number of available cpu's and
4309 * there is no imbalance between this and busiest group
4310 * wrt to idle cpu's, it is balanced.
4312 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
4313 sds.busiest_nr_running <= sds.busiest_group_weight)
4317 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
4318 * imbalance_pct to be conservative.
4320 if (100 * sds.max_load <= env->sd->imbalance_pct * sds.this_load)
4325 /* Looks like there is an imbalance. Compute it */
4326 calculate_imbalance(env, &sds);
4336 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4338 static struct rq *find_busiest_queue(struct lb_env *env,
4339 struct sched_group *group)
4341 struct rq *busiest = NULL, *rq;
4342 unsigned long max_load = 0;
4345 for_each_cpu(i, sched_group_cpus(group)) {
4346 unsigned long power = power_of(i);
4347 unsigned long capacity = DIV_ROUND_CLOSEST(power,
4352 capacity = fix_small_capacity(env->sd, group);
4354 if (!cpumask_test_cpu(i, env->cpus))
4358 wl = weighted_cpuload(i);
4361 * When comparing with imbalance, use weighted_cpuload()
4362 * which is not scaled with the cpu power.
4364 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
4368 * For the load comparisons with the other cpu's, consider
4369 * the weighted_cpuload() scaled with the cpu power, so that
4370 * the load can be moved away from the cpu that is potentially
4371 * running at a lower capacity.
4373 wl = (wl * SCHED_POWER_SCALE) / power;
4375 if (wl > max_load) {
4385 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4386 * so long as it is large enough.
4388 #define MAX_PINNED_INTERVAL 512
4390 /* Working cpumask for load_balance and load_balance_newidle. */
4391 DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4393 static int need_active_balance(struct lb_env *env)
4395 struct sched_domain *sd = env->sd;
4397 if (env->idle == CPU_NEWLY_IDLE) {
4400 * ASYM_PACKING needs to force migrate tasks from busy but
4401 * higher numbered CPUs in order to pack all tasks in the
4402 * lowest numbered CPUs.
4404 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
4408 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
4411 static int active_load_balance_cpu_stop(void *data);
4414 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4415 * tasks if there is an imbalance.
4417 static int load_balance(int this_cpu, struct rq *this_rq,
4418 struct sched_domain *sd, enum cpu_idle_type idle,
4421 int ld_moved, cur_ld_moved, active_balance = 0;
4422 int lb_iterations, max_lb_iterations;
4423 struct sched_group *group;
4425 unsigned long flags;
4426 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4428 struct lb_env env = {
4430 .dst_cpu = this_cpu,
4432 .dst_grpmask = sched_group_cpus(sd->groups),
4434 .loop_break = sched_nr_migrate_break,
4438 cpumask_copy(cpus, cpu_active_mask);
4439 max_lb_iterations = cpumask_weight(env.dst_grpmask);
4441 schedstat_inc(sd, lb_count[idle]);
4444 group = find_busiest_group(&env, balance);
4450 schedstat_inc(sd, lb_nobusyg[idle]);
4454 busiest = find_busiest_queue(&env, group);
4456 schedstat_inc(sd, lb_nobusyq[idle]);
4460 BUG_ON(busiest == env.dst_rq);
4462 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
4466 if (busiest->nr_running > 1) {
4468 * Attempt to move tasks. If find_busiest_group has found
4469 * an imbalance but busiest->nr_running <= 1, the group is
4470 * still unbalanced. ld_moved simply stays zero, so it is
4471 * correctly treated as an imbalance.
4473 env.flags |= LBF_ALL_PINNED;
4474 env.src_cpu = busiest->cpu;
4475 env.src_rq = busiest;
4476 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
4478 update_h_load(env.src_cpu);
4480 local_irq_save(flags);
4481 double_rq_lock(env.dst_rq, busiest);
4484 * cur_ld_moved - load moved in current iteration
4485 * ld_moved - cumulative load moved across iterations
4487 cur_ld_moved = move_tasks(&env);
4488 ld_moved += cur_ld_moved;
4489 double_rq_unlock(env.dst_rq, busiest);
4490 local_irq_restore(flags);
4492 if (env.flags & LBF_NEED_BREAK) {
4493 env.flags &= ~LBF_NEED_BREAK;
4498 * some other cpu did the load balance for us.
