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 max_vruntime, u64 vruntime)
436 s64 delta = (s64)(vruntime - max_vruntime);
438 max_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 /* ensure we never gain time by being placed backwards. */
477 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
480 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
485 * Enqueue an entity into the rb-tree:
487 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
489 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
490 struct rb_node *parent = NULL;
491 struct sched_entity *entry;
495 * Find the right place in the rbtree:
499 entry = rb_entry(parent, struct sched_entity, run_node);
501 * We dont care about collisions. Nodes with
502 * the same key stay together.
504 if (entity_before(se, entry)) {
505 link = &parent->rb_left;
507 link = &parent->rb_right;
513 * Maintain a cache of leftmost tree entries (it is frequently
517 cfs_rq->rb_leftmost = &se->run_node;
519 rb_link_node(&se->run_node, parent, link);
520 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
523 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
525 if (cfs_rq->rb_leftmost == &se->run_node) {
526 struct rb_node *next_node;
528 next_node = rb_next(&se->run_node);
529 cfs_rq->rb_leftmost = next_node;
532 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
535 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
537 struct rb_node *left = cfs_rq->rb_leftmost;
542 return rb_entry(left, struct sched_entity, run_node);
545 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
547 struct rb_node *next = rb_next(&se->run_node);
552 return rb_entry(next, struct sched_entity, run_node);
555 #ifdef CONFIG_SCHED_DEBUG
556 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
558 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
563 return rb_entry(last, struct sched_entity, run_node);
566 /**************************************************************
567 * Scheduling class statistics methods:
570 int sched_proc_update_handler(struct ctl_table *table, int write,
571 void __user *buffer, size_t *lenp,
574 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
575 int factor = get_update_sysctl_factor();
580 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
581 sysctl_sched_min_granularity);
583 #define WRT_SYSCTL(name) \
584 (normalized_sysctl_##name = sysctl_##name / (factor))
585 WRT_SYSCTL(sched_min_granularity);
586 WRT_SYSCTL(sched_latency);
587 WRT_SYSCTL(sched_wakeup_granularity);
597 static inline unsigned long
598 calc_delta_fair(unsigned long delta, struct sched_entity *se)
600 if (unlikely(se->load.weight != NICE_0_LOAD))
601 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
607 * The idea is to set a period in which each task runs once.
609 * When there are too many tasks (sched_nr_latency) we have to stretch
610 * this period because otherwise the slices get too small.
612 * p = (nr <= nl) ? l : l*nr/nl
614 static u64 __sched_period(unsigned long nr_running)
616 u64 period = sysctl_sched_latency;
617 unsigned long nr_latency = sched_nr_latency;
619 if (unlikely(nr_running > nr_latency)) {
620 period = sysctl_sched_min_granularity;
621 period *= nr_running;
628 * We calculate the wall-time slice from the period by taking a part
629 * proportional to the weight.
633 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
635 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
637 for_each_sched_entity(se) {
638 struct load_weight *load;
639 struct load_weight lw;
641 cfs_rq = cfs_rq_of(se);
642 load = &cfs_rq->load;
644 if (unlikely(!se->on_rq)) {
647 update_load_add(&lw, se->load.weight);
650 slice = calc_delta_mine(slice, se->load.weight, load);
656 * We calculate the vruntime slice of a to-be-inserted task.
660 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
662 return calc_delta_fair(sched_slice(cfs_rq, se), se);
666 * Update the current task's runtime statistics. Skip current tasks that
667 * are not in our scheduling class.
670 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
671 unsigned long delta_exec)
673 unsigned long delta_exec_weighted;
675 schedstat_set(curr->statistics.exec_max,
676 max((u64)delta_exec, curr->statistics.exec_max));
678 curr->sum_exec_runtime += delta_exec;
679 schedstat_add(cfs_rq, exec_clock, delta_exec);
680 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
682 curr->vruntime += delta_exec_weighted;
683 update_min_vruntime(cfs_rq);
686 static void update_curr(struct cfs_rq *cfs_rq)
688 struct sched_entity *curr = cfs_rq->curr;
689 u64 now = rq_of(cfs_rq)->clock_task;
690 unsigned long delta_exec;
696 * Get the amount of time the current task was running
697 * since the last time we changed load (this cannot
698 * overflow on 32 bits):
700 delta_exec = (unsigned long)(now - curr->exec_start);
704 __update_curr(cfs_rq, curr, delta_exec);
705 curr->exec_start = now;
707 if (entity_is_task(curr)) {
708 struct task_struct *curtask = task_of(curr);
710 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
711 cpuacct_charge(curtask, delta_exec);
712 account_group_exec_runtime(curtask, delta_exec);
715 account_cfs_rq_runtime(cfs_rq, delta_exec);
719 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
721 schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock);
725 * Task is being enqueued - update stats:
727 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
730 * Are we enqueueing a waiting task? (for current tasks
731 * a dequeue/enqueue event is a NOP)
733 if (se != cfs_rq->curr)
734 update_stats_wait_start(cfs_rq, se);
738 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
740 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
741 rq_of(cfs_rq)->clock - se->statistics.wait_start));
742 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
743 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
744 rq_of(cfs_rq)->clock - se->statistics.wait_start);
745 #ifdef CONFIG_SCHEDSTATS
746 if (entity_is_task(se)) {
747 trace_sched_stat_wait(task_of(se),
748 rq_of(cfs_rq)->clock - se->statistics.wait_start);
751 schedstat_set(se->statistics.wait_start, 0);
755 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
758 * Mark the end of the wait period if dequeueing a
761 if (se != cfs_rq->curr)
762 update_stats_wait_end(cfs_rq, se);
766 * We are picking a new current task - update its stats:
769 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
772 * We are starting a new run period:
774 se->exec_start = rq_of(cfs_rq)->clock_task;
777 /**************************************************
778 * Scheduling class queueing methods:
781 #ifdef CONFIG_NUMA_BALANCING
783 * numa task sample period in ms
785 unsigned int sysctl_numa_balancing_scan_period_min = 100;
786 unsigned int sysctl_numa_balancing_scan_period_max = 100*50;
787 unsigned int sysctl_numa_balancing_scan_period_reset = 100*600;
789 /* Portion of address space to scan in MB */
790 unsigned int sysctl_numa_balancing_scan_size = 256;
792 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
793 unsigned int sysctl_numa_balancing_scan_delay = 1000;
795 static void task_numa_placement(struct task_struct *p)
799 if (!p->mm) /* for example, ksmd faulting in a user's mm */
801 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
802 if (p->numa_scan_seq == seq)
804 p->numa_scan_seq = seq;
806 /* FIXME: Scheduling placement policy hints go here */
810 * Got a PROT_NONE fault for a page on @node.
