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 */
815 task_numa_placement(p);
818 static void reset_ptenuma_scan(struct task_struct *p)
820 ACCESS_ONCE(p->mm->numa_scan_seq)++;
821 p->mm->numa_scan_offset = 0;
825 * The expensive part of numa migration is done from task_work context.
826 * Triggered from task_tick_numa().
828 void task_numa_work(struct callback_head *work)
830 unsigned long migrate, next_scan, now = jiffies;
831 struct task_struct *p = current;
832 struct mm_struct *mm = p->mm;
833 struct vm_area_struct *vma;
834 unsigned long start, end;
837 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
839 work->next = work; /* protect against double add */
841 * Who cares about NUMA placement when they're dying.
843 * NOTE: make sure not to dereference p->mm before this check,
844 * exit_task_work() happens _after_ exit_mm() so we could be called
845 * without p->mm even though we still had it when we enqueued this
848 if (p->flags & PF_EXITING)
852 * Enforce maximal scan/migration frequency..
854 migrate = mm->numa_next_scan;
855 if (time_before(now, migrate))
858 if (p->numa_scan_period == 0)
859 p->numa_scan_period = sysctl_numa_balancing_scan_period_min;
861 next_scan = now + 2*msecs_to_jiffies(p->numa_scan_period);
862 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
866 * Do not set pte_numa if the current running node is rate-limited.
867 * This loses statistics on the fault but if we are unwilling to
868 * migrate to this node, it is less likely we can do useful work
870 if (migrate_ratelimited(numa_node_id()))
873 start = mm->numa_scan_offset;
874 pages = sysctl_numa_balancing_scan_size;
875 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
879 down_read(&mm->mmap_sem);
880 vma = find_vma(mm, start);
882 reset_ptenuma_scan(p);
886 for (; vma; vma = vma->vm_next) {
887 if (!vma_migratable(vma))
890 /* Skip small VMAs. They are not likely to be of relevance */
891 if (((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) < HPAGE_PMD_NR)
895 start = max(start, vma->vm_start);
896 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
897 end = min(end, vma->vm_end);
898 pages -= change_prot_numa(vma, start, end);
903 } while (end != vma->vm_end);
908 * It is possible to reach the end of the VMA list but the last few VMAs are
909 * not guaranteed to the vma_migratable. If they are not, we would find the
910 * !migratable VMA on the next scan but not reset the scanner to the start
914 mm->numa_scan_offset = start;
916 reset_ptenuma_scan(p);
917 up_read(&mm->mmap_sem);
921 * Drive the periodic memory faults..
923 void task_tick_numa(struct rq *rq, struct task_struct *curr)
925 struct callback_head *work = &curr->numa_work;
929 * We don't care about NUMA placement if we don't have memory.
931 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
935 * Using runtime rather than walltime has the dual advantage that
936 * we (mostly) drive the selection from busy threads and that the
937 * task needs to have done some actual work before we bother with
940 now = curr->se.sum_exec_runtime;
941 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
943 if (now - curr->node_stamp > period) {
944 if (!curr->node_stamp)
945 curr->numa_scan_period = sysctl_numa_balancing_scan_period_min;
946 curr->node_stamp = now;
948 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
949 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
950 task_work_add(curr, work, true);
955 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
958 #endif /* CONFIG_NUMA_BALANCING */
961 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
963 update_load_add(&cfs_rq->load, se->load.weight);
964 if (!parent_entity(se))
965 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
967 if (entity_is_task(se))
968 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
970 cfs_rq->nr_running++;
974 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
976 update_load_sub(&cfs_rq->load, se->load.weight);
977 if (!parent_entity(se))
978 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
979 if (entity_is_task(se))
980 list_del_init(&se->group_node);
981 cfs_rq->nr_running--;
984 #ifdef CONFIG_FAIR_GROUP_SCHED
985 /* we need this in update_cfs_load and load-balance functions below */
986 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
988 static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq,
991 struct task_group *tg = cfs_rq->tg;
994 load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1);
995 load_avg -= cfs_rq->load_contribution;
997 if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) {
998 atomic_add(load_avg, &tg->load_weight);
999 cfs_rq->load_contribution += load_avg;
1003 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
1005 u64 period = sysctl_sched_shares_window;
1007 unsigned long load = cfs_rq->load.weight;
1009 if (cfs_rq->tg == &root_task_group || throttled_hierarchy(cfs_rq))
1012 now = rq_of(cfs_rq)->clock_task;
1013 delta = now - cfs_rq->load_stamp;
1015 /* truncate load history at 4 idle periods */
1016 if (cfs_rq->load_stamp > cfs_rq->load_last &&
1017 now - cfs_rq->load_last > 4 * period) {
1018 cfs_rq->load_period = 0;
1019 cfs_rq->load_avg = 0;
1023 cfs_rq->load_stamp = now;
1024 cfs_rq->load_unacc_exec_time = 0;
1025 cfs_rq->load_period += delta;
1027 cfs_rq->load_last = now;
1028 cfs_rq->load_avg += delta * load;
1031 /* consider updating load contribution on each fold or truncate */
1032 if (global_update || cfs_rq->load_period > period
1033 || !cfs_rq->load_period)
1034 update_cfs_rq_load_contribution(cfs_rq, global_update);
1036 while (cfs_rq->load_period > period) {
1038 * Inline assembly required to prevent the compiler
1039 * optimising this loop into a divmod call.
1040 * See __iter_div_u64_rem() for another example of this.
1042 asm("" : "+rm" (cfs_rq->load_period));
1043 cfs_rq->load_period /= 2;
1044 cfs_rq->load_avg /= 2;
1047 if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg)
1048 list_del_leaf_cfs_rq(cfs_rq);
1051 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1056 * Use this CPU's actual weight instead of the last load_contribution
1057 * to gain a more accurate current total weight. See
1058 * update_cfs_rq_load_contribution().
1060 tg_weight = atomic_read(&tg->load_weight);
1061 tg_weight -= cfs_rq->load_contribution;
1062 tg_weight += cfs_rq->load.weight;
1067 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1069 long tg_weight, load, shares;
1071 tg_weight = calc_tg_weight(tg, cfs_rq);
1072 load = cfs_rq->load.weight;
1074 shares = (tg->shares * load);
1076 shares /= tg_weight;
1078 if (shares < MIN_SHARES)
1079 shares = MIN_SHARES;
1080 if (shares > tg->shares)
1081 shares = tg->shares;
1086 static void update_entity_shares_tick(struct cfs_rq *cfs_rq)
1088 if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) {
1089 update_cfs_load(cfs_rq, 0);
1090 update_cfs_shares(cfs_rq);
1093 # else /* CONFIG_SMP */
1094 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
1098 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1103 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
1106 # endif /* CONFIG_SMP */
1107 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1108 unsigned long weight)
1111 /* commit outstanding execution time */
1112 if (cfs_rq->curr == se)
1113 update_curr(cfs_rq);
1114 account_entity_dequeue(cfs_rq, se);
1117 update_load_set(&se->load, weight);
1120 account_entity_enqueue(cfs_rq, se);
1123 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1125 struct task_group *tg;
1126 struct sched_entity *se;
1130 se = tg->se[cpu_of(rq_of(cfs_rq))];
1131 if (!se || throttled_hierarchy(cfs_rq))
1134 if (likely(se->load.weight == tg->shares))
1137 shares = calc_cfs_shares(cfs_rq, tg);
1139 reweight_entity(cfs_rq_of(se), se, shares);
1141 #else /* CONFIG_FAIR_GROUP_SCHED */
1142 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
1146 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1150 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
1153 #endif /* CONFIG_FAIR_GROUP_SCHED */
1155 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1157 #ifdef CONFIG_SCHEDSTATS
1158 struct task_struct *tsk = NULL;
1160 if (entity_is_task(se))
1163 if (se->statistics.sleep_start) {
1164 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
1169 if (unlikely(delta > se->statistics.sleep_max))
1170 se->statistics.sleep_max = delta;
1172 se->statistics.sleep_start = 0;
1173 se->statistics.sum_sleep_runtime += delta;
1176 account_scheduler_latency(tsk, delta >> 10, 1);
1177 trace_sched_stat_sleep(tsk, delta);
1180 if (se->statistics.block_start) {
1181 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
1186 if (unlikely(delta > se->statistics.block_max))
1187 se->statistics.block_max = delta;
1189 se->statistics.block_start = 0;
1190 se->statistics.sum_sleep_runtime += delta;
1193 if (tsk->in_iowait) {
1194 se->statistics.iowait_sum += delta;
1195 se->statistics.iowait_count++;
1196 trace_sched_stat_iowait(tsk, delta);
1199 trace_sched_stat_blocked(tsk, delta);
1202 * Blocking time is in units of nanosecs, so shift by
1203 * 20 to get a milliseconds-range estimation of the
1204 * amount of time that the task spent sleeping:
1206 if (unlikely(prof_on == SLEEP_PROFILING)) {
1207 profile_hits(SLEEP_PROFILING,
1208 (void *)get_wchan(tsk),
1211 account_scheduler_latency(tsk, delta >> 10, 0);
1217 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1219 #ifdef CONFIG_SCHED_DEBUG
1220 s64 d = se->vruntime - cfs_rq->min_vruntime;
1225 if (d > 3*sysctl_sched_latency)
1226 schedstat_inc(cfs_rq, nr_spread_over);
1231 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1233 u64 vruntime = cfs_rq->min_vruntime;
1236 * The 'current' period is already promised to the current tasks,
1237 * however the extra weight of the new task will slow them down a
1238 * little, place the new task so that it fits in the slot that
1239 * stays open at the end.
1241 if (initial && sched_feat(START_DEBIT))
1242 vruntime += sched_vslice(cfs_rq, se);
1244 /* sleeps up to a single latency don't count. */
1246 unsigned long thresh = sysctl_sched_latency;
1249 * Halve their sleep time's effect, to allow
1250 * for a gentler effect of sleepers:
1252 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1258 /* ensure we never gain time by being placed backwards. */
1259 vruntime = max_vruntime(se->vruntime, vruntime);
1261 se->vruntime = vruntime;
1264 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1267 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1270 * Update the normalized vruntime before updating min_vruntime
1271 * through callig update_curr().
1273 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1274 se->vruntime += cfs_rq->min_vruntime;
1277 * Update run-time statistics of the 'current'.
1279 update_curr(cfs_rq);
1280 update_cfs_load(cfs_rq, 0);
1281 account_entity_enqueue(cfs_rq, se);
1282 update_cfs_shares(cfs_rq);
1284 if (flags & ENQUEUE_WAKEUP) {
1285 place_entity(cfs_rq, se, 0);
1286 enqueue_sleeper(cfs_rq, se);
1289 update_stats_enqueue(cfs_rq, se);
1290 check_spread(cfs_rq, se);
1291 if (se != cfs_rq->curr)
1292 __enqueue_entity(cfs_rq, se);
1295 if (cfs_rq->nr_running == 1) {
1296 list_add_leaf_cfs_rq(cfs_rq);
1297 check_enqueue_throttle(cfs_rq);
1301 static void __clear_buddies_last(struct sched_entity *se)
1303 for_each_sched_entity(se) {
1304 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1305 if (cfs_rq->last == se)
1306 cfs_rq->last = NULL;
1312 static void __clear_buddies_next(struct sched_entity *se)
1314 for_each_sched_entity(se) {
1315 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1316 if (cfs_rq->next == se)
1317 cfs_rq->next = NULL;
1323 static void __clear_buddies_skip(struct sched_entity *se)
1325 for_each_sched_entity(se) {
1326 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1327 if (cfs_rq->skip == se)
1328 cfs_rq->skip = NULL;
1334 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1336 if (cfs_rq->last == se)
1337 __clear_buddies_last(se);
1339 if (cfs_rq->next == se)
1340 __clear_buddies_next(se);
1342 if (cfs_rq->skip == se)
1343 __clear_buddies_skip(se);
1346 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1349 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1352 * Update run-time statistics of the 'current'.
1354 update_curr(cfs_rq);
1356 update_stats_dequeue(cfs_rq, se);
1357 if (flags & DEQUEUE_SLEEP) {
1358 #ifdef CONFIG_SCHEDSTATS
1359 if (entity_is_task(se)) {
1360 struct task_struct *tsk = task_of(se);
1362 if (tsk->state & TASK_INTERRUPTIBLE)
1363 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1364 if (tsk->state & TASK_UNINTERRUPTIBLE)
1365 se->statistics.block_start = rq_of(cfs_rq)->clock;
1370 clear_buddies(cfs_rq, se);
1372 if (se != cfs_rq->curr)
1373 __dequeue_entity(cfs_rq, se);
1375 update_cfs_load(cfs_rq, 0);
1376 account_entity_dequeue(cfs_rq, se);
1379 * Normalize the entity after updating the min_vruntime because the
1380 * update can refer to the ->curr item and we need to reflect this
1381 * movement in our normalized position.
1383 if (!(flags & DEQUEUE_SLEEP))
1384 se->vruntime -= cfs_rq->min_vruntime;
1386 /* return excess runtime on last dequeue */
1387 return_cfs_rq_runtime(cfs_rq);
1389 update_min_vruntime(cfs_rq);
1390 update_cfs_shares(cfs_rq);
1394 * Preempt the current task with a newly woken task if needed:
1397 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1399 unsigned long ideal_runtime, delta_exec;
1400 struct sched_entity *se;
1403 ideal_runtime = sched_slice(cfs_rq, curr);
1404 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1405 if (delta_exec > ideal_runtime) {
1406 resched_task(rq_of(cfs_rq)->curr);
1408 * The current task ran long enough, ensure it doesn't get
1409 * re-elected due to buddy favours.
1411 clear_buddies(cfs_rq, curr);
1416 * Ensure that a task that missed wakeup preemption by a
1417 * narrow margin doesn't have to wait for a full slice.