4500 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
4501 resched_cpu(env.dst_cpu);
4504 * Revisit (affine) tasks on src_cpu that couldn't be moved to
4505 * us and move them to an alternate dst_cpu in our sched_group
4506 * where they can run. The upper limit on how many times we
4507 * iterate on same src_cpu is dependent on number of cpus in our
4510 * This changes load balance semantics a bit on who can move
4511 * load to a given_cpu. In addition to the given_cpu itself
4512 * (or a ilb_cpu acting on its behalf where given_cpu is
4513 * nohz-idle), we now have balance_cpu in a position to move
4514 * load to given_cpu. In rare situations, this may cause
4515 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
4516 * _independently_ and at _same_ time to move some load to
4517 * given_cpu) causing exceess load to be moved to given_cpu.
4518 * This however should not happen so much in practice and
4519 * moreover subsequent load balance cycles should correct the
4520 * excess load moved.
4522 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0 &&
4523 lb_iterations++ < max_lb_iterations) {
4525 env.dst_rq = cpu_rq(env.new_dst_cpu);
4526 env.dst_cpu = env.new_dst_cpu;
4527 env.flags &= ~LBF_SOME_PINNED;
4529 env.loop_break = sched_nr_migrate_break;
4531 * Go back to "more_balance" rather than "redo" since we
4532 * need to continue with same src_cpu.
4537 /* All tasks on this runqueue were pinned by CPU affinity */
4538 if (unlikely(env.flags & LBF_ALL_PINNED)) {
4539 cpumask_clear_cpu(cpu_of(busiest), cpus);
4540 if (!cpumask_empty(cpus)) {
4542 env.loop_break = sched_nr_migrate_break;
4550 schedstat_inc(sd, lb_failed[idle]);
4552 * Increment the failure counter only on periodic balance.
4553 * We do not want newidle balance, which can be very
4554 * frequent, pollute the failure counter causing
4555 * excessive cache_hot migrations and active balances.
4557 if (idle != CPU_NEWLY_IDLE)
4558 sd->nr_balance_failed++;
4560 if (need_active_balance(&env)) {
4561 raw_spin_lock_irqsave(&busiest->lock, flags);
4563 /* don't kick the active_load_balance_cpu_stop,
4564 * if the curr task on busiest cpu can't be
4567 if (!cpumask_test_cpu(this_cpu,
4568 tsk_cpus_allowed(busiest->curr))) {
4569 raw_spin_unlock_irqrestore(&busiest->lock,
4571 env.flags |= LBF_ALL_PINNED;
4572 goto out_one_pinned;
4576 * ->active_balance synchronizes accesses to
4577 * ->active_balance_work. Once set, it's cleared
4578 * only after active load balance is finished.
4580 if (!busiest->active_balance) {
4581 busiest->active_balance = 1;
4582 busiest->push_cpu = this_cpu;
4585 raw_spin_unlock_irqrestore(&busiest->lock, flags);
4587 if (active_balance) {
4588 stop_one_cpu_nowait(cpu_of(busiest),
4589 active_load_balance_cpu_stop, busiest,
4590 &busiest->active_balance_work);
4594 * We've kicked active balancing, reset the failure
4597 sd->nr_balance_failed = sd->cache_nice_tries+1;
4600 sd->nr_balance_failed = 0;
4602 if (likely(!active_balance)) {
4603 /* We were unbalanced, so reset the balancing interval */
4604 sd->balance_interval = sd->min_interval;
4607 * If we've begun active balancing, start to back off. This
4608 * case may not be covered by the all_pinned logic if there
4609 * is only 1 task on the busy runqueue (because we don't call
4612 if (sd->balance_interval < sd->max_interval)
4613 sd->balance_interval *= 2;
4619 schedstat_inc(sd, lb_balanced[idle]);
4621 sd->nr_balance_failed = 0;
4624 /* tune up the balancing interval */
4625 if (((env.flags & LBF_ALL_PINNED) &&
4626 sd->balance_interval < MAX_PINNED_INTERVAL) ||
4627 (sd->balance_interval < sd->max_interval))
4628 sd->balance_interval *= 2;
4636 * idle_balance is called by schedule() if this_cpu is about to become
4637 * idle. Attempts to pull tasks from other CPUs.