812 void task_numa_fault(int node, int pages, bool migrated)
814 struct task_struct *p = current;
816 if (!sched_feat_numa(NUMA))
819 /* FIXME: Allocate task-specific structure for placement policy here */
822 * If pages are properly placed (did not migrate) then scan slower.
823 * This is reset periodically in case of phase changes
826 p->numa_scan_period = min(sysctl_numa_balancing_scan_period_max,
827 p->numa_scan_period + jiffies_to_msecs(10));
829 task_numa_placement(p);
832 static void reset_ptenuma_scan(struct task_struct *p)
834 ACCESS_ONCE(p->mm->numa_scan_seq)++;
835 p->mm->numa_scan_offset = 0;
839 * The expensive part of numa migration is done from task_work context.
840 * Triggered from task_tick_numa().
842 void task_numa_work(struct callback_head *work)
844 unsigned long migrate, next_scan, now = jiffies;
845 struct task_struct *p = current;
846 struct mm_struct *mm = p->mm;
847 struct vm_area_struct *vma;
848 unsigned long start, end;
851 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
853 work->next = work; /* protect against double add */
855 * Who cares about NUMA placement when they're dying.
857 * NOTE: make sure not to dereference p->mm before this check,
858 * exit_task_work() happens _after_ exit_mm() so we could be called
859 * without p->mm even though we still had it when we enqueued this
862 if (p->flags & PF_EXITING)
866 * We do not care about task placement until a task runs on a node
867 * other than the first one used by the address space. This is
868 * largely because migrations are driven by what CPU the task
869 * is running on. If it's never scheduled on another node, it'll
870 * not migrate so why bother trapping the fault.
872 if (mm->first_nid == NUMA_PTE_SCAN_INIT)
873 mm->first_nid = numa_node_id();
874 if (mm->first_nid != NUMA_PTE_SCAN_ACTIVE) {
875 /* Are we running on a new node yet? */
876 if (numa_node_id() == mm->first_nid &&
877 !sched_feat_numa(NUMA_FORCE))
880 mm->first_nid = NUMA_PTE_SCAN_ACTIVE;
884 * Reset the scan period if enough time has gone by. Objective is that
885 * scanning will be reduced if pages are properly placed. As tasks
886 * can enter different phases this needs to be re-examined. Lacking
887 * proper tracking of reference behaviour, this blunt hammer is used.
889 migrate = mm->numa_next_reset;
890 if (time_after(now, migrate)) {
891 p->numa_scan_period = sysctl_numa_balancing_scan_period_min;
892 next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
893 xchg(&mm->numa_next_reset, next_scan);
897 * Enforce maximal scan/migration frequency..
899 migrate = mm->numa_next_scan;
900 if (time_before(now, migrate))
903 if (p->numa_scan_period == 0)
904 p->numa_scan_period = sysctl_numa_balancing_scan_period_min;
906 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
907 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
911 * Do not set pte_numa if the current running node is rate-limited.
912 * This loses statistics on the fault but if we are unwilling to
913 * migrate to this node, it is less likely we can do useful work
915 if (migrate_ratelimited(numa_node_id()))
918 start = mm->numa_scan_offset;
919 pages = sysctl_numa_balancing_scan_size;
920 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
924 down_read(&mm->mmap_sem);
925 vma = find_vma(mm, start);
927 reset_ptenuma_scan(p);
931 for (; vma; vma = vma->vm_next) {
932 if (!vma_migratable(vma))
935 /* Skip small VMAs. They are not likely to be of relevance */
936 if (vma->vm_end - vma->vm_start < HPAGE_SIZE)
940 start = max(start, vma->vm_start);
941 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
942 end = min(end, vma->vm_end);
943 pages -= change_prot_numa(vma, start, end);
948 } while (end != vma->vm_end);
953 * It is possible to reach the end of the VMA list but the last few VMAs are
954 * not guaranteed to the vma_migratable. If they are not, we would find the
955 * !migratable VMA on the next scan but not reset the scanner to the start
959 mm->numa_scan_offset = start;
961 reset_ptenuma_scan(p);
962 up_read(&mm->mmap_sem);
966 * Drive the periodic memory faults..
968 void task_tick_numa(struct rq *rq, struct task_struct *curr)
970 struct callback_head *work = &curr->numa_work;
974 * We don't care about NUMA placement if we don't have memory.
976 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
980 * Using runtime rather than walltime has the dual advantage that
981 * we (mostly) drive the selection from busy threads and that the
982 * task needs to have done some actual work before we bother with
985 now = curr->se.sum_exec_runtime;
986 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
988 if (now - curr->node_stamp > period) {
989 if (!curr->node_stamp)
990 curr->numa_scan_period = sysctl_numa_balancing_scan_period_min;
991 curr->node_stamp = now;
993 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
994 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
995 task_work_add(curr, work, true);
1000 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1003 #endif /* CONFIG_NUMA_BALANCING */
1006 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1008 update_load_add(&cfs_rq->load, se->load.weight);
1009 if (!parent_entity(se))
1010 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1012 if (entity_is_task(se))
1013 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1015 cfs_rq->nr_running++;
1019 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1021 update_load_sub(&cfs_rq->load, se->load.weight);
1022 if (!parent_entity(se))
1023 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1024 if (entity_is_task(se))
1025 list_del_init(&se->group_node);
1026 cfs_rq->nr_running--;
1029 #ifdef CONFIG_FAIR_GROUP_SCHED
1031 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1036 * Use this CPU's actual weight instead of the last load_contribution
1037 * to gain a more accurate current total weight. See
1038 * update_cfs_rq_load_contribution().