1418 * This also mitigates buddy induced latencies under load.
1420 if (delta_exec < sysctl_sched_min_granularity)
1423 se = __pick_first_entity(cfs_rq);
1424 delta = curr->vruntime - se->vruntime;
1429 if (delta > ideal_runtime)
1430 resched_task(rq_of(cfs_rq)->curr);
1434 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1436 /* 'current' is not kept within the tree. */
1439 * Any task has to be enqueued before it get to execute on
1440 * a CPU. So account for the time it spent waiting on the
1443 update_stats_wait_end(cfs_rq, se);
1444 __dequeue_entity(cfs_rq, se);
1447 update_stats_curr_start(cfs_rq, se);
1449 #ifdef CONFIG_SCHEDSTATS
1451 * Track our maximum slice length, if the CPU's load is at
1452 * least twice that of our own weight (i.e. dont track it
1453 * when there are only lesser-weight tasks around):
1455 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1456 se->statistics.slice_max = max(se->statistics.slice_max,
1457 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1460 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1464 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1467 * Pick the next process, keeping these things in mind, in this order:
1468 * 1) keep things fair between processes/task groups
1469 * 2) pick the "next" process, since someone really wants that to run
1470 * 3) pick the "last" process, for cache locality
1471 * 4) do not run the "skip" process, if something else is available
1473 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1475 struct sched_entity *se = __pick_first_entity(cfs_rq);
1476 struct sched_entity *left = se;
1479 * Avoid running the skip buddy, if running something else can
1480 * be done without getting too unfair.
1482 if (cfs_rq->skip == se) {
1483 struct sched_entity *second = __pick_next_entity(se);
1484 if (second && wakeup_preempt_entity(second, left) < 1)
1489 * Prefer last buddy, try to return the CPU to a preempted task.
1491 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1495 * Someone really wants this to run. If it's not unfair, run it.
1497 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1500 clear_buddies(cfs_rq, se);
1505 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1507 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1510 * If still on the runqueue then deactivate_task()
1511 * was not called and update_curr() has to be done:
1514 update_curr(cfs_rq);
1516 /* throttle cfs_rqs exceeding runtime */
1517 check_cfs_rq_runtime(cfs_rq);
1519 check_spread(cfs_rq, prev);
1521 update_stats_wait_start(cfs_rq, prev);
1522 /* Put 'current' back into the tree. */
1523 __enqueue_entity(cfs_rq, prev);
1525 cfs_rq->curr = NULL;
1529 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1532 * Update run-time statistics of the 'current'.
1534 update_curr(cfs_rq);
1537 * Update share accounting for long-running entities.
1539 update_entity_shares_tick(cfs_rq);
1541 #ifdef CONFIG_SCHED_HRTICK
1543 * queued ticks are scheduled to match the slice, so don't bother
1544 * validating it and just reschedule.
1547 resched_task(rq_of(cfs_rq)->curr);
1551 * don't let the period tick interfere with the hrtick preemption
1553 if (!sched_feat(DOUBLE_TICK) &&
1554 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1558 if (cfs_rq->nr_running > 1)
1559 check_preempt_tick(cfs_rq, curr);
1563 /**************************************************
1564 * CFS bandwidth control machinery
1567 #ifdef CONFIG_CFS_BANDWIDTH
1569 #ifdef HAVE_JUMP_LABEL
1570 static struct static_key __cfs_bandwidth_used;
1572 static inline bool cfs_bandwidth_used(void)
1574 return static_key_false(&__cfs_bandwidth_used);
1577 void account_cfs_bandwidth_used(int enabled, int was_enabled)
1579 /* only need to count groups transitioning between enabled/!enabled */
1580 if (enabled && !was_enabled)
1581 static_key_slow_inc(&__cfs_bandwidth_used);
1582 else if (!enabled && was_enabled)
1583 static_key_slow_dec(&__cfs_bandwidth_used);
1585 #else /* HAVE_JUMP_LABEL */
1586 static bool cfs_bandwidth_used(void)
1591 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
1592 #endif /* HAVE_JUMP_LABEL */
1595 * default period for cfs group bandwidth.
1596 * default: 0.1s, units: nanoseconds
1598 static inline u64 default_cfs_period(void)
1600 return 100000000ULL;
1603 static inline u64 sched_cfs_bandwidth_slice(void)
1605 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
1609 * Replenish runtime according to assigned quota and update expiration time.
1610 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
1611 * additional synchronization around rq->lock.
1613 * requires cfs_b->lock
1615 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
1619 if (cfs_b->quota == RUNTIME_INF)
1622 now = sched_clock_cpu(smp_processor_id());
1623 cfs_b->runtime = cfs_b->quota;
1624 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
1627 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
1629 return &tg->cfs_bandwidth;
1632 /* returns 0 on failure to allocate runtime */
1633 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1635 struct task_group *tg = cfs_rq->tg;
1636 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
1637 u64 amount = 0, min_amount, expires;
1639 /* note: this is a positive sum as runtime_remaining <= 0 */
1640 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
1642 raw_spin_lock(&cfs_b->lock);
1643 if (cfs_b->quota == RUNTIME_INF)
1644 amount = min_amount;
1647 * If the bandwidth pool has become inactive, then at least one
1648 * period must have elapsed since the last consumption.
1649 * Refresh the global state and ensure bandwidth timer becomes
1652 if (!cfs_b->timer_active) {
1653 __refill_cfs_bandwidth_runtime(cfs_b);
1654 __start_cfs_bandwidth(cfs_b);
1657 if (cfs_b->runtime > 0) {
1658 amount = min(cfs_b->runtime, min_amount);
1659 cfs_b->runtime -= amount;
1663 expires = cfs_b->runtime_expires;
1664 raw_spin_unlock(&cfs_b->lock);
1666 cfs_rq->runtime_remaining += amount;
1668 * we may have advanced our local expiration to account for allowed
1669 * spread between our sched_clock and the one on which runtime was
1672 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
1673 cfs_rq->runtime_expires = expires;
1675 return cfs_rq->runtime_remaining > 0;
1679 * Note: This depends on the synchronization provided by sched_clock and the
1680 * fact that rq->clock snapshots this value.
1682 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1684 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1685 struct rq *rq = rq_of(cfs_rq);
1687 /* if the deadline is ahead of our clock, nothing to do */
1688 if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0))
1691 if (cfs_rq->runtime_remaining < 0)
1695 * If the local deadline has passed we have to consider the
1696 * possibility that our sched_clock is 'fast' and the global deadline
1697 * has not truly expired.
1699 * Fortunately we can check determine whether this the case by checking
1700 * whether the global deadline has advanced.
1703 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
1704 /* extend local deadline, drift is bounded above by 2 ticks */
1705 cfs_rq->runtime_expires += TICK_NSEC;
1707 /* global deadline is ahead, expiration has passed */
1708 cfs_rq->runtime_remaining = 0;
1712 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1713 unsigned long delta_exec)
1715 /* dock delta_exec before expiring quota (as it could span periods) */
1716 cfs_rq->runtime_remaining -= delta_exec;
1717 expire_cfs_rq_runtime(cfs_rq);
1719 if (likely(cfs_rq->runtime_remaining > 0))
1723 * if we're unable to extend our runtime we resched so that the active
1724 * hierarchy can be throttled
1726 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
1727 resched_task(rq_of(cfs_rq)->curr);
1730 static __always_inline
1731 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
1733 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
1736 __account_cfs_rq_runtime(cfs_rq, delta_exec);
1739 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
1741 return cfs_bandwidth_used() && cfs_rq->throttled;
1744 /* check whether cfs_rq, or any parent, is throttled */
1745 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
1747 return cfs_bandwidth_used() && cfs_rq->throttle_count;
1751 * Ensure that neither of the group entities corresponding to src_cpu or
1752 * dest_cpu are members of a throttled hierarchy when performing group
1753 * load-balance operations.
1755 static inline int throttled_lb_pair(struct task_group *tg,
1756 int src_cpu, int dest_cpu)
1758 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
1760 src_cfs_rq = tg->cfs_rq[src_cpu];
1761 dest_cfs_rq = tg->cfs_rq[dest_cpu];
1763 return throttled_hierarchy(src_cfs_rq) ||
1764 throttled_hierarchy(dest_cfs_rq);
1767 /* updated child weight may affect parent so we have to do this bottom up */
1768 static int tg_unthrottle_up(struct task_group *tg, void *data)
1770 struct rq *rq = data;
1771 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1773 cfs_rq->throttle_count--;
1775 if (!cfs_rq->throttle_count) {
1776 u64 delta = rq->clock_task - cfs_rq->load_stamp;
1778 /* leaving throttled state, advance shares averaging windows */
1779 cfs_rq->load_stamp += delta;
1780 cfs_rq->load_last += delta;
1782 /* update entity weight now that we are on_rq again */
1783 update_cfs_shares(cfs_rq);
1790 static int tg_throttle_down(struct task_group *tg, void *data)
1792 struct rq *rq = data;
1793 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1795 /* group is entering throttled state, record last load */
1796 if (!cfs_rq->throttle_count)
1797 update_cfs_load(cfs_rq, 0);
1798 cfs_rq->throttle_count++;
1803 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
1805 struct rq *rq = rq_of(cfs_rq);
1806 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1807 struct sched_entity *se;
1808 long task_delta, dequeue = 1;
1810 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1812 /* account load preceding throttle */
1814 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
1817 task_delta = cfs_rq->h_nr_running;
1818 for_each_sched_entity(se) {
1819 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
1820 /* throttled entity or throttle-on-deactivate */
1825 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
1826 qcfs_rq->h_nr_running -= task_delta;
1828 if (qcfs_rq->load.weight)
1833 rq->nr_running -= task_delta;
1835 cfs_rq->throttled = 1;
1836 cfs_rq->throttled_timestamp = rq->clock;
1837 raw_spin_lock(&cfs_b->lock);
1838 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
1839 raw_spin_unlock(&cfs_b->lock);
1842 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
1844 struct rq *rq = rq_of(cfs_rq);
1845 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1846 struct sched_entity *se;
1850 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1852 cfs_rq->throttled = 0;
1853 raw_spin_lock(&cfs_b->lock);
1854 cfs_b->throttled_time += rq->clock - cfs_rq->throttled_timestamp;
1855 list_del_rcu(&cfs_rq->throttled_list);
1856 raw_spin_unlock(&cfs_b->lock);
1857 cfs_rq->throttled_timestamp = 0;
1859 update_rq_clock(rq);
1860 /* update hierarchical throttle state */
1861 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
1863 if (!cfs_rq->load.weight)
1866 task_delta = cfs_rq->h_nr_running;
1867 for_each_sched_entity(se) {
1871 cfs_rq = cfs_rq_of(se);
1873 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
1874 cfs_rq->h_nr_running += task_delta;
1876 if (cfs_rq_throttled(cfs_rq))
1881 rq->nr_running += task_delta;
1883 /* determine whether we need to wake up potentially idle cpu */
1884 if (rq->curr == rq->idle && rq->cfs.nr_running)
1885 resched_task(rq->curr);
1888 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
1889 u64 remaining, u64 expires)
1891 struct cfs_rq *cfs_rq;
1892 u64 runtime = remaining;
1895 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
1897 struct rq *rq = rq_of(cfs_rq);
1899 raw_spin_lock(&rq->lock);
1900 if (!cfs_rq_throttled(cfs_rq))
1903 runtime = -cfs_rq->runtime_remaining + 1;
1904 if (runtime > remaining)
1905 runtime = remaining;
1906 remaining -= runtime;
1908 cfs_rq->runtime_remaining += runtime;
1909 cfs_rq->runtime_expires = expires;
1911 /* we check whether we're throttled above */
1912 if (cfs_rq->runtime_remaining > 0)
1913 unthrottle_cfs_rq(cfs_rq);
1916 raw_spin_unlock(&rq->lock);
1927 * Responsible for refilling a task_group's bandwidth and unthrottling its
1928 * cfs_rqs as appropriate. If there has been no activity within the last
1929 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
1930 * used to track this state.
1932 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
1934 u64 runtime, runtime_expires;
1935 int idle = 1, throttled;
1937 raw_spin_lock(&cfs_b->lock);
1938 /* no need to continue the timer with no bandwidth constraint */
1939 if (cfs_b->quota == RUNTIME_INF)
1942 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
1943 /* idle depends on !throttled (for the case of a large deficit) */
1944 idle = cfs_b->idle && !throttled;
1945 cfs_b->nr_periods += overrun;
1947 /* if we're going inactive then everything else can be deferred */
1951 __refill_cfs_bandwidth_runtime(cfs_b);
1954 /* mark as potentially idle for the upcoming period */
1959 /* account preceding periods in which throttling occurred */
1960 cfs_b->nr_throttled += overrun;
1963 * There are throttled entities so we must first use the new bandwidth
1964 * to unthrottle them before making it generally available. This
1965 * ensures that all existing debts will be paid before a new cfs_rq is
1968 runtime = cfs_b->runtime;
1969 runtime_expires = cfs_b->runtime_expires;
1973 * This check is repeated as we are holding onto the new bandwidth
1974 * while we unthrottle. This can potentially race with an unthrottled
1975 * group trying to acquire new bandwidth from the global pool.
1977 while (throttled && runtime > 0) {
1978 raw_spin_unlock(&cfs_b->lock);
1979 /* we can't nest cfs_b->lock while distributing bandwidth */
1980 runtime = distribute_cfs_runtime(cfs_b, runtime,
1982 raw_spin_lock(&cfs_b->lock);
1984 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
1987 /* return (any) remaining runtime */
1988 cfs_b->runtime = runtime;
1990 * While we are ensured activity in the period following an
1991 * unthrottle, this also covers the case in which the new bandwidth is
1992 * insufficient to cover the existing bandwidth deficit. (Forcing the
1993 * timer to remain active while there are any throttled entities.)