4639 void idle_balance(int this_cpu, struct rq *this_rq)
4641 struct sched_domain *sd;
4642 int pulled_task = 0;
4643 unsigned long next_balance = jiffies + HZ;
4645 this_rq->idle_stamp = this_rq->clock;
4647 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4651 * Drop the rq->lock, but keep IRQ/preempt disabled.
4653 raw_spin_unlock(&this_rq->lock);
4655 update_shares(this_cpu);
4657 for_each_domain(this_cpu, sd) {
4658 unsigned long interval;
4661 if (!(sd->flags & SD_LOAD_BALANCE))
4664 if (sd->flags & SD_BALANCE_NEWIDLE) {
4665 /* If we've pulled tasks over stop searching: */
4666 pulled_task = load_balance(this_cpu, this_rq,
4667 sd, CPU_NEWLY_IDLE, &balance);
4670 interval = msecs_to_jiffies(sd->balance_interval);
4671 if (time_after(next_balance, sd->last_balance + interval))
4672 next_balance = sd->last_balance + interval;
4674 this_rq->idle_stamp = 0;
4680 raw_spin_lock(&this_rq->lock);
4682 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4684 * We are going idle. next_balance may be set based on
4685 * a busy processor. So reset next_balance.
4687 this_rq->next_balance = next_balance;
4692 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
4693 * running tasks off the busiest CPU onto idle CPUs. It requires at
4694 * least 1 task to be running on each physical CPU where possible, and
4695 * avoids physical / logical imbalances.
4697 static int active_load_balance_cpu_stop(void *data)
4699 struct rq *busiest_rq = data;
4700 int busiest_cpu = cpu_of(busiest_rq);
4701 int target_cpu = busiest_rq->push_cpu;
4702 struct rq *target_rq = cpu_rq(target_cpu);
4703 struct sched_domain *sd;
4705 raw_spin_lock_irq(&busiest_rq->lock);
4707 /* make sure the requested cpu hasn't gone down in the meantime */
4708 if (unlikely(busiest_cpu != smp_processor_id() ||
4709 !busiest_rq->active_balance))
4712 /* Is there any task to move? */
4713 if (busiest_rq->nr_running <= 1)
4717 * This condition is "impossible", if it occurs
4718 * we need to fix it. Originally reported by
4719 * Bjorn Helgaas on a 128-cpu setup.
4721 BUG_ON(busiest_rq == target_rq);
4723 /* move a task from busiest_rq to target_rq */
4724 double_lock_balance(busiest_rq, target_rq);
4726 /* Search for an sd spanning us and the target CPU. */
4728 for_each_domain(target_cpu, sd) {
4729 if ((sd->flags & SD_LOAD_BALANCE) &&
4730 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4735 struct lb_env env = {
4737 .dst_cpu = target_cpu,
4738 .dst_rq = target_rq,
4739 .src_cpu = busiest_rq->cpu,
4740 .src_rq = busiest_rq,
4744 schedstat_inc(sd, alb_count);
4746 if (move_one_task(&env))
4747 schedstat_inc(sd, alb_pushed);
4749 schedstat_inc(sd, alb_failed);
4752 double_unlock_balance(busiest_rq, target_rq);
4754 busiest_rq->active_balance = 0;
4755 raw_spin_unlock_irq(&busiest_rq->lock);
4761 * idle load balancing details
4762 * - When one of the busy CPUs notice that there may be an idle rebalancing
4763 * needed, they will kick the idle load balancer, which then does idle
4764 * load balancing for all the idle CPUs.
4767 cpumask_var_t idle_cpus_mask;
4769 unsigned long next_balance; /* in jiffy units */
4770 } nohz ____cacheline_aligned;
4772 static inline int find_new_ilb(int call_cpu)
4774 int ilb = cpumask_first(nohz.idle_cpus_mask);
4776 if (ilb < nr_cpu_ids && idle_cpu(ilb))
4783 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
4784 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
4785 * CPU (if there is one).
4787 static void nohz_balancer_kick(int cpu)
4791 nohz.next_balance++;
4793 ilb_cpu = find_new_ilb(cpu);
4795 if (ilb_cpu >= nr_cpu_ids)
4798 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
4801 * Use smp_send_reschedule() instead of resched_cpu().
4802 * This way we generate a sched IPI on the target cpu which
4803 * is idle. And the softirq performing nohz idle load balance
4804 * will be run before returning from the IPI.