1040 tg_weight = atomic64_read(&tg->load_avg);
1041 tg_weight -= cfs_rq->tg_load_contrib;
1042 tg_weight += cfs_rq->load.weight;
1047 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1049 long tg_weight, load, shares;
1051 tg_weight = calc_tg_weight(tg, cfs_rq);
1052 load = cfs_rq->load.weight;
1054 shares = (tg->shares * load);
1056 shares /= tg_weight;
1058 if (shares < MIN_SHARES)
1059 shares = MIN_SHARES;
1060 if (shares > tg->shares)
1061 shares = tg->shares;
1065 # else /* CONFIG_SMP */
1066 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1070 # endif /* CONFIG_SMP */
1071 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1072 unsigned long weight)
1075 /* commit outstanding execution time */
1076 if (cfs_rq->curr == se)
1077 update_curr(cfs_rq);
1078 account_entity_dequeue(cfs_rq, se);
1081 update_load_set(&se->load, weight);
1084 account_entity_enqueue(cfs_rq, se);
1087 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1089 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1091 struct task_group *tg;
1092 struct sched_entity *se;
1096 se = tg->se[cpu_of(rq_of(cfs_rq))];
1097 if (!se || throttled_hierarchy(cfs_rq))
1100 if (likely(se->load.weight == tg->shares))
1103 shares = calc_cfs_shares(cfs_rq, tg);
1105 reweight_entity(cfs_rq_of(se), se, shares);
1107 #else /* CONFIG_FAIR_GROUP_SCHED */
1108 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1111 #endif /* CONFIG_FAIR_GROUP_SCHED */
1113 /* Only depends on SMP, FAIR_GROUP_SCHED may be removed when useful in lb */
1114 #if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)
1116 * We choose a half-life close to 1 scheduling period.
1117 * Note: The tables below are dependent on this value.
1119 #define LOAD_AVG_PERIOD 32
1120 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1121 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1123 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1124 static const u32 runnable_avg_yN_inv[] = {
1125 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1126 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1127 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1128 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1129 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1130 0x85aac367, 0x82cd8698,
1134 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1135 * over-estimates when re-combining.
1137 static const u32 runnable_avg_yN_sum[] = {
1138 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1139 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1140 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1145 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1147 static __always_inline u64 decay_load(u64 val, u64 n)
1149 unsigned int local_n;
1153 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1156 /* after bounds checking we can collapse to 32-bit */
1160 * As y^PERIOD = 1/2, we can combine
1161 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1162 * With a look-up table which covers k^n (n<PERIOD)
1164 * To achieve constant time decay_load.
1166 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1167 val >>= local_n / LOAD_AVG_PERIOD;
1168 local_n %= LOAD_AVG_PERIOD;
1171 val *= runnable_avg_yN_inv[local_n];
1172 /* We don't use SRR here since we always want to round down. */
1177 * For updates fully spanning n periods, the contribution to runnable
1178 * average will be: \Sum 1024*y^n
1180 * We can compute this reasonably efficiently by combining:
1181 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1183 static u32 __compute_runnable_contrib(u64 n)
1187 if (likely(n <= LOAD_AVG_PERIOD))
1188 return runnable_avg_yN_sum[n];
1189 else if (unlikely(n >= LOAD_AVG_MAX_N))
1190 return LOAD_AVG_MAX;
1192 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1194 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1195 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1197 n -= LOAD_AVG_PERIOD;
1198 } while (n > LOAD_AVG_PERIOD);
1200 contrib = decay_load(contrib, n);
1201 return contrib + runnable_avg_yN_sum[n];
1205 * We can represent the historical contribution to runnable average as the
1206 * coefficients of a geometric series. To do this we sub-divide our runnable
1207 * history into segments of approximately 1ms (1024us); label the segment that
1208 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1210 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1212 * (now) (~1ms ago) (~2ms ago)
1214 * Let u_i denote the fraction of p_i that the entity was runnable.
1216 * We then designate the fractions u_i as our co-efficients, yielding the
1217 * following representation of historical load:
1218 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1220 * We choose y based on the with of a reasonably scheduling period, fixing:
1223 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1224 * approximately half as much as the contribution to load within the last ms
1227 * When a period "rolls over" and we have new u_0`, multiplying the previous
1228 * sum again by y is sufficient to update:
1229 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1230 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1232 static __always_inline int __update_entity_runnable_avg(u64 now,
1233 struct sched_avg *sa,
1237 u32 runnable_contrib;
1238 int delta_w, decayed = 0;
1240 delta = now - sa->last_runnable_update;
1242 * This should only happen when time goes backwards, which it
1243 * unfortunately does during sched clock init when we swap over to TSC.
1245 if ((s64)delta < 0) {
1246 sa->last_runnable_update = now;
1251 * Use 1024ns as the unit of measurement since it's a reasonable
1252 * approximation of 1us and fast to compute.
1257 sa->last_runnable_update = now;
1259 /* delta_w is the amount already accumulated against our next period */
1260 delta_w = sa->runnable_avg_period % 1024;
1261 if (delta + delta_w >= 1024) {
1262 /* period roll-over */
1266 * Now that we know we're crossing a period boundary, figure
1267 * out how much from delta we need to complete the current
1268 * period and accrue it.