1998 cfs_b->timer_active = 0;
1999 raw_spin_unlock(&cfs_b->lock);
2004 /* a cfs_rq won't donate quota below this amount */
2005 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2006 /* minimum remaining period time to redistribute slack quota */
2007 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2008 /* how long we wait to gather additional slack before distributing */
2009 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2011 /* are we near the end of the current quota period? */
2012 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2014 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2017 /* if the call-back is running a quota refresh is already occurring */
2018 if (hrtimer_callback_running(refresh_timer))
2021 /* is a quota refresh about to occur? */
2022 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2023 if (remaining < min_expire)
2029 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2031 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2033 /* if there's a quota refresh soon don't bother with slack */
2034 if (runtime_refresh_within(cfs_b, min_left))
2037 start_bandwidth_timer(&cfs_b->slack_timer,
2038 ns_to_ktime(cfs_bandwidth_slack_period));
2041 /* we know any runtime found here is valid as update_curr() precedes return */
2042 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2044 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2045 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2047 if (slack_runtime <= 0)
2050 raw_spin_lock(&cfs_b->lock);
2051 if (cfs_b->quota != RUNTIME_INF &&
2052 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2053 cfs_b->runtime += slack_runtime;
2055 /* we are under rq->lock, defer unthrottling using a timer */
2056 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2057 !list_empty(&cfs_b->throttled_cfs_rq))
2058 start_cfs_slack_bandwidth(cfs_b);
2060 raw_spin_unlock(&cfs_b->lock);
2062 /* even if it's not valid for return we don't want to try again */
2063 cfs_rq->runtime_remaining -= slack_runtime;
2066 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2068 if (!cfs_bandwidth_used())
2071 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2074 __return_cfs_rq_runtime(cfs_rq);
2078 * This is done with a timer (instead of inline with bandwidth return) since
2079 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2081 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2083 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2086 /* confirm we're still not at a refresh boundary */
2087 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2090 raw_spin_lock(&cfs_b->lock);
2091 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2092 runtime = cfs_b->runtime;
2095 expires = cfs_b->runtime_expires;
2096 raw_spin_unlock(&cfs_b->lock);
2101 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2103 raw_spin_lock(&cfs_b->lock);
2104 if (expires == cfs_b->runtime_expires)
2105 cfs_b->runtime = runtime;
2106 raw_spin_unlock(&cfs_b->lock);
2110 * When a group wakes up we want to make sure that its quota is not already
2111 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2112 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2114 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2116 if (!cfs_bandwidth_used())
2119 /* an active group must be handled by the update_curr()->put() path */
2120 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2123 /* ensure the group is not already throttled */
2124 if (cfs_rq_throttled(cfs_rq))
2127 /* update runtime allocation */
2128 account_cfs_rq_runtime(cfs_rq, 0);
2129 if (cfs_rq->runtime_remaining <= 0)
2130 throttle_cfs_rq(cfs_rq);
2133 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2134 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2136 if (!cfs_bandwidth_used())
2139 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2143 * it's possible for a throttled entity to be forced into a running
2144 * state (e.g. set_curr_task), in this case we're finished.
2146 if (cfs_rq_throttled(cfs_rq))
2149 throttle_cfs_rq(cfs_rq);
2152 static inline u64 default_cfs_period(void);
2153 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun);
2154 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b);
2156 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2158 struct cfs_bandwidth *cfs_b =
2159 container_of(timer, struct cfs_bandwidth, slack_timer);
2160 do_sched_cfs_slack_timer(cfs_b);
2162 return HRTIMER_NORESTART;
2165 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2167 struct cfs_bandwidth *cfs_b =
2168 container_of(timer, struct cfs_bandwidth, period_timer);
2174 now = hrtimer_cb_get_time(timer);
2175 overrun = hrtimer_forward(timer, now, cfs_b->period);
2180 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2183 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2186 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2188 raw_spin_lock_init(&cfs_b->lock);
2190 cfs_b->quota = RUNTIME_INF;
2191 cfs_b->period = ns_to_ktime(default_cfs_period());
2193 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2194 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2195 cfs_b->period_timer.function = sched_cfs_period_timer;
2196 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2197 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2200 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2202 cfs_rq->runtime_enabled = 0;
2203 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2206 /* requires cfs_b->lock, may release to reprogram timer */
2207 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2210 * The timer may be active because we're trying to set a new bandwidth
2211 * period or because we're racing with the tear-down path
2212 * (timer_active==0 becomes visible before the hrtimer call-back
2213 * terminates). In either case we ensure that it's re-programmed
2215 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2216 raw_spin_unlock(&cfs_b->lock);
2217 /* ensure cfs_b->lock is available while we wait */
2218 hrtimer_cancel(&cfs_b->period_timer);
2220 raw_spin_lock(&cfs_b->lock);
2221 /* if someone else restarted the timer then we're done */
2222 if (cfs_b->timer_active)
2226 cfs_b->timer_active = 1;
2227 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2230 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2232 hrtimer_cancel(&cfs_b->period_timer);
2233 hrtimer_cancel(&cfs_b->slack_timer);
2236 static void unthrottle_offline_cfs_rqs(struct rq *rq)
2238 struct cfs_rq *cfs_rq;
2240 for_each_leaf_cfs_rq(rq, cfs_rq) {
2241 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2243 if (!cfs_rq->runtime_enabled)
2247 * clock_task is not advancing so we just need to make sure
2248 * there's some valid quota amount
2250 cfs_rq->runtime_remaining = cfs_b->quota;
2251 if (cfs_rq_throttled(cfs_rq))
2252 unthrottle_cfs_rq(cfs_rq);
2256 #else /* CONFIG_CFS_BANDWIDTH */
2257 static __always_inline
2258 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec) {}
2259 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2260 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2261 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2263 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2268 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2273 static inline int throttled_lb_pair(struct task_group *tg,
2274 int src_cpu, int dest_cpu)
2279 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2281 #ifdef CONFIG_FAIR_GROUP_SCHED
2282 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2285 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2289 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2290 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2292 #endif /* CONFIG_CFS_BANDWIDTH */
2294 /**************************************************
2295 * CFS operations on tasks:
2298 #ifdef CONFIG_SCHED_HRTICK
2299 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2301 struct sched_entity *se = &p->se;
2302 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2304 WARN_ON(task_rq(p) != rq);
2306 if (cfs_rq->nr_running > 1) {
2307 u64 slice = sched_slice(cfs_rq, se);
2308 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2309 s64 delta = slice - ran;
2318 * Don't schedule slices shorter than 10000ns, that just
2319 * doesn't make sense. Rely on vruntime for fairness.
2322 delta = max_t(s64, 10000LL, delta);
2324 hrtick_start(rq, delta);
2329 * called from enqueue/dequeue and updates the hrtick when the
2330 * current task is from our class and nr_running is low enough
2333 static void hrtick_update(struct rq *rq)
2335 struct task_struct *curr = rq->curr;
2337 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2340 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2341 hrtick_start_fair(rq, curr);
2343 #else /* !CONFIG_SCHED_HRTICK */
2345 hrtick_start_fair(struct rq *rq, struct task_struct *p)
2349 static inline void hrtick_update(struct rq *rq)
2355 * The enqueue_task method is called before nr_running is
2356 * increased. Here we update the fair scheduling stats and
2357 * then put the task into the rbtree:
2360 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2362 struct cfs_rq *cfs_rq;
2363 struct sched_entity *se = &p->se;
2365 for_each_sched_entity(se) {
2368 cfs_rq = cfs_rq_of(se);
2369 enqueue_entity(cfs_rq, se, flags);
2372 * end evaluation on encountering a throttled cfs_rq
2374 * note: in the case of encountering a throttled cfs_rq we will
2375 * post the final h_nr_running increment below.
2377 if (cfs_rq_throttled(cfs_rq))
2379 cfs_rq->h_nr_running++;
2381 flags = ENQUEUE_WAKEUP;
2384 for_each_sched_entity(se) {
2385 cfs_rq = cfs_rq_of(se);
2386 cfs_rq->h_nr_running++;
2388 if (cfs_rq_throttled(cfs_rq))
2391 update_cfs_load(cfs_rq, 0);
2392 update_cfs_shares(cfs_rq);
2400 static void set_next_buddy(struct sched_entity *se);
2403 * The dequeue_task method is called before nr_running is
2404 * decreased. We remove the task from the rbtree and
2405 * update the fair scheduling stats:
2407 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2409 struct cfs_rq *cfs_rq;
2410 struct sched_entity *se = &p->se;
2411 int task_sleep = flags & DEQUEUE_SLEEP;
2413 for_each_sched_entity(se) {
2414 cfs_rq = cfs_rq_of(se);
2415 dequeue_entity(cfs_rq, se, flags);
2418 * end evaluation on encountering a throttled cfs_rq
2420 * note: in the case of encountering a throttled cfs_rq we will
2421 * post the final h_nr_running decrement below.
2423 if (cfs_rq_throttled(cfs_rq))
2425 cfs_rq->h_nr_running--;
2427 /* Don't dequeue parent if it has other entities besides us */
2428 if (cfs_rq->load.weight) {
2430 * Bias pick_next to pick a task from this cfs_rq, as
2431 * p is sleeping when it is within its sched_slice.
2433 if (task_sleep && parent_entity(se))
2434 set_next_buddy(parent_entity(se));
2436 /* avoid re-evaluating load for this entity */
2437 se = parent_entity(se);
2440 flags |= DEQUEUE_SLEEP;
2443 for_each_sched_entity(se) {
2444 cfs_rq = cfs_rq_of(se);
2445 cfs_rq->h_nr_running--;
2447 if (cfs_rq_throttled(cfs_rq))
2450 update_cfs_load(cfs_rq, 0);
2451 update_cfs_shares(cfs_rq);
2460 /* Used instead of source_load when we know the type == 0 */
2461 static unsigned long weighted_cpuload(const int cpu)
2463 return cpu_rq(cpu)->load.weight;
2467 * Return a low guess at the load of a migration-source cpu weighted
2468 * according to the scheduling class and "nice" value.
2470 * We want to under-estimate the load of migration sources, to
2471 * balance conservatively.
2473 static unsigned long source_load(int cpu, int type)
2475 struct rq *rq = cpu_rq(cpu);
2476 unsigned long total = weighted_cpuload(cpu);
2478 if (type == 0 || !sched_feat(LB_BIAS))
2481 return min(rq->cpu_load[type-1], total);
2485 * Return a high guess at the load of a migration-target cpu weighted
2486 * according to the scheduling class and "nice" value.
2488 static unsigned long target_load(int cpu, int type)
2490 struct rq *rq = cpu_rq(cpu);
2491 unsigned long total = weighted_cpuload(cpu);
2493 if (type == 0 || !sched_feat(LB_BIAS))
2496 return max(rq->cpu_load[type-1], total);
2499 static unsigned long power_of(int cpu)
2501 return cpu_rq(cpu)->cpu_power;
2504 static unsigned long cpu_avg_load_per_task(int cpu)
2506 struct rq *rq = cpu_rq(cpu);
2507 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
2510 return rq->load.weight / nr_running;
2516 static void task_waking_fair(struct task_struct *p)
2518 struct sched_entity *se = &p->se;
2519 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2522 #ifndef CONFIG_64BIT
2523 u64 min_vruntime_copy;
2526 min_vruntime_copy = cfs_rq->min_vruntime_copy;
2528 min_vruntime = cfs_rq->min_vruntime;
2529 } while (min_vruntime != min_vruntime_copy);
2531 min_vruntime = cfs_rq->min_vruntime;
2534 se->vruntime -= min_vruntime;
2537 #ifdef CONFIG_FAIR_GROUP_SCHED
2539 * effective_load() calculates the load change as seen from the root_task_group
2541 * Adding load to a group doesn't make a group heavier, but can cause movement
2542 * of group shares between cpus. Assuming the shares were perfectly aligned one
2543 * can calculate the shift in shares.
2545 * Calculate the effective load difference if @wl is added (subtracted) to @tg
2546 * on this @cpu and results in a total addition (subtraction) of @wg to the
2547 * total group weight.
2549 * Given a runqueue weight distribution (rw_i) we can compute a shares
2550 * distribution (s_i) using:
2552 * s_i = rw_i / \Sum rw_j (1)
2554 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
2555 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
2556 * shares distribution (s_i):
2558 * rw_i = { 2, 4, 1, 0 }
2559 * s_i = { 2/7, 4/7, 1/7, 0 }
2561 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
2562 * task used to run on and the CPU the waker is running on), we need to
2563 * compute the effect of waking a task on either CPU and, in case of a sync
2564 * wakeup, compute the effect of the current task going to sleep.
2566 * So for a change of @wl to the local @cpu with an overall group weight change
2567 * of @wl we can compute the new shares distribution (s'_i) using:
2569 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
2571 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
2572 * differences in waking a task to CPU 0. The additional task changes the
2573 * weight and shares distributions like:
2575 * rw'_i = { 3, 4, 1, 0 }
2576 * s'_i = { 3/8, 4/8, 1/8, 0 }
2578 * We can then compute the difference in effective weight by using:
2580 * dw_i = S * (s'_i - s_i) (3)
2582 * Where 'S' is the group weight as seen by its parent.
2584 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
2585 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
2586 * 4/7) times the weight of the group.
2588 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
2590 struct sched_entity *se = tg->se[cpu];
2592 if (!tg->parent) /* the trivial, non-cgroup case */
2595 for_each_sched_entity(se) {
2601 * W = @wg + \Sum rw_j
2603 W = wg + calc_tg_weight(tg, se->my_q);
2608 w = se->my_q->load.weight + wl;
2611 * wl = S * s'_i; see (2)
2614 wl = (w * tg->shares) / W;
2619 * Per the above, wl is the new se->load.weight value; since
2620 * those are clipped to [MIN_SHARES, ...) do so now. See
2621 * calc_cfs_shares().