4806 smp_send_reschedule(ilb_cpu);
4810 static inline void nohz_balance_exit_idle(int cpu)
4812 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
4813 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
4814 atomic_dec(&nohz.nr_cpus);
4815 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
4819 static inline void set_cpu_sd_state_busy(void)
4821 struct sched_domain *sd;
4822 int cpu = smp_processor_id();
4824 if (!test_bit(NOHZ_IDLE, nohz_flags(cpu)))
4826 clear_bit(NOHZ_IDLE, nohz_flags(cpu));
4829 for_each_domain(cpu, sd)
4830 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
4834 void set_cpu_sd_state_idle(void)
4836 struct sched_domain *sd;
4837 int cpu = smp_processor_id();
4839 if (test_bit(NOHZ_IDLE, nohz_flags(cpu)))
4841 set_bit(NOHZ_IDLE, nohz_flags(cpu));
4844 for_each_domain(cpu, sd)
4845 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
4850 * This routine will record that the cpu is going idle with tick stopped.
4851 * This info will be used in performing idle load balancing in the future.
4853 void nohz_balance_enter_idle(int cpu)
4856 * If this cpu is going down, then nothing needs to be done.
4858 if (!cpu_active(cpu))
4861 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
4864 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
4865 atomic_inc(&nohz.nr_cpus);
4866 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
4869 static int __cpuinit sched_ilb_notifier(struct notifier_block *nfb,
4870 unsigned long action, void *hcpu)
4872 switch (action & ~CPU_TASKS_FROZEN) {
4874 nohz_balance_exit_idle(smp_processor_id());
4882 static DEFINE_SPINLOCK(balancing);
4885 * Scale the max load_balance interval with the number of CPUs in the system.
4886 * This trades load-balance latency on larger machines for less cross talk.
4888 void update_max_interval(void)
4890 max_load_balance_interval = HZ*num_online_cpus()/10;
4894 * It checks each scheduling domain to see if it is due to be balanced,
4895 * and initiates a balancing operation if so.
4897 * Balancing parameters are set up in arch_init_sched_domains.
4899 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4902 struct rq *rq = cpu_rq(cpu);
4903 unsigned long interval;
4904 struct sched_domain *sd;
4905 /* Earliest time when we have to do rebalance again */
4906 unsigned long next_balance = jiffies + 60*HZ;
4907 int update_next_balance = 0;
4913 for_each_domain(cpu, sd) {
4914 if (!(sd->flags & SD_LOAD_BALANCE))
4917 interval = sd->balance_interval;
4918 if (idle != CPU_IDLE)
4919 interval *= sd->busy_factor;
4921 /* scale ms to jiffies */
4922 interval = msecs_to_jiffies(interval);
4923 interval = clamp(interval, 1UL, max_load_balance_interval);
4925 need_serialize = sd->flags & SD_SERIALIZE;
4927 if (need_serialize) {
4928 if (!spin_trylock(&balancing))
4932 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4933 if (load_balance(cpu, rq, sd, idle, &balance)) {
4935 * We've pulled tasks over so either we're no
4938 idle = CPU_NOT_IDLE;
4940 sd->last_balance = jiffies;
4943 spin_unlock(&balancing);
4945 if (time_after(next_balance, sd->last_balance + interval)) {
4946 next_balance = sd->last_balance + interval;
4947 update_next_balance = 1;
4951 * Stop the load balance at this level. There is another
4952 * CPU in our sched group which is doing load balancing more
4961 * next_balance will be updated only when there is a need.
4962 * When the cpu is attached to null domain for ex, it will not be
4965 if (likely(update_next_balance))
4966 rq->next_balance = next_balance;
4971 * In CONFIG_NO_HZ case, the idle balance kickee will do the
4972 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4974 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
4976 struct rq *this_rq = cpu_rq(this_cpu);
4980 if (idle != CPU_IDLE ||
4981 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
4984 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
4985 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
4989 * If this cpu gets work to do, stop the load balancing
4990 * work being done for other cpus. Next load
4991 * balancing owner will pick it up.