1270 delta_w = 1024 - delta_w;
1272 sa->runnable_avg_sum += delta_w;
1273 sa->runnable_avg_period += delta_w;
1277 /* Figure out how many additional periods this update spans */
1278 periods = delta / 1024;
1281 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1283 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1286 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1287 runnable_contrib = __compute_runnable_contrib(periods);
1289 sa->runnable_avg_sum += runnable_contrib;
1290 sa->runnable_avg_period += runnable_contrib;
1293 /* Remainder of delta accrued against u_0` */
1295 sa->runnable_avg_sum += delta;
1296 sa->runnable_avg_period += delta;
1301 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1302 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1304 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1305 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1307 decays -= se->avg.decay_count;
1311 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1312 se->avg.decay_count = 0;
1317 #ifdef CONFIG_FAIR_GROUP_SCHED
1318 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1321 struct task_group *tg = cfs_rq->tg;
1324 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1325 tg_contrib -= cfs_rq->tg_load_contrib;
1327 if (force_update || abs64(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1328 atomic64_add(tg_contrib, &tg->load_avg);
1329 cfs_rq->tg_load_contrib += tg_contrib;
1334 * Aggregate cfs_rq runnable averages into an equivalent task_group
1335 * representation for computing load contributions.
1337 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1338 struct cfs_rq *cfs_rq)
1340 struct task_group *tg = cfs_rq->tg;
1343 /* The fraction of a cpu used by this cfs_rq */
1344 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1345 sa->runnable_avg_period + 1);
1346 contrib -= cfs_rq->tg_runnable_contrib;
1348 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
1349 atomic_add(contrib, &tg->runnable_avg);
1350 cfs_rq->tg_runnable_contrib += contrib;
1354 static inline void __update_group_entity_contrib(struct sched_entity *se)
1356 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1357 struct task_group *tg = cfs_rq->tg;
1362 contrib = cfs_rq->tg_load_contrib * tg->shares;
1363 se->avg.load_avg_contrib = div64_u64(contrib,
1364 atomic64_read(&tg->load_avg) + 1);
1367 * For group entities we need to compute a correction term in the case
1368 * that they are consuming <1 cpu so that we would contribute the same
1369 * load as a task of equal weight.
1371 * Explicitly co-ordinating this measurement would be expensive, but
1372 * fortunately the sum of each cpus contribution forms a usable
1373 * lower-bound on the true value.
1375 * Consider the aggregate of 2 contributions. Either they are disjoint
1376 * (and the sum represents true value) or they are disjoint and we are
1377 * understating by the aggregate of their overlap.
1379 * Extending this to N cpus, for a given overlap, the maximum amount we
1380 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1381 * cpus that overlap for this interval and w_i is the interval width.
1383 * On a small machine; the first term is well-bounded which bounds the
1384 * total error since w_i is a subset of the period. Whereas on a
1385 * larger machine, while this first term can be larger, if w_i is the
1386 * of consequential size guaranteed to see n_i*w_i quickly converge to
1387 * our upper bound of 1-cpu.
1389 runnable_avg = atomic_read(&tg->runnable_avg);
1390 if (runnable_avg < NICE_0_LOAD) {
1391 se->avg.load_avg_contrib *= runnable_avg;
1392 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1396 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1397 int force_update) {}
1398 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1399 struct cfs_rq *cfs_rq) {}
1400 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1403 static inline void __update_task_entity_contrib(struct sched_entity *se)
1407 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1408 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1409 contrib /= (se->avg.runnable_avg_period + 1);
1410 se->avg.load_avg_contrib = scale_load(contrib);
1413 /* Compute the current contribution to load_avg by se, return any delta */
1414 static long __update_entity_load_avg_contrib(struct sched_entity *se)
1416 long old_contrib = se->avg.load_avg_contrib;
1418 if (entity_is_task(se)) {
1419 __update_task_entity_contrib(se);
1421 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1422 __update_group_entity_contrib(se);
1425 return se->avg.load_avg_contrib - old_contrib;
1428 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1431 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1432 cfs_rq->blocked_load_avg -= load_contrib;
1434 cfs_rq->blocked_load_avg = 0;
1437 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1439 /* Update a sched_entity's runnable average */
1440 static inline void update_entity_load_avg(struct sched_entity *se,
1443 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1448 * For a group entity we need to use their owned cfs_rq_clock_task() in
1449 * case they are the parent of a throttled hierarchy.
1451 if (entity_is_task(se))
1452 now = cfs_rq_clock_task(cfs_rq);
1454 now = cfs_rq_clock_task(group_cfs_rq(se));
1456 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
1459 contrib_delta = __update_entity_load_avg_contrib(se);
1465 cfs_rq->runnable_load_avg += contrib_delta;
1467 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1471 * Decay the load contributed by all blocked children and account this so that
1472 * their contribution may appropriately discounted when they wake up.
1474 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1476 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1479 decays = now - cfs_rq->last_decay;
1480 if (!decays && !force_update)
1483 if (atomic64_read(&cfs_rq->removed_load)) {
1484 u64 removed_load = atomic64_xchg(&cfs_rq->removed_load, 0);
1485 subtract_blocked_load_contrib(cfs_rq, removed_load);
1489 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1491 atomic64_add(decays, &cfs_rq->decay_counter);
1492 cfs_rq->last_decay = now;
1495 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1498 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1500 __update_entity_runnable_avg(rq->clock_task, &rq->avg, runnable);
1501 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1504 /* Add the load generated by se into cfs_rq's child load-average */
1505 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1506 struct sched_entity *se,
1510 * We track migrations using entity decay_count <= 0, on a wake-up
1511 * migration we use a negative decay count to track the remote decays
1512 * accumulated while sleeping.
1514 if (unlikely(se->avg.decay_count <= 0)) {
1515 se->avg.last_runnable_update = rq_of(cfs_rq)->clock_task;
1516 if (se->avg.decay_count) {
1518 * In a wake-up migration we have to approximate the
1519 * time sleeping. This is because we can't synchronize
1520 * clock_task between the two cpus, and it is not
1521 * guaranteed to be read-safe. Instead, we can
1522 * approximate this using our carried decays, which are
1523 * explicitly atomically readable.
