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 void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
268 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
270 if (!cfs_rq->on_list) {
272 * Ensure we either appear before our parent (if already
273 * enqueued) or force our parent to appear after us when it is
274 * enqueued. The fact that we always enqueue bottom-up
275 * reduces this to two cases.
277 if (cfs_rq->tg->parent &&
278 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
279 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
280 &rq_of(cfs_rq)->leaf_cfs_rq_list);
282 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
283 &rq_of(cfs_rq)->leaf_cfs_rq_list);
287 /* We should have no load, but we need to update last_decay. */
288 update_cfs_rq_blocked_load(cfs_rq, 0);
292 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
294 if (cfs_rq->on_list) {
295 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
300 /* Iterate thr' all leaf cfs_rq's on a runqueue */
301 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
302 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
304 /* Do the two (enqueued) entities belong to the same group ? */
306 is_same_group(struct sched_entity *se, struct sched_entity *pse)
308 if (se->cfs_rq == pse->cfs_rq)
314 static inline struct sched_entity *parent_entity(struct sched_entity *se)
319 /* return depth at which a sched entity is present in the hierarchy */
320 static inline int depth_se(struct sched_entity *se)
324 for_each_sched_entity(se)
331 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
333 int se_depth, pse_depth;
336 * preemption test can be made between sibling entities who are in the
337 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
338 * both tasks until we find their ancestors who are siblings of common
342 /* First walk up until both entities are at same depth */
343 se_depth = depth_se(*se);
344 pse_depth = depth_se(*pse);
346 while (se_depth > pse_depth) {
348 *se = parent_entity(*se);
351 while (pse_depth > se_depth) {
353 *pse = parent_entity(*pse);
356 while (!is_same_group(*se, *pse)) {
357 *se = parent_entity(*se);
358 *pse = parent_entity(*pse);
362 #else /* !CONFIG_FAIR_GROUP_SCHED */
364 static inline struct task_struct *task_of(struct sched_entity *se)
366 return container_of(se, struct task_struct, se);
369 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
371 return container_of(cfs_rq, struct rq, cfs);
374 #define entity_is_task(se) 1
376 #define for_each_sched_entity(se) \
377 for (; se; se = NULL)
379 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
381 return &task_rq(p)->cfs;
384 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
386 struct task_struct *p = task_of(se);
387 struct rq *rq = task_rq(p);
392 /* runqueue "owned" by this group */
393 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
398 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
402 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
406 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
407 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
410 is_same_group(struct sched_entity *se, struct sched_entity *pse)
415 static inline struct sched_entity *parent_entity(struct sched_entity *se)
421 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
425 #endif /* CONFIG_FAIR_GROUP_SCHED */
427 static __always_inline
428 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec);
430 /**************************************************************
431 * Scheduling class tree data structure manipulation methods:
434 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
436 s64 delta = (s64)(vruntime - min_vruntime);
438 min_vruntime = vruntime;
443 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
445 s64 delta = (s64)(vruntime - min_vruntime);
447 min_vruntime = vruntime;
452 static inline int entity_before(struct sched_entity *a,
453 struct sched_entity *b)
455 return (s64)(a->vruntime - b->vruntime) < 0;
458 static void update_min_vruntime(struct cfs_rq *cfs_rq)
460 u64 vruntime = cfs_rq->min_vruntime;
463 vruntime = cfs_rq->curr->vruntime;
465 if (cfs_rq->rb_leftmost) {
466 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
471 vruntime = se->vruntime;
473 vruntime = min_vruntime(vruntime, se->vruntime);
476 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
479 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
484 * Enqueue an entity into the rb-tree:
486 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
488 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
489 struct rb_node *parent = NULL;
490 struct sched_entity *entry;
494 * Find the right place in the rbtree:
498 entry = rb_entry(parent, struct sched_entity, run_node);
500 * We dont care about collisions. Nodes with
501 * the same key stay together.
503 if (entity_before(se, entry)) {
504 link = &parent->rb_left;
506 link = &parent->rb_right;
512 * Maintain a cache of leftmost tree entries (it is frequently
516 cfs_rq->rb_leftmost = &se->run_node;
518 rb_link_node(&se->run_node, parent, link);
519 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
522 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
524 if (cfs_rq->rb_leftmost == &se->run_node) {
525 struct rb_node *next_node;
527 next_node = rb_next(&se->run_node);
528 cfs_rq->rb_leftmost = next_node;
531 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
534 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
536 struct rb_node *left = cfs_rq->rb_leftmost;
541 return rb_entry(left, struct sched_entity, run_node);
544 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
546 struct rb_node *next = rb_next(&se->run_node);
551 return rb_entry(next, struct sched_entity, run_node);
554 #ifdef CONFIG_SCHED_DEBUG
555 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
557 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
562 return rb_entry(last, struct sched_entity, run_node);
565 /**************************************************************
566 * Scheduling class statistics methods:
569 int sched_proc_update_handler(struct ctl_table *table, int write,
570 void __user *buffer, size_t *lenp,
573 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
574 int factor = get_update_sysctl_factor();
579 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
580 sysctl_sched_min_granularity);
582 #define WRT_SYSCTL(name) \
583 (normalized_sysctl_##name = sysctl_##name / (factor))
584 WRT_SYSCTL(sched_min_granularity);
585 WRT_SYSCTL(sched_latency);
586 WRT_SYSCTL(sched_wakeup_granularity);
596 static inline unsigned long
597 calc_delta_fair(unsigned long delta, struct sched_entity *se)
599 if (unlikely(se->load.weight != NICE_0_LOAD))
600 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
606 * The idea is to set a period in which each task runs once.
608 * When there are too many tasks (sched_nr_latency) we have to stretch
609 * this period because otherwise the slices get too small.
611 * p = (nr <= nl) ? l : l*nr/nl
613 static u64 __sched_period(unsigned long nr_running)
615 u64 period = sysctl_sched_latency;
616 unsigned long nr_latency = sched_nr_latency;
618 if (unlikely(nr_running > nr_latency)) {
619 period = sysctl_sched_min_granularity;
620 period *= nr_running;
627 * We calculate the wall-time slice from the period by taking a part
628 * proportional to the weight.
632 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
634 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
636 for_each_sched_entity(se) {
637 struct load_weight *load;
638 struct load_weight lw;
640 cfs_rq = cfs_rq_of(se);
641 load = &cfs_rq->load;
643 if (unlikely(!se->on_rq)) {
646 update_load_add(&lw, se->load.weight);
649 slice = calc_delta_mine(slice, se->load.weight, load);
655 * We calculate the vruntime slice of a to be inserted task
659 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
661 return calc_delta_fair(sched_slice(cfs_rq, se), se);
665 * Update the current task's runtime statistics. Skip current tasks that
666 * are not in our scheduling class.
669 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
670 unsigned long delta_exec)
672 unsigned long delta_exec_weighted;
674 schedstat_set(curr->statistics.exec_max,
675 max((u64)delta_exec, curr->statistics.exec_max));
677 curr->sum_exec_runtime += delta_exec;
678 schedstat_add(cfs_rq, exec_clock, delta_exec);
679 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
681 curr->vruntime += delta_exec_weighted;
682 update_min_vruntime(cfs_rq);
685 static void update_curr(struct cfs_rq *cfs_rq)
687 struct sched_entity *curr = cfs_rq->curr;
688 u64 now = rq_of(cfs_rq)->clock_task;
689 unsigned long delta_exec;
695 * Get the amount of time the current task was running
696 * since the last time we changed load (this cannot
697 * overflow on 32 bits):
699 delta_exec = (unsigned long)(now - curr->exec_start);
703 __update_curr(cfs_rq, curr, delta_exec);
704 curr->exec_start = now;
706 if (entity_is_task(curr)) {
707 struct task_struct *curtask = task_of(curr);
709 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
710 cpuacct_charge(curtask, delta_exec);
711 account_group_exec_runtime(curtask, delta_exec);
714 account_cfs_rq_runtime(cfs_rq, delta_exec);
718 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
720 schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock);
724 * Task is being enqueued - update stats:
726 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
729 * Are we enqueueing a waiting task? (for current tasks
730 * a dequeue/enqueue event is a NOP)
732 if (se != cfs_rq->curr)
733 update_stats_wait_start(cfs_rq, se);
737 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
739 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
740 rq_of(cfs_rq)->clock - se->statistics.wait_start));
741 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
742 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
743 rq_of(cfs_rq)->clock - se->statistics.wait_start);
744 #ifdef CONFIG_SCHEDSTATS
745 if (entity_is_task(se)) {
746 trace_sched_stat_wait(task_of(se),
747 rq_of(cfs_rq)->clock - se->statistics.wait_start);
750 schedstat_set(se->statistics.wait_start, 0);
754 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
757 * Mark the end of the wait period if dequeueing a
760 if (se != cfs_rq->curr)
761 update_stats_wait_end(cfs_rq, se);
765 * We are picking a new current task - update its stats:
768 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
771 * We are starting a new run period:
773 se->exec_start = rq_of(cfs_rq)->clock_task;
776 /**************************************************
777 * Scheduling class queueing methods:
780 #ifdef CONFIG_NUMA_BALANCING
782 * numa task sample period in ms
784 unsigned int sysctl_numa_balancing_scan_period_min = 100;
785 unsigned int sysctl_numa_balancing_scan_period_max = 100*50;
786 unsigned int sysctl_numa_balancing_scan_period_reset = 100*600;
788 /* Portion of address space to scan in MB */
789 unsigned int sysctl_numa_balancing_scan_size = 256;
791 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
792 unsigned int sysctl_numa_balancing_scan_delay = 1000;
794 static void task_numa_placement(struct task_struct *p)
796 int seq = ACCESS_ONCE(p->mm->numa_scan_seq);
798 if (p->numa_scan_seq == seq)
800 p->numa_scan_seq = seq;
802 /* FIXME: Scheduling placement policy hints go here */
806 * Got a PROT_NONE fault for a page on @node.
808 void task_numa_fault(int node, int pages, bool migrated)
810 struct task_struct *p = current;
812 if (!sched_feat_numa(NUMA))
815 /* FIXME: Allocate task-specific structure for placement policy here */
818 * If pages are properly placed (did not migrate) then scan slower.
819 * This is reset periodically in case of phase changes
822 p->numa_scan_period = min(sysctl_numa_balancing_scan_period_max,
823 p->numa_scan_period + jiffies_to_msecs(10));
825 task_numa_placement(p);
828 static void reset_ptenuma_scan(struct task_struct *p)
830 ACCESS_ONCE(p->mm->numa_scan_seq)++;
831 p->mm->numa_scan_offset = 0;
835 * The expensive part of numa migration is done from task_work context.
836 * Triggered from task_tick_numa().
838 void task_numa_work(struct callback_head *work)
840 unsigned long migrate, next_scan, now = jiffies;
841 struct task_struct *p = current;
842 struct mm_struct *mm = p->mm;
843 struct vm_area_struct *vma;
844 unsigned long start, end;
847 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
849 work->next = work; /* protect against double add */
851 * Who cares about NUMA placement when they're dying.
853 * NOTE: make sure not to dereference p->mm before this check,
854 * exit_task_work() happens _after_ exit_mm() so we could be called
855 * without p->mm even though we still had it when we enqueued this
858 if (p->flags & PF_EXITING)
862 * We do not care about task placement until a task runs on a node
863 * other than the first one used by the address space. This is
864 * largely because migrations are driven by what CPU the task
865 * is running on. If it's never scheduled on another node, it'll
866 * not migrate so why bother trapping the fault.
868 if (mm->first_nid == NUMA_PTE_SCAN_INIT)
869 mm->first_nid = numa_node_id();
870 if (mm->first_nid != NUMA_PTE_SCAN_ACTIVE) {
871 /* Are we running on a new node yet? */
872 if (numa_node_id() == mm->first_nid &&
873 !sched_feat_numa(NUMA_FORCE))
876 mm->first_nid = NUMA_PTE_SCAN_ACTIVE;
880 * Reset the scan period if enough time has gone by. Objective is that
881 * scanning will be reduced if pages are properly placed. As tasks
882 * can enter different phases this needs to be re-examined. Lacking
883 * proper tracking of reference behaviour, this blunt hammer is used.
885 migrate = mm->numa_next_reset;
886 if (time_after(now, migrate)) {
887 p->numa_scan_period = sysctl_numa_balancing_scan_period_min;
888 next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
889 xchg(&mm->numa_next_reset, next_scan);
893 * Enforce maximal scan/migration frequency..
895 migrate = mm->numa_next_scan;
896 if (time_before(now, migrate))
899 if (p->numa_scan_period == 0)
900 p->numa_scan_period = sysctl_numa_balancing_scan_period_min;
902 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
903 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
907 * Do not set pte_numa if the current running node is rate-limited.
908 * This loses statistics on the fault but if we are unwilling to
909 * migrate to this node, it is less likely we can do useful work
911 if (migrate_ratelimited(numa_node_id()))
914 start = mm->numa_scan_offset;
915 pages = sysctl_numa_balancing_scan_size;
916 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
920 down_read(&mm->mmap_sem);
921 vma = find_vma(mm, start);
923 reset_ptenuma_scan(p);
927 for (; vma; vma = vma->vm_next) {
928 if (!vma_migratable(vma))
931 /* Skip small VMAs. They are not likely to be of relevance */
932 if (vma->vm_end - vma->vm_start < HPAGE_SIZE)
936 start = max(start, vma->vm_start);
937 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
938 end = min(end, vma->vm_end);
939 pages -= change_prot_numa(vma, start, end);
944 } while (end != vma->vm_end);
949 * It is possible to reach the end of the VMA list but the last few VMAs are
950 * not guaranteed to the vma_migratable. If they are not, we would find the
951 * !migratable VMA on the next scan but not reset the scanner to the start
955 mm->numa_scan_offset = start;
957 reset_ptenuma_scan(p);
958 up_read(&mm->mmap_sem);
962 * Drive the periodic memory faults..
964 void task_tick_numa(struct rq *rq, struct task_struct *curr)
966 struct callback_head *work = &curr->numa_work;
970 * We don't care about NUMA placement if we don't have memory.
972 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
976 * Using runtime rather than walltime has the dual advantage that
977 * we (mostly) drive the selection from busy threads and that the
978 * task needs to have done some actual work before we bother with
981 now = curr->se.sum_exec_runtime;
982 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
984 if (now - curr->node_stamp > period) {
985 if (!curr->node_stamp)
986 curr->numa_scan_period = sysctl_numa_balancing_scan_period_min;
987 curr->node_stamp = now;
989 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
990 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
991 task_work_add(curr, work, true);
996 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
999 #endif /* CONFIG_NUMA_BALANCING */
1002 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1004 update_load_add(&cfs_rq->load, se->load.weight);
1005 if (!parent_entity(se))
1006 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1008 if (entity_is_task(se))
1009 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1011 cfs_rq->nr_running++;
1015 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1017 update_load_sub(&cfs_rq->load, se->load.weight);
1018 if (!parent_entity(se))
1019 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1020 if (entity_is_task(se))
1021 list_del_init(&se->group_node);
1022 cfs_rq->nr_running--;
1025 #ifdef CONFIG_FAIR_GROUP_SCHED
1027 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1032 * Use this CPU's actual weight instead of the last load_contribution
1033 * to gain a more accurate current total weight. See
1034 * update_cfs_rq_load_contribution().
1036 tg_weight = atomic64_read(&tg->load_avg);
1037 tg_weight -= cfs_rq->tg_load_contrib;
1038 tg_weight += cfs_rq->load.weight;
1043 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1045 long tg_weight, load, shares;
1047 tg_weight = calc_tg_weight(tg, cfs_rq);
1048 load = cfs_rq->load.weight;
1050 shares = (tg->shares * load);
1052 shares /= tg_weight;
1054 if (shares < MIN_SHARES)
1055 shares = MIN_SHARES;
1056 if (shares > tg->shares)
1057 shares = tg->shares;
1061 # else /* CONFIG_SMP */
1062 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1066 # endif /* CONFIG_SMP */
1067 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1068 unsigned long weight)
1071 /* commit outstanding execution time */
1072 if (cfs_rq->curr == se)
1073 update_curr(cfs_rq);
1074 account_entity_dequeue(cfs_rq, se);
1077 update_load_set(&se->load, weight);
1080 account_entity_enqueue(cfs_rq, se);
1083 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1085 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1087 struct task_group *tg;
1088 struct sched_entity *se;
1092 se = tg->se[cpu_of(rq_of(cfs_rq))];
1093 if (!se || throttled_hierarchy(cfs_rq))
1096 if (likely(se->load.weight == tg->shares))
1099 shares = calc_cfs_shares(cfs_rq, tg);
1101 reweight_entity(cfs_rq_of(se), se, shares);
1103 #else /* CONFIG_FAIR_GROUP_SCHED */
1104 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1107 #endif /* CONFIG_FAIR_GROUP_SCHED */
1109 /* Only depends on SMP, FAIR_GROUP_SCHED may be removed when useful in lb */
1110 #if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)
1112 * We choose a half-life close to 1 scheduling period.