4996 rq = cpu_rq(balance_cpu);
4998 raw_spin_lock_irq(&rq->lock);
4999 update_rq_clock(rq);
5000 update_idle_cpu_load(rq);
5001 raw_spin_unlock_irq(&rq->lock);
5003 rebalance_domains(balance_cpu, CPU_IDLE);
5005 if (time_after(this_rq->next_balance, rq->next_balance))
5006 this_rq->next_balance = rq->next_balance;
5008 nohz.next_balance = this_rq->next_balance;
5010 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
5014 * Current heuristic for kicking the idle load balancer in the presence
5015 * of an idle cpu is the system.
5016 * - This rq has more than one task.
5017 * - At any scheduler domain level, this cpu's scheduler group has multiple
5018 * busy cpu's exceeding the group's power.
5019 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
5020 * domain span are idle.
5022 static inline int nohz_kick_needed(struct rq *rq, int cpu)
5024 unsigned long now = jiffies;
5025 struct sched_domain *sd;
5027 if (unlikely(idle_cpu(cpu)))
5031 * We may be recently in ticked or tickless idle mode. At the first
5032 * busy tick after returning from idle, we will update the busy stats.
5034 set_cpu_sd_state_busy();
5035 nohz_balance_exit_idle(cpu);
5038 * None are in tickless mode and hence no need for NOHZ idle load
5041 if (likely(!atomic_read(&nohz.nr_cpus)))
5044 if (time_before(now, nohz.next_balance))
5047 if (rq->nr_running >= 2)
5051 for_each_domain(cpu, sd) {
5052 struct sched_group *sg = sd->groups;
5053 struct sched_group_power *sgp = sg->sgp;
5054 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
5056 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
5057 goto need_kick_unlock;
5059 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
5060 && (cpumask_first_and(nohz.idle_cpus_mask,
5061 sched_domain_span(sd)) < cpu))
5062 goto need_kick_unlock;
5064 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
5076 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
5080 * run_rebalance_domains is triggered when needed from the scheduler tick.
5081 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
5083 static void run_rebalance_domains(struct softirq_action *h)
5085 int this_cpu = smp_processor_id();
5086 struct rq *this_rq = cpu_rq(this_cpu);
5087 enum cpu_idle_type idle = this_rq->idle_balance ?
5088 CPU_IDLE : CPU_NOT_IDLE;
5090 rebalance_domains(this_cpu, idle);
5093 * If this cpu has a pending nohz_balance_kick, then do the
5094 * balancing on behalf of the other idle cpus whose ticks are
5097 nohz_idle_balance(this_cpu, idle);
5100 static inline int on_null_domain(int cpu)
5102 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
5106 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
5108 void trigger_load_balance(struct rq *rq, int cpu)
5110 /* Don't need to rebalance while attached to NULL domain */
5111 if (time_after_eq(jiffies, rq->next_balance) &&
5112 likely(!on_null_domain(cpu)))
5113 raise_softirq(SCHED_SOFTIRQ);
5115 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
5116 nohz_balancer_kick(cpu);
5120 static void rq_online_fair(struct rq *rq)
5125 static void rq_offline_fair(struct rq *rq)
5129 /* Ensure any throttled groups are reachable by pick_next_task */
5130 unthrottle_offline_cfs_rqs(rq);
5133 #endif /* CONFIG_SMP */
5136 * scheduler tick hitting a task of our scheduling class:
5138 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
5140 struct cfs_rq *cfs_rq;
5141 struct sched_entity *se = &curr->se;
5143 for_each_sched_entity(se) {
5144 cfs_rq = cfs_rq_of(se);
5145 entity_tick(cfs_rq, se, queued);
5148 if (sched_feat_numa(NUMA))
5149 task_tick_numa(rq, curr);
5153 * called on fork with the child task as argument from the parent's context
5154 * - child not yet on the tasklist
5155 * - preemption disabled
5157 static void task_fork_fair(struct task_struct *p)
5159 struct cfs_rq *cfs_rq;
5160 struct sched_entity *se = &p->se, *curr;
5161 int this_cpu = smp_processor_id();
5162 struct rq *rq = this_rq();
5163 unsigned long flags;
5165 raw_spin_lock_irqsave(&rq->lock, flags);
5167 update_rq_clock(rq);
5169 cfs_rq = task_cfs_rq(current);
5170 curr = cfs_rq->curr;
5172 if (unlikely(task_cpu(p) != this_cpu)) {
5174 __set_task_cpu(p, this_cpu);
5178 update_curr(cfs_rq);
5181 se->vruntime = curr->vruntime;
5182 place_entity(cfs_rq, se, 1);
5184 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
5186 * Upon rescheduling, sched_class::put_prev_task() will place
5187 * 'current' within the tree based on its new key value.