2623 if (wl < MIN_SHARES)
2627 * wl = dw_i = S * (s'_i - s_i); see (3)
2629 wl -= se->load.weight;
2632 * Recursively apply this logic to all parent groups to compute
2633 * the final effective load change on the root group. Since
2634 * only the @tg group gets extra weight, all parent groups can
2635 * only redistribute existing shares. @wl is the shift in shares
2636 * resulting from this level per the above.
2645 static inline unsigned long effective_load(struct task_group *tg, int cpu,
2646 unsigned long wl, unsigned long wg)
2653 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
2655 s64 this_load, load;
2656 int idx, this_cpu, prev_cpu;
2657 unsigned long tl_per_task;
2658 struct task_group *tg;
2659 unsigned long weight;
2663 this_cpu = smp_processor_id();
2664 prev_cpu = task_cpu(p);
2665 load = source_load(prev_cpu, idx);
2666 this_load = target_load(this_cpu, idx);
2669 * If sync wakeup then subtract the (maximum possible)
2670 * effect of the currently running task from the load
2671 * of the current CPU:
2674 tg = task_group(current);
2675 weight = current->se.load.weight;
2677 this_load += effective_load(tg, this_cpu, -weight, -weight);
2678 load += effective_load(tg, prev_cpu, 0, -weight);
2682 weight = p->se.load.weight;
2685 * In low-load situations, where prev_cpu is idle and this_cpu is idle
2686 * due to the sync cause above having dropped this_load to 0, we'll
2687 * always have an imbalance, but there's really nothing you can do
2688 * about that, so that's good too.
2690 * Otherwise check if either cpus are near enough in load to allow this
2691 * task to be woken on this_cpu.
2693 if (this_load > 0) {
2694 s64 this_eff_load, prev_eff_load;
2696 this_eff_load = 100;
2697 this_eff_load *= power_of(prev_cpu);
2698 this_eff_load *= this_load +
2699 effective_load(tg, this_cpu, weight, weight);
2701 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
2702 prev_eff_load *= power_of(this_cpu);
2703 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
2705 balanced = this_eff_load <= prev_eff_load;
2710 * If the currently running task will sleep within
2711 * a reasonable amount of time then attract this newly
2714 if (sync && balanced)
2717 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
2718 tl_per_task = cpu_avg_load_per_task(this_cpu);
2721 (this_load <= load &&
2722 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
2724 * This domain has SD_WAKE_AFFINE and
2725 * p is cache cold in this domain, and
2726 * there is no bad imbalance.
2728 schedstat_inc(sd, ttwu_move_affine);
2729 schedstat_inc(p, se.statistics.nr_wakeups_affine);
2737 * find_idlest_group finds and returns the least busy CPU group within the
2740 static struct sched_group *
2741 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
2742 int this_cpu, int load_idx)
2744 struct sched_group *idlest = NULL, *group = sd->groups;
2745 unsigned long min_load = ULONG_MAX, this_load = 0;
2746 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2749 unsigned long load, avg_load;
2753 /* Skip over this group if it has no CPUs allowed */
2754 if (!cpumask_intersects(sched_group_cpus(group),
2755 tsk_cpus_allowed(p)))
2758 local_group = cpumask_test_cpu(this_cpu,
2759 sched_group_cpus(group));
2761 /* Tally up the load of all CPUs in the group */
2764 for_each_cpu(i, sched_group_cpus(group)) {
2765 /* Bias balancing toward cpus of our domain */
2767 load = source_load(i, load_idx);
2769 load = target_load(i, load_idx);
2774 /* Adjust by relative CPU power of the group */
2775 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
2778 this_load = avg_load;
2779 } else if (avg_load < min_load) {
2780 min_load = avg_load;
2783 } while (group = group->next, group != sd->groups);
2785 if (!idlest || 100*this_load < imbalance*min_load)
2791 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2794 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2796 unsigned long load, min_load = ULONG_MAX;
2800 /* Traverse only the allowed CPUs */
2801 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
2802 load = weighted_cpuload(i);
2804 if (load < min_load || (load == min_load && i == this_cpu)) {
2814 * Try and locate an idle CPU in the sched_domain.
2816 static int select_idle_sibling(struct task_struct *p, int target)
2818 int cpu = smp_processor_id();
2819 int prev_cpu = task_cpu(p);
2820 struct sched_domain *sd;
2821 struct sched_group *sg;
2825 * If the task is going to be woken-up on this cpu and if it is
2826 * already idle, then it is the right target.
2828 if (target == cpu && idle_cpu(cpu))
2832 * If the task is going to be woken-up on the cpu where it previously
2833 * ran and if it is currently idle, then it the right target.
2835 if (target == prev_cpu && idle_cpu(prev_cpu))
2839 * Otherwise, iterate the domains and find an elegible idle cpu.
2841 sd = rcu_dereference(per_cpu(sd_llc, target));
2842 for_each_lower_domain(sd) {
2845 if (!cpumask_intersects(sched_group_cpus(sg),
2846 tsk_cpus_allowed(p)))
2849 for_each_cpu(i, sched_group_cpus(sg)) {
2854 target = cpumask_first_and(sched_group_cpus(sg),
2855 tsk_cpus_allowed(p));
2859 } while (sg != sd->groups);
2866 * sched_balance_self: balance the current task (running on cpu) in domains
2867 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2870 * Balance, ie. select the least loaded group.
2872 * Returns the target CPU number, or the same CPU if no balancing is needed.
2874 * preempt must be disabled.
2877 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
2879 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
2880 int cpu = smp_processor_id();
2881 int prev_cpu = task_cpu(p);
2883 int want_affine = 0;
2884 int sync = wake_flags & WF_SYNC;
2886 if (p->nr_cpus_allowed == 1)
2889 if (sd_flag & SD_BALANCE_WAKE) {
2890 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
2896 for_each_domain(cpu, tmp) {
2897 if (!(tmp->flags & SD_LOAD_BALANCE))
2901 * If both cpu and prev_cpu are part of this domain,
2902 * cpu is a valid SD_WAKE_AFFINE target.
2904 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
2905 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
2910 if (tmp->flags & sd_flag)
2915 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
2918 new_cpu = select_idle_sibling(p, prev_cpu);
2923 int load_idx = sd->forkexec_idx;
2924 struct sched_group *group;
2927 if (!(sd->flags & sd_flag)) {
2932 if (sd_flag & SD_BALANCE_WAKE)
2933 load_idx = sd->wake_idx;
2935 group = find_idlest_group(sd, p, cpu, load_idx);
2941 new_cpu = find_idlest_cpu(group, p, cpu);
2942 if (new_cpu == -1 || new_cpu == cpu) {
2943 /* Now try balancing at a lower domain level of cpu */
2948 /* Now try balancing at a lower domain level of new_cpu */
2950 weight = sd->span_weight;
2952 for_each_domain(cpu, tmp) {
2953 if (weight <= tmp->span_weight)
2955 if (tmp->flags & sd_flag)
2958 /* while loop will break here if sd == NULL */
2965 #endif /* CONFIG_SMP */
2967 static unsigned long
2968 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
2970 unsigned long gran = sysctl_sched_wakeup_granularity;
2973 * Since its curr running now, convert the gran from real-time
2974 * to virtual-time in his units.
2976 * By using 'se' instead of 'curr' we penalize light tasks, so
2977 * they get preempted easier. That is, if 'se' < 'curr' then
2978 * the resulting gran will be larger, therefore penalizing the
2979 * lighter, if otoh 'se' > 'curr' then the resulting gran will
2980 * be smaller, again penalizing the lighter task.
2982 * This is especially important for buddies when the leftmost
2983 * task is higher priority than the buddy.
2985 return calc_delta_fair(gran, se);
2989 * Should 'se' preempt 'curr'.
3003 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3005 s64 gran, vdiff = curr->vruntime - se->vruntime;
3010 gran = wakeup_gran(curr, se);
3017 static void set_last_buddy(struct sched_entity *se)
3019 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3022 for_each_sched_entity(se)
3023 cfs_rq_of(se)->last = se;
3026 static void set_next_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)->next = se;
3035 static void set_skip_buddy(struct sched_entity *se)
3037 for_each_sched_entity(se)
3038 cfs_rq_of(se)->skip = se;
3042 * Preempt the current task with a newly woken task if needed:
3044 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3046 struct task_struct *curr = rq->curr;
3047 struct sched_entity *se = &curr->se, *pse = &p->se;
3048 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3049 int scale = cfs_rq->nr_running >= sched_nr_latency;
3050 int next_buddy_marked = 0;
3052 if (unlikely(se == pse))
3056 * This is possible from callers such as move_task(), in which we
3057 * unconditionally check_prempt_curr() after an enqueue (which may have
3058 * lead to a throttle). This both saves work and prevents false
3059 * next-buddy nomination below.
3061 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3064 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3065 set_next_buddy(pse);
3066 next_buddy_marked = 1;
3070 * We can come here with TIF_NEED_RESCHED already set from new task
3073 * Note: this also catches the edge-case of curr being in a throttled
3074 * group (e.g. via set_curr_task), since update_curr() (in the
3075 * enqueue of curr) will have resulted in resched being set. This
3076 * prevents us from potentially nominating it as a false LAST_BUDDY
3079 if (test_tsk_need_resched(curr))
3082 /* Idle tasks are by definition preempted by non-idle tasks. */
3083 if (unlikely(curr->policy == SCHED_IDLE) &&
3084 likely(p->policy != SCHED_IDLE))
3088 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3089 * is driven by the tick):
3091 if (unlikely(p->policy != SCHED_NORMAL))
3094 find_matching_se(&se, &pse);
3095 update_curr(cfs_rq_of(se));
3097 if (wakeup_preempt_entity(se, pse) == 1) {
3099 * Bias pick_next to pick the sched entity that is
3100 * triggering this preemption.
3102 if (!next_buddy_marked)
3103 set_next_buddy(pse);
3112 * Only set the backward buddy when the current task is still
3113 * on the rq. This can happen when a wakeup gets interleaved
3114 * with schedule on the ->pre_schedule() or idle_balance()
3115 * point, either of which can * drop the rq lock.
3117 * Also, during early boot the idle thread is in the fair class,
3118 * for obvious reasons its a bad idea to schedule back to it.
3120 if (unlikely(!se->on_rq || curr == rq->idle))
3123 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3127 static struct task_struct *pick_next_task_fair(struct rq *rq)
3129 struct task_struct *p;
3130 struct cfs_rq *cfs_rq = &rq->cfs;
3131 struct sched_entity *se;
3133 if (!cfs_rq->nr_running)
3137 se = pick_next_entity(cfs_rq);
3138 set_next_entity(cfs_rq, se);
3139 cfs_rq = group_cfs_rq(se);
3143 if (hrtick_enabled(rq))
3144 hrtick_start_fair(rq, p);
3150 * Account for a descheduled task:
3152 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3154 struct sched_entity *se = &prev->se;
3155 struct cfs_rq *cfs_rq;
3157 for_each_sched_entity(se) {
3158 cfs_rq = cfs_rq_of(se);
3159 put_prev_entity(cfs_rq, se);
3164 * sched_yield() is very simple
3166 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3168 static void yield_task_fair(struct rq *rq)
3170 struct task_struct *curr = rq->curr;
3171 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3172 struct sched_entity *se = &curr->se;
3175 * Are we the only task in the tree?
3177 if (unlikely(rq->nr_running == 1))
3180 clear_buddies(cfs_rq, se);
3182 if (curr->policy != SCHED_BATCH) {
3183 update_rq_clock(rq);
3185 * Update run-time statistics of the 'current'.
3187 update_curr(cfs_rq);
3189 * Tell update_rq_clock() that we've just updated,
3190 * so we don't do microscopic update in schedule()
3191 * and double the fastpath cost.
3193 rq->skip_clock_update = 1;
3199 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3201 struct sched_entity *se = &p->se;
3203 /* throttled hierarchies are not runnable */
3204 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3207 /* Tell the scheduler that we'd really like pse to run next. */
3210 yield_task_fair(rq);
3216 /**************************************************
3217 * Fair scheduling class load-balancing methods:
3220 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
3222 #define LBF_ALL_PINNED 0x01
3223 #define LBF_NEED_BREAK 0x02
3224 #define LBF_SOME_PINNED 0x04
3227 struct sched_domain *sd;
3235 struct cpumask *dst_grpmask;
3237 enum cpu_idle_type idle;
3239 /* The set of CPUs under consideration for load-balancing */
3240 struct cpumask *cpus;
3245 unsigned int loop_break;
3246 unsigned int loop_max;
3250 * move_task - move a task from one runqueue to another runqueue.
3251 * Both runqueues must be locked.
3253 static void move_task(struct task_struct *p, struct lb_env *env)
3255 deactivate_task(env->src_rq, p, 0);
3256 set_task_cpu(p, env->dst_cpu);
3257 activate_task(env->dst_rq, p, 0);
3258 check_preempt_curr(env->dst_rq, p, 0);
3262 * Is this task likely cache-hot:
3265 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
3269 if (p->sched_class != &fair_sched_class)
3272 if (unlikely(p->policy == SCHED_IDLE))
3276 * Buddy candidates are cache hot:
3278 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
3279 (&p->se == cfs_rq_of(&p->se)->next ||
3280 &p->se == cfs_rq_of(&p->se)->last))
3283 if (sysctl_sched_migration_cost == -1)
3285 if (sysctl_sched_migration_cost == 0)
3288 delta = now - p->se.exec_start;
3290 return delta < (s64)sysctl_sched_migration_cost;
3294 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3297 int can_migrate_task(struct task_struct *p, struct lb_env *env)
3299 int tsk_cache_hot = 0;
3301 * We do not migrate tasks that are:
3302 * 1) running (obviously), or
3303 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3304 * 3) are cache-hot on their current CPU.