1113 * Note: The tables below are dependent on this value.
1115 #define LOAD_AVG_PERIOD 32
1116 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1117 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1119 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1120 static const u32 runnable_avg_yN_inv[] = {
1121 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1122 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1123 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1124 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1125 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1126 0x85aac367, 0x82cd8698,
1130 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1131 * over-estimates when re-combining.
1133 static const u32 runnable_avg_yN_sum[] = {
1134 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1135 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1136 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1141 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1143 static __always_inline u64 decay_load(u64 val, u64 n)
1145 unsigned int local_n;
1149 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1152 /* after bounds checking we can collapse to 32-bit */
1156 * As y^PERIOD = 1/2, we can combine
1157 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1158 * With a look-up table which covers k^n (n<PERIOD)
1160 * To achieve constant time decay_load.
1162 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1163 val >>= local_n / LOAD_AVG_PERIOD;
1164 local_n %= LOAD_AVG_PERIOD;
1167 val *= runnable_avg_yN_inv[local_n];
1168 /* We don't use SRR here since we always want to round down. */
1173 * For updates fully spanning n periods, the contribution to runnable
1174 * average will be: \Sum 1024*y^n
1176 * We can compute this reasonably efficiently by combining:
1177 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1179 static u32 __compute_runnable_contrib(u64 n)
1183 if (likely(n <= LOAD_AVG_PERIOD))
1184 return runnable_avg_yN_sum[n];
1185 else if (unlikely(n >= LOAD_AVG_MAX_N))
1186 return LOAD_AVG_MAX;
1188 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1190 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1191 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1193 n -= LOAD_AVG_PERIOD;
1194 } while (n > LOAD_AVG_PERIOD);
1196 contrib = decay_load(contrib, n);
1197 return contrib + runnable_avg_yN_sum[n];
1201 * We can represent the historical contribution to runnable average as the
1202 * coefficients of a geometric series. To do this we sub-divide our runnable
1203 * history into segments of approximately 1ms (1024us); label the segment that
1204 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1206 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1208 * (now) (~1ms ago) (~2ms ago)
1210 * Let u_i denote the fraction of p_i that the entity was runnable.
1212 * We then designate the fractions u_i as our co-efficients, yielding the
1213 * following representation of historical load:
1214 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1216 * We choose y based on the with of a reasonably scheduling period, fixing:
1219 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1220 * approximately half as much as the contribution to load within the last ms
1223 * When a period "rolls over" and we have new u_0`, multiplying the previous
1224 * sum again by y is sufficient to update:
1225 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1226 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1228 static __always_inline int __update_entity_runnable_avg(u64 now,
1229 struct sched_avg *sa,
1233 u32 runnable_contrib;
1234 int delta_w, decayed = 0;
1236 delta = now - sa->last_runnable_update;
1238 * This should only happen when time goes backwards, which it
1239 * unfortunately does during sched clock init when we swap over to TSC.
1241 if ((s64)delta < 0) {
1242 sa->last_runnable_update = now;
1247 * Use 1024ns as the unit of measurement since it's a reasonable
1248 * approximation of 1us and fast to compute.
1253 sa->last_runnable_update = now;
1255 /* delta_w is the amount already accumulated against our next period */
1256 delta_w = sa->runnable_avg_period % 1024;
1257 if (delta + delta_w >= 1024) {
1258 /* period roll-over */
1262 * Now that we know we're crossing a period boundary, figure
1263 * out how much from delta we need to complete the current
1264 * period and accrue it.
1266 delta_w = 1024 - delta_w;
1268 sa->runnable_avg_sum += delta_w;
1269 sa->runnable_avg_period += delta_w;
1273 /* Figure out how many additional periods this update spans */
1274 periods = delta / 1024;
1277 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1279 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1282 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1283 runnable_contrib = __compute_runnable_contrib(periods);
1285 sa->runnable_avg_sum += runnable_contrib;
1286 sa->runnable_avg_period += runnable_contrib;
1289 /* Remainder of delta accrued against u_0` */
1291 sa->runnable_avg_sum += delta;
1292 sa->runnable_avg_period += delta;
1297 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1298 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1300 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1301 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1303 decays -= se->avg.decay_count;
1307 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1308 se->avg.decay_count = 0;
1313 #ifdef CONFIG_FAIR_GROUP_SCHED
1314 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1317 struct task_group *tg = cfs_rq->tg;
1320 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1321 tg_contrib -= cfs_rq->tg_load_contrib;
1323 if (force_update || abs64(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1324 atomic64_add(tg_contrib, &tg->load_avg);
1325 cfs_rq->tg_load_contrib += tg_contrib;
1330 * Aggregate cfs_rq runnable averages into an equivalent task_group
1331 * representation for computing load contributions.
1333 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1334 struct cfs_rq *cfs_rq)
1336 struct task_group *tg = cfs_rq->tg;
1339 /* The fraction of a cpu used by this cfs_rq */
1340 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1341 sa->runnable_avg_period + 1);
1342 contrib -= cfs_rq->tg_runnable_contrib;
1344 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
1345 atomic_add(contrib, &tg->runnable_avg);
1346 cfs_rq->tg_runnable_contrib += contrib;
1350 static inline void __update_group_entity_contrib(struct sched_entity *se)
1352 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1353 struct task_group *tg = cfs_rq->tg;
1358 contrib = cfs_rq->tg_load_contrib * tg->shares;
1359 se->avg.load_avg_contrib = div64_u64(contrib,
1360 atomic64_read(&tg->load_avg) + 1);
1363 * For group entities we need to compute a correction term in the case
1364 * that they are consuming <1 cpu so that we would contribute the same
1365 * load as a task of equal weight.
1367 * Explicitly co-ordinating this measurement would be expensive, but
1368 * fortunately the sum of each cpus contribution forms a usable
1369 * lower-bound on the true value.
1371 * Consider the aggregate of 2 contributions. Either they are disjoint
1372 * (and the sum represents true value) or they are disjoint and we are
1373 * understating by the aggregate of their overlap.
1375 * Extending this to N cpus, for a given overlap, the maximum amount we
1376 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1377 * cpus that overlap for this interval and w_i is the interval width.
1379 * On a small machine; the first term is well-bounded which bounds the
1380 * total error since w_i is a subset of the period. Whereas on a
1381 * larger machine, while this first term can be larger, if w_i is the
1382 * of consequential size guaranteed to see n_i*w_i quickly converge to
1383 * our upper bound of 1-cpu.
1385 runnable_avg = atomic_read(&tg->runnable_avg);
1386 if (runnable_avg < NICE_0_LOAD) {
1387 se->avg.load_avg_contrib *= runnable_avg;
1388 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1392 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1393 int force_update) {}
1394 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1395 struct cfs_rq *cfs_rq) {}
1396 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1399 static inline void __update_task_entity_contrib(struct sched_entity *se)
1403 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1404 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1405 contrib /= (se->avg.runnable_avg_period + 1);
1406 se->avg.load_avg_contrib = scale_load(contrib);
1409 /* Compute the current contribution to load_avg by se, return any delta */
1410 static long __update_entity_load_avg_contrib(struct sched_entity *se)
1412 long old_contrib = se->avg.load_avg_contrib;
1414 if (entity_is_task(se)) {
1415 __update_task_entity_contrib(se);
1417 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1418 __update_group_entity_contrib(se);
1421 return se->avg.load_avg_contrib - old_contrib;
1424 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1427 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1428 cfs_rq->blocked_load_avg -= load_contrib;
1430 cfs_rq->blocked_load_avg = 0;
1433 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1435 /* Update a sched_entity's runnable average */
1436 static inline void update_entity_load_avg(struct sched_entity *se,
1439 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1444 * For a group entity we need to use their owned cfs_rq_clock_task() in
1445 * case they are the parent of a throttled hierarchy.
1447 if (entity_is_task(se))
1448 now = cfs_rq_clock_task(cfs_rq);
1450 now = cfs_rq_clock_task(group_cfs_rq(se));
1452 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
1455 contrib_delta = __update_entity_load_avg_contrib(se);
1461 cfs_rq->runnable_load_avg += contrib_delta;
1463 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1467 * Decay the load contributed by all blocked children and account this so that
1468 * their contribution may appropriately discounted when they wake up.
1470 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1472 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1475 decays = now - cfs_rq->last_decay;
1476 if (!decays && !force_update)
1479 if (atomic64_read(&cfs_rq->removed_load)) {
1480 u64 removed_load = atomic64_xchg(&cfs_rq->removed_load, 0);
1481 subtract_blocked_load_contrib(cfs_rq, removed_load);
1485 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1487 atomic64_add(decays, &cfs_rq->decay_counter);
1488 cfs_rq->last_decay = now;
1491 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1494 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1496 __update_entity_runnable_avg(rq->clock_task, &rq->avg, runnable);
1497 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1500 /* Add the load generated by se into cfs_rq's child load-average */
1501 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1502 struct sched_entity *se,
1506 * We track migrations using entity decay_count <= 0, on a wake-up
1507 * migration we use a negative decay count to track the remote decays
1508 * accumulated while sleeping.
1510 if (unlikely(se->avg.decay_count <= 0)) {
1511 se->avg.last_runnable_update = rq_of(cfs_rq)->clock_task;
1512 if (se->avg.decay_count) {
1514 * In a wake-up migration we have to approximate the
1515 * time sleeping. This is because we can't synchronize
1516 * clock_task between the two cpus, and it is not
1517 * guaranteed to be read-safe. Instead, we can
1518 * approximate this using our carried decays, which are
1519 * explicitly atomically readable.
1521 se->avg.last_runnable_update -= (-se->avg.decay_count)
1523 update_entity_load_avg(se, 0);
1524 /* Indicate that we're now synchronized and on-rq */
1525 se->avg.decay_count = 0;
1529 __synchronize_entity_decay(se);
1532 /* migrated tasks did not contribute to our blocked load */
1534 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1535 update_entity_load_avg(se, 0);
1538 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1539 /* we force update consideration on load-balancer moves */
1540 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1544 * Remove se's load from this cfs_rq child load-average, if the entity is
1545 * transitioning to a blocked state we track its projected decay using
1548 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1549 struct sched_entity *se,
1552 update_entity_load_avg(se, 1);
1553 /* we force update consideration on load-balancer moves */
1554 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1556 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1558 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1559 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1560 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1563 static inline void update_entity_load_avg(struct sched_entity *se,
1564 int update_cfs_rq) {}
1565 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1566 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1567 struct sched_entity *se,
1569 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1570 struct sched_entity *se,
1572 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1573 int force_update) {}
1576 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1578 #ifdef CONFIG_SCHEDSTATS
1579 struct task_struct *tsk = NULL;
1581 if (entity_is_task(se))
1584 if (se->statistics.sleep_start) {
1585 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
1590 if (unlikely(delta > se->statistics.sleep_max))
1591 se->statistics.sleep_max = delta;
1593 se->statistics.sleep_start = 0;
1594 se->statistics.sum_sleep_runtime += delta;
1597 account_scheduler_latency(tsk, delta >> 10, 1);
1598 trace_sched_stat_sleep(tsk, delta);
1601 if (se->statistics.block_start) {
1602 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
1607 if (unlikely(delta > se->statistics.block_max))
1608 se->statistics.block_max = delta;
1610 se->statistics.block_start = 0;
1611 se->statistics.sum_sleep_runtime += delta;
1614 if (tsk->in_iowait) {
1615 se->statistics.iowait_sum += delta;
1616 se->statistics.iowait_count++;
1617 trace_sched_stat_iowait(tsk, delta);
1620 trace_sched_stat_blocked(tsk, delta);
1623 * Blocking time is in units of nanosecs, so shift by
1624 * 20 to get a milliseconds-range estimation of the
1625 * amount of time that the task spent sleeping:
1627 if (unlikely(prof_on == SLEEP_PROFILING)) {
1628 profile_hits(SLEEP_PROFILING,
1629 (void *)get_wchan(tsk),
1632 account_scheduler_latency(tsk, delta >> 10, 0);
1638 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1640 #ifdef CONFIG_SCHED_DEBUG
1641 s64 d = se->vruntime - cfs_rq->min_vruntime;
1646 if (d > 3*sysctl_sched_latency)
1647 schedstat_inc(cfs_rq, nr_spread_over);
1652 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1654 u64 vruntime = cfs_rq->min_vruntime;
1657 * The 'current' period is already promised to the current tasks,
1658 * however the extra weight of the new task will slow them down a
1659 * little, place the new task so that it fits in the slot that
1660 * stays open at the end.
1662 if (initial && sched_feat(START_DEBIT))
1663 vruntime += sched_vslice(cfs_rq, se);
1665 /* sleeps up to a single latency don't count. */
1667 unsigned long thresh = sysctl_sched_latency;
1670 * Halve their sleep time's effect, to allow
1671 * for a gentler effect of sleepers:
1673 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1679 /* ensure we never gain time by being placed backwards. */
1680 vruntime = max_vruntime(se->vruntime, vruntime);
1682 se->vruntime = vruntime;
1685 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1688 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1691 * Update the normalized vruntime before updating min_vruntime
1692 * through callig update_curr().
1694 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1695 se->vruntime += cfs_rq->min_vruntime;
1698 * Update run-time statistics of the 'current'.
1700 update_curr(cfs_rq);
1701 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
1702 account_entity_enqueue(cfs_rq, se);
1703 update_cfs_shares(cfs_rq);
1705 if (flags & ENQUEUE_WAKEUP) {
1706 place_entity(cfs_rq, se, 0);
1707 enqueue_sleeper(cfs_rq, se);
1710 update_stats_enqueue(cfs_rq, se);
1711 check_spread(cfs_rq, se);
1712 if (se != cfs_rq->curr)
1713 __enqueue_entity(cfs_rq, se);
1716 if (cfs_rq->nr_running == 1) {
1717 list_add_leaf_cfs_rq(cfs_rq);
1718 check_enqueue_throttle(cfs_rq);
1722 static void __clear_buddies_last(struct sched_entity *se)
1724 for_each_sched_entity(se) {
1725 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1726 if (cfs_rq->last == se)
1727 cfs_rq->last = NULL;
1733 static void __clear_buddies_next(struct sched_entity *se)
1735 for_each_sched_entity(se) {
1736 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1737 if (cfs_rq->next == se)
1738 cfs_rq->next = NULL;
1744 static void __clear_buddies_skip(struct sched_entity *se)
1746 for_each_sched_entity(se) {
1747 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1748 if (cfs_rq->skip == se)
1749 cfs_rq->skip = NULL;
1755 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1757 if (cfs_rq->last == se)
1758 __clear_buddies_last(se);
1760 if (cfs_rq->next == se)
1761 __clear_buddies_next(se);
1763 if (cfs_rq->skip == se)
1764 __clear_buddies_skip(se);
1767 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1770 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1773 * Update run-time statistics of the 'current'.
1775 update_curr(cfs_rq);
1776 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
1778 update_stats_dequeue(cfs_rq, se);
1779 if (flags & DEQUEUE_SLEEP) {
1780 #ifdef CONFIG_SCHEDSTATS
1781 if (entity_is_task(se)) {
1782 struct task_struct *tsk = task_of(se);
1784 if (tsk->state & TASK_INTERRUPTIBLE)
1785 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1786 if (tsk->state & TASK_UNINTERRUPTIBLE)
1787 se->statistics.block_start = rq_of(cfs_rq)->clock;
1792 clear_buddies(cfs_rq, se);
1794 if (se != cfs_rq->curr)
1795 __dequeue_entity(cfs_rq, se);
1797 account_entity_dequeue(cfs_rq, se);
1800 * Normalize the entity after updating the min_vruntime because the
1801 * update can refer to the ->curr item and we need to reflect this
1802 * movement in our normalized position.
1804 if (!(flags & DEQUEUE_SLEEP))
1805 se->vruntime -= cfs_rq->min_vruntime;
1807 /* return excess runtime on last dequeue */
1808 return_cfs_rq_runtime(cfs_rq);
1810 update_min_vruntime(cfs_rq);
1811 update_cfs_shares(cfs_rq);
1815 * Preempt the current task with a newly woken task if needed:
1818 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1820 unsigned long ideal_runtime, delta_exec;
1821 struct sched_entity *se;
1824 ideal_runtime = sched_slice(cfs_rq, curr);
1825 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1826 if (delta_exec > ideal_runtime) {
1827 resched_task(rq_of(cfs_rq)->curr);
1829 * The current task ran long enough, ensure it doesn't get
1830 * re-elected due to buddy favours.
1832 clear_buddies(cfs_rq, curr);
1837 * Ensure that a task that missed wakeup preemption by a
1838 * narrow margin doesn't have to wait for a full slice.
1839 * This also mitigates buddy induced latencies under load.