1525 se->avg.last_runnable_update -= (-se->avg.decay_count)
1527 update_entity_load_avg(se, 0);
1528 /* Indicate that we're now synchronized and on-rq */
1529 se->avg.decay_count = 0;
1533 __synchronize_entity_decay(se);
1536 /* migrated tasks did not contribute to our blocked load */
1538 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1539 update_entity_load_avg(se, 0);
1542 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1543 /* we force update consideration on load-balancer moves */
1544 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1548 * Remove se's load from this cfs_rq child load-average, if the entity is
1549 * transitioning to a blocked state we track its projected decay using
1552 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1553 struct sched_entity *se,
1556 update_entity_load_avg(se, 1);
1557 /* we force update consideration on load-balancer moves */
1558 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1560 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1562 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1563 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1564 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1567 static inline void update_entity_load_avg(struct sched_entity *se,
1568 int update_cfs_rq) {}
1569 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1570 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1571 struct sched_entity *se,
1573 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1574 struct sched_entity *se,
1576 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1577 int force_update) {}
1580 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1582 #ifdef CONFIG_SCHEDSTATS
1583 struct task_struct *tsk = NULL;
1585 if (entity_is_task(se))
1588 if (se->statistics.sleep_start) {
1589 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
1594 if (unlikely(delta > se->statistics.sleep_max))
1595 se->statistics.sleep_max = delta;
1597 se->statistics.sleep_start = 0;
1598 se->statistics.sum_sleep_runtime += delta;
1601 account_scheduler_latency(tsk, delta >> 10, 1);
1602 trace_sched_stat_sleep(tsk, delta);
1605 if (se->statistics.block_start) {
1606 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
1611 if (unlikely(delta > se->statistics.block_max))
1612 se->statistics.block_max = delta;
1614 se->statistics.block_start = 0;
1615 se->statistics.sum_sleep_runtime += delta;
1618 if (tsk->in_iowait) {
1619 se->statistics.iowait_sum += delta;
1620 se->statistics.iowait_count++;
1621 trace_sched_stat_iowait(tsk, delta);
1624 trace_sched_stat_blocked(tsk, delta);
1627 * Blocking time is in units of nanosecs, so shift by
1628 * 20 to get a milliseconds-range estimation of the
1629 * amount of time that the task spent sleeping:
1631 if (unlikely(prof_on == SLEEP_PROFILING)) {
1632 profile_hits(SLEEP_PROFILING,
1633 (void *)get_wchan(tsk),
1636 account_scheduler_latency(tsk, delta >> 10, 0);
1642 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1644 #ifdef CONFIG_SCHED_DEBUG
1645 s64 d = se->vruntime - cfs_rq->min_vruntime;
1650 if (d > 3*sysctl_sched_latency)
1651 schedstat_inc(cfs_rq, nr_spread_over);
1656 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1658 u64 vruntime = cfs_rq->min_vruntime;
1661 * The 'current' period is already promised to the current tasks,
1662 * however the extra weight of the new task will slow them down a
1663 * little, place the new task so that it fits in the slot that
1664 * stays open at the end.
1666 if (initial && sched_feat(START_DEBIT))
1667 vruntime += sched_vslice(cfs_rq, se);
1669 /* sleeps up to a single latency don't count. */
1671 unsigned long thresh = sysctl_sched_latency;
1674 * Halve their sleep time's effect, to allow
1675 * for a gentler effect of sleepers:
1677 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1683 /* ensure we never gain time by being placed backwards. */
1684 se->vruntime = max_vruntime(se->vruntime, vruntime);
1687 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1690 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1693 * Update the normalized vruntime before updating min_vruntime
1694 * through callig update_curr().
1696 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1697 se->vruntime += cfs_rq->min_vruntime;
1700 * Update run-time statistics of the 'current'.
1702 update_curr(cfs_rq);
1703 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
1704 account_entity_enqueue(cfs_rq, se);
1705 update_cfs_shares(cfs_rq);
1707 if (flags & ENQUEUE_WAKEUP) {
1708 place_entity(cfs_rq, se, 0);
1709 enqueue_sleeper(cfs_rq, se);
1712 update_stats_enqueue(cfs_rq, se);
1713 check_spread(cfs_rq, se);
1714 if (se != cfs_rq->curr)
1715 __enqueue_entity(cfs_rq, se);
1718 if (cfs_rq->nr_running == 1) {
1719 list_add_leaf_cfs_rq(cfs_rq);
1720 check_enqueue_throttle(cfs_rq);
1724 static void __clear_buddies_last(struct sched_entity *se)
1726 for_each_sched_entity(se) {
1727 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1728 if (cfs_rq->last == se)
1729 cfs_rq->last = NULL;
1735 static void __clear_buddies_next(struct sched_entity *se)
1737 for_each_sched_entity(se) {
1738 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1739 if (cfs_rq->next == se)
1740 cfs_rq->next = NULL;
1746 static void __clear_buddies_skip(struct sched_entity *se)
1748 for_each_sched_entity(se) {
1749 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1750 if (cfs_rq->skip == se)
1751 cfs_rq->skip = NULL;
1757 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1759 if (cfs_rq->last == se)
1760 __clear_buddies_last(se);
1762 if (cfs_rq->next == se)
1763 __clear_buddies_next(se);
1765 if (cfs_rq->skip == se)
1766 __clear_buddies_skip(se);
1769 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1772 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1775 * Update run-time statistics of the 'current'.
1777 update_curr(cfs_rq);
1778 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
1780 update_stats_dequeue(cfs_rq, se);
1781 if (flags & DEQUEUE_SLEEP) {
1782 #ifdef CONFIG_SCHEDSTATS
1783 if (entity_is_task(se)) {
1784 struct task_struct *tsk = task_of(se);
1786 if (tsk->state & TASK_INTERRUPTIBLE)
1787 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1788 if (tsk->state & TASK_UNINTERRUPTIBLE)
1789 se->statistics.block_start = rq_of(cfs_rq)->clock;
1794 clear_buddies(cfs_rq, se);
1796 if (se != cfs_rq->curr)
1797 __dequeue_entity(cfs_rq, se);
1799 account_entity_dequeue(cfs_rq, se);
1802 * Normalize the entity after updating the min_vruntime because the
1803 * update can refer to the ->curr item and we need to reflect this
1804 * movement in our normalized position.