5189 swap(curr->vruntime, se->vruntime);
5190 resched_task(rq->curr);
5193 se->vruntime -= cfs_rq->min_vruntime;
5195 raw_spin_unlock_irqrestore(&rq->lock, flags);
5199 * Priority of the task has changed. Check to see if we preempt
5203 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
5209 * Reschedule if we are currently running on this runqueue and
5210 * our priority decreased, or if we are not currently running on
5211 * this runqueue and our priority is higher than the current's
5213 if (rq->curr == p) {
5214 if (p->prio > oldprio)
5215 resched_task(rq->curr);
5217 check_preempt_curr(rq, p, 0);
5220 static void switched_from_fair(struct rq *rq, struct task_struct *p)
5222 struct sched_entity *se = &p->se;
5223 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5226 * Ensure the task's vruntime is normalized, so that when its
5227 * switched back to the fair class the enqueue_entity(.flags=0) will
5228 * do the right thing.
5230 * If it was on_rq, then the dequeue_entity(.flags=0) will already
5231 * have normalized the vruntime, if it was !on_rq, then only when
5232 * the task is sleeping will it still have non-normalized vruntime.
5234 if (!se->on_rq && p->state != TASK_RUNNING) {
5236 * Fix up our vruntime so that the current sleep doesn't
5237 * cause 'unlimited' sleep bonus.
5239 place_entity(cfs_rq, se, 0);
5240 se->vruntime -= cfs_rq->min_vruntime;
5245 * We switched to the sched_fair class.
5247 static void switched_to_fair(struct rq *rq, struct task_struct *p)
5253 * We were most likely switched from sched_rt, so
5254 * kick off the schedule if running, otherwise just see
5255 * if we can still preempt the current task.
5258 resched_task(rq->curr);
5260 check_preempt_curr(rq, p, 0);
5263 /* Account for a task changing its policy or group.
5265 * This routine is mostly called to set cfs_rq->curr field when a task
5266 * migrates between groups/classes.
5268 static void set_curr_task_fair(struct rq *rq)
5270 struct sched_entity *se = &rq->curr->se;
5272 for_each_sched_entity(se) {
5273 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5275 set_next_entity(cfs_rq, se);
5276 /* ensure bandwidth has been allocated on our new cfs_rq */
5277 account_cfs_rq_runtime(cfs_rq, 0);
5281 void init_cfs_rq(struct cfs_rq *cfs_rq)
5283 cfs_rq->tasks_timeline = RB_ROOT;
5284 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
5285 #ifndef CONFIG_64BIT
5286 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
5290 #ifdef CONFIG_FAIR_GROUP_SCHED
5291 static void task_move_group_fair(struct task_struct *p, int on_rq)
5294 * If the task was not on the rq at the time of this cgroup movement
5295 * it must have been asleep, sleeping tasks keep their ->vruntime
5296 * absolute on their old rq until wakeup (needed for the fair sleeper
5297 * bonus in place_entity()).
5299 * If it was on the rq, we've just 'preempted' it, which does convert
5300 * ->vruntime to a relative base.
5302 * Make sure both cases convert their relative position when migrating
5303 * to another cgroup's rq. This does somewhat interfere with the
5304 * fair sleeper stuff for the first placement, but who cares.
5307 * When !on_rq, vruntime of the task has usually NOT been normalized.
5308 * But there are some cases where it has already been normalized:
5310 * - Moving a forked child which is waiting for being woken up by
5311 * wake_up_new_task().
5312 * - Moving a task which has been woken up by try_to_wake_up() and
5313 * waiting for actually being woken up by sched_ttwu_pending().