1841 if (delta_exec < sysctl_sched_min_granularity)
1844 se = __pick_first_entity(cfs_rq);
1845 delta = curr->vruntime - se->vruntime;
1850 if (delta > ideal_runtime)
1851 resched_task(rq_of(cfs_rq)->curr);
1855 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1857 /* 'current' is not kept within the tree. */
1860 * Any task has to be enqueued before it get to execute on
1861 * a CPU. So account for the time it spent waiting on the
1864 update_stats_wait_end(cfs_rq, se);
1865 __dequeue_entity(cfs_rq, se);
1868 update_stats_curr_start(cfs_rq, se);
1870 #ifdef CONFIG_SCHEDSTATS
1872 * Track our maximum slice length, if the CPU's load is at
1873 * least twice that of our own weight (i.e. dont track it
1874 * when there are only lesser-weight tasks around):
1876 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1877 se->statistics.slice_max = max(se->statistics.slice_max,
1878 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1881 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1885 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1888 * Pick the next process, keeping these things in mind, in this order:
1889 * 1) keep things fair between processes/task groups
1890 * 2) pick the "next" process, since someone really wants that to run
1891 * 3) pick the "last" process, for cache locality
1892 * 4) do not run the "skip" process, if something else is available
1894 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1896 struct sched_entity *se = __pick_first_entity(cfs_rq);
1897 struct sched_entity *left = se;
1900 * Avoid running the skip buddy, if running something else can
1901 * be done without getting too unfair.
1903 if (cfs_rq->skip == se) {
1904 struct sched_entity *second = __pick_next_entity(se);
1905 if (second && wakeup_preempt_entity(second, left) < 1)
1910 * Prefer last buddy, try to return the CPU to a preempted task.
1912 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1916 * Someone really wants this to run. If it's not unfair, run it.
1918 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1921 clear_buddies(cfs_rq, se);
1926 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1928 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1931 * If still on the runqueue then deactivate_task()
1932 * was not called and update_curr() has to be done:
1935 update_curr(cfs_rq);
1937 /* throttle cfs_rqs exceeding runtime */
1938 check_cfs_rq_runtime(cfs_rq);
1940 check_spread(cfs_rq, prev);
1942 update_stats_wait_start(cfs_rq, prev);
1943 /* Put 'current' back into the tree. */
1944 __enqueue_entity(cfs_rq, prev);
1945 /* in !on_rq case, update occurred at dequeue */
1946 update_entity_load_avg(prev, 1);
1948 cfs_rq->curr = NULL;
1952 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1955 * Update run-time statistics of the 'current'.
1957 update_curr(cfs_rq);
1960 * Ensure that runnable average is periodically updated.
1962 update_entity_load_avg(curr, 1);
1963 update_cfs_rq_blocked_load(cfs_rq, 1);
1965 #ifdef CONFIG_SCHED_HRTICK
1967 * queued ticks are scheduled to match the slice, so don't bother
1968 * validating it and just reschedule.
1971 resched_task(rq_of(cfs_rq)->curr);
1975 * don't let the period tick interfere with the hrtick preemption
1977 if (!sched_feat(DOUBLE_TICK) &&
1978 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1982 if (cfs_rq->nr_running > 1)
1983 check_preempt_tick(cfs_rq, curr);
1987 /**************************************************
1988 * CFS bandwidth control machinery
1991 #ifdef CONFIG_CFS_BANDWIDTH
1993 #ifdef HAVE_JUMP_LABEL
1994 static struct static_key __cfs_bandwidth_used;
1996 static inline bool cfs_bandwidth_used(void)
1998 return static_key_false(&__cfs_bandwidth_used);
2001 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2003 /* only need to count groups transitioning between enabled/!enabled */
2004 if (enabled && !was_enabled)
2005 static_key_slow_inc(&__cfs_bandwidth_used);
2006 else if (!enabled && was_enabled)
2007 static_key_slow_dec(&__cfs_bandwidth_used);
2009 #else /* HAVE_JUMP_LABEL */
2010 static bool cfs_bandwidth_used(void)
2015 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2016 #endif /* HAVE_JUMP_LABEL */
2019 * default period for cfs group bandwidth.
2020 * default: 0.1s, units: nanoseconds
2022 static inline u64 default_cfs_period(void)
2024 return 100000000ULL;
2027 static inline u64 sched_cfs_bandwidth_slice(void)
2029 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2033 * Replenish runtime according to assigned quota and update expiration time.
2034 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2035 * additional synchronization around rq->lock.
2037 * requires cfs_b->lock
2039 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2043 if (cfs_b->quota == RUNTIME_INF)
2046 now = sched_clock_cpu(smp_processor_id());
2047 cfs_b->runtime = cfs_b->quota;
2048 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2051 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2053 return &tg->cfs_bandwidth;
2056 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2057 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2059 if (unlikely(cfs_rq->throttle_count))
2060 return cfs_rq->throttled_clock_task;
2062 return rq_of(cfs_rq)->clock_task - cfs_rq->throttled_clock_task_time;
2065 /* returns 0 on failure to allocate runtime */
2066 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2068 struct task_group *tg = cfs_rq->tg;
2069 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2070 u64 amount = 0, min_amount, expires;
2072 /* note: this is a positive sum as runtime_remaining <= 0 */
2073 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2075 raw_spin_lock(&cfs_b->lock);
2076 if (cfs_b->quota == RUNTIME_INF)
2077 amount = min_amount;
2080 * If the bandwidth pool has become inactive, then at least one
2081 * period must have elapsed since the last consumption.
2082 * Refresh the global state and ensure bandwidth timer becomes
2085 if (!cfs_b->timer_active) {
2086 __refill_cfs_bandwidth_runtime(cfs_b);
2087 __start_cfs_bandwidth(cfs_b);
2090 if (cfs_b->runtime > 0) {
2091 amount = min(cfs_b->runtime, min_amount);
2092 cfs_b->runtime -= amount;
2096 expires = cfs_b->runtime_expires;
2097 raw_spin_unlock(&cfs_b->lock);
2099 cfs_rq->runtime_remaining += amount;
2101 * we may have advanced our local expiration to account for allowed
2102 * spread between our sched_clock and the one on which runtime was
2105 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2106 cfs_rq->runtime_expires = expires;
2108 return cfs_rq->runtime_remaining > 0;
2112 * Note: This depends on the synchronization provided by sched_clock and the
2113 * fact that rq->clock snapshots this value.
2115 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2117 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2118 struct rq *rq = rq_of(cfs_rq);
2120 /* if the deadline is ahead of our clock, nothing to do */
2121 if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0))
2124 if (cfs_rq->runtime_remaining < 0)
2128 * If the local deadline has passed we have to consider the
2129 * possibility that our sched_clock is 'fast' and the global deadline
2130 * has not truly expired.
2132 * Fortunately we can check determine whether this the case by checking
2133 * whether the global deadline has advanced.
2136 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2137 /* extend local deadline, drift is bounded above by 2 ticks */
2138 cfs_rq->runtime_expires += TICK_NSEC;
2140 /* global deadline is ahead, expiration has passed */
2141 cfs_rq->runtime_remaining = 0;
2145 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2146 unsigned long delta_exec)
2148 /* dock delta_exec before expiring quota (as it could span periods) */
2149 cfs_rq->runtime_remaining -= delta_exec;
2150 expire_cfs_rq_runtime(cfs_rq);
2152 if (likely(cfs_rq->runtime_remaining > 0))
2156 * if we're unable to extend our runtime we resched so that the active
2157 * hierarchy can be throttled
2159 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2160 resched_task(rq_of(cfs_rq)->curr);
2163 static __always_inline
2164 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2166 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2169 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2172 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2174 return cfs_bandwidth_used() && cfs_rq->throttled;
2177 /* check whether cfs_rq, or any parent, is throttled */
2178 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2180 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2184 * Ensure that neither of the group entities corresponding to src_cpu or
2185 * dest_cpu are members of a throttled hierarchy when performing group
2186 * load-balance operations.
2188 static inline int throttled_lb_pair(struct task_group *tg,
2189 int src_cpu, int dest_cpu)
2191 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2193 src_cfs_rq = tg->cfs_rq[src_cpu];
2194 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2196 return throttled_hierarchy(src_cfs_rq) ||
2197 throttled_hierarchy(dest_cfs_rq);
2200 /* updated child weight may affect parent so we have to do this bottom up */
2201 static int tg_unthrottle_up(struct task_group *tg, void *data)
2203 struct rq *rq = data;
2204 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2206 cfs_rq->throttle_count--;
2208 if (!cfs_rq->throttle_count) {
2209 /* adjust cfs_rq_clock_task() */
2210 cfs_rq->throttled_clock_task_time += rq->clock_task -
2211 cfs_rq->throttled_clock_task;
2218 static int tg_throttle_down(struct task_group *tg, void *data)
2220 struct rq *rq = data;
2221 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2223 /* group is entering throttled state, stop time */
2224 if (!cfs_rq->throttle_count)
2225 cfs_rq->throttled_clock_task = rq->clock_task;
2226 cfs_rq->throttle_count++;
2231 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2233 struct rq *rq = rq_of(cfs_rq);
2234 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2235 struct sched_entity *se;
2236 long task_delta, dequeue = 1;
2238 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2240 /* freeze hierarchy runnable averages while throttled */
2242 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2245 task_delta = cfs_rq->h_nr_running;
2246 for_each_sched_entity(se) {
2247 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2248 /* throttled entity or throttle-on-deactivate */
2253 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2254 qcfs_rq->h_nr_running -= task_delta;
2256 if (qcfs_rq->load.weight)
2261 rq->nr_running -= task_delta;
2263 cfs_rq->throttled = 1;
2264 cfs_rq->throttled_clock = rq->clock;
2265 raw_spin_lock(&cfs_b->lock);
2266 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2267 raw_spin_unlock(&cfs_b->lock);
2270 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2272 struct rq *rq = rq_of(cfs_rq);
2273 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2274 struct sched_entity *se;
2278 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2280 cfs_rq->throttled = 0;
2281 raw_spin_lock(&cfs_b->lock);
2282 cfs_b->throttled_time += rq->clock - cfs_rq->throttled_clock;
2283 list_del_rcu(&cfs_rq->throttled_list);
2284 raw_spin_unlock(&cfs_b->lock);
2286 update_rq_clock(rq);
2287 /* update hierarchical throttle state */
2288 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2290 if (!cfs_rq->load.weight)
2293 task_delta = cfs_rq->h_nr_running;
2294 for_each_sched_entity(se) {
2298 cfs_rq = cfs_rq_of(se);
2300 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2301 cfs_rq->h_nr_running += task_delta;
2303 if (cfs_rq_throttled(cfs_rq))
2308 rq->nr_running += task_delta;
2310 /* determine whether we need to wake up potentially idle cpu */
2311 if (rq->curr == rq->idle && rq->cfs.nr_running)
2312 resched_task(rq->curr);
2315 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2316 u64 remaining, u64 expires)
2318 struct cfs_rq *cfs_rq;
2319 u64 runtime = remaining;
2322 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2324 struct rq *rq = rq_of(cfs_rq);
2326 raw_spin_lock(&rq->lock);
2327 if (!cfs_rq_throttled(cfs_rq))
2330 runtime = -cfs_rq->runtime_remaining + 1;
2331 if (runtime > remaining)
2332 runtime = remaining;
2333 remaining -= runtime;
2335 cfs_rq->runtime_remaining += runtime;
2336 cfs_rq->runtime_expires = expires;
2338 /* we check whether we're throttled above */
2339 if (cfs_rq->runtime_remaining > 0)
2340 unthrottle_cfs_rq(cfs_rq);
2343 raw_spin_unlock(&rq->lock);
2354 * Responsible for refilling a task_group's bandwidth and unthrottling its
2355 * cfs_rqs as appropriate. If there has been no activity within the last
2356 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2357 * used to track this state.
2359 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2361 u64 runtime, runtime_expires;
2362 int idle = 1, throttled;
2364 raw_spin_lock(&cfs_b->lock);
2365 /* no need to continue the timer with no bandwidth constraint */
2366 if (cfs_b->quota == RUNTIME_INF)
2369 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2370 /* idle depends on !throttled (for the case of a large deficit) */
2371 idle = cfs_b->idle && !throttled;
2372 cfs_b->nr_periods += overrun;
2374 /* if we're going inactive then everything else can be deferred */
2378 __refill_cfs_bandwidth_runtime(cfs_b);
2381 /* mark as potentially idle for the upcoming period */
2386 /* account preceding periods in which throttling occurred */
2387 cfs_b->nr_throttled += overrun;
2390 * There are throttled entities so we must first use the new bandwidth
2391 * to unthrottle them before making it generally available. This
2392 * ensures that all existing debts will be paid before a new cfs_rq is
2395 runtime = cfs_b->runtime;
2396 runtime_expires = cfs_b->runtime_expires;
2400 * This check is repeated as we are holding onto the new bandwidth
2401 * while we unthrottle. This can potentially race with an unthrottled
2402 * group trying to acquire new bandwidth from the global pool.
2404 while (throttled && runtime > 0) {
2405 raw_spin_unlock(&cfs_b->lock);
2406 /* we can't nest cfs_b->lock while distributing bandwidth */
2407 runtime = distribute_cfs_runtime(cfs_b, runtime,
2409 raw_spin_lock(&cfs_b->lock);
2411 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2414 /* return (any) remaining runtime */
2415 cfs_b->runtime = runtime;
2417 * While we are ensured activity in the period following an
2418 * unthrottle, this also covers the case in which the new bandwidth is
2419 * insufficient to cover the existing bandwidth deficit. (Forcing the
2420 * timer to remain active while there are any throttled entities.)
2425 cfs_b->timer_active = 0;
2426 raw_spin_unlock(&cfs_b->lock);
2431 /* a cfs_rq won't donate quota below this amount */
2432 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2433 /* minimum remaining period time to redistribute slack quota */
2434 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2435 /* how long we wait to gather additional slack before distributing */
2436 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2438 /* are we near the end of the current quota period? */
2439 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2441 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2444 /* if the call-back is running a quota refresh is already occurring */
2445 if (hrtimer_callback_running(refresh_timer))
2448 /* is a quota refresh about to occur? */
2449 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2450 if (remaining < min_expire)
2456 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2458 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2460 /* if there's a quota refresh soon don't bother with slack */
2461 if (runtime_refresh_within(cfs_b, min_left))
2464 start_bandwidth_timer(&cfs_b->slack_timer,
2465 ns_to_ktime(cfs_bandwidth_slack_period));
2468 /* we know any runtime found here is valid as update_curr() precedes return */
2469 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2471 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2472 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2474 if (slack_runtime <= 0)
2477 raw_spin_lock(&cfs_b->lock);
2478 if (cfs_b->quota != RUNTIME_INF &&
2479 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2480 cfs_b->runtime += slack_runtime;
2482 /* we are under rq->lock, defer unthrottling using a timer */
2483 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2484 !list_empty(&cfs_b->throttled_cfs_rq))
2485 start_cfs_slack_bandwidth(cfs_b);
2487 raw_spin_unlock(&cfs_b->lock);
2489 /* even if it's not valid for return we don't want to try again */
2490 cfs_rq->runtime_remaining -= slack_runtime;
2493 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2495 if (!cfs_bandwidth_used())
2498 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2501 __return_cfs_rq_runtime(cfs_rq);
2505 * This is done with a timer (instead of inline with bandwidth return) since
2506 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2508 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2510 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2513 /* confirm we're still not at a refresh boundary */
2514 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2517 raw_spin_lock(&cfs_b->lock);
2518 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2519 runtime = cfs_b->runtime;
2522 expires = cfs_b->runtime_expires;
2523 raw_spin_unlock(&cfs_b->lock);
2528 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2530 raw_spin_lock(&cfs_b->lock);
2531 if (expires == cfs_b->runtime_expires)
2532 cfs_b->runtime = runtime;
2533 raw_spin_unlock(&cfs_b->lock);
2537 * When a group wakes up we want to make sure that its quota is not already
2538 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2539 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2541 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2543 if (!cfs_bandwidth_used())
2546 /* an active group must be handled by the update_curr()->put() path */
2547 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2550 /* ensure the group is not already throttled */
2551 if (cfs_rq_throttled(cfs_rq))
2554 /* update runtime allocation */
2555 account_cfs_rq_runtime(cfs_rq, 0);
2556 if (cfs_rq->runtime_remaining <= 0)
2557 throttle_cfs_rq(cfs_rq);
2560 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2561 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2563 if (!cfs_bandwidth_used())
2566 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2570 * it's possible for a throttled entity to be forced into a running
2571 * state (e.g. set_curr_task), in this case we're finished.