1806 if (!(flags & DEQUEUE_SLEEP))
1807 se->vruntime -= cfs_rq->min_vruntime;
1809 /* return excess runtime on last dequeue */
1810 return_cfs_rq_runtime(cfs_rq);
1812 update_min_vruntime(cfs_rq);
1813 update_cfs_shares(cfs_rq);
1817 * Preempt the current task with a newly woken task if needed:
1820 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1822 unsigned long ideal_runtime, delta_exec;
1823 struct sched_entity *se;
1826 ideal_runtime = sched_slice(cfs_rq, curr);
1827 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1828 if (delta_exec > ideal_runtime) {
1829 resched_task(rq_of(cfs_rq)->curr);
1831 * The current task ran long enough, ensure it doesn't get
1832 * re-elected due to buddy favours.
1834 clear_buddies(cfs_rq, curr);
1839 * Ensure that a task that missed wakeup preemption by a
1840 * narrow margin doesn't have to wait for a full slice.
1841 * This also mitigates buddy induced latencies under load.
1843 if (delta_exec < sysctl_sched_min_granularity)
1846 se = __pick_first_entity(cfs_rq);
1847 delta = curr->vruntime - se->vruntime;
1852 if (delta > ideal_runtime)
1853 resched_task(rq_of(cfs_rq)->curr);
1857 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1859 /* 'current' is not kept within the tree. */
1862 * Any task has to be enqueued before it get to execute on
1863 * a CPU. So account for the time it spent waiting on the
1866 update_stats_wait_end(cfs_rq, se);
1867 __dequeue_entity(cfs_rq, se);
1870 update_stats_curr_start(cfs_rq, se);
1872 #ifdef CONFIG_SCHEDSTATS
1874 * Track our maximum slice length, if the CPU's load is at
1875 * least twice that of our own weight (i.e. dont track it
1876 * when there are only lesser-weight tasks around):
1878 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1879 se->statistics.slice_max = max(se->statistics.slice_max,
1880 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1883 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1887 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1890 * Pick the next process, keeping these things in mind, in this order:
1891 * 1) keep things fair between processes/task groups
1892 * 2) pick the "next" process, since someone really wants that to run
1893 * 3) pick the "last" process, for cache locality
1894 * 4) do not run the "skip" process, if something else is available
1896 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1898 struct sched_entity *se = __pick_first_entity(cfs_rq);
1899 struct sched_entity *left = se;
1902 * Avoid running the skip buddy, if running something else can
1903 * be done without getting too unfair.
1905 if (cfs_rq->skip == se) {
1906 struct sched_entity *second = __pick_next_entity(se);
1907 if (second && wakeup_preempt_entity(second, left) < 1)
1912 * Prefer last buddy, try to return the CPU to a preempted task.
1914 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1918 * Someone really wants this to run. If it's not unfair, run it.
1920 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1923 clear_buddies(cfs_rq, se);
1928 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1930 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1933 * If still on the runqueue then deactivate_task()
1934 * was not called and update_curr() has to be done:
1937 update_curr(cfs_rq);
1939 /* throttle cfs_rqs exceeding runtime */
1940 check_cfs_rq_runtime(cfs_rq);
1942 check_spread(cfs_rq, prev);
1944 update_stats_wait_start(cfs_rq, prev);
1945 /* Put 'current' back into the tree. */
1946 __enqueue_entity(cfs_rq, prev);
1947 /* in !on_rq case, update occurred at dequeue */
1948 update_entity_load_avg(prev, 1);
1950 cfs_rq->curr = NULL;
1954 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1957 * Update run-time statistics of the 'current'.
1959 update_curr(cfs_rq);
1962 * Ensure that runnable average is periodically updated.
1964 update_entity_load_avg(curr, 1);
1965 update_cfs_rq_blocked_load(cfs_rq, 1);
1967 #ifdef CONFIG_SCHED_HRTICK
1969 * queued ticks are scheduled to match the slice, so don't bother
1970 * validating it and just reschedule.
1973 resched_task(rq_of(cfs_rq)->curr);
1977 * don't let the period tick interfere with the hrtick preemption
1979 if (!sched_feat(DOUBLE_TICK) &&
1980 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1984 if (cfs_rq->nr_running > 1)
1985 check_preempt_tick(cfs_rq, curr);
1989 /**************************************************
1990 * CFS bandwidth control machinery
1993 #ifdef CONFIG_CFS_BANDWIDTH
1995 #ifdef HAVE_JUMP_LABEL
1996 static struct static_key __cfs_bandwidth_used;
1998 static inline bool cfs_bandwidth_used(void)
2000 return static_key_false(&__cfs_bandwidth_used);
2003 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2005 /* only need to count groups transitioning between enabled/!enabled */
2006 if (enabled && !was_enabled)
2007 static_key_slow_inc(&__cfs_bandwidth_used);
2008 else if (!enabled && was_enabled)
2009 static_key_slow_dec(&__cfs_bandwidth_used);
2011 #else /* HAVE_JUMP_LABEL */
2012 static bool cfs_bandwidth_used(void)
2017 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2018 #endif /* HAVE_JUMP_LABEL */
2021 * default period for cfs group bandwidth.
2022 * default: 0.1s, units: nanoseconds
2024 static inline u64 default_cfs_period(void)
2026 return 100000000ULL;
2029 static inline u64 sched_cfs_bandwidth_slice(void)
2031 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2035 * Replenish runtime according to assigned quota and update expiration time.
2036 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2037 * additional synchronization around rq->lock.