3306 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
3309 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
3312 * Remember if this task can be migrated to any other cpu in
3313 * our sched_group. We may want to revisit it if we couldn't
3314 * meet load balance goals by pulling other tasks on src_cpu.
3316 * Also avoid computing new_dst_cpu if we have already computed
3317 * one in current iteration.
3319 if (!env->dst_grpmask || (env->flags & LBF_SOME_PINNED))
3322 new_dst_cpu = cpumask_first_and(env->dst_grpmask,
3323 tsk_cpus_allowed(p));
3324 if (new_dst_cpu < nr_cpu_ids) {
3325 env->flags |= LBF_SOME_PINNED;
3326 env->new_dst_cpu = new_dst_cpu;
3331 /* Record that we found atleast one task that could run on dst_cpu */
3332 env->flags &= ~LBF_ALL_PINNED;
3334 if (task_running(env->src_rq, p)) {
3335 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
3340 * Aggressive migration if:
3341 * 1) task is cache cold, or
3342 * 2) too many balance attempts have failed.
3345 tsk_cache_hot = task_hot(p, env->src_rq->clock_task, env->sd);
3346 if (!tsk_cache_hot ||
3347 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
3348 #ifdef CONFIG_SCHEDSTATS
3349 if (tsk_cache_hot) {
3350 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
3351 schedstat_inc(p, se.statistics.nr_forced_migrations);
3357 if (tsk_cache_hot) {
3358 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
3365 * move_one_task tries to move exactly one task from busiest to this_rq, as
3366 * part of active balancing operations within "domain".
3367 * Returns 1 if successful and 0 otherwise.
3369 * Called with both runqueues locked.
3371 static int move_one_task(struct lb_env *env)
3373 struct task_struct *p, *n;
3375 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
3376 if (throttled_lb_pair(task_group(p), env->src_rq->cpu, env->dst_cpu))
3379 if (!can_migrate_task(p, env))
3384 * Right now, this is only the second place move_task()
3385 * is called, so we can safely collect move_task()
3386 * stats here rather than inside move_task().
3388 schedstat_inc(env->sd, lb_gained[env->idle]);
3394 static unsigned long task_h_load(struct task_struct *p);
3396 static const unsigned int sched_nr_migrate_break = 32;
3399 * move_tasks tries to move up to imbalance weighted load from busiest to
3400 * this_rq, as part of a balancing operation within domain "sd".
3401 * Returns 1 if successful and 0 otherwise.
3403 * Called with both runqueues locked.
3405 static int move_tasks(struct lb_env *env)
3407 struct list_head *tasks = &env->src_rq->cfs_tasks;
3408 struct task_struct *p;
3412 if (env->imbalance <= 0)
3415 while (!list_empty(tasks)) {
3416 p = list_first_entry(tasks, struct task_struct, se.group_node);
3419 /* We've more or less seen every task there is, call it quits */
3420 if (env->loop > env->loop_max)
3423 /* take a breather every nr_migrate tasks */
3424 if (env->loop > env->loop_break) {
3425 env->loop_break += sched_nr_migrate_break;
3426 env->flags |= LBF_NEED_BREAK;
3430 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
3433 load = task_h_load(p);
3435 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
3438 if ((load / 2) > env->imbalance)
3441 if (!can_migrate_task(p, env))
3446 env->imbalance -= load;
3448 #ifdef CONFIG_PREEMPT
3450 * NEWIDLE balancing is a source of latency, so preemptible
3451 * kernels will stop after the first task is pulled to minimize
3452 * the critical section.
3454 if (env->idle == CPU_NEWLY_IDLE)
3459 * We only want to steal up to the prescribed amount of
3462 if (env->imbalance <= 0)
3467 list_move_tail(&p->se.group_node, tasks);
3471 * Right now, this is one of only two places move_task() is called,
3472 * so we can safely collect move_task() stats here rather than
3473 * inside move_task().
3475 schedstat_add(env->sd, lb_gained[env->idle], pulled);
3480 #ifdef CONFIG_FAIR_GROUP_SCHED
3482 * update tg->load_weight by folding this cpu's load_avg
3484 static int update_shares_cpu(struct task_group *tg, int cpu)
3486 struct cfs_rq *cfs_rq;
3487 unsigned long flags;
3494 cfs_rq = tg->cfs_rq[cpu];
3496 raw_spin_lock_irqsave(&rq->lock, flags);
3498 update_rq_clock(rq);
3499 update_cfs_load(cfs_rq, 1);
3502 * We need to update shares after updating tg->load_weight in
3503 * order to adjust the weight of groups with long running tasks.
3505 update_cfs_shares(cfs_rq);
3507 raw_spin_unlock_irqrestore(&rq->lock, flags);
3512 static void update_shares(int cpu)
3514 struct cfs_rq *cfs_rq;
3515 struct rq *rq = cpu_rq(cpu);
3519 * Iterates the task_group tree in a bottom up fashion, see
3520 * list_add_leaf_cfs_rq() for details.
3522 for_each_leaf_cfs_rq(rq, cfs_rq) {
3523 /* throttled entities do not contribute to load */
3524 if (throttled_hierarchy(cfs_rq))
3527 update_shares_cpu(cfs_rq->tg, cpu);
3533 * Compute the cpu's hierarchical load factor for each task group.
3534 * This needs to be done in a top-down fashion because the load of a child
3535 * group is a fraction of its parents load.
3537 static int tg_load_down(struct task_group *tg, void *data)
3540 long cpu = (long)data;
3543 load = cpu_rq(cpu)->load.weight;
3545 load = tg->parent->cfs_rq[cpu]->h_load;
3546 load *= tg->se[cpu]->load.weight;
3547 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
3550 tg->cfs_rq[cpu]->h_load = load;
3555 static void update_h_load(long cpu)
3557 struct rq *rq = cpu_rq(cpu);
3558 unsigned long now = jiffies;
3560 if (rq->h_load_throttle == now)
3563 rq->h_load_throttle = now;
3566 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
3570 static unsigned long task_h_load(struct task_struct *p)
3572 struct cfs_rq *cfs_rq = task_cfs_rq(p);
3575 load = p->se.load.weight;
3576 load = div_u64(load * cfs_rq->h_load, cfs_rq->load.weight + 1);
3581 static inline void update_shares(int cpu)
3585 static inline void update_h_load(long cpu)
3589 static unsigned long task_h_load(struct task_struct *p)
3591 return p->se.load.weight;
3595 /********** Helpers for find_busiest_group ************************/
3597 * sd_lb_stats - Structure to store the statistics of a sched_domain
3598 * during load balancing.
3600 struct sd_lb_stats {
3601 struct sched_group *busiest; /* Busiest group in this sd */
3602 struct sched_group *this; /* Local group in this sd */
3603 unsigned long total_load; /* Total load of all groups in sd */
3604 unsigned long total_pwr; /* Total power of all groups in sd */
3605 unsigned long avg_load; /* Average load across all groups in sd */
3607 /** Statistics of this group */
3608 unsigned long this_load;
3609 unsigned long this_load_per_task;
3610 unsigned long this_nr_running;
3611 unsigned long this_has_capacity;
3612 unsigned int this_idle_cpus;
3614 /* Statistics of the busiest group */
3615 unsigned int busiest_idle_cpus;
3616 unsigned long max_load;
3617 unsigned long busiest_load_per_task;
3618 unsigned long busiest_nr_running;
3619 unsigned long busiest_group_capacity;
3620 unsigned long busiest_has_capacity;
3621 unsigned int busiest_group_weight;
3623 int group_imb; /* Is there imbalance in this sd */
3627 * sg_lb_stats - stats of a sched_group required for load_balancing
3629 struct sg_lb_stats {
3630 unsigned long avg_load; /*Avg load across the CPUs of the group */
3631 unsigned long group_load; /* Total load over the CPUs of the group */
3632 unsigned long sum_nr_running; /* Nr tasks running in the group */
3633 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3634 unsigned long group_capacity;
3635 unsigned long idle_cpus;
3636 unsigned long group_weight;
3637 int group_imb; /* Is there an imbalance in the group ? */
3638 int group_has_capacity; /* Is there extra capacity in the group? */
3642 * get_sd_load_idx - Obtain the load index for a given sched domain.
3643 * @sd: The sched_domain whose load_idx is to be obtained.
3644 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3646 static inline int get_sd_load_idx(struct sched_domain *sd,
3647 enum cpu_idle_type idle)
3653 load_idx = sd->busy_idx;
3656 case CPU_NEWLY_IDLE:
3657 load_idx = sd->newidle_idx;
3660 load_idx = sd->idle_idx;
3667 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3669 return SCHED_POWER_SCALE;
3672 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3674 return default_scale_freq_power(sd, cpu);
3677 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3679 unsigned long weight = sd->span_weight;
3680 unsigned long smt_gain = sd->smt_gain;
3687 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3689 return default_scale_smt_power(sd, cpu);
3692 unsigned long scale_rt_power(int cpu)
3694 struct rq *rq = cpu_rq(cpu);
3695 u64 total, available, age_stamp, avg;
3698 * Since we're reading these variables without serialization make sure
3699 * we read them once before doing sanity checks on them.
3701 age_stamp = ACCESS_ONCE(rq->age_stamp);
3702 avg = ACCESS_ONCE(rq->rt_avg);
3704 total = sched_avg_period() + (rq->clock - age_stamp);
3706 if (unlikely(total < avg)) {
3707 /* Ensures that power won't end up being negative */
3710 available = total - avg;
3713 if (unlikely((s64)total < SCHED_POWER_SCALE))
3714 total = SCHED_POWER_SCALE;
3716 total >>= SCHED_POWER_SHIFT;
3718 return div_u64(available, total);
3721 static void update_cpu_power(struct sched_domain *sd, int cpu)
3723 unsigned long weight = sd->span_weight;
3724 unsigned long power = SCHED_POWER_SCALE;
3725 struct sched_group *sdg = sd->groups;
3727 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3728 if (sched_feat(ARCH_POWER))
3729 power *= arch_scale_smt_power(sd, cpu);
3731 power *= default_scale_smt_power(sd, cpu);
3733 power >>= SCHED_POWER_SHIFT;
3736 sdg->sgp->power_orig = power;
3738 if (sched_feat(ARCH_POWER))
3739 power *= arch_scale_freq_power(sd, cpu);
3741 power *= default_scale_freq_power(sd, cpu);
3743 power >>= SCHED_POWER_SHIFT;
3745 power *= scale_rt_power(cpu);
3746 power >>= SCHED_POWER_SHIFT;
3751 cpu_rq(cpu)->cpu_power = power;
3752 sdg->sgp->power = power;
3755 void update_group_power(struct sched_domain *sd, int cpu)
3757 struct sched_domain *child = sd->child;
3758 struct sched_group *group, *sdg = sd->groups;
3759 unsigned long power;
3760 unsigned long interval;
3762 interval = msecs_to_jiffies(sd->balance_interval);
3763 interval = clamp(interval, 1UL, max_load_balance_interval);
3764 sdg->sgp->next_update = jiffies + interval;
3767 update_cpu_power(sd, cpu);
3773 if (child->flags & SD_OVERLAP) {
3775 * SD_OVERLAP domains cannot assume that child groups
3776 * span the current group.
3779 for_each_cpu(cpu, sched_group_cpus(sdg))
3780 power += power_of(cpu);
3783 * !SD_OVERLAP domains can assume that child groups
3784 * span the current group.
3787 group = child->groups;
3789 power += group->sgp->power;
3790 group = group->next;
3791 } while (group != child->groups);
3794 sdg->sgp->power_orig = sdg->sgp->power = power;
3798 * Try and fix up capacity for tiny siblings, this is needed when
3799 * things like SD_ASYM_PACKING need f_b_g to select another sibling
3800 * which on its own isn't powerful enough.
3802 * See update_sd_pick_busiest() and check_asym_packing().
3805 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
3808 * Only siblings can have significantly less than SCHED_POWER_SCALE
3810 if (!(sd->flags & SD_SHARE_CPUPOWER))
3814 * If ~90% of the cpu_power is still there, we're good.
3816 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
3823 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3824 * @env: The load balancing environment.
3825 * @group: sched_group whose statistics are to be updated.
3826 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3827 * @local_group: Does group contain this_cpu.
3828 * @balance: Should we balance.
3829 * @sgs: variable to hold the statistics for this group.
3831 static inline void update_sg_lb_stats(struct lb_env *env,
3832 struct sched_group *group, int load_idx,
3833 int local_group, int *balance, struct sg_lb_stats *sgs)
3835 unsigned long nr_running, max_nr_running, min_nr_running;
3836 unsigned long load, max_cpu_load, min_cpu_load;
3837 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3838 unsigned long avg_load_per_task = 0;
3842 balance_cpu = group_balance_cpu(group);
3844 /* Tally up the load of all CPUs in the group */
3846 min_cpu_load = ~0UL;
3848 min_nr_running = ~0UL;
3850 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
3851 struct rq *rq = cpu_rq(i);
3853 nr_running = rq->nr_running;
3855 /* Bias balancing toward cpus of our domain */
3857 if (idle_cpu(i) && !first_idle_cpu &&
3858 cpumask_test_cpu(i, sched_group_mask(group))) {
3863 load = target_load(i, load_idx);
3865 load = source_load(i, load_idx);
3866 if (load > max_cpu_load)
3867 max_cpu_load = load;
3868 if (min_cpu_load > load)
3869 min_cpu_load = load;
3871 if (nr_running > max_nr_running)
3872 max_nr_running = nr_running;
3873 if (min_nr_running > nr_running)
3874 min_nr_running = nr_running;
3877 sgs->group_load += load;
3878 sgs->sum_nr_running += nr_running;
3879 sgs->sum_weighted_load += weighted_cpuload(i);
3885 * First idle cpu or the first cpu(busiest) in this sched group
3886 * is eligible for doing load balancing at this and above
3887 * domains. In the newly idle case, we will allow all the cpu's
3888 * to do the newly idle load balance.