2573 if (cfs_rq_throttled(cfs_rq))
2576 throttle_cfs_rq(cfs_rq);
2579 static inline u64 default_cfs_period(void);
2580 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun);
2581 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b);
2583 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2585 struct cfs_bandwidth *cfs_b =
2586 container_of(timer, struct cfs_bandwidth, slack_timer);
2587 do_sched_cfs_slack_timer(cfs_b);
2589 return HRTIMER_NORESTART;
2592 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2594 struct cfs_bandwidth *cfs_b =
2595 container_of(timer, struct cfs_bandwidth, period_timer);
2601 now = hrtimer_cb_get_time(timer);
2602 overrun = hrtimer_forward(timer, now, cfs_b->period);
2607 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2610 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2613 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2615 raw_spin_lock_init(&cfs_b->lock);
2617 cfs_b->quota = RUNTIME_INF;
2618 cfs_b->period = ns_to_ktime(default_cfs_period());
2620 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2621 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2622 cfs_b->period_timer.function = sched_cfs_period_timer;
2623 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2624 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2627 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2629 cfs_rq->runtime_enabled = 0;
2630 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2633 /* requires cfs_b->lock, may release to reprogram timer */
2634 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2637 * The timer may be active because we're trying to set a new bandwidth
2638 * period or because we're racing with the tear-down path
2639 * (timer_active==0 becomes visible before the hrtimer call-back
2640 * terminates). In either case we ensure that it's re-programmed
2642 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2643 raw_spin_unlock(&cfs_b->lock);
2644 /* ensure cfs_b->lock is available while we wait */
2645 hrtimer_cancel(&cfs_b->period_timer);
2647 raw_spin_lock(&cfs_b->lock);
2648 /* if someone else restarted the timer then we're done */
2649 if (cfs_b->timer_active)
2653 cfs_b->timer_active = 1;
2654 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2657 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2659 hrtimer_cancel(&cfs_b->period_timer);
2660 hrtimer_cancel(&cfs_b->slack_timer);
2663 static void unthrottle_offline_cfs_rqs(struct rq *rq)
2665 struct cfs_rq *cfs_rq;
2667 for_each_leaf_cfs_rq(rq, cfs_rq) {
2668 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2670 if (!cfs_rq->runtime_enabled)
2674 * clock_task is not advancing so we just need to make sure
2675 * there's some valid quota amount
2677 cfs_rq->runtime_remaining = cfs_b->quota;
2678 if (cfs_rq_throttled(cfs_rq))
2679 unthrottle_cfs_rq(cfs_rq);
2683 #else /* CONFIG_CFS_BANDWIDTH */
2684 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2686 return rq_of(cfs_rq)->clock_task;
2689 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2690 unsigned long delta_exec) {}
2691 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2692 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2693 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2695 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2700 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2705 static inline int throttled_lb_pair(struct task_group *tg,
2706 int src_cpu, int dest_cpu)
2711 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2713 #ifdef CONFIG_FAIR_GROUP_SCHED
2714 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2717 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2721 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2722 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2724 #endif /* CONFIG_CFS_BANDWIDTH */
2726 /**************************************************
2727 * CFS operations on tasks:
2730 #ifdef CONFIG_SCHED_HRTICK
2731 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2733 struct sched_entity *se = &p->se;
2734 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2736 WARN_ON(task_rq(p) != rq);
2738 if (cfs_rq->nr_running > 1) {
2739 u64 slice = sched_slice(cfs_rq, se);
2740 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2741 s64 delta = slice - ran;
2750 * Don't schedule slices shorter than 10000ns, that just
2751 * doesn't make sense. Rely on vruntime for fairness.
2754 delta = max_t(s64, 10000LL, delta);
2756 hrtick_start(rq, delta);
2761 * called from enqueue/dequeue and updates the hrtick when the
2762 * current task is from our class and nr_running is low enough
2765 static void hrtick_update(struct rq *rq)
2767 struct task_struct *curr = rq->curr;
2769 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2772 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2773 hrtick_start_fair(rq, curr);
2775 #else /* !CONFIG_SCHED_HRTICK */
2777 hrtick_start_fair(struct rq *rq, struct task_struct *p)
2781 static inline void hrtick_update(struct rq *rq)
2787 * The enqueue_task method is called before nr_running is
2788 * increased. Here we update the fair scheduling stats and
2789 * then put the task into the rbtree:
2792 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2794 struct cfs_rq *cfs_rq;
2795 struct sched_entity *se = &p->se;
2797 for_each_sched_entity(se) {
2800 cfs_rq = cfs_rq_of(se);
2801 enqueue_entity(cfs_rq, se, flags);
2804 * end evaluation on encountering a throttled cfs_rq
2806 * note: in the case of encountering a throttled cfs_rq we will
2807 * post the final h_nr_running increment below.
2809 if (cfs_rq_throttled(cfs_rq))
2811 cfs_rq->h_nr_running++;
2813 flags = ENQUEUE_WAKEUP;
2816 for_each_sched_entity(se) {
2817 cfs_rq = cfs_rq_of(se);
2818 cfs_rq->h_nr_running++;
2820 if (cfs_rq_throttled(cfs_rq))
2823 update_cfs_shares(cfs_rq);
2824 update_entity_load_avg(se, 1);
2828 update_rq_runnable_avg(rq, rq->nr_running);
2834 static void set_next_buddy(struct sched_entity *se);
2837 * The dequeue_task method is called before nr_running is
2838 * decreased. We remove the task from the rbtree and
2839 * update the fair scheduling stats:
2841 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2843 struct cfs_rq *cfs_rq;
2844 struct sched_entity *se = &p->se;
2845 int task_sleep = flags & DEQUEUE_SLEEP;
2847 for_each_sched_entity(se) {
2848 cfs_rq = cfs_rq_of(se);
2849 dequeue_entity(cfs_rq, se, flags);
2852 * end evaluation on encountering a throttled cfs_rq
2854 * note: in the case of encountering a throttled cfs_rq we will
2855 * post the final h_nr_running decrement below.
2857 if (cfs_rq_throttled(cfs_rq))
2859 cfs_rq->h_nr_running--;
2861 /* Don't dequeue parent if it has other entities besides us */
2862 if (cfs_rq->load.weight) {
2864 * Bias pick_next to pick a task from this cfs_rq, as
2865 * p is sleeping when it is within its sched_slice.
2867 if (task_sleep && parent_entity(se))
2868 set_next_buddy(parent_entity(se));
2870 /* avoid re-evaluating load for this entity */
2871 se = parent_entity(se);
2874 flags |= DEQUEUE_SLEEP;
2877 for_each_sched_entity(se) {
2878 cfs_rq = cfs_rq_of(se);
2879 cfs_rq->h_nr_running--;
2881 if (cfs_rq_throttled(cfs_rq))
2884 update_cfs_shares(cfs_rq);
2885 update_entity_load_avg(se, 1);
2890 update_rq_runnable_avg(rq, 1);
2896 /* Used instead of source_load when we know the type == 0 */
2897 static unsigned long weighted_cpuload(const int cpu)
2899 return cpu_rq(cpu)->load.weight;
2903 * Return a low guess at the load of a migration-source cpu weighted
2904 * according to the scheduling class and "nice" value.
2906 * We want to under-estimate the load of migration sources, to
2907 * balance conservatively.
2909 static unsigned long source_load(int cpu, int type)
2911 struct rq *rq = cpu_rq(cpu);
2912 unsigned long total = weighted_cpuload(cpu);
2914 if (type == 0 || !sched_feat(LB_BIAS))
2917 return min(rq->cpu_load[type-1], total);
2921 * Return a high guess at the load of a migration-target cpu weighted
2922 * according to the scheduling class and "nice" value.
2924 static unsigned long target_load(int cpu, int type)
2926 struct rq *rq = cpu_rq(cpu);
2927 unsigned long total = weighted_cpuload(cpu);
2929 if (type == 0 || !sched_feat(LB_BIAS))
2932 return max(rq->cpu_load[type-1], total);
2935 static unsigned long power_of(int cpu)
2937 return cpu_rq(cpu)->cpu_power;
2940 static unsigned long cpu_avg_load_per_task(int cpu)
2942 struct rq *rq = cpu_rq(cpu);
2943 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
2946 return rq->load.weight / nr_running;
2952 static void task_waking_fair(struct task_struct *p)
2954 struct sched_entity *se = &p->se;
2955 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2958 #ifndef CONFIG_64BIT
2959 u64 min_vruntime_copy;
2962 min_vruntime_copy = cfs_rq->min_vruntime_copy;
2964 min_vruntime = cfs_rq->min_vruntime;
2965 } while (min_vruntime != min_vruntime_copy);
2967 min_vruntime = cfs_rq->min_vruntime;
2970 se->vruntime -= min_vruntime;
2973 #ifdef CONFIG_FAIR_GROUP_SCHED
2975 * effective_load() calculates the load change as seen from the root_task_group
2977 * Adding load to a group doesn't make a group heavier, but can cause movement
2978 * of group shares between cpus. Assuming the shares were perfectly aligned one
2979 * can calculate the shift in shares.
2981 * Calculate the effective load difference if @wl is added (subtracted) to @tg
2982 * on this @cpu and results in a total addition (subtraction) of @wg to the
2983 * total group weight.
2985 * Given a runqueue weight distribution (rw_i) we can compute a shares
2986 * distribution (s_i) using:
2988 * s_i = rw_i / \Sum rw_j (1)
2990 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
2991 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
2992 * shares distribution (s_i):
2994 * rw_i = { 2, 4, 1, 0 }
2995 * s_i = { 2/7, 4/7, 1/7, 0 }
2997 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
2998 * task used to run on and the CPU the waker is running on), we need to
2999 * compute the effect of waking a task on either CPU and, in case of a sync
3000 * wakeup, compute the effect of the current task going to sleep.
3002 * So for a change of @wl to the local @cpu with an overall group weight change
3003 * of @wl we can compute the new shares distribution (s'_i) using:
3005 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3007 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3008 * differences in waking a task to CPU 0. The additional task changes the
3009 * weight and shares distributions like:
3011 * rw'_i = { 3, 4, 1, 0 }
3012 * s'_i = { 3/8, 4/8, 1/8, 0 }
3014 * We can then compute the difference in effective weight by using:
3016 * dw_i = S * (s'_i - s_i) (3)
3018 * Where 'S' is the group weight as seen by its parent.
3020 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3021 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3022 * 4/7) times the weight of the group.
3024 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3026 struct sched_entity *se = tg->se[cpu];
3028 if (!tg->parent) /* the trivial, non-cgroup case */
3031 for_each_sched_entity(se) {
3037 * W = @wg + \Sum rw_j
3039 W = wg + calc_tg_weight(tg, se->my_q);
3044 w = se->my_q->load.weight + wl;
3047 * wl = S * s'_i; see (2)
3050 wl = (w * tg->shares) / W;
3055 * Per the above, wl is the new se->load.weight value; since
3056 * those are clipped to [MIN_SHARES, ...) do so now. See
3057 * calc_cfs_shares().
3059 if (wl < MIN_SHARES)
3063 * wl = dw_i = S * (s'_i - s_i); see (3)
3065 wl -= se->load.weight;
3068 * Recursively apply this logic to all parent groups to compute
3069 * the final effective load change on the root group. Since
3070 * only the @tg group gets extra weight, all parent groups can
3071 * only redistribute existing shares. @wl is the shift in shares
3072 * resulting from this level per the above.
3081 static inline unsigned long effective_load(struct task_group *tg, int cpu,
3082 unsigned long wl, unsigned long wg)
3089 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3091 s64 this_load, load;
3092 int idx, this_cpu, prev_cpu;
3093 unsigned long tl_per_task;
3094 struct task_group *tg;
3095 unsigned long weight;
3099 this_cpu = smp_processor_id();
3100 prev_cpu = task_cpu(p);
3101 load = source_load(prev_cpu, idx);
3102 this_load = target_load(this_cpu, idx);
3105 * If sync wakeup then subtract the (maximum possible)
3106 * effect of the currently running task from the load
3107 * of the current CPU:
3110 tg = task_group(current);
3111 weight = current->se.load.weight;
3113 this_load += effective_load(tg, this_cpu, -weight, -weight);
3114 load += effective_load(tg, prev_cpu, 0, -weight);
3118 weight = p->se.load.weight;
3121 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3122 * due to the sync cause above having dropped this_load to 0, we'll
3123 * always have an imbalance, but there's really nothing you can do
3124 * about that, so that's good too.
3126 * Otherwise check if either cpus are near enough in load to allow this
3127 * task to be woken on this_cpu.
3129 if (this_load > 0) {
3130 s64 this_eff_load, prev_eff_load;
3132 this_eff_load = 100;
3133 this_eff_load *= power_of(prev_cpu);
3134 this_eff_load *= this_load +
3135 effective_load(tg, this_cpu, weight, weight);
3137 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3138 prev_eff_load *= power_of(this_cpu);
3139 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3141 balanced = this_eff_load <= prev_eff_load;
3146 * If the currently running task will sleep within
3147 * a reasonable amount of time then attract this newly
3150 if (sync && balanced)
3153 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3154 tl_per_task = cpu_avg_load_per_task(this_cpu);
3157 (this_load <= load &&
3158 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3160 * This domain has SD_WAKE_AFFINE and
3161 * p is cache cold in this domain, and
3162 * there is no bad imbalance.
3164 schedstat_inc(sd, ttwu_move_affine);
3165 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3173 * find_idlest_group finds and returns the least busy CPU group within the
3176 static struct sched_group *
3177 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3178 int this_cpu, int load_idx)
3180 struct sched_group *idlest = NULL, *group = sd->groups;
3181 unsigned long min_load = ULONG_MAX, this_load = 0;
3182 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3185 unsigned long load, avg_load;
3189 /* Skip over this group if it has no CPUs allowed */
3190 if (!cpumask_intersects(sched_group_cpus(group),
3191 tsk_cpus_allowed(p)))
3194 local_group = cpumask_test_cpu(this_cpu,
3195 sched_group_cpus(group));
3197 /* Tally up the load of all CPUs in the group */
3200 for_each_cpu(i, sched_group_cpus(group)) {
3201 /* Bias balancing toward cpus of our domain */
3203 load = source_load(i, load_idx);
3205 load = target_load(i, load_idx);
3210 /* Adjust by relative CPU power of the group */
3211 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3214 this_load = avg_load;
3215 } else if (avg_load < min_load) {
3216 min_load = avg_load;
3219 } while (group = group->next, group != sd->groups);
3221 if (!idlest || 100*this_load < imbalance*min_load)
3227 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3230 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3232 unsigned long load, min_load = ULONG_MAX;
3236 /* Traverse only the allowed CPUs */
3237 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3238 load = weighted_cpuload(i);
3240 if (load < min_load || (load == min_load && i == this_cpu)) {
3250 * Try and locate an idle CPU in the sched_domain.
3252 static int select_idle_sibling(struct task_struct *p, int target)
3254 int cpu = smp_processor_id();
3255 int prev_cpu = task_cpu(p);
3256 struct sched_domain *sd;
3257 struct sched_group *sg;
3261 * If the task is going to be woken-up on this cpu and if it is
3262 * already idle, then it is the right target.
3264 if (target == cpu && idle_cpu(cpu))
3268 * If the task is going to be woken-up on the cpu where it previously
3269 * ran and if it is currently idle, then it the right target.
3271 if (target == prev_cpu && idle_cpu(prev_cpu))
3275 * Otherwise, iterate the domains and find an elegible idle cpu.
3277 sd = rcu_dereference(per_cpu(sd_llc, target));
3278 for_each_lower_domain(sd) {
3281 if (!cpumask_intersects(sched_group_cpus(sg),
3282 tsk_cpus_allowed(p)))
3285 for_each_cpu(i, sched_group_cpus(sg)) {
3290 target = cpumask_first_and(sched_group_cpus(sg),
3291 tsk_cpus_allowed(p));
3295 } while (sg != sd->groups);
3302 * sched_balance_self: balance the current task (running on cpu) in domains
3303 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3306 * Balance, ie. select the least loaded group.
3308 * Returns the target CPU number, or the same CPU if no balancing is needed.
3310 * preempt must be disabled.
3313 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
3315 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3316 int cpu = smp_processor_id();
3317 int prev_cpu = task_cpu(p);
3319 int want_affine = 0;
3320 int sync = wake_flags & WF_SYNC;
3322 if (p->nr_cpus_allowed == 1)
3325 if (sd_flag & SD_BALANCE_WAKE) {
3326 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
3332 for_each_domain(cpu, tmp) {
3333 if (!(tmp->flags & SD_LOAD_BALANCE))
3337 * If both cpu and prev_cpu are part of this domain,
3338 * cpu is a valid SD_WAKE_AFFINE target.