2039 * requires cfs_b->lock
2041 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2045 if (cfs_b->quota == RUNTIME_INF)
2048 now = sched_clock_cpu(smp_processor_id());
2049 cfs_b->runtime = cfs_b->quota;
2050 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2053 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2055 return &tg->cfs_bandwidth;
2058 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2059 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2061 if (unlikely(cfs_rq->throttle_count))
2062 return cfs_rq->throttled_clock_task;
2064 return rq_of(cfs_rq)->clock_task - cfs_rq->throttled_clock_task_time;
2067 /* returns 0 on failure to allocate runtime */
2068 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2070 struct task_group *tg = cfs_rq->tg;
2071 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2072 u64 amount = 0, min_amount, expires;
2074 /* note: this is a positive sum as runtime_remaining <= 0 */
2075 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2077 raw_spin_lock(&cfs_b->lock);
2078 if (cfs_b->quota == RUNTIME_INF)
2079 amount = min_amount;
2082 * If the bandwidth pool has become inactive, then at least one
2083 * period must have elapsed since the last consumption.
2084 * Refresh the global state and ensure bandwidth timer becomes
2087 if (!cfs_b->timer_active) {
2088 __refill_cfs_bandwidth_runtime(cfs_b);
2089 __start_cfs_bandwidth(cfs_b);
2092 if (cfs_b->runtime > 0) {
2093 amount = min(cfs_b->runtime, min_amount);
2094 cfs_b->runtime -= amount;
2098 expires = cfs_b->runtime_expires;
2099 raw_spin_unlock(&cfs_b->lock);
2101 cfs_rq->runtime_remaining += amount;
2103 * we may have advanced our local expiration to account for allowed
2104 * spread between our sched_clock and the one on which runtime was
2107 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2108 cfs_rq->runtime_expires = expires;
2110 return cfs_rq->runtime_remaining > 0;
2114 * Note: This depends on the synchronization provided by sched_clock and the
2115 * fact that rq->clock snapshots this value.
2117 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2119 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2120 struct rq *rq = rq_of(cfs_rq);
2122 /* if the deadline is ahead of our clock, nothing to do */
2123 if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0))
2126 if (cfs_rq->runtime_remaining < 0)
2130 * If the local deadline has passed we have to consider the
2131 * possibility that our sched_clock is 'fast' and the global deadline
2132 * has not truly expired.
2134 * Fortunately we can check determine whether this the case by checking
2135 * whether the global deadline has advanced.
2138 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2139 /* extend local deadline, drift is bounded above by 2 ticks */
2140 cfs_rq->runtime_expires += TICK_NSEC;
2142 /* global deadline is ahead, expiration has passed */
2143 cfs_rq->runtime_remaining = 0;
2147 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2148 unsigned long delta_exec)
2150 /* dock delta_exec before expiring quota (as it could span periods) */
2151 cfs_rq->runtime_remaining -= delta_exec;
2152 expire_cfs_rq_runtime(cfs_rq);
2154 if (likely(cfs_rq->runtime_remaining > 0))
2158 * if we're unable to extend our runtime we resched so that the active
2159 * hierarchy can be throttled
2161 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2162 resched_task(rq_of(cfs_rq)->curr);
2165 static __always_inline
2166 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2168 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2171 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2174 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2176 return cfs_bandwidth_used() && cfs_rq->throttled;
2179 /* check whether cfs_rq, or any parent, is throttled */
2180 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2182 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2186 * Ensure that neither of the group entities corresponding to src_cpu or
2187 * dest_cpu are members of a throttled hierarchy when performing group
2188 * load-balance operations.
2190 static inline int throttled_lb_pair(struct task_group *tg,
2191 int src_cpu, int dest_cpu)
2193 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2195 src_cfs_rq = tg->cfs_rq[src_cpu];
2196 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2198 return throttled_hierarchy(src_cfs_rq) ||
2199 throttled_hierarchy(dest_cfs_rq);
2202 /* updated child weight may affect parent so we have to do this bottom up */
2203 static int tg_unthrottle_up(struct task_group *tg, void *data)
2205 struct rq *rq = data;
2206 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2208 cfs_rq->throttle_count--;
2210 if (!cfs_rq->throttle_count) {
2211 /* adjust cfs_rq_clock_task() */
2212 cfs_rq->throttled_clock_task_time += rq->clock_task -
2213 cfs_rq->throttled_clock_task;
2220 static int tg_throttle_down(struct task_group *tg, void *data)
2222 struct rq *rq = data;
2223 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2225 /* group is entering throttled state, stop time */
2226 if (!cfs_rq->throttle_count)
2227 cfs_rq->throttled_clock_task = rq->clock_task;
2228 cfs_rq->throttle_count++;
2233 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2235 struct rq *rq = rq_of(cfs_rq);
2236 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2237 struct sched_entity *se;
2238 long task_delta, dequeue = 1;
2240 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2242 /* freeze hierarchy runnable averages while throttled */
2244 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2247 task_delta = cfs_rq->h_nr_running;
2248 for_each_sched_entity(se) {
2249 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2250 /* throttled entity or throttle-on-deactivate */
2255 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2256 qcfs_rq->h_nr_running -= task_delta;
2258 if (qcfs_rq->load.weight)
2263 rq->nr_running -= task_delta;
2265 cfs_rq->throttled = 1;
2266 cfs_rq->throttled_clock = rq->clock;
2267 raw_spin_lock(&cfs_b->lock);
2268 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2269 raw_spin_unlock(&cfs_b->lock);
2272 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2274 struct rq *rq = rq_of(cfs_rq);
2275 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2276 struct sched_entity *se;
2280 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2282 cfs_rq->throttled = 0;
2283 raw_spin_lock(&cfs_b->lock);
2284 cfs_b->throttled_time += rq->clock - cfs_rq->throttled_clock;
2285 list_del_rcu(&cfs_rq->throttled_list);
2286 raw_spin_unlock(&cfs_b->lock);
2288 update_rq_clock(rq);
2289 /* update hierarchical throttle state */
2290 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2292 if (!cfs_rq->load.weight)
2295 task_delta = cfs_rq->h_nr_running;
2296 for_each_sched_entity(se) {
2300 cfs_rq = cfs_rq_of(se);
2302 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2303 cfs_rq->h_nr_running += task_delta;
2305 if (cfs_rq_throttled(cfs_rq))
2310 rq->nr_running += task_delta;
2312 /* determine whether we need to wake up potentially idle cpu */
2313 if (rq->curr == rq->idle && rq->cfs.nr_running)
2314 resched_task(rq->curr);
2317 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2318 u64 remaining, u64 expires)
2320 struct cfs_rq *cfs_rq;
2321 u64 runtime = remaining;
2324 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2326 struct rq *rq = rq_of(cfs_rq);
2328 raw_spin_lock(&rq->lock);
2329 if (!cfs_rq_throttled(cfs_rq))
2332 runtime = -cfs_rq->runtime_remaining + 1;
2333 if (runtime > remaining)
2334 runtime = remaining;
2335 remaining -= runtime;
2337 cfs_rq->runtime_remaining += runtime;
2338 cfs_rq->runtime_expires = expires;
2340 /* we check whether we're throttled above */
2341 if (cfs_rq->runtime_remaining > 0)
2342 unthrottle_cfs_rq(cfs_rq);
2345 raw_spin_unlock(&rq->lock);
2356 * Responsible for refilling a task_group's bandwidth and unthrottling its
2357 * cfs_rqs as appropriate. If there has been no activity within the last
2358 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2359 * used to track this state.