5315 * To prevent boost or penalty in the new cfs_rq caused by delta
5316 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
5318 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
5322 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
5323 set_task_rq(p, task_cpu(p));
5325 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
5328 void free_fair_sched_group(struct task_group *tg)
5332 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
5334 for_each_possible_cpu(i) {
5336 kfree(tg->cfs_rq[i]);
5345 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
5347 struct cfs_rq *cfs_rq;
5348 struct sched_entity *se;
5351 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
5354 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
5358 tg->shares = NICE_0_LOAD;
5360 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
5362 for_each_possible_cpu(i) {
5363 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
5364 GFP_KERNEL, cpu_to_node(i));
5368 se = kzalloc_node(sizeof(struct sched_entity),
5369 GFP_KERNEL, cpu_to_node(i));
5373 init_cfs_rq(cfs_rq);
5374 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
5385 void unregister_fair_sched_group(struct task_group *tg, int cpu)
5387 struct rq *rq = cpu_rq(cpu);
5388 unsigned long flags;
5391 * Only empty task groups can be destroyed; so we can speculatively
5392 * check on_list without danger of it being re-added.
5394 if (!tg->cfs_rq[cpu]->on_list)
5397 raw_spin_lock_irqsave(&rq->lock, flags);
5398 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
5399 raw_spin_unlock_irqrestore(&rq->lock, flags);
5402 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
5403 struct sched_entity *se, int cpu,
5404 struct sched_entity *parent)
5406 struct rq *rq = cpu_rq(cpu);
5411 /* allow initial update_cfs_load() to truncate */
5412 cfs_rq->load_stamp = 1;
5414 init_cfs_rq_runtime(cfs_rq);
5416 tg->cfs_rq[cpu] = cfs_rq;
5419 /* se could be NULL for root_task_group */
5424 se->cfs_rq = &rq->cfs;
5426 se->cfs_rq = parent->my_q;
5429 update_load_set(&se->load, 0);
5430 se->parent = parent;
5433 static DEFINE_MUTEX(shares_mutex);
5435 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
5438 unsigned long flags;
5441 * We can't change the weight of the root cgroup.
5446 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
5448 mutex_lock(&shares_mutex);
5449 if (tg->shares == shares)
5452 tg->shares = shares;
5453 for_each_possible_cpu(i) {
5454 struct rq *rq = cpu_rq(i);
5455 struct sched_entity *se;
5458 /* Propagate contribution to hierarchy */
5459 raw_spin_lock_irqsave(&rq->lock, flags);
5460 for_each_sched_entity(se)
5461 update_cfs_shares(group_cfs_rq(se));
5462 raw_spin_unlock_irqrestore(&rq->lock, flags);
5466 mutex_unlock(&shares_mutex);
5469 #else /* CONFIG_FAIR_GROUP_SCHED */
5471 void free_fair_sched_group(struct task_group *tg) { }
5473 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
5478 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
5480 #endif /* CONFIG_FAIR_GROUP_SCHED */
5483 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
5485 struct sched_entity *se = &task->se;
5486 unsigned int rr_interval = 0;
5489 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
5492 if (rq->cfs.load.weight)
5493 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5499 * All the scheduling class methods:
5501 const struct sched_class fair_sched_class = {
5502 .next = &idle_sched_class,
5503 .enqueue_task = enqueue_task_fair,
5504 .dequeue_task = dequeue_task_fair,
5505 .yield_task = yield_task_fair,
5506 .yield_to_task = yield_to_task_fair,
5508 .check_preempt_curr = check_preempt_wakeup,
5510 .pick_next_task = pick_next_task_fair,
5511 .put_prev_task = put_prev_task_fair,
5514 .select_task_rq = select_task_rq_fair,
5516 .rq_online = rq_online_fair,
5517 .rq_offline = rq_offline_fair,
5519 .task_waking = task_waking_fair,
5522 .set_curr_task = set_curr_task_fair,
5523 .task_tick = task_tick_fair,
5524 .task_fork = task_fork_fair,
5526 .prio_changed = prio_changed_fair,
5527 .switched_from = switched_from_fair,
5528 .switched_to = switched_to_fair,
5530 .get_rr_interval = get_rr_interval_fair,
5532 #ifdef CONFIG_FAIR_GROUP_SCHED
5533 .task_move_group = task_move_group_fair,
5537 #ifdef CONFIG_SCHED_DEBUG
5538 void print_cfs_stats(struct seq_file *m, int cpu)
5540 struct cfs_rq *cfs_rq;
5543 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
5544 print_cfs_rq(m, cpu, cfs_rq);
5549 __init void init_sched_fair_class(void)
5552 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
5555 nohz.next_balance = jiffies;
5556 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
5557 cpu_notifier(sched_ilb_notifier, 0);