3340 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3341 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3346 if (tmp->flags & sd_flag)
3351 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3354 new_cpu = select_idle_sibling(p, prev_cpu);
3359 int load_idx = sd->forkexec_idx;
3360 struct sched_group *group;
3363 if (!(sd->flags & sd_flag)) {
3368 if (sd_flag & SD_BALANCE_WAKE)
3369 load_idx = sd->wake_idx;
3371 group = find_idlest_group(sd, p, cpu, load_idx);
3377 new_cpu = find_idlest_cpu(group, p, cpu);
3378 if (new_cpu == -1 || new_cpu == cpu) {
3379 /* Now try balancing at a lower domain level of cpu */
3384 /* Now try balancing at a lower domain level of new_cpu */
3386 weight = sd->span_weight;
3388 for_each_domain(cpu, tmp) {
3389 if (weight <= tmp->span_weight)
3391 if (tmp->flags & sd_flag)
3394 /* while loop will break here if sd == NULL */
3403 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
3404 * removed when useful for applications beyond shares distribution (e.g.
3407 #ifdef CONFIG_FAIR_GROUP_SCHED
3409 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3410 * cfs_rq_of(p) references at time of call are still valid and identify the
3411 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3412 * other assumptions, including the state of rq->lock, should be made.
3415 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
3417 struct sched_entity *se = &p->se;
3418 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3421 * Load tracking: accumulate removed load so that it can be processed
3422 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3423 * to blocked load iff they have a positive decay-count. It can never
3424 * be negative here since on-rq tasks have decay-count == 0.
3426 if (se->avg.decay_count) {
3427 se->avg.decay_count = -__synchronize_entity_decay(se);
3428 atomic64_add(se->avg.load_avg_contrib, &cfs_rq->removed_load);
3432 #endif /* CONFIG_SMP */
3434 static unsigned long
3435 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3437 unsigned long gran = sysctl_sched_wakeup_granularity;
3440 * Since its curr running now, convert the gran from real-time
3441 * to virtual-time in his units.
3443 * By using 'se' instead of 'curr' we penalize light tasks, so
3444 * they get preempted easier. That is, if 'se' < 'curr' then
3445 * the resulting gran will be larger, therefore penalizing the
3446 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3447 * be smaller, again penalizing the lighter task.
3449 * This is especially important for buddies when the leftmost
3450 * task is higher priority than the buddy.
3452 return calc_delta_fair(gran, se);
3456 * Should 'se' preempt 'curr'.
3470 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3472 s64 gran, vdiff = curr->vruntime - se->vruntime;
3477 gran = wakeup_gran(curr, se);
3484 static void set_last_buddy(struct sched_entity *se)
3486 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3489 for_each_sched_entity(se)
3490 cfs_rq_of(se)->last = se;
3493 static void set_next_buddy(struct sched_entity *se)
3495 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3498 for_each_sched_entity(se)
3499 cfs_rq_of(se)->next = se;
3502 static void set_skip_buddy(struct sched_entity *se)
3504 for_each_sched_entity(se)
3505 cfs_rq_of(se)->skip = se;
3509 * Preempt the current task with a newly woken task if needed:
3511 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3513 struct task_struct *curr = rq->curr;
3514 struct sched_entity *se = &curr->se, *pse = &p->se;
3515 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3516 int scale = cfs_rq->nr_running >= sched_nr_latency;
3517 int next_buddy_marked = 0;
3519 if (unlikely(se == pse))
3523 * This is possible from callers such as move_task(), in which we
3524 * unconditionally check_prempt_curr() after an enqueue (which may have
3525 * lead to a throttle). This both saves work and prevents false
3526 * next-buddy nomination below.
3528 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3531 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3532 set_next_buddy(pse);
3533 next_buddy_marked = 1;
3537 * We can come here with TIF_NEED_RESCHED already set from new task
3540 * Note: this also catches the edge-case of curr being in a throttled
3541 * group (e.g. via set_curr_task), since update_curr() (in the
3542 * enqueue of curr) will have resulted in resched being set. This
3543 * prevents us from potentially nominating it as a false LAST_BUDDY
3546 if (test_tsk_need_resched(curr))
3549 /* Idle tasks are by definition preempted by non-idle tasks. */
3550 if (unlikely(curr->policy == SCHED_IDLE) &&
3551 likely(p->policy != SCHED_IDLE))
3555 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3556 * is driven by the tick):
3558 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
3561 find_matching_se(&se, &pse);
3562 update_curr(cfs_rq_of(se));
3564 if (wakeup_preempt_entity(se, pse) == 1) {
3566 * Bias pick_next to pick the sched entity that is
3567 * triggering this preemption.
3569 if (!next_buddy_marked)
3570 set_next_buddy(pse);
3579 * Only set the backward buddy when the current task is still
3580 * on the rq. This can happen when a wakeup gets interleaved
3581 * with schedule on the ->pre_schedule() or idle_balance()
3582 * point, either of which can * drop the rq lock.
3584 * Also, during early boot the idle thread is in the fair class,
3585 * for obvious reasons its a bad idea to schedule back to it.
3587 if (unlikely(!se->on_rq || curr == rq->idle))
3590 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3594 static struct task_struct *pick_next_task_fair(struct rq *rq)
3596 struct task_struct *p;
3597 struct cfs_rq *cfs_rq = &rq->cfs;
3598 struct sched_entity *se;
3600 if (!cfs_rq->nr_running)
3604 se = pick_next_entity(cfs_rq);
3605 set_next_entity(cfs_rq, se);
3606 cfs_rq = group_cfs_rq(se);
3610 if (hrtick_enabled(rq))
3611 hrtick_start_fair(rq, p);
3617 * Account for a descheduled task:
3619 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3621 struct sched_entity *se = &prev->se;
3622 struct cfs_rq *cfs_rq;
3624 for_each_sched_entity(se) {
3625 cfs_rq = cfs_rq_of(se);
3626 put_prev_entity(cfs_rq, se);
3631 * sched_yield() is very simple
3633 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3635 static void yield_task_fair(struct rq *rq)
3637 struct task_struct *curr = rq->curr;
3638 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3639 struct sched_entity *se = &curr->se;
3642 * Are we the only task in the tree?
3644 if (unlikely(rq->nr_running == 1))
3647 clear_buddies(cfs_rq, se);
3649 if (curr->policy != SCHED_BATCH) {
3650 update_rq_clock(rq);
3652 * Update run-time statistics of the 'current'.
3654 update_curr(cfs_rq);
3656 * Tell update_rq_clock() that we've just updated,
3657 * so we don't do microscopic update in schedule()
3658 * and double the fastpath cost.
3660 rq->skip_clock_update = 1;
3666 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3668 struct sched_entity *se = &p->se;
3670 /* throttled hierarchies are not runnable */
3671 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3674 /* Tell the scheduler that we'd really like pse to run next. */
3677 yield_task_fair(rq);
3683 /**************************************************
3684 * Fair scheduling class load-balancing methods.
3688 * The purpose of load-balancing is to achieve the same basic fairness the
3689 * per-cpu scheduler provides, namely provide a proportional amount of compute
3690 * time to each task. This is expressed in the following equation:
3692 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
3694 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
3695 * W_i,0 is defined as:
3697 * W_i,0 = \Sum_j w_i,j (2)
3699 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
3700 * is derived from the nice value as per prio_to_weight[].
3702 * The weight average is an exponential decay average of the instantaneous
3705 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
3707 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
3708 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
3709 * can also include other factors [XXX].
3711 * To achieve this balance we define a measure of imbalance which follows
3712 * directly from (1):
3714 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
3716 * We them move tasks around to minimize the imbalance. In the continuous
3717 * function space it is obvious this converges, in the discrete case we get
3718 * a few fun cases generally called infeasible weight scenarios.
3721 * - infeasible weights;
3722 * - local vs global optima in the discrete case. ]
3727 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
3728 * for all i,j solution, we create a tree of cpus that follows the hardware
3729 * topology where each level pairs two lower groups (or better). This results
3730 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
3731 * tree to only the first of the previous level and we decrease the frequency
3732 * of load-balance at each level inv. proportional to the number of cpus in
3738 * \Sum { --- * --- * 2^i } = O(n) (5)
3740 * `- size of each group
3741 * | | `- number of cpus doing load-balance
3743 * `- sum over all levels
3745 * Coupled with a limit on how many tasks we can migrate every balance pass,
3746 * this makes (5) the runtime complexity of the balancer.
3748 * An important property here is that each CPU is still (indirectly) connected
3749 * to every other cpu in at most O(log n) steps:
3751 * The adjacency matrix of the resulting graph is given by:
3754 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
3757 * And you'll find that:
3759 * A^(log_2 n)_i,j != 0 for all i,j (7)
3761 * Showing there's indeed a path between every cpu in at most O(log n) steps.
3762 * The task movement gives a factor of O(m), giving a convergence complexity
3765 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
3770 * In order to avoid CPUs going idle while there's still work to do, new idle
3771 * balancing is more aggressive and has the newly idle cpu iterate up the domain
3772 * tree itself instead of relying on other CPUs to bring it work.
3774 * This adds some complexity to both (5) and (8) but it reduces the total idle
3782 * Cgroups make a horror show out of (2), instead of a simple sum we get:
3785 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
3790 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
3792 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
3794 * The big problem is S_k, its a global sum needed to compute a local (W_i)
3797 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
3798 * rewrite all of this once again.]
3801 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
3803 #define LBF_ALL_PINNED 0x01
3804 #define LBF_NEED_BREAK 0x02
3805 #define LBF_SOME_PINNED 0x04
3808 struct sched_domain *sd;
3816 struct cpumask *dst_grpmask;
3818 enum cpu_idle_type idle;
3820 /* The set of CPUs under consideration for load-balancing */
3821 struct cpumask *cpus;
3826 unsigned int loop_break;
3827 unsigned int loop_max;
3831 * move_task - move a task from one runqueue to another runqueue.
3832 * Both runqueues must be locked.
3834 static void move_task(struct task_struct *p, struct lb_env *env)
3836 deactivate_task(env->src_rq, p, 0);
3837 set_task_cpu(p, env->dst_cpu);
3838 activate_task(env->dst_rq, p, 0);
3839 check_preempt_curr(env->dst_rq, p, 0);
3843 * Is this task likely cache-hot:
3846 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
3850 if (p->sched_class != &fair_sched_class)
3853 if (unlikely(p->policy == SCHED_IDLE))
3857 * Buddy candidates are cache hot:
3859 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
3860 (&p->se == cfs_rq_of(&p->se)->next ||
3861 &p->se == cfs_rq_of(&p->se)->last))
3864 if (sysctl_sched_migration_cost == -1)
3866 if (sysctl_sched_migration_cost == 0)
3869 delta = now - p->se.exec_start;
3871 return delta < (s64)sysctl_sched_migration_cost;
3875 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3878 int can_migrate_task(struct task_struct *p, struct lb_env *env)
3880 int tsk_cache_hot = 0;
3882 * We do not migrate tasks that are:
3883 * 1) running (obviously), or
3884 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3885 * 3) are cache-hot on their current CPU.
3887 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
3890 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
3893 * Remember if this task can be migrated to any other cpu in
3894 * our sched_group. We may want to revisit it if we couldn't
3895 * meet load balance goals by pulling other tasks on src_cpu.
3897 * Also avoid computing new_dst_cpu if we have already computed
3898 * one in current iteration.
3900 if (!env->dst_grpmask || (env->flags & LBF_SOME_PINNED))
3903 new_dst_cpu = cpumask_first_and(env->dst_grpmask,
3904 tsk_cpus_allowed(p));
3905 if (new_dst_cpu < nr_cpu_ids) {
3906 env->flags |= LBF_SOME_PINNED;
3907 env->new_dst_cpu = new_dst_cpu;
3912 /* Record that we found atleast one task that could run on dst_cpu */
3913 env->flags &= ~LBF_ALL_PINNED;
3915 if (task_running(env->src_rq, p)) {
3916 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
3921 * Aggressive migration if:
3922 * 1) task is cache cold, or
3923 * 2) too many balance attempts have failed.
3926 tsk_cache_hot = task_hot(p, env->src_rq->clock_task, env->sd);
3927 if (!tsk_cache_hot ||
3928 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
3929 #ifdef CONFIG_SCHEDSTATS
3930 if (tsk_cache_hot) {
3931 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
3932 schedstat_inc(p, se.statistics.nr_forced_migrations);
3938 if (tsk_cache_hot) {
3939 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
3946 * move_one_task tries to move exactly one task from busiest to this_rq, as
3947 * part of active balancing operations within "domain".
3948 * Returns 1 if successful and 0 otherwise.
3950 * Called with both runqueues locked.
3952 static int move_one_task(struct lb_env *env)
3954 struct task_struct *p, *n;
3956 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
3957 if (throttled_lb_pair(task_group(p), env->src_rq->cpu, env->dst_cpu))
3960 if (!can_migrate_task(p, env))
3965 * Right now, this is only the second place move_task()
3966 * is called, so we can safely collect move_task()
3967 * stats here rather than inside move_task().
3969 schedstat_inc(env->sd, lb_gained[env->idle]);
3975 static unsigned long task_h_load(struct task_struct *p);
3977 static const unsigned int sched_nr_migrate_break = 32;
3980 * move_tasks tries to move up to imbalance weighted load from busiest to
3981 * this_rq, as part of a balancing operation within domain "sd".
3982 * Returns 1 if successful and 0 otherwise.
3984 * Called with both runqueues locked.
3986 static int move_tasks(struct lb_env *env)
3988 struct list_head *tasks = &env->src_rq->cfs_tasks;
3989 struct task_struct *p;
3993 if (env->imbalance <= 0)
3996 while (!list_empty(tasks)) {
3997 p = list_first_entry(tasks, struct task_struct, se.group_node);
4000 /* We've more or less seen every task there is, call it quits */
4001 if (env->loop > env->loop_max)
4004 /* take a breather every nr_migrate tasks */
4005 if (env->loop > env->loop_break) {
4006 env->loop_break += sched_nr_migrate_break;
4007 env->flags |= LBF_NEED_BREAK;
4011 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4014 load = task_h_load(p);
4016 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4019 if ((load / 2) > env->imbalance)
4022 if (!can_migrate_task(p, env))
4027 env->imbalance -= load;
4029 #ifdef CONFIG_PREEMPT
4031 * NEWIDLE balancing is a source of latency, so preemptible
4032 * kernels will stop after the first task is pulled to minimize
4033 * the critical section.
4035 if (env->idle == CPU_NEWLY_IDLE)
4040 * We only want to steal up to the prescribed amount of
4043 if (env->imbalance <= 0)
4048 list_move_tail(&p->se.group_node, tasks);
4052 * Right now, this is one of only two places move_task() is called,
4053 * so we can safely collect move_task() stats here rather than
4054 * inside move_task().
4056 schedstat_add(env->sd, lb_gained[env->idle], pulled);
4061 #ifdef CONFIG_FAIR_GROUP_SCHED
4063 * update tg->load_weight by folding this cpu's load_avg
4065 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
4067 struct sched_entity *se = tg->se[cpu];
4068 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
4070 /* throttled entities do not contribute to load */
4071 if (throttled_hierarchy(cfs_rq))
4074 update_cfs_rq_blocked_load(cfs_rq, 1);
4077 update_entity_load_avg(se, 1);
4079 * We pivot on our runnable average having decayed to zero for
4080 * list removal. This generally implies that all our children
4081 * have also been removed (modulo rounding error or bandwidth
4082 * control); however, such cases are rare and we can fix these
4085 * TODO: fix up out-of-order children on enqueue.
4087 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
4088 list_del_leaf_cfs_rq(cfs_rq);
4090 struct rq *rq = rq_of(cfs_rq);
4091 update_rq_runnable_avg(rq, rq->nr_running);
4095 static void update_blocked_averages(int cpu)
4097 struct rq *rq = cpu_rq(cpu);
4098 struct cfs_rq *cfs_rq;
4099 unsigned long flags;
4101 raw_spin_lock_irqsave(&rq->lock, flags);
4102 update_rq_clock(rq);
4104 * Iterates the task_group tree in a bottom up fashion, see
4105 * list_add_leaf_cfs_rq() for details.
4107 for_each_leaf_cfs_rq(rq, cfs_rq) {
4109 * Note: We may want to consider periodically releasing
4110 * rq->lock about these updates so that creating many task
4111 * groups does not result in continually extending hold time.
4113 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
4116 raw_spin_unlock_irqrestore(&rq->lock, flags);
4120 * Compute the cpu's hierarchical load factor for each task group.
4121 * This needs to be done in a top-down fashion because the load of a child
4122 * group is a fraction of its parents load.