3891 if (env->idle != CPU_NEWLY_IDLE) {
3892 if (balance_cpu != env->dst_cpu) {
3896 update_group_power(env->sd, env->dst_cpu);
3897 } else if (time_after_eq(jiffies, group->sgp->next_update))
3898 update_group_power(env->sd, env->dst_cpu);
3901 /* Adjust by relative CPU power of the group */
3902 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
3905 * Consider the group unbalanced when the imbalance is larger
3906 * than the average weight of a task.
3908 * APZ: with cgroup the avg task weight can vary wildly and
3909 * might not be a suitable number - should we keep a
3910 * normalized nr_running number somewhere that negates
3913 if (sgs->sum_nr_running)
3914 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
3916 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task &&
3917 (max_nr_running - min_nr_running) > 1)
3920 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
3922 if (!sgs->group_capacity)
3923 sgs->group_capacity = fix_small_capacity(env->sd, group);
3924 sgs->group_weight = group->group_weight;
3926 if (sgs->group_capacity > sgs->sum_nr_running)
3927 sgs->group_has_capacity = 1;
3931 * update_sd_pick_busiest - return 1 on busiest group
3932 * @env: The load balancing environment.
3933 * @sds: sched_domain statistics
3934 * @sg: sched_group candidate to be checked for being the busiest
3935 * @sgs: sched_group statistics
3937 * Determine if @sg is a busier group than the previously selected
3940 static bool update_sd_pick_busiest(struct lb_env *env,
3941 struct sd_lb_stats *sds,
3942 struct sched_group *sg,
3943 struct sg_lb_stats *sgs)
3945 if (sgs->avg_load <= sds->max_load)
3948 if (sgs->sum_nr_running > sgs->group_capacity)
3955 * ASYM_PACKING needs to move all the work to the lowest
3956 * numbered CPUs in the group, therefore mark all groups
3957 * higher than ourself as busy.
3959 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
3960 env->dst_cpu < group_first_cpu(sg)) {
3964 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
3972 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
3973 * @env: The load balancing environment.
3974 * @balance: Should we balance.
3975 * @sds: variable to hold the statistics for this sched_domain.
3977 static inline void update_sd_lb_stats(struct lb_env *env,
3978 int *balance, struct sd_lb_stats *sds)
3980 struct sched_domain *child = env->sd->child;
3981 struct sched_group *sg = env->sd->groups;
3982 struct sg_lb_stats sgs;
3983 int load_idx, prefer_sibling = 0;
3985 if (child && child->flags & SD_PREFER_SIBLING)
3988 load_idx = get_sd_load_idx(env->sd, env->idle);
3993 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
3994 memset(&sgs, 0, sizeof(sgs));
3995 update_sg_lb_stats(env, sg, load_idx, local_group, balance, &sgs);
3997 if (local_group && !(*balance))
4000 sds->total_load += sgs.group_load;
4001 sds->total_pwr += sg->sgp->power;
4004 * In case the child domain prefers tasks go to siblings
4005 * first, lower the sg capacity to one so that we'll try
4006 * and move all the excess tasks away. We lower the capacity
4007 * of a group only if the local group has the capacity to fit
4008 * these excess tasks, i.e. nr_running < group_capacity. The
4009 * extra check prevents the case where you always pull from the
4010 * heaviest group when it is already under-utilized (possible
4011 * with a large weight task outweighs the tasks on the system).
4013 if (prefer_sibling && !local_group && sds->this_has_capacity)
4014 sgs.group_capacity = min(sgs.group_capacity, 1UL);
4017 sds->this_load = sgs.avg_load;
4019 sds->this_nr_running = sgs.sum_nr_running;
4020 sds->this_load_per_task = sgs.sum_weighted_load;
4021 sds->this_has_capacity = sgs.group_has_capacity;
4022 sds->this_idle_cpus = sgs.idle_cpus;
4023 } else if (update_sd_pick_busiest(env, sds, sg, &sgs)) {
4024 sds->max_load = sgs.avg_load;
4026 sds->busiest_nr_running = sgs.sum_nr_running;
4027 sds->busiest_idle_cpus = sgs.idle_cpus;
4028 sds->busiest_group_capacity = sgs.group_capacity;
4029 sds->busiest_load_per_task = sgs.sum_weighted_load;
4030 sds->busiest_has_capacity = sgs.group_has_capacity;
4031 sds->busiest_group_weight = sgs.group_weight;
4032 sds->group_imb = sgs.group_imb;
4036 } while (sg != env->sd->groups);
4040 * check_asym_packing - Check to see if the group is packed into the
4043 * This is primarily intended to used at the sibling level. Some
4044 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4045 * case of POWER7, it can move to lower SMT modes only when higher
4046 * threads are idle. When in lower SMT modes, the threads will
4047 * perform better since they share less core resources. Hence when we
4048 * have idle threads, we want them to be the higher ones.
4050 * This packing function is run on idle threads. It checks to see if
4051 * the busiest CPU in this domain (core in the P7 case) has a higher
4052 * CPU number than the packing function is being run on. Here we are
4053 * assuming lower CPU number will be equivalent to lower a SMT thread
4056 * Returns 1 when packing is required and a task should be moved to
4057 * this CPU. The amount of the imbalance is returned in *imbalance.
4059 * @env: The load balancing environment.
4060 * @sds: Statistics of the sched_domain which is to be packed
4062 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
4066 if (!(env->sd->flags & SD_ASYM_PACKING))
4072 busiest_cpu = group_first_cpu(sds->busiest);
4073 if (env->dst_cpu > busiest_cpu)
4076 env->imbalance = DIV_ROUND_CLOSEST(
4077 sds->max_load * sds->busiest->sgp->power, SCHED_POWER_SCALE);
4083 * fix_small_imbalance - Calculate the minor imbalance that exists
4084 * amongst the groups of a sched_domain, during
4086 * @env: The load balancing environment.
4087 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4090 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4092 unsigned long tmp, pwr_now = 0, pwr_move = 0;
4093 unsigned int imbn = 2;
4094 unsigned long scaled_busy_load_per_task;
4096 if (sds->this_nr_running) {
4097 sds->this_load_per_task /= sds->this_nr_running;
4098 if (sds->busiest_load_per_task >
4099 sds->this_load_per_task)
4102 sds->this_load_per_task =
4103 cpu_avg_load_per_task(env->dst_cpu);
4106 scaled_busy_load_per_task = sds->busiest_load_per_task
4107 * SCHED_POWER_SCALE;
4108 scaled_busy_load_per_task /= sds->busiest->sgp->power;
4110 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
4111 (scaled_busy_load_per_task * imbn)) {
4112 env->imbalance = sds->busiest_load_per_task;
4117 * OK, we don't have enough imbalance to justify moving tasks,
4118 * however we may be able to increase total CPU power used by
4122 pwr_now += sds->busiest->sgp->power *
4123 min(sds->busiest_load_per_task, sds->max_load);
4124 pwr_now += sds->this->sgp->power *
4125 min(sds->this_load_per_task, sds->this_load);
4126 pwr_now /= SCHED_POWER_SCALE;
4128 /* Amount of load we'd subtract */
4129 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4130 sds->busiest->sgp->power;
4131 if (sds->max_load > tmp)
4132 pwr_move += sds->busiest->sgp->power *
4133 min(sds->busiest_load_per_task, sds->max_load - tmp);
4135 /* Amount of load we'd add */
4136 if (sds->max_load * sds->busiest->sgp->power <
4137 sds->busiest_load_per_task * SCHED_POWER_SCALE)
4138 tmp = (sds->max_load * sds->busiest->sgp->power) /
4139 sds->this->sgp->power;
4141 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4142 sds->this->sgp->power;
4143 pwr_move += sds->this->sgp->power *
4144 min(sds->this_load_per_task, sds->this_load + tmp);
4145 pwr_move /= SCHED_POWER_SCALE;
4147 /* Move if we gain throughput */
4148 if (pwr_move > pwr_now)
4149 env->imbalance = sds->busiest_load_per_task;
4153 * calculate_imbalance - Calculate the amount of imbalance present within the
4154 * groups of a given sched_domain during load balance.
4155 * @env: load balance environment
4156 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4158 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4160 unsigned long max_pull, load_above_capacity = ~0UL;
4162 sds->busiest_load_per_task /= sds->busiest_nr_running;
4163 if (sds->group_imb) {
4164 sds->busiest_load_per_task =
4165 min(sds->busiest_load_per_task, sds->avg_load);
4169 * In the presence of smp nice balancing, certain scenarios can have
4170 * max load less than avg load(as we skip the groups at or below
4171 * its cpu_power, while calculating max_load..)
4173 if (sds->max_load < sds->avg_load) {
4175 return fix_small_imbalance(env, sds);
4178 if (!sds->group_imb) {
4180 * Don't want to pull so many tasks that a group would go idle.
4182 load_above_capacity = (sds->busiest_nr_running -
4183 sds->busiest_group_capacity);
4185 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
4187 load_above_capacity /= sds->busiest->sgp->power;
4191 * We're trying to get all the cpus to the average_load, so we don't
4192 * want to push ourselves above the average load, nor do we wish to
4193 * reduce the max loaded cpu below the average load. At the same time,
4194 * we also don't want to reduce the group load below the group capacity
4195 * (so that we can implement power-savings policies etc). Thus we look
4196 * for the minimum possible imbalance.
4197 * Be careful of negative numbers as they'll appear as very large values
4198 * with unsigned longs.
4200 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
4202 /* How much load to actually move to equalise the imbalance */
4203 env->imbalance = min(max_pull * sds->busiest->sgp->power,
4204 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
4205 / SCHED_POWER_SCALE;
4208 * if *imbalance is less than the average load per runnable task
4209 * there is no guarantee that any tasks will be moved so we'll have
4210 * a think about bumping its value to force at least one task to be
4213 if (env->imbalance < sds->busiest_load_per_task)
4214 return fix_small_imbalance(env, sds);
4218 /******* find_busiest_group() helpers end here *********************/
4221 * find_busiest_group - Returns the busiest group within the sched_domain
4222 * if there is an imbalance. If there isn't an imbalance, and
4223 * the user has opted for power-savings, it returns a group whose
4224 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4225 * such a group exists.
4227 * Also calculates the amount of weighted load which should be moved
4228 * to restore balance.
4230 * @env: The load balancing environment.
4231 * @balance: Pointer to a variable indicating if this_cpu
4232 * is the appropriate cpu to perform load balancing at this_level.
4234 * Returns: - the busiest group if imbalance exists.
4235 * - If no imbalance and user has opted for power-savings balance,
4236 * return the least loaded group whose CPUs can be
4237 * put to idle by rebalancing its tasks onto our group.
4239 static struct sched_group *
4240 find_busiest_group(struct lb_env *env, int *balance)
4242 struct sd_lb_stats sds;
4244 memset(&sds, 0, sizeof(sds));
4247 * Compute the various statistics relavent for load balancing at
4250 update_sd_lb_stats(env, balance, &sds);
4253 * this_cpu is not the appropriate cpu to perform load balancing at
4259 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
4260 check_asym_packing(env, &sds))
4263 /* There is no busy sibling group to pull tasks from */
4264 if (!sds.busiest || sds.busiest_nr_running == 0)
4267 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
4270 * If the busiest group is imbalanced the below checks don't
4271 * work because they assumes all things are equal, which typically
4272 * isn't true due to cpus_allowed constraints and the like.
4277 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
4278 if (env->idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
4279 !sds.busiest_has_capacity)
4283 * If the local group is more busy than the selected busiest group
4284 * don't try and pull any tasks.
4286 if (sds.this_load >= sds.max_load)
4290 * Don't pull any tasks if this group is already above the domain
4293 if (sds.this_load >= sds.avg_load)
4296 if (env->idle == CPU_IDLE) {
4298 * This cpu is idle. If the busiest group load doesn't
4299 * have more tasks than the number of available cpu's and
4300 * there is no imbalance between this and busiest group
4301 * wrt to idle cpu's, it is balanced.
4303 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
4304 sds.busiest_nr_running <= sds.busiest_group_weight)
4308 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
4309 * imbalance_pct to be conservative.
4311 if (100 * sds.max_load <= env->sd->imbalance_pct * sds.this_load)
4316 /* Looks like there is an imbalance. Compute it */
4317 calculate_imbalance(env, &sds);
4327 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4329 static struct rq *find_busiest_queue(struct lb_env *env,
4330 struct sched_group *group)
4332 struct rq *busiest = NULL, *rq;
4333 unsigned long max_load = 0;
4336 for_each_cpu(i, sched_group_cpus(group)) {
4337 unsigned long power = power_of(i);
4338 unsigned long capacity = DIV_ROUND_CLOSEST(power,
4343 capacity = fix_small_capacity(env->sd, group);
4345 if (!cpumask_test_cpu(i, env->cpus))
4349 wl = weighted_cpuload(i);
4352 * When comparing with imbalance, use weighted_cpuload()
4353 * which is not scaled with the cpu power.
4355 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
4359 * For the load comparisons with the other cpu's, consider
4360 * the weighted_cpuload() scaled with the cpu power, so that
4361 * the load can be moved away from the cpu that is potentially
4362 * running at a lower capacity.
4364 wl = (wl * SCHED_POWER_SCALE) / power;
4366 if (wl > max_load) {
4376 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4377 * so long as it is large enough.
4379 #define MAX_PINNED_INTERVAL 512
4381 /* Working cpumask for load_balance and load_balance_newidle. */
4382 DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4384 static int need_active_balance(struct lb_env *env)
4386 struct sched_domain *sd = env->sd;
4388 if (env->idle == CPU_NEWLY_IDLE) {
4391 * ASYM_PACKING needs to force migrate tasks from busy but
4392 * higher numbered CPUs in order to pack all tasks in the
4393 * lowest numbered CPUs.