4124 static int tg_load_down(struct task_group *tg, void *data)
4127 long cpu = (long)data;
4130 load = cpu_rq(cpu)->load.weight;
4132 load = tg->parent->cfs_rq[cpu]->h_load;
4133 load *= tg->se[cpu]->load.weight;
4134 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
4137 tg->cfs_rq[cpu]->h_load = load;
4142 static void update_h_load(long cpu)
4144 struct rq *rq = cpu_rq(cpu);
4145 unsigned long now = jiffies;
4147 if (rq->h_load_throttle == now)
4150 rq->h_load_throttle = now;
4153 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
4157 static unsigned long task_h_load(struct task_struct *p)
4159 struct cfs_rq *cfs_rq = task_cfs_rq(p);
4162 load = p->se.load.weight;
4163 load = div_u64(load * cfs_rq->h_load, cfs_rq->load.weight + 1);
4168 static inline void update_blocked_averages(int cpu)
4172 static inline void update_h_load(long cpu)
4176 static unsigned long task_h_load(struct task_struct *p)
4178 return p->se.load.weight;
4182 /********** Helpers for find_busiest_group ************************/
4184 * sd_lb_stats - Structure to store the statistics of a sched_domain
4185 * during load balancing.
4187 struct sd_lb_stats {
4188 struct sched_group *busiest; /* Busiest group in this sd */
4189 struct sched_group *this; /* Local group in this sd */
4190 unsigned long total_load; /* Total load of all groups in sd */
4191 unsigned long total_pwr; /* Total power of all groups in sd */
4192 unsigned long avg_load; /* Average load across all groups in sd */
4194 /** Statistics of this group */
4195 unsigned long this_load;
4196 unsigned long this_load_per_task;
4197 unsigned long this_nr_running;
4198 unsigned long this_has_capacity;
4199 unsigned int this_idle_cpus;
4201 /* Statistics of the busiest group */
4202 unsigned int busiest_idle_cpus;
4203 unsigned long max_load;
4204 unsigned long busiest_load_per_task;
4205 unsigned long busiest_nr_running;
4206 unsigned long busiest_group_capacity;
4207 unsigned long busiest_has_capacity;
4208 unsigned int busiest_group_weight;
4210 int group_imb; /* Is there imbalance in this sd */
4214 * sg_lb_stats - stats of a sched_group required for load_balancing
4216 struct sg_lb_stats {
4217 unsigned long avg_load; /*Avg load across the CPUs of the group */
4218 unsigned long group_load; /* Total load over the CPUs of the group */
4219 unsigned long sum_nr_running; /* Nr tasks running in the group */
4220 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
4221 unsigned long group_capacity;
4222 unsigned long idle_cpus;
4223 unsigned long group_weight;
4224 int group_imb; /* Is there an imbalance in the group ? */
4225 int group_has_capacity; /* Is there extra capacity in the group? */
4229 * get_sd_load_idx - Obtain the load index for a given sched domain.
4230 * @sd: The sched_domain whose load_idx is to be obtained.
4231 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4233 static inline int get_sd_load_idx(struct sched_domain *sd,
4234 enum cpu_idle_type idle)
4240 load_idx = sd->busy_idx;
4243 case CPU_NEWLY_IDLE:
4244 load_idx = sd->newidle_idx;
4247 load_idx = sd->idle_idx;
4254 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
4256 return SCHED_POWER_SCALE;
4259 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
4261 return default_scale_freq_power(sd, cpu);
4264 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
4266 unsigned long weight = sd->span_weight;
4267 unsigned long smt_gain = sd->smt_gain;
4274 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
4276 return default_scale_smt_power(sd, cpu);
4279 unsigned long scale_rt_power(int cpu)
4281 struct rq *rq = cpu_rq(cpu);
4282 u64 total, available, age_stamp, avg;
4285 * Since we're reading these variables without serialization make sure
4286 * we read them once before doing sanity checks on them.
4288 age_stamp = ACCESS_ONCE(rq->age_stamp);
4289 avg = ACCESS_ONCE(rq->rt_avg);
4291 total = sched_avg_period() + (rq->clock - age_stamp);
4293 if (unlikely(total < avg)) {
4294 /* Ensures that power won't end up being negative */
4297 available = total - avg;
4300 if (unlikely((s64)total < SCHED_POWER_SCALE))
4301 total = SCHED_POWER_SCALE;
4303 total >>= SCHED_POWER_SHIFT;
4305 return div_u64(available, total);
4308 static void update_cpu_power(struct sched_domain *sd, int cpu)
4310 unsigned long weight = sd->span_weight;
4311 unsigned long power = SCHED_POWER_SCALE;
4312 struct sched_group *sdg = sd->groups;
4314 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
4315 if (sched_feat(ARCH_POWER))
4316 power *= arch_scale_smt_power(sd, cpu);
4318 power *= default_scale_smt_power(sd, cpu);
4320 power >>= SCHED_POWER_SHIFT;
4323 sdg->sgp->power_orig = power;
4325 if (sched_feat(ARCH_POWER))
4326 power *= arch_scale_freq_power(sd, cpu);
4328 power *= default_scale_freq_power(sd, cpu);
4330 power >>= SCHED_POWER_SHIFT;
4332 power *= scale_rt_power(cpu);
4333 power >>= SCHED_POWER_SHIFT;
4338 cpu_rq(cpu)->cpu_power = power;
4339 sdg->sgp->power = power;
4342 void update_group_power(struct sched_domain *sd, int cpu)
4344 struct sched_domain *child = sd->child;
4345 struct sched_group *group, *sdg = sd->groups;
4346 unsigned long power;
4347 unsigned long interval;
4349 interval = msecs_to_jiffies(sd->balance_interval);
4350 interval = clamp(interval, 1UL, max_load_balance_interval);
4351 sdg->sgp->next_update = jiffies + interval;
4354 update_cpu_power(sd, cpu);
4360 if (child->flags & SD_OVERLAP) {
4362 * SD_OVERLAP domains cannot assume that child groups
4363 * span the current group.
4366 for_each_cpu(cpu, sched_group_cpus(sdg))
4367 power += power_of(cpu);
4370 * !SD_OVERLAP domains can assume that child groups
4371 * span the current group.
4374 group = child->groups;
4376 power += group->sgp->power;
4377 group = group->next;
4378 } while (group != child->groups);
4381 sdg->sgp->power_orig = sdg->sgp->power = power;
4385 * Try and fix up capacity for tiny siblings, this is needed when
4386 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4387 * which on its own isn't powerful enough.
4389 * See update_sd_pick_busiest() and check_asym_packing().
4392 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
4395 * Only siblings can have significantly less than SCHED_POWER_SCALE
4397 if (!(sd->flags & SD_SHARE_CPUPOWER))
4401 * If ~90% of the cpu_power is still there, we're good.
4403 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
4410 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4411 * @env: The load balancing environment.
4412 * @group: sched_group whose statistics are to be updated.
4413 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4414 * @local_group: Does group contain this_cpu.
4415 * @balance: Should we balance.
4416 * @sgs: variable to hold the statistics for this group.
4418 static inline void update_sg_lb_stats(struct lb_env *env,
4419 struct sched_group *group, int load_idx,
4420 int local_group, int *balance, struct sg_lb_stats *sgs)
4422 unsigned long nr_running, max_nr_running, min_nr_running;
4423 unsigned long load, max_cpu_load, min_cpu_load;
4424 unsigned int balance_cpu = -1, first_idle_cpu = 0;
4425 unsigned long avg_load_per_task = 0;
4429 balance_cpu = group_balance_cpu(group);
4431 /* Tally up the load of all CPUs in the group */
4433 min_cpu_load = ~0UL;
4435 min_nr_running = ~0UL;
4437 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
4438 struct rq *rq = cpu_rq(i);
4440 nr_running = rq->nr_running;
4442 /* Bias balancing toward cpus of our domain */
4444 if (idle_cpu(i) && !first_idle_cpu &&
4445 cpumask_test_cpu(i, sched_group_mask(group))) {
4450 load = target_load(i, load_idx);
4452 load = source_load(i, load_idx);
4453 if (load > max_cpu_load)
4454 max_cpu_load = load;
4455 if (min_cpu_load > load)
4456 min_cpu_load = load;
4458 if (nr_running > max_nr_running)
4459 max_nr_running = nr_running;
4460 if (min_nr_running > nr_running)
4461 min_nr_running = nr_running;
4464 sgs->group_load += load;
4465 sgs->sum_nr_running += nr_running;
4466 sgs->sum_weighted_load += weighted_cpuload(i);
4472 * First idle cpu or the first cpu(busiest) in this sched group
4473 * is eligible for doing load balancing at this and above
4474 * domains. In the newly idle case, we will allow all the cpu's
4475 * to do the newly idle load balance.
4478 if (env->idle != CPU_NEWLY_IDLE) {
4479 if (balance_cpu != env->dst_cpu) {
4483 update_group_power(env->sd, env->dst_cpu);
4484 } else if (time_after_eq(jiffies, group->sgp->next_update))
4485 update_group_power(env->sd, env->dst_cpu);
4488 /* Adjust by relative CPU power of the group */
4489 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
4492 * Consider the group unbalanced when the imbalance is larger
4493 * than the average weight of a task.
4495 * APZ: with cgroup the avg task weight can vary wildly and
4496 * might not be a suitable number - should we keep a
4497 * normalized nr_running number somewhere that negates
4500 if (sgs->sum_nr_running)
4501 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
4503 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task &&
4504 (max_nr_running - min_nr_running) > 1)
4507 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
4509 if (!sgs->group_capacity)
4510 sgs->group_capacity = fix_small_capacity(env->sd, group);
4511 sgs->group_weight = group->group_weight;
4513 if (sgs->group_capacity > sgs->sum_nr_running)
4514 sgs->group_has_capacity = 1;
4518 * update_sd_pick_busiest - return 1 on busiest group
4519 * @env: The load balancing environment.
4520 * @sds: sched_domain statistics
4521 * @sg: sched_group candidate to be checked for being the busiest
4522 * @sgs: sched_group statistics
4524 * Determine if @sg is a busier group than the previously selected
4527 static bool update_sd_pick_busiest(struct lb_env *env,
4528 struct sd_lb_stats *sds,
4529 struct sched_group *sg,
4530 struct sg_lb_stats *sgs)
4532 if (sgs->avg_load <= sds->max_load)
4535 if (sgs->sum_nr_running > sgs->group_capacity)
4542 * ASYM_PACKING needs to move all the work to the lowest
4543 * numbered CPUs in the group, therefore mark all groups
4544 * higher than ourself as busy.
4546 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
4547 env->dst_cpu < group_first_cpu(sg)) {
4551 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
4559 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
4560 * @env: The load balancing environment.
4561 * @balance: Should we balance.
4562 * @sds: variable to hold the statistics for this sched_domain.
4564 static inline void update_sd_lb_stats(struct lb_env *env,
4565 int *balance, struct sd_lb_stats *sds)
4567 struct sched_domain *child = env->sd->child;
4568 struct sched_group *sg = env->sd->groups;
4569 struct sg_lb_stats sgs;
4570 int load_idx, prefer_sibling = 0;
4572 if (child && child->flags & SD_PREFER_SIBLING)
4575 load_idx = get_sd_load_idx(env->sd, env->idle);
4580 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
4581 memset(&sgs, 0, sizeof(sgs));
4582 update_sg_lb_stats(env, sg, load_idx, local_group, balance, &sgs);
4584 if (local_group && !(*balance))
4587 sds->total_load += sgs.group_load;
4588 sds->total_pwr += sg->sgp->power;
4591 * In case the child domain prefers tasks go to siblings
4592 * first, lower the sg capacity to one so that we'll try
4593 * and move all the excess tasks away. We lower the capacity
4594 * of a group only if the local group has the capacity to fit
4595 * these excess tasks, i.e. nr_running < group_capacity. The
4596 * extra check prevents the case where you always pull from the
4597 * heaviest group when it is already under-utilized (possible
4598 * with a large weight task outweighs the tasks on the system).
4600 if (prefer_sibling && !local_group && sds->this_has_capacity)
4601 sgs.group_capacity = min(sgs.group_capacity, 1UL);
4604 sds->this_load = sgs.avg_load;
4606 sds->this_nr_running = sgs.sum_nr_running;
4607 sds->this_load_per_task = sgs.sum_weighted_load;
4608 sds->this_has_capacity = sgs.group_has_capacity;
4609 sds->this_idle_cpus = sgs.idle_cpus;
4610 } else if (update_sd_pick_busiest(env, sds, sg, &sgs)) {
4611 sds->max_load = sgs.avg_load;
4613 sds->busiest_nr_running = sgs.sum_nr_running;
4614 sds->busiest_idle_cpus = sgs.idle_cpus;
4615 sds->busiest_group_capacity = sgs.group_capacity;
4616 sds->busiest_load_per_task = sgs.sum_weighted_load;
4617 sds->busiest_has_capacity = sgs.group_has_capacity;
4618 sds->busiest_group_weight = sgs.group_weight;
4619 sds->group_imb = sgs.group_imb;
4623 } while (sg != env->sd->groups);
4627 * check_asym_packing - Check to see if the group is packed into the
4630 * This is primarily intended to used at the sibling level. Some
4631 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4632 * case of POWER7, it can move to lower SMT modes only when higher
4633 * threads are idle. When in lower SMT modes, the threads will
4634 * perform better since they share less core resources. Hence when we
4635 * have idle threads, we want them to be the higher ones.
4637 * This packing function is run on idle threads. It checks to see if
4638 * the busiest CPU in this domain (core in the P7 case) has a higher
4639 * CPU number than the packing function is being run on. Here we are
4640 * assuming lower CPU number will be equivalent to lower a SMT thread
4643 * Returns 1 when packing is required and a task should be moved to
4644 * this CPU. The amount of the imbalance is returned in *imbalance.
4646 * @env: The load balancing environment.
4647 * @sds: Statistics of the sched_domain which is to be packed
4649 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
4653 if (!(env->sd->flags & SD_ASYM_PACKING))
4659 busiest_cpu = group_first_cpu(sds->busiest);
4660 if (env->dst_cpu > busiest_cpu)
4663 env->imbalance = DIV_ROUND_CLOSEST(
4664 sds->max_load * sds->busiest->sgp->power, SCHED_POWER_SCALE);
4670 * fix_small_imbalance - Calculate the minor imbalance that exists
4671 * amongst the groups of a sched_domain, during
4673 * @env: The load balancing environment.
4674 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4677 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4679 unsigned long tmp, pwr_now = 0, pwr_move = 0;
4680 unsigned int imbn = 2;
4681 unsigned long scaled_busy_load_per_task;
4683 if (sds->this_nr_running) {
4684 sds->this_load_per_task /= sds->this_nr_running;
4685 if (sds->busiest_load_per_task >
4686 sds->this_load_per_task)
4689 sds->this_load_per_task =
4690 cpu_avg_load_per_task(env->dst_cpu);
4693 scaled_busy_load_per_task = sds->busiest_load_per_task
4694 * SCHED_POWER_SCALE;
4695 scaled_busy_load_per_task /= sds->busiest->sgp->power;
4697 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
4698 (scaled_busy_load_per_task * imbn)) {
4699 env->imbalance = sds->busiest_load_per_task;
4704 * OK, we don't have enough imbalance to justify moving tasks,
4705 * however we may be able to increase total CPU power used by
4709 pwr_now += sds->busiest->sgp->power *
4710 min(sds->busiest_load_per_task, sds->max_load);
4711 pwr_now += sds->this->sgp->power *
4712 min(sds->this_load_per_task, sds->this_load);
4713 pwr_now /= SCHED_POWER_SCALE;
4715 /* Amount of load we'd subtract */
4716 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4717 sds->busiest->sgp->power;
4718 if (sds->max_load > tmp)
4719 pwr_move += sds->busiest->sgp->power *
4720 min(sds->busiest_load_per_task, sds->max_load - tmp);
4722 /* Amount of load we'd add */
4723 if (sds->max_load * sds->busiest->sgp->power <
4724 sds->busiest_load_per_task * SCHED_POWER_SCALE)
4725 tmp = (sds->max_load * sds->busiest->sgp->power) /
4726 sds->this->sgp->power;
4728 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4729 sds->this->sgp->power;
4730 pwr_move += sds->this->sgp->power *
4731 min(sds->this_load_per_task, sds->this_load + tmp);
4732 pwr_move /= SCHED_POWER_SCALE;
4734 /* Move if we gain throughput */
4735 if (pwr_move > pwr_now)
4736 env->imbalance = sds->busiest_load_per_task;
4740 * calculate_imbalance - Calculate the amount of imbalance present within the
4741 * groups of a given sched_domain during load balance.
4742 * @env: load balance environment
4743 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4745 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4747 unsigned long max_pull, load_above_capacity = ~0UL;
4749 sds->busiest_load_per_task /= sds->busiest_nr_running;
4750 if (sds->group_imb) {
4751 sds->busiest_load_per_task =
4752 min(sds->busiest_load_per_task, sds->avg_load);
4756 * In the presence of smp nice balancing, certain scenarios can have
4757 * max load less than avg load(as we skip the groups at or below
4758 * its cpu_power, while calculating max_load..)
4760 if (sds->max_load < sds->avg_load) {
4762 return fix_small_imbalance(env, sds);
4765 if (!sds->group_imb) {
4767 * Don't want to pull so many tasks that a group would go idle.
4769 load_above_capacity = (sds->busiest_nr_running -
4770 sds->busiest_group_capacity);
4772 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
4774 load_above_capacity /= sds->busiest->sgp->power;
4778 * We're trying to get all the cpus to the average_load, so we don't
4779 * want to push ourselves above the average load, nor do we wish to
4780 * reduce the max loaded cpu below the average load. At the same time,
4781 * we also don't want to reduce the group load below the group capacity
4782 * (so that we can implement power-savings policies etc). Thus we look
4783 * for the minimum possible imbalance.
4784 * Be careful of negative numbers as they'll appear as very large values
4785 * with unsigned longs.
4787 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
4789 /* How much load to actually move to equalise the imbalance */
4790 env->imbalance = min(max_pull * sds->busiest->sgp->power,
4791 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
4792 / SCHED_POWER_SCALE;
4795 * if *imbalance is less than the average load per runnable task
4796 * there is no guarantee that any tasks will be moved so we'll have
4797 * a think about bumping its value to force at least one task to be
4800 if (env->imbalance < sds->busiest_load_per_task)
4801 return fix_small_imbalance(env, sds);
4805 /******* find_busiest_group() helpers end here *********************/
4808 * find_busiest_group - Returns the busiest group within the sched_domain
4809 * if there is an imbalance. If there isn't an imbalance, and
4810 * the user has opted for power-savings, it returns a group whose
4811 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4812 * such a group exists.
4814 * Also calculates the amount of weighted load which should be moved
4815 * to restore balance.
4817 * @env: The load balancing environment.
4818 * @balance: Pointer to a variable indicating if this_cpu
4819 * is the appropriate cpu to perform load balancing at this_level.
4821 * Returns: - the busiest group if imbalance exists.
4822 * - If no imbalance and user has opted for power-savings balance,
4823 * return the least loaded group whose CPUs can be
4824 * put to idle by rebalancing its tasks onto our group.
4826 static struct sched_group *
4827 find_busiest_group(struct lb_env *env, int *balance)
4829 struct sd_lb_stats sds;
4831 memset(&sds, 0, sizeof(sds));
4834 * Compute the various statistics relavent for load balancing at
4837 update_sd_lb_stats(env, balance, &sds);
4840 * this_cpu is not the appropriate cpu to perform load balancing at
4846 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
4847 check_asym_packing(env, &sds))
4850 /* There is no busy sibling group to pull tasks from */
4851 if (!sds.busiest || sds.busiest_nr_running == 0)
4854 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
4857 * If the busiest group is imbalanced the below checks don't
4858 * work because they assumes all things are equal, which typically
4859 * isn't true due to cpus_allowed constraints and the like.
4864 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
4865 if (env->idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
4866 !sds.busiest_has_capacity)
4870 * If the local group is more busy than the selected busiest group
4871 * don't try and pull any tasks.
4873 if (sds.this_load >= sds.max_load)
4877 * Don't pull any tasks if this group is already above the domain
4880 if (sds.this_load >= sds.avg_load)
4883 if (env->idle == CPU_IDLE) {
4885 * This cpu is idle. If the busiest group load doesn't
4886 * have more tasks than the number of available cpu's and
4887 * there is no imbalance between this and busiest group
4888 * wrt to idle cpu's, it is balanced.
4890 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
4891 sds.busiest_nr_running <= sds.busiest_group_weight)
4895 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
4896 * imbalance_pct to be conservative.
4898 if (100 * sds.max_load <= env->sd->imbalance_pct * sds.this_load)
4903 /* Looks like there is an imbalance. Compute it */
4904 calculate_imbalance(env, &sds);
4914 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4916 static struct rq *find_busiest_queue(struct lb_env *env,
4917 struct sched_group *group)
4919 struct rq *busiest = NULL, *rq;
4920 unsigned long max_load = 0;
4923 for_each_cpu(i, sched_group_cpus(group)) {
4924 unsigned long power = power_of(i);
4925 unsigned long capacity = DIV_ROUND_CLOSEST(power,
4930 capacity = fix_small_capacity(env->sd, group);
4932 if (!cpumask_test_cpu(i, env->cpus))
4936 wl = weighted_cpuload(i);
4939 * When comparing with imbalance, use weighted_cpuload()
4940 * which is not scaled with the cpu power.
4942 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
4946 * For the load comparisons with the other cpu's, consider
4947 * the weighted_cpuload() scaled with the cpu power, so that
4948 * the load can be moved away from the cpu that is potentially
4949 * running at a lower capacity.
4951 wl = (wl * SCHED_POWER_SCALE) / power;
4953 if (wl > max_load) {
4963 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4964 * so long as it is large enough.
4966 #define MAX_PINNED_INTERVAL 512
4968 /* Working cpumask for load_balance and load_balance_newidle. */
4969 DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4971 static int need_active_balance(struct lb_env *env)
4973 struct sched_domain *sd = env->sd;
4975 if (env->idle == CPU_NEWLY_IDLE) {
4978 * ASYM_PACKING needs to force migrate tasks from busy but
4979 * higher numbered CPUs in order to pack all tasks in the
4980 * lowest numbered CPUs.
4982 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
4986 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
4989 static int active_load_balance_cpu_stop(void *data);
4992 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4993 * tasks if there is an imbalance.
4995 static int load_balance(int this_cpu, struct rq *this_rq,
4996 struct sched_domain *sd, enum cpu_idle_type idle,
4999 int ld_moved, cur_ld_moved, active_balance = 0;
5000 int lb_iterations, max_lb_iterations;
5001 struct sched_group *group;
5003 unsigned long flags;
5004 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
5006 struct lb_env env = {
5008 .dst_cpu = this_cpu,
5010 .dst_grpmask = sched_group_cpus(sd->groups),
5012 .loop_break = sched_nr_migrate_break,
5016 cpumask_copy(cpus, cpu_active_mask);
5017 max_lb_iterations = cpumask_weight(env.dst_grpmask);
5019 schedstat_inc(sd, lb_count[idle]);
5022 group = find_busiest_group(&env, balance);
5028 schedstat_inc(sd, lb_nobusyg[idle]);
5032 busiest = find_busiest_queue(&env, group);
5034 schedstat_inc(sd, lb_nobusyq[idle]);
5038 BUG_ON(busiest == env.dst_rq);
5040 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
5044 if (busiest->nr_running > 1) {
5046 * Attempt to move tasks. If find_busiest_group has found
5047 * an imbalance but busiest->nr_running <= 1, the group is
5048 * still unbalanced. ld_moved simply stays zero, so it is
5049 * correctly treated as an imbalance.
5051 env.flags |= LBF_ALL_PINNED;
5052 env.src_cpu = busiest->cpu;
5053 env.src_rq = busiest;
5054 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
5056 update_h_load(env.src_cpu);
5058 local_irq_save(flags);
5059 double_rq_lock(env.dst_rq, busiest);
5062 * cur_ld_moved - load moved in current iteration
5063 * ld_moved - cumulative load moved across iterations
5065 cur_ld_moved = move_tasks(&env);
5066 ld_moved += cur_ld_moved;
5067 double_rq_unlock(env.dst_rq, busiest);
5068 local_irq_restore(flags);
5070 if (env.flags & LBF_NEED_BREAK) {
5071 env.flags &= ~LBF_NEED_BREAK;
5076 * some other cpu did the load balance for us.
5078 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
5079 resched_cpu(env.dst_cpu);
5082 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5083 * us and move them to an alternate dst_cpu in our sched_group
5084 * where they can run. The upper limit on how many times we
5085 * iterate on same src_cpu is dependent on number of cpus in our
5088 * This changes load balance semantics a bit on who can move
5089 * load to a given_cpu. In addition to the given_cpu itself
5090 * (or a ilb_cpu acting on its behalf where given_cpu is
5091 * nohz-idle), we now have balance_cpu in a position to move
5092 * load to given_cpu. In rare situations, this may cause
5093 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5094 * _independently_ and at _same_ time to move some load to
5095 * given_cpu) causing exceess load to be moved to given_cpu.
5096 * This however should not happen so much in practice and
5097 * moreover subsequent load balance cycles should correct the
5098 * excess load moved.
5100 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0 &&
5101 lb_iterations++ < max_lb_iterations) {
5103 env.dst_rq = cpu_rq(env.new_dst_cpu);
5104 env.dst_cpu = env.new_dst_cpu;
5105 env.flags &= ~LBF_SOME_PINNED;
5107 env.loop_break = sched_nr_migrate_break;
5109 * Go back to "more_balance" rather than "redo" since we
5110 * need to continue with same src_cpu.
5115 /* All tasks on this runqueue were pinned by CPU affinity */
5116 if (unlikely(env.flags & LBF_ALL_PINNED)) {
5117 cpumask_clear_cpu(cpu_of(busiest), cpus);
5118 if (!cpumask_empty(cpus)) {
5120 env.loop_break = sched_nr_migrate_break;
5128 schedstat_inc(sd, lb_failed[idle]);
5130 * Increment the failure counter only on periodic balance.
5131 * We do not want newidle balance, which can be very
5132 * frequent, pollute the failure counter causing
5133 * excessive cache_hot migrations and active balances.
5135 if (idle != CPU_NEWLY_IDLE)
5136 sd->nr_balance_failed++;
5138 if (need_active_balance(&env)) {
5139 raw_spin_lock_irqsave(&busiest->lock, flags);
5141 /* don't kick the active_load_balance_cpu_stop,
5142 * if the curr task on busiest cpu can't be
5145 if (!cpumask_test_cpu(this_cpu,
5146 tsk_cpus_allowed(busiest->curr))) {
5147 raw_spin_unlock_irqrestore(&busiest->lock,
5149 env.flags |= LBF_ALL_PINNED;
5150 goto out_one_pinned;
5154 * ->active_balance synchronizes accesses to
5155 * ->active_balance_work. Once set, it's cleared
5156 * only after active load balance is finished.
5158 if (!busiest->active_balance) {
5159 busiest->active_balance = 1;
5160 busiest->push_cpu = this_cpu;
5163 raw_spin_unlock_irqrestore(&busiest->lock, flags);
5165 if (active_balance) {
5166 stop_one_cpu_nowait(cpu_of(busiest),
5167 active_load_balance_cpu_stop, busiest,
5168 &busiest->active_balance_work);
5172 * We've kicked active balancing, reset the failure
5175 sd->nr_balance_failed = sd->cache_nice_tries+1;
5178 sd->nr_balance_failed = 0;
5180 if (likely(!active_balance)) {
5181 /* We were unbalanced, so reset the balancing interval */
5182 sd->balance_interval = sd->min_interval;
5185 * If we've begun active balancing, start to back off. This
5186 * case may not be covered by the all_pinned logic if there
5187 * is only 1 task on the busy runqueue (because we don't call
5190 if (sd->balance_interval < sd->max_interval)
5191 sd->balance_interval *= 2;
5197 schedstat_inc(sd, lb_balanced[idle]);
5199 sd->nr_balance_failed = 0;
5202 /* tune up the balancing interval */
5203 if (((env.flags & LBF_ALL_PINNED) &&
5204 sd->balance_interval < MAX_PINNED_INTERVAL) ||
5205 (sd->balance_interval < sd->max_interval))
5206 sd->balance_interval *= 2;
5214 * idle_balance is called by schedule() if this_cpu is about to become
5215 * idle. Attempts to pull tasks from other CPUs.
5217 void idle_balance(int this_cpu, struct rq *this_rq)
5219 struct sched_domain *sd;
5220 int pulled_task = 0;
5221 unsigned long next_balance = jiffies + HZ;
5223 this_rq->idle_stamp = this_rq->clock;
5225 if (this_rq->avg_idle < sysctl_sched_migration_cost)
5228 update_rq_runnable_avg(this_rq, 1);
5231 * Drop the rq->lock, but keep IRQ/preempt disabled.
5233 raw_spin_unlock(&this_rq->lock);
5235 update_blocked_averages(this_cpu);
5237 for_each_domain(this_cpu, sd) {
5238 unsigned long interval;
5241 if (!(sd->flags & SD_LOAD_BALANCE))
5244 if (sd->flags & SD_BALANCE_NEWIDLE) {
5245 /* If we've pulled tasks over stop searching: */
5246 pulled_task = load_balance(this_cpu, this_rq,
5247 sd, CPU_NEWLY_IDLE, &balance);
5250 interval = msecs_to_jiffies(sd->balance_interval);
5251 if (time_after(next_balance, sd->last_balance + interval))
5252 next_balance = sd->last_balance + interval;
5254 this_rq->idle_stamp = 0;
5260 raw_spin_lock(&this_rq->lock);
5262 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
5264 * We are going idle. next_balance may be set based on
5265 * a busy processor. So reset next_balance.
5267 this_rq->next_balance = next_balance;
5272 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5273 * running tasks off the busiest CPU onto idle CPUs. It requires at
5274 * least 1 task to be running on each physical CPU where possible, and
5275 * avoids physical / logical imbalances.
5277 static int active_load_balance_cpu_stop(void *data)
5279 struct rq *busiest_rq = data;
5280 int busiest_cpu = cpu_of(busiest_rq);
5281 int target_cpu = busiest_rq->push_cpu;
5282 struct rq *target_rq = cpu_rq(target_cpu);
5283 struct sched_domain *sd;
5285 raw_spin_lock_irq(&busiest_rq->lock);
5287 /* make sure the requested cpu hasn't gone down in the meantime */
5288 if (unlikely(busiest_cpu != smp_processor_id() ||
5289 !busiest_rq->active_balance))
5292 /* Is there any task to move? */
5293 if (busiest_rq->nr_running <= 1)
5297 * This condition is "impossible", if it occurs
5298 * we need to fix it. Originally reported by
5299 * Bjorn Helgaas on a 128-cpu setup.
5301 BUG_ON(busiest_rq == target_rq);
5303 /* move a task from busiest_rq to target_rq */
5304 double_lock_balance(busiest_rq, target_rq);
5306 /* Search for an sd spanning us and the target CPU. */
5308 for_each_domain(target_cpu, sd) {
5309 if ((sd->flags & SD_LOAD_BALANCE) &&
5310 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
5315 struct lb_env env = {
5317 .dst_cpu = target_cpu,
5318 .dst_rq = target_rq,
5319 .src_cpu = busiest_rq->cpu,
5320 .src_rq = busiest_rq,
5324 schedstat_inc(sd, alb_count);
5326 if (move_one_task(&env))
5327 schedstat_inc(sd, alb_pushed);
5329 schedstat_inc(sd, alb_failed);
5332 double_unlock_balance(busiest_rq, target_rq);
5334 busiest_rq->active_balance = 0;
5335 raw_spin_unlock_irq(&busiest_rq->lock);
5341 * idle load balancing details
5342 * - When one of the busy CPUs notice that there may be an idle rebalancing
5343 * needed, they will kick the idle load balancer, which then does idle
5344 * load balancing for all the idle CPUs.
5347 cpumask_var_t idle_cpus_mask;
5349 unsigned long next_balance; /* in jiffy units */
5350 } nohz ____cacheline_aligned;
5352 static inline int find_new_ilb(int call_cpu)
5354 int ilb = cpumask_first(nohz.idle_cpus_mask);
5356 if (ilb < nr_cpu_ids && idle_cpu(ilb))
5363 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5364 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5365 * CPU (if there is one).
5367 static void nohz_balancer_kick(int cpu)
5371 nohz.next_balance++;
5373 ilb_cpu = find_new_ilb(cpu);
5375 if (ilb_cpu >= nr_cpu_ids)
5378 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
5381 * Use smp_send_reschedule() instead of resched_cpu().
5382 * This way we generate a sched IPI on the target cpu which
5383 * is idle. And the softirq performing nohz idle load balance
5384 * will be run before returning from the IPI.
5386 smp_send_reschedule(ilb_cpu);
5390 static inline void nohz_balance_exit_idle(int cpu)
5392 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5393 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5394 atomic_dec(&nohz.nr_cpus);
5395 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5399 static inline void set_cpu_sd_state_busy(void)
5401 struct sched_domain *sd;
5402 int cpu = smp_processor_id();
5404 if (!test_bit(NOHZ_IDLE, nohz_flags(cpu)))
5406 clear_bit(NOHZ_IDLE, nohz_flags(cpu));
5409 for_each_domain(cpu, sd)
5410 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
5414 void set_cpu_sd_state_idle(void)
5416 struct sched_domain *sd;
5417 int cpu = smp_processor_id();
5419 if (test_bit(NOHZ_IDLE, nohz_flags(cpu)))
5421 set_bit(NOHZ_IDLE, nohz_flags(cpu));
5424 for_each_domain(cpu, sd)
5425 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
5430 * This routine will record that the cpu is going idle with tick stopped.
5431 * This info will be used in performing idle load balancing in the future.
5433 void nohz_balance_enter_idle(int cpu)
5436 * If this cpu is going down, then nothing needs to be done.