4395 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
4399 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
4402 static int active_load_balance_cpu_stop(void *data);
4405 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4406 * tasks if there is an imbalance.
4408 static int load_balance(int this_cpu, struct rq *this_rq,
4409 struct sched_domain *sd, enum cpu_idle_type idle,
4412 int ld_moved, cur_ld_moved, active_balance = 0;
4413 int lb_iterations, max_lb_iterations;
4414 struct sched_group *group;
4416 unsigned long flags;
4417 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4419 struct lb_env env = {
4421 .dst_cpu = this_cpu,
4423 .dst_grpmask = sched_group_cpus(sd->groups),
4425 .loop_break = sched_nr_migrate_break,
4429 cpumask_copy(cpus, cpu_active_mask);
4430 max_lb_iterations = cpumask_weight(env.dst_grpmask);
4432 schedstat_inc(sd, lb_count[idle]);
4435 group = find_busiest_group(&env, balance);
4441 schedstat_inc(sd, lb_nobusyg[idle]);
4445 busiest = find_busiest_queue(&env, group);
4447 schedstat_inc(sd, lb_nobusyq[idle]);
4451 BUG_ON(busiest == env.dst_rq);
4453 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
4457 if (busiest->nr_running > 1) {
4459 * Attempt to move tasks. If find_busiest_group has found
4460 * an imbalance but busiest->nr_running <= 1, the group is
4461 * still unbalanced. ld_moved simply stays zero, so it is
4462 * correctly treated as an imbalance.
4464 env.flags |= LBF_ALL_PINNED;
4465 env.src_cpu = busiest->cpu;
4466 env.src_rq = busiest;
4467 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
4469 update_h_load(env.src_cpu);
4471 local_irq_save(flags);
4472 double_rq_lock(env.dst_rq, busiest);
4475 * cur_ld_moved - load moved in current iteration
4476 * ld_moved - cumulative load moved across iterations
4478 cur_ld_moved = move_tasks(&env);
4479 ld_moved += cur_ld_moved;
4480 double_rq_unlock(env.dst_rq, busiest);
4481 local_irq_restore(flags);
4483 if (env.flags & LBF_NEED_BREAK) {
4484 env.flags &= ~LBF_NEED_BREAK;
4489 * some other cpu did the load balance for us.
4491 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
4492 resched_cpu(env.dst_cpu);
4495 * Revisit (affine) tasks on src_cpu that couldn't be moved to
4496 * us and move them to an alternate dst_cpu in our sched_group
4497 * where they can run. The upper limit on how many times we
4498 * iterate on same src_cpu is dependent on number of cpus in our
4501 * This changes load balance semantics a bit on who can move
4502 * load to a given_cpu. In addition to the given_cpu itself
4503 * (or a ilb_cpu acting on its behalf where given_cpu is
4504 * nohz-idle), we now have balance_cpu in a position to move
4505 * load to given_cpu. In rare situations, this may cause
4506 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
4507 * _independently_ and at _same_ time to move some load to
4508 * given_cpu) causing exceess load to be moved to given_cpu.
4509 * This however should not happen so much in practice and
4510 * moreover subsequent load balance cycles should correct the
4511 * excess load moved.
4513 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0 &&
4514 lb_iterations++ < max_lb_iterations) {
4516 env.dst_rq = cpu_rq(env.new_dst_cpu);
4517 env.dst_cpu = env.new_dst_cpu;
4518 env.flags &= ~LBF_SOME_PINNED;
4520 env.loop_break = sched_nr_migrate_break;
4522 * Go back to "more_balance" rather than "redo" since we
4523 * need to continue with same src_cpu.
4528 /* All tasks on this runqueue were pinned by CPU affinity */
4529 if (unlikely(env.flags & LBF_ALL_PINNED)) {
4530 cpumask_clear_cpu(cpu_of(busiest), cpus);
4531 if (!cpumask_empty(cpus)) {
4533 env.loop_break = sched_nr_migrate_break;
4541 schedstat_inc(sd, lb_failed[idle]);
4543 * Increment the failure counter only on periodic balance.
4544 * We do not want newidle balance, which can be very
4545 * frequent, pollute the failure counter causing
4546 * excessive cache_hot migrations and active balances.
4548 if (idle != CPU_NEWLY_IDLE)
4549 sd->nr_balance_failed++;
4551 if (need_active_balance(&env)) {
4552 raw_spin_lock_irqsave(&busiest->lock, flags);
4554 /* don't kick the active_load_balance_cpu_stop,
4555 * if the curr task on busiest cpu can't be
4558 if (!cpumask_test_cpu(this_cpu,
4559 tsk_cpus_allowed(busiest->curr))) {
4560 raw_spin_unlock_irqrestore(&busiest->lock,
4562 env.flags |= LBF_ALL_PINNED;
4563 goto out_one_pinned;
4567 * ->active_balance synchronizes accesses to
4568 * ->active_balance_work. Once set, it's cleared
4569 * only after active load balance is finished.
4571 if (!busiest->active_balance) {
4572 busiest->active_balance = 1;
4573 busiest->push_cpu = this_cpu;
4576 raw_spin_unlock_irqrestore(&busiest->lock, flags);
4578 if (active_balance) {
4579 stop_one_cpu_nowait(cpu_of(busiest),
4580 active_load_balance_cpu_stop, busiest,
4581 &busiest->active_balance_work);
4585 * We've kicked active balancing, reset the failure
4588 sd->nr_balance_failed = sd->cache_nice_tries+1;
4591 sd->nr_balance_failed = 0;
4593 if (likely(!active_balance)) {
4594 /* We were unbalanced, so reset the balancing interval */
4595 sd->balance_interval = sd->min_interval;
4598 * If we've begun active balancing, start to back off. This
4599 * case may not be covered by the all_pinned logic if there
4600 * is only 1 task on the busy runqueue (because we don't call
4603 if (sd->balance_interval < sd->max_interval)
4604 sd->balance_interval *= 2;
4610 schedstat_inc(sd, lb_balanced[idle]);
4612 sd->nr_balance_failed = 0;
4615 /* tune up the balancing interval */
4616 if (((env.flags & LBF_ALL_PINNED) &&
4617 sd->balance_interval < MAX_PINNED_INTERVAL) ||
4618 (sd->balance_interval < sd->max_interval))
4619 sd->balance_interval *= 2;
4627 * idle_balance is called by schedule() if this_cpu is about to become
4628 * idle. Attempts to pull tasks from other CPUs.
4630 void idle_balance(int this_cpu, struct rq *this_rq)
4632 struct sched_domain *sd;
4633 int pulled_task = 0;
4634 unsigned long next_balance = jiffies + HZ;
4636 this_rq->idle_stamp = this_rq->clock;
4638 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4642 * Drop the rq->lock, but keep IRQ/preempt disabled.
4644 raw_spin_unlock(&this_rq->lock);
4646 update_shares(this_cpu);
4648 for_each_domain(this_cpu, sd) {
4649 unsigned long interval;
4652 if (!(sd->flags & SD_LOAD_BALANCE))
4655 if (sd->flags & SD_BALANCE_NEWIDLE) {
4656 /* If we've pulled tasks over stop searching: */
4657 pulled_task = load_balance(this_cpu, this_rq,
4658 sd, CPU_NEWLY_IDLE, &balance);
4661 interval = msecs_to_jiffies(sd->balance_interval);
4662 if (time_after(next_balance, sd->last_balance + interval))
4663 next_balance = sd->last_balance + interval;
4665 this_rq->idle_stamp = 0;
4671 raw_spin_lock(&this_rq->lock);
4673 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4675 * We are going idle. next_balance may be set based on
4676 * a busy processor. So reset next_balance.
4678 this_rq->next_balance = next_balance;
4683 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
4684 * running tasks off the busiest CPU onto idle CPUs. It requires at
4685 * least 1 task to be running on each physical CPU where possible, and
4686 * avoids physical / logical imbalances.
4688 static int active_load_balance_cpu_stop(void *data)
4690 struct rq *busiest_rq = data;
4691 int busiest_cpu = cpu_of(busiest_rq);
4692 int target_cpu = busiest_rq->push_cpu;
4693 struct rq *target_rq = cpu_rq(target_cpu);
4694 struct sched_domain *sd;
4696 raw_spin_lock_irq(&busiest_rq->lock);
4698 /* make sure the requested cpu hasn't gone down in the meantime */
4699 if (unlikely(busiest_cpu != smp_processor_id() ||
4700 !busiest_rq->active_balance))
4703 /* Is there any task to move? */
4704 if (busiest_rq->nr_running <= 1)
4708 * This condition is "impossible", if it occurs
4709 * we need to fix it. Originally reported by
4710 * Bjorn Helgaas on a 128-cpu setup.
4712 BUG_ON(busiest_rq == target_rq);
4714 /* move a task from busiest_rq to target_rq */
4715 double_lock_balance(busiest_rq, target_rq);
4717 /* Search for an sd spanning us and the target CPU. */
4719 for_each_domain(target_cpu, sd) {
4720 if ((sd->flags & SD_LOAD_BALANCE) &&
4721 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4726 struct lb_env env = {
4728 .dst_cpu = target_cpu,
4729 .dst_rq = target_rq,
4730 .src_cpu = busiest_rq->cpu,
4731 .src_rq = busiest_rq,
4735 schedstat_inc(sd, alb_count);
4737 if (move_one_task(&env))
4738 schedstat_inc(sd, alb_pushed);
4740 schedstat_inc(sd, alb_failed);
4743 double_unlock_balance(busiest_rq, target_rq);
4745 busiest_rq->active_balance = 0;
4746 raw_spin_unlock_irq(&busiest_rq->lock);
4752 * idle load balancing details
4753 * - When one of the busy CPUs notice that there may be an idle rebalancing
4754 * needed, they will kick the idle load balancer, which then does idle
4755 * load balancing for all the idle CPUs.
4758 cpumask_var_t idle_cpus_mask;
4760 unsigned long next_balance; /* in jiffy units */
4761 } nohz ____cacheline_aligned;
4763 static inline int find_new_ilb(int call_cpu)
4765 int ilb = cpumask_first(nohz.idle_cpus_mask);
4767 if (ilb < nr_cpu_ids && idle_cpu(ilb))
4774 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
4775 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
4776 * CPU (if there is one).
4778 static void nohz_balancer_kick(int cpu)
4782 nohz.next_balance++;
4784 ilb_cpu = find_new_ilb(cpu);
4786 if (ilb_cpu >= nr_cpu_ids)
4789 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
4792 * Use smp_send_reschedule() instead of resched_cpu().
4793 * This way we generate a sched IPI on the target cpu which
4794 * is idle. And the softirq performing nohz idle load balance
4795 * will be run before returning from the IPI.
4797 smp_send_reschedule(ilb_cpu);
4801 static inline void nohz_balance_exit_idle(int cpu)
4803 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
4804 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
4805 atomic_dec(&nohz.nr_cpus);
4806 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
4810 static inline void set_cpu_sd_state_busy(void)
4812 struct sched_domain *sd;
4813 int cpu = smp_processor_id();
4815 if (!test_bit(NOHZ_IDLE, nohz_flags(cpu)))
4817 clear_bit(NOHZ_IDLE, nohz_flags(cpu));
4820 for_each_domain(cpu, sd)
4821 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
4825 void set_cpu_sd_state_idle(void)
4827 struct sched_domain *sd;
4828 int cpu = smp_processor_id();
4830 if (test_bit(NOHZ_IDLE, nohz_flags(cpu)))
4832 set_bit(NOHZ_IDLE, nohz_flags(cpu));
4835 for_each_domain(cpu, sd)
4836 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
4841 * This routine will record that the cpu is going idle with tick stopped.
4842 * This info will be used in performing idle load balancing in the future.
4844 void nohz_balance_enter_idle(int cpu)
4847 * If this cpu is going down, then nothing needs to be done.
4849 if (!cpu_active(cpu))
4852 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
4855 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
4856 atomic_inc(&nohz.nr_cpus);
4857 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
4860 static int __cpuinit sched_ilb_notifier(struct notifier_block *nfb,
4861 unsigned long action, void *hcpu)
4863 switch (action & ~CPU_TASKS_FROZEN) {
4865 nohz_balance_exit_idle(smp_processor_id());
4873 static DEFINE_SPINLOCK(balancing);
4876 * Scale the max load_balance interval with the number of CPUs in the system.
4877 * This trades load-balance latency on larger machines for less cross talk.
4879 void update_max_interval(void)
4881 max_load_balance_interval = HZ*num_online_cpus()/10;
4885 * It checks each scheduling domain to see if it is due to be balanced,
4886 * and initiates a balancing operation if so.
4888 * Balancing parameters are set up in arch_init_sched_domains.
4890 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4893 struct rq *rq = cpu_rq(cpu);
4894 unsigned long interval;
4895 struct sched_domain *sd;
4896 /* Earliest time when we have to do rebalance again */
4897 unsigned long next_balance = jiffies + 60*HZ;
4898 int update_next_balance = 0;
4904 for_each_domain(cpu, sd) {
4905 if (!(sd->flags & SD_LOAD_BALANCE))
4908 interval = sd->balance_interval;
4909 if (idle != CPU_IDLE)
4910 interval *= sd->busy_factor;
4912 /* scale ms to jiffies */
4913 interval = msecs_to_jiffies(interval);
4914 interval = clamp(interval, 1UL, max_load_balance_interval);
4916 need_serialize = sd->flags & SD_SERIALIZE;
4918 if (need_serialize) {
4919 if (!spin_trylock(&balancing))
4923 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4924 if (load_balance(cpu, rq, sd, idle, &balance)) {
4926 * We've pulled tasks over so either we're no
4929 idle = CPU_NOT_IDLE;
4931 sd->last_balance = jiffies;
4934 spin_unlock(&balancing);
4936 if (time_after(next_balance, sd->last_balance + interval)) {
4937 next_balance = sd->last_balance + interval;
4938 update_next_balance = 1;
4942 * Stop the load balance at this level. There is another
4943 * CPU in our sched group which is doing load balancing more
4952 * next_balance will be updated only when there is a need.