5438 if (!cpu_active(cpu))
5441 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
5444 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
5445 atomic_inc(&nohz.nr_cpus);
5446 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5449 static int __cpuinit sched_ilb_notifier(struct notifier_block *nfb,
5450 unsigned long action, void *hcpu)
5452 switch (action & ~CPU_TASKS_FROZEN) {
5454 nohz_balance_exit_idle(smp_processor_id());
5462 static DEFINE_SPINLOCK(balancing);
5465 * Scale the max load_balance interval with the number of CPUs in the system.
5466 * This trades load-balance latency on larger machines for less cross talk.
5468 void update_max_interval(void)
5470 max_load_balance_interval = HZ*num_online_cpus()/10;
5474 * It checks each scheduling domain to see if it is due to be balanced,
5475 * and initiates a balancing operation if so.
5477 * Balancing parameters are set up in arch_init_sched_domains.
5479 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
5482 struct rq *rq = cpu_rq(cpu);
5483 unsigned long interval;
5484 struct sched_domain *sd;
5485 /* Earliest time when we have to do rebalance again */
5486 unsigned long next_balance = jiffies + 60*HZ;
5487 int update_next_balance = 0;
5490 update_blocked_averages(cpu);
5493 for_each_domain(cpu, sd) {
5494 if (!(sd->flags & SD_LOAD_BALANCE))
5497 interval = sd->balance_interval;
5498 if (idle != CPU_IDLE)
5499 interval *= sd->busy_factor;
5501 /* scale ms to jiffies */
5502 interval = msecs_to_jiffies(interval);
5503 interval = clamp(interval, 1UL, max_load_balance_interval);
5505 need_serialize = sd->flags & SD_SERIALIZE;
5507 if (need_serialize) {
5508 if (!spin_trylock(&balancing))
5512 if (time_after_eq(jiffies, sd->last_balance + interval)) {
5513 if (load_balance(cpu, rq, sd, idle, &balance)) {
5515 * We've pulled tasks over so either we're no
5518 idle = CPU_NOT_IDLE;
5520 sd->last_balance = jiffies;
5523 spin_unlock(&balancing);
5525 if (time_after(next_balance, sd->last_balance + interval)) {
5526 next_balance = sd->last_balance + interval;
5527 update_next_balance = 1;
5531 * Stop the load balance at this level. There is another
5532 * CPU in our sched group which is doing load balancing more
5541 * next_balance will be updated only when there is a need.
5542 * When the cpu is attached to null domain for ex, it will not be
5545 if (likely(update_next_balance))
5546 rq->next_balance = next_balance;
5551 * In CONFIG_NO_HZ case, the idle balance kickee will do the
5552 * rebalancing for all the cpus for whom scheduler ticks are stopped.
5554 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
5556 struct rq *this_rq = cpu_rq(this_cpu);
5560 if (idle != CPU_IDLE ||
5561 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
5564 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
5565 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
5569 * If this cpu gets work to do, stop the load balancing
5570 * work being done for other cpus. Next load
5571 * balancing owner will pick it up.
5576 rq = cpu_rq(balance_cpu);
5578 raw_spin_lock_irq(&rq->lock);
5579 update_rq_clock(rq);
5580 update_idle_cpu_load(rq);
5581 raw_spin_unlock_irq(&rq->lock);
5583 rebalance_domains(balance_cpu, CPU_IDLE);
5585 if (time_after(this_rq->next_balance, rq->next_balance))
5586 this_rq->next_balance = rq->next_balance;
5588 nohz.next_balance = this_rq->next_balance;
5590 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
5594 * Current heuristic for kicking the idle load balancer in the presence
5595 * of an idle cpu is the system.
5596 * - This rq has more than one task.
5597 * - At any scheduler domain level, this cpu's scheduler group has multiple
5598 * busy cpu's exceeding the group's power.
5599 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
5600 * domain span are idle.
5602 static inline int nohz_kick_needed(struct rq *rq, int cpu)
5604 unsigned long now = jiffies;
5605 struct sched_domain *sd;
5607 if (unlikely(idle_cpu(cpu)))
5611 * We may be recently in ticked or tickless idle mode. At the first
5612 * busy tick after returning from idle, we will update the busy stats.
5614 set_cpu_sd_state_busy();
5615 nohz_balance_exit_idle(cpu);
5618 * None are in tickless mode and hence no need for NOHZ idle load
5621 if (likely(!atomic_read(&nohz.nr_cpus)))
5624 if (time_before(now, nohz.next_balance))
5627 if (rq->nr_running >= 2)
5631 for_each_domain(cpu, sd) {
5632 struct sched_group *sg = sd->groups;
5633 struct sched_group_power *sgp = sg->sgp;
5634 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
5636 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
5637 goto need_kick_unlock;
5639 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
5640 && (cpumask_first_and(nohz.idle_cpus_mask,
5641 sched_domain_span(sd)) < cpu))
5642 goto need_kick_unlock;
5644 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
5656 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
5660 * run_rebalance_domains is triggered when needed from the scheduler tick.
5661 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
5663 static void run_rebalance_domains(struct softirq_action *h)
5665 int this_cpu = smp_processor_id();
5666 struct rq *this_rq = cpu_rq(this_cpu);
5667 enum cpu_idle_type idle = this_rq->idle_balance ?
5668 CPU_IDLE : CPU_NOT_IDLE;
5670 rebalance_domains(this_cpu, idle);
5673 * If this cpu has a pending nohz_balance_kick, then do the
5674 * balancing on behalf of the other idle cpus whose ticks are
5677 nohz_idle_balance(this_cpu, idle);
5680 static inline int on_null_domain(int cpu)
5682 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
5686 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
5688 void trigger_load_balance(struct rq *rq, int cpu)
5690 /* Don't need to rebalance while attached to NULL domain */
5691 if (time_after_eq(jiffies, rq->next_balance) &&
5692 likely(!on_null_domain(cpu)))
5693 raise_softirq(SCHED_SOFTIRQ);
5695 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
5696 nohz_balancer_kick(cpu);
5700 static void rq_online_fair(struct rq *rq)
5705 static void rq_offline_fair(struct rq *rq)
5709 /* Ensure any throttled groups are reachable by pick_next_task */
5710 unthrottle_offline_cfs_rqs(rq);
5713 #endif /* CONFIG_SMP */
5716 * scheduler tick hitting a task of our scheduling class:
5718 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
5720 struct cfs_rq *cfs_rq;
5721 struct sched_entity *se = &curr->se;
5723 for_each_sched_entity(se) {
5724 cfs_rq = cfs_rq_of(se);
5725 entity_tick(cfs_rq, se, queued);
5728 if (sched_feat_numa(NUMA))
5729 task_tick_numa(rq, curr);
5731 update_rq_runnable_avg(rq, 1);
5735 * called on fork with the child task as argument from the parent's context
5736 * - child not yet on the tasklist
5737 * - preemption disabled
5739 static void task_fork_fair(struct task_struct *p)
5741 struct cfs_rq *cfs_rq;
5742 struct sched_entity *se = &p->se, *curr;
5743 int this_cpu = smp_processor_id();
5744 struct rq *rq = this_rq();
5745 unsigned long flags;
5747 raw_spin_lock_irqsave(&rq->lock, flags);
5749 update_rq_clock(rq);
5751 cfs_rq = task_cfs_rq(current);
5752 curr = cfs_rq->curr;
5754 if (unlikely(task_cpu(p) != this_cpu)) {
5756 __set_task_cpu(p, this_cpu);
5760 update_curr(cfs_rq);
5763 se->vruntime = curr->vruntime;
5764 place_entity(cfs_rq, se, 1);
5766 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
5768 * Upon rescheduling, sched_class::put_prev_task() will place
5769 * 'current' within the tree based on its new key value.
5771 swap(curr->vruntime, se->vruntime);
5772 resched_task(rq->curr);
5775 se->vruntime -= cfs_rq->min_vruntime;
5777 raw_spin_unlock_irqrestore(&rq->lock, flags);
5781 * Priority of the task has changed. Check to see if we preempt
5785 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
5791 * Reschedule if we are currently running on this runqueue and
5792 * our priority decreased, or if we are not currently running on
5793 * this runqueue and our priority is higher than the current's
5795 if (rq->curr == p) {
5796 if (p->prio > oldprio)
5797 resched_task(rq->curr);
5799 check_preempt_curr(rq, p, 0);
5802 static void switched_from_fair(struct rq *rq, struct task_struct *p)
5804 struct sched_entity *se = &p->se;
5805 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5808 * Ensure the task's vruntime is normalized, so that when its
5809 * switched back to the fair class the enqueue_entity(.flags=0) will
5810 * do the right thing.
5812 * If it was on_rq, then the dequeue_entity(.flags=0) will already
5813 * have normalized the vruntime, if it was !on_rq, then only when
5814 * the task is sleeping will it still have non-normalized vruntime.
5816 if (!se->on_rq && p->state != TASK_RUNNING) {
5818 * Fix up our vruntime so that the current sleep doesn't
5819 * cause 'unlimited' sleep bonus.
5821 place_entity(cfs_rq, se, 0);
5822 se->vruntime -= cfs_rq->min_vruntime;
5825 #if defined(CONFIG_FAIR_GROUP_SCHED) && defined(CONFIG_SMP)
5827 * Remove our load from contribution when we leave sched_fair
5828 * and ensure we don't carry in an old decay_count if we
5831 if (p->se.avg.decay_count) {
5832 struct cfs_rq *cfs_rq = cfs_rq_of(&p->se);
5833 __synchronize_entity_decay(&p->se);
5834 subtract_blocked_load_contrib(cfs_rq,
5835 p->se.avg.load_avg_contrib);
5841 * We switched to the sched_fair class.
5843 static void switched_to_fair(struct rq *rq, struct task_struct *p)
5849 * We were most likely switched from sched_rt, so
5850 * kick off the schedule if running, otherwise just see
5851 * if we can still preempt the current task.
5854 resched_task(rq->curr);
5856 check_preempt_curr(rq, p, 0);
5859 /* Account for a task changing its policy or group.
5861 * This routine is mostly called to set cfs_rq->curr field when a task
5862 * migrates between groups/classes.
5864 static void set_curr_task_fair(struct rq *rq)
5866 struct sched_entity *se = &rq->curr->se;
5868 for_each_sched_entity(se) {
5869 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5871 set_next_entity(cfs_rq, se);
5872 /* ensure bandwidth has been allocated on our new cfs_rq */
5873 account_cfs_rq_runtime(cfs_rq, 0);
5877 void init_cfs_rq(struct cfs_rq *cfs_rq)
5879 cfs_rq->tasks_timeline = RB_ROOT;
5880 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
5881 #ifndef CONFIG_64BIT
5882 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
5884 #if defined(CONFIG_FAIR_GROUP_SCHED) && defined(CONFIG_SMP)
5885 atomic64_set(&cfs_rq->decay_counter, 1);
5886 atomic64_set(&cfs_rq->removed_load, 0);
5890 #ifdef CONFIG_FAIR_GROUP_SCHED
5891 static void task_move_group_fair(struct task_struct *p, int on_rq)
5893 struct cfs_rq *cfs_rq;
5895 * If the task was not on the rq at the time of this cgroup movement
5896 * it must have been asleep, sleeping tasks keep their ->vruntime
5897 * absolute on their old rq until wakeup (needed for the fair sleeper
5898 * bonus in place_entity()).
5900 * If it was on the rq, we've just 'preempted' it, which does convert
5901 * ->vruntime to a relative base.
5903 * Make sure both cases convert their relative position when migrating
5904 * to another cgroup's rq. This does somewhat interfere with the
5905 * fair sleeper stuff for the first placement, but who cares.
5908 * When !on_rq, vruntime of the task has usually NOT been normalized.
5909 * But there are some cases where it has already been normalized:
5911 * - Moving a forked child which is waiting for being woken up by
5912 * wake_up_new_task().
5913 * - Moving a task which has been woken up by try_to_wake_up() and
5914 * waiting for actually being woken up by sched_ttwu_pending().
5916 * To prevent boost or penalty in the new cfs_rq caused by delta
5917 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
5919 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
5923 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
5924 set_task_rq(p, task_cpu(p));
5926 cfs_rq = cfs_rq_of(&p->se);
5927 p->se.vruntime += cfs_rq->min_vruntime;
5930 * migrate_task_rq_fair() will have removed our previous
5931 * contribution, but we must synchronize for ongoing future
5934 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
5935 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
5940 void free_fair_sched_group(struct task_group *tg)
5944 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
5946 for_each_possible_cpu(i) {
5948 kfree(tg->cfs_rq[i]);
5957 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
5959 struct cfs_rq *cfs_rq;
5960 struct sched_entity *se;
5963 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
5966 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
5970 tg->shares = NICE_0_LOAD;
5972 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
5974 for_each_possible_cpu(i) {
5975 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
5976 GFP_KERNEL, cpu_to_node(i));
5980 se = kzalloc_node(sizeof(struct sched_entity),
5981 GFP_KERNEL, cpu_to_node(i));
5985 init_cfs_rq(cfs_rq);
5986 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
5997 void unregister_fair_sched_group(struct task_group *tg, int cpu)
5999 struct rq *rq = cpu_rq(cpu);
6000 unsigned long flags;
6003 * Only empty task groups can be destroyed; so we can speculatively
6004 * check on_list without danger of it being re-added.
6006 if (!tg->cfs_rq[cpu]->on_list)
6009 raw_spin_lock_irqsave(&rq->lock, flags);
6010 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
6011 raw_spin_unlock_irqrestore(&rq->lock, flags);
6014 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
6015 struct sched_entity *se, int cpu,
6016 struct sched_entity *parent)
6018 struct rq *rq = cpu_rq(cpu);
6022 init_cfs_rq_runtime(cfs_rq);
6024 tg->cfs_rq[cpu] = cfs_rq;
6027 /* se could be NULL for root_task_group */
6032 se->cfs_rq = &rq->cfs;
6034 se->cfs_rq = parent->my_q;
6037 update_load_set(&se->load, 0);
6038 se->parent = parent;
6041 static DEFINE_MUTEX(shares_mutex);
6043 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6046 unsigned long flags;
6049 * We can't change the weight of the root cgroup.
6054 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
6056 mutex_lock(&shares_mutex);
6057 if (tg->shares == shares)
6060 tg->shares = shares;
6061 for_each_possible_cpu(i) {
6062 struct rq *rq = cpu_rq(i);
6063 struct sched_entity *se;
6066 /* Propagate contribution to hierarchy */
6067 raw_spin_lock_irqsave(&rq->lock, flags);
6068 for_each_sched_entity(se)
6069 update_cfs_shares(group_cfs_rq(se));
6070 raw_spin_unlock_irqrestore(&rq->lock, flags);
6074 mutex_unlock(&shares_mutex);
6077 #else /* CONFIG_FAIR_GROUP_SCHED */
6079 void free_fair_sched_group(struct task_group *tg) { }
6081 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6086 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
6088 #endif /* CONFIG_FAIR_GROUP_SCHED */
6091 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
6093 struct sched_entity *se = &task->se;
6094 unsigned int rr_interval = 0;
6097 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6100 if (rq->cfs.load.weight)
6101 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
6107 * All the scheduling class methods:
6109 const struct sched_class fair_sched_class = {
6110 .next = &idle_sched_class,
6111 .enqueue_task = enqueue_task_fair,
6112 .dequeue_task = dequeue_task_fair,
6113 .yield_task = yield_task_fair,
6114 .yield_to_task = yield_to_task_fair,
6116 .check_preempt_curr = check_preempt_wakeup,
6118 .pick_next_task = pick_next_task_fair,
6119 .put_prev_task = put_prev_task_fair,
6122 .select_task_rq = select_task_rq_fair,
6123 #ifdef CONFIG_FAIR_GROUP_SCHED
6124 .migrate_task_rq = migrate_task_rq_fair,
6126 .rq_online = rq_online_fair,
6127 .rq_offline = rq_offline_fair,
6129 .task_waking = task_waking_fair,
6132 .set_curr_task = set_curr_task_fair,
6133 .task_tick = task_tick_fair,
6134 .task_fork = task_fork_fair,
6136 .prio_changed = prio_changed_fair,
6137 .switched_from = switched_from_fair,
6138 .switched_to = switched_to_fair,
6140 .get_rr_interval = get_rr_interval_fair,
6142 #ifdef CONFIG_FAIR_GROUP_SCHED
6143 .task_move_group = task_move_group_fair,
6147 #ifdef CONFIG_SCHED_DEBUG
6148 void print_cfs_stats(struct seq_file *m, int cpu)
6150 struct cfs_rq *cfs_rq;
6153 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
6154 print_cfs_rq(m, cpu, cfs_rq);
6159 __init void init_sched_fair_class(void)
6162 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
6165 nohz.next_balance = jiffies;
6166 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
6167 cpu_notifier(sched_ilb_notifier, 0);