4953 * When the cpu is attached to null domain for ex, it will not be
4956 if (likely(update_next_balance))
4957 rq->next_balance = next_balance;
4962 * In CONFIG_NO_HZ case, the idle balance kickee will do the
4963 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4965 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
4967 struct rq *this_rq = cpu_rq(this_cpu);
4971 if (idle != CPU_IDLE ||
4972 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
4975 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
4976 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
4980 * If this cpu gets work to do, stop the load balancing
4981 * work being done for other cpus. Next load
4982 * balancing owner will pick it up.
4987 rq = cpu_rq(balance_cpu);
4989 raw_spin_lock_irq(&rq->lock);
4990 update_rq_clock(rq);
4991 update_idle_cpu_load(rq);
4992 raw_spin_unlock_irq(&rq->lock);
4994 rebalance_domains(balance_cpu, CPU_IDLE);
4996 if (time_after(this_rq->next_balance, rq->next_balance))
4997 this_rq->next_balance = rq->next_balance;
4999 nohz.next_balance = this_rq->next_balance;
5001 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
5005 * Current heuristic for kicking the idle load balancer in the presence
5006 * of an idle cpu is the system.
5007 * - This rq has more than one task.
5008 * - At any scheduler domain level, this cpu's scheduler group has multiple
5009 * busy cpu's exceeding the group's power.
5010 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
5011 * domain span are idle.
5013 static inline int nohz_kick_needed(struct rq *rq, int cpu)
5015 unsigned long now = jiffies;
5016 struct sched_domain *sd;
5018 if (unlikely(idle_cpu(cpu)))
5022 * We may be recently in ticked or tickless idle mode. At the first
5023 * busy tick after returning from idle, we will update the busy stats.
5025 set_cpu_sd_state_busy();
5026 nohz_balance_exit_idle(cpu);
5029 * None are in tickless mode and hence no need for NOHZ idle load
5032 if (likely(!atomic_read(&nohz.nr_cpus)))
5035 if (time_before(now, nohz.next_balance))
5038 if (rq->nr_running >= 2)
5042 for_each_domain(cpu, sd) {
5043 struct sched_group *sg = sd->groups;
5044 struct sched_group_power *sgp = sg->sgp;
5045 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
5047 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
5048 goto need_kick_unlock;
5050 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
5051 && (cpumask_first_and(nohz.idle_cpus_mask,
5052 sched_domain_span(sd)) < cpu))
5053 goto need_kick_unlock;
5055 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
5067 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
5071 * run_rebalance_domains is triggered when needed from the scheduler tick.
5072 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
5074 static void run_rebalance_domains(struct softirq_action *h)
5076 int this_cpu = smp_processor_id();
5077 struct rq *this_rq = cpu_rq(this_cpu);
5078 enum cpu_idle_type idle = this_rq->idle_balance ?
5079 CPU_IDLE : CPU_NOT_IDLE;
5081 rebalance_domains(this_cpu, idle);
5084 * If this cpu has a pending nohz_balance_kick, then do the
5085 * balancing on behalf of the other idle cpus whose ticks are
5088 nohz_idle_balance(this_cpu, idle);
5091 static inline int on_null_domain(int cpu)
5093 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
5097 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
5099 void trigger_load_balance(struct rq *rq, int cpu)
5101 /* Don't need to rebalance while attached to NULL domain */
5102 if (time_after_eq(jiffies, rq->next_balance) &&
5103 likely(!on_null_domain(cpu)))
5104 raise_softirq(SCHED_SOFTIRQ);
5106 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
5107 nohz_balancer_kick(cpu);
5111 static void rq_online_fair(struct rq *rq)
5116 static void rq_offline_fair(struct rq *rq)
5120 /* Ensure any throttled groups are reachable by pick_next_task */
5121 unthrottle_offline_cfs_rqs(rq);
5124 #endif /* CONFIG_SMP */
5127 * scheduler tick hitting a task of our scheduling class:
5129 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
5131 struct cfs_rq *cfs_rq;
5132 struct sched_entity *se = &curr->se;
5134 for_each_sched_entity(se) {
5135 cfs_rq = cfs_rq_of(se);
5136 entity_tick(cfs_rq, se, queued);
5139 if (sched_feat_numa(NUMA))
5140 task_tick_numa(rq, curr);
5144 * called on fork with the child task as argument from the parent's context
5145 * - child not yet on the tasklist
5146 * - preemption disabled
5148 static void task_fork_fair(struct task_struct *p)
5150 struct cfs_rq *cfs_rq;
5151 struct sched_entity *se = &p->se, *curr;
5152 int this_cpu = smp_processor_id();
5153 struct rq *rq = this_rq();
5154 unsigned long flags;
5156 raw_spin_lock_irqsave(&rq->lock, flags);
5158 update_rq_clock(rq);
5160 cfs_rq = task_cfs_rq(current);
5161 curr = cfs_rq->curr;
5163 if (unlikely(task_cpu(p) != this_cpu)) {
5165 __set_task_cpu(p, this_cpu);
5169 update_curr(cfs_rq);
5172 se->vruntime = curr->vruntime;
5173 place_entity(cfs_rq, se, 1);
5175 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
5177 * Upon rescheduling, sched_class::put_prev_task() will place
5178 * 'current' within the tree based on its new key value.
5180 swap(curr->vruntime, se->vruntime);
5181 resched_task(rq->curr);
5184 se->vruntime -= cfs_rq->min_vruntime;
5186 raw_spin_unlock_irqrestore(&rq->lock, flags);
5190 * Priority of the task has changed. Check to see if we preempt
5194 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
5200 * Reschedule if we are currently running on this runqueue and
5201 * our priority decreased, or if we are not currently running on
5202 * this runqueue and our priority is higher than the current's
5204 if (rq->curr == p) {
5205 if (p->prio > oldprio)
5206 resched_task(rq->curr);
5208 check_preempt_curr(rq, p, 0);
5211 static void switched_from_fair(struct rq *rq, struct task_struct *p)
5213 struct sched_entity *se = &p->se;
5214 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5217 * Ensure the task's vruntime is normalized, so that when its
5218 * switched back to the fair class the enqueue_entity(.flags=0) will
5219 * do the right thing.
5221 * If it was on_rq, then the dequeue_entity(.flags=0) will already
5222 * have normalized the vruntime, if it was !on_rq, then only when
5223 * the task is sleeping will it still have non-normalized vruntime.
5225 if (!se->on_rq && p->state != TASK_RUNNING) {
5227 * Fix up our vruntime so that the current sleep doesn't
5228 * cause 'unlimited' sleep bonus.
5230 place_entity(cfs_rq, se, 0);
5231 se->vruntime -= cfs_rq->min_vruntime;
5236 * We switched to the sched_fair class.
5238 static void switched_to_fair(struct rq *rq, struct task_struct *p)
5244 * We were most likely switched from sched_rt, so
5245 * kick off the schedule if running, otherwise just see
5246 * if we can still preempt the current task.
5249 resched_task(rq->curr);
5251 check_preempt_curr(rq, p, 0);
5254 /* Account for a task changing its policy or group.
5256 * This routine is mostly called to set cfs_rq->curr field when a task
5257 * migrates between groups/classes.
5259 static void set_curr_task_fair(struct rq *rq)
5261 struct sched_entity *se = &rq->curr->se;
5263 for_each_sched_entity(se) {
5264 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5266 set_next_entity(cfs_rq, se);
5267 /* ensure bandwidth has been allocated on our new cfs_rq */
5268 account_cfs_rq_runtime(cfs_rq, 0);
5272 void init_cfs_rq(struct cfs_rq *cfs_rq)
5274 cfs_rq->tasks_timeline = RB_ROOT;
5275 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
5276 #ifndef CONFIG_64BIT
5277 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
5281 #ifdef CONFIG_FAIR_GROUP_SCHED
5282 static void task_move_group_fair(struct task_struct *p, int on_rq)
5285 * If the task was not on the rq at the time of this cgroup movement
5286 * it must have been asleep, sleeping tasks keep their ->vruntime
5287 * absolute on their old rq until wakeup (needed for the fair sleeper
5288 * bonus in place_entity()).
5290 * If it was on the rq, we've just 'preempted' it, which does convert
5291 * ->vruntime to a relative base.
5293 * Make sure both cases convert their relative position when migrating
5294 * to another cgroup's rq. This does somewhat interfere with the
5295 * fair sleeper stuff for the first placement, but who cares.
5298 * When !on_rq, vruntime of the task has usually NOT been normalized.
5299 * But there are some cases where it has already been normalized:
5301 * - Moving a forked child which is waiting for being woken up by
5302 * wake_up_new_task().
5303 * - Moving a task which has been woken up by try_to_wake_up() and
5304 * waiting for actually being woken up by sched_ttwu_pending().
5306 * To prevent boost or penalty in the new cfs_rq caused by delta
5307 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
5309 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
5313 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
5314 set_task_rq(p, task_cpu(p));
5316 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
5319 void free_fair_sched_group(struct task_group *tg)
5323 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
5325 for_each_possible_cpu(i) {
5327 kfree(tg->cfs_rq[i]);
5336 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
5338 struct cfs_rq *cfs_rq;
5339 struct sched_entity *se;
5342 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
5345 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
5349 tg->shares = NICE_0_LOAD;
5351 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
5353 for_each_possible_cpu(i) {
5354 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
5355 GFP_KERNEL, cpu_to_node(i));
5359 se = kzalloc_node(sizeof(struct sched_entity),
5360 GFP_KERNEL, cpu_to_node(i));
5364 init_cfs_rq(cfs_rq);
5365 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
5376 void unregister_fair_sched_group(struct task_group *tg, int cpu)
5378 struct rq *rq = cpu_rq(cpu);
5379 unsigned long flags;
5382 * Only empty task groups can be destroyed; so we can speculatively
5383 * check on_list without danger of it being re-added.
5385 if (!tg->cfs_rq[cpu]->on_list)
5388 raw_spin_lock_irqsave(&rq->lock, flags);
5389 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
5390 raw_spin_unlock_irqrestore(&rq->lock, flags);
5393 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
5394 struct sched_entity *se, int cpu,
5395 struct sched_entity *parent)
5397 struct rq *rq = cpu_rq(cpu);
5402 /* allow initial update_cfs_load() to truncate */
5403 cfs_rq->load_stamp = 1;
5405 init_cfs_rq_runtime(cfs_rq);
5407 tg->cfs_rq[cpu] = cfs_rq;
5410 /* se could be NULL for root_task_group */
5415 se->cfs_rq = &rq->cfs;
5417 se->cfs_rq = parent->my_q;
5420 update_load_set(&se->load, 0);
5421 se->parent = parent;
5424 static DEFINE_MUTEX(shares_mutex);
5426 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
5429 unsigned long flags;
5432 * We can't change the weight of the root cgroup.
5437 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
5439 mutex_lock(&shares_mutex);
5440 if (tg->shares == shares)
5443 tg->shares = shares;
5444 for_each_possible_cpu(i) {
5445 struct rq *rq = cpu_rq(i);
5446 struct sched_entity *se;
5449 /* Propagate contribution to hierarchy */
5450 raw_spin_lock_irqsave(&rq->lock, flags);
5451 for_each_sched_entity(se)
5452 update_cfs_shares(group_cfs_rq(se));
5453 raw_spin_unlock_irqrestore(&rq->lock, flags);
5457 mutex_unlock(&shares_mutex);
5460 #else /* CONFIG_FAIR_GROUP_SCHED */
5462 void free_fair_sched_group(struct task_group *tg) { }
5464 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
5469 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
5471 #endif /* CONFIG_FAIR_GROUP_SCHED */
5474 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
5476 struct sched_entity *se = &task->se;
5477 unsigned int rr_interval = 0;
5480 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
5483 if (rq->cfs.load.weight)
5484 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5490 * All the scheduling class methods:
5492 const struct sched_class fair_sched_class = {
5493 .next = &idle_sched_class,
5494 .enqueue_task = enqueue_task_fair,
5495 .dequeue_task = dequeue_task_fair,
5496 .yield_task = yield_task_fair,
5497 .yield_to_task = yield_to_task_fair,
5499 .check_preempt_curr = check_preempt_wakeup,
5501 .pick_next_task = pick_next_task_fair,
5502 .put_prev_task = put_prev_task_fair,
5505 .select_task_rq = select_task_rq_fair,
5507 .rq_online = rq_online_fair,
5508 .rq_offline = rq_offline_fair,
5510 .task_waking = task_waking_fair,
5513 .set_curr_task = set_curr_task_fair,
5514 .task_tick = task_tick_fair,
5515 .task_fork = task_fork_fair,
5517 .prio_changed = prio_changed_fair,
5518 .switched_from = switched_from_fair,
5519 .switched_to = switched_to_fair,
5521 .get_rr_interval = get_rr_interval_fair,
5523 #ifdef CONFIG_FAIR_GROUP_SCHED
5524 .task_move_group = task_move_group_fair,
5528 #ifdef CONFIG_SCHED_DEBUG
5529 void print_cfs_stats(struct seq_file *m, int cpu)
5531 struct cfs_rq *cfs_rq;
5534 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
5535 print_cfs_rq(m, cpu, cfs_rq);
5540 __init void init_sched_fair_class(void)
5543 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
5546 nohz.next_balance = jiffies;
5547 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
5548 cpu_notifier(sched_ilb_notifier, 0);