2361 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2363 u64 runtime, runtime_expires;
2364 int idle = 1, throttled;
2366 raw_spin_lock(&cfs_b->lock);
2367 /* no need to continue the timer with no bandwidth constraint */
2368 if (cfs_b->quota == RUNTIME_INF)
2371 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2372 /* idle depends on !throttled (for the case of a large deficit) */
2373 idle = cfs_b->idle && !throttled;
2374 cfs_b->nr_periods += overrun;
2376 /* if we're going inactive then everything else can be deferred */
2380 __refill_cfs_bandwidth_runtime(cfs_b);
2383 /* mark as potentially idle for the upcoming period */
2388 /* account preceding periods in which throttling occurred */
2389 cfs_b->nr_throttled += overrun;
2392 * There are throttled entities so we must first use the new bandwidth
2393 * to unthrottle them before making it generally available. This
2394 * ensures that all existing debts will be paid before a new cfs_rq is
2397 runtime = cfs_b->runtime;
2398 runtime_expires = cfs_b->runtime_expires;
2402 * This check is repeated as we are holding onto the new bandwidth
2403 * while we unthrottle. This can potentially race with an unthrottled
2404 * group trying to acquire new bandwidth from the global pool.
2406 while (throttled && runtime > 0) {
2407 raw_spin_unlock(&cfs_b->lock);
2408 /* we can't nest cfs_b->lock while distributing bandwidth */
2409 runtime = distribute_cfs_runtime(cfs_b, runtime,
2411 raw_spin_lock(&cfs_b->lock);
2413 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2416 /* return (any) remaining runtime */
2417 cfs_b->runtime = runtime;
2419 * While we are ensured activity in the period following an
2420 * unthrottle, this also covers the case in which the new bandwidth is
2421 * insufficient to cover the existing bandwidth deficit. (Forcing the
2422 * timer to remain active while there are any throttled entities.)
2427 cfs_b->timer_active = 0;
2428 raw_spin_unlock(&cfs_b->lock);
2433 /* a cfs_rq won't donate quota below this amount */
2434 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2435 /* minimum remaining period time to redistribute slack quota */
2436 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2437 /* how long we wait to gather additional slack before distributing */
2438 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2440 /* are we near the end of the current quota period? */
2441 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2443 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2446 /* if the call-back is running a quota refresh is already occurring */
2447 if (hrtimer_callback_running(refresh_timer))
2450 /* is a quota refresh about to occur? */
2451 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2452 if (remaining < min_expire)
2458 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2460 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2462 /* if there's a quota refresh soon don't bother with slack */
2463 if (runtime_refresh_within(cfs_b, min_left))
2466 start_bandwidth_timer(&cfs_b->slack_timer,
2467 ns_to_ktime(cfs_bandwidth_slack_period));
2470 /* we know any runtime found here is valid as update_curr() precedes return */
2471 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2473 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2474 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2476 if (slack_runtime <= 0)
2479 raw_spin_lock(&cfs_b->lock);
2480 if (cfs_b->quota != RUNTIME_INF &&
2481 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2482 cfs_b->runtime += slack_runtime;
2484 /* we are under rq->lock, defer unthrottling using a timer */
2485 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2486 !list_empty(&cfs_b->throttled_cfs_rq))
2487 start_cfs_slack_bandwidth(cfs_b);
2489 raw_spin_unlock(&cfs_b->lock);
2491 /* even if it's not valid for return we don't want to try again */
2492 cfs_rq->runtime_remaining -= slack_runtime;
2495 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2497 if (!cfs_bandwidth_used())
2500 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2503 __return_cfs_rq_runtime(cfs_rq);
2507 * This is done with a timer (instead of inline with bandwidth return) since
2508 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2510 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2512 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2515 /* confirm we're still not at a refresh boundary */
2516 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2519 raw_spin_lock(&cfs_b->lock);
2520 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2521 runtime = cfs_b->runtime;
2524 expires = cfs_b->runtime_expires;
2525 raw_spin_unlock(&cfs_b->lock);
2530 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2532 raw_spin_lock(&cfs_b->lock);
2533 if (expires == cfs_b->runtime_expires)
2534 cfs_b->runtime = runtime;
2535 raw_spin_unlock(&cfs_b->lock);
2539 * When a group wakes up we want to make sure that its quota is not already
2540 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2541 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2543 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2545 if (!cfs_bandwidth_used())
2548 /* an active group must be handled by the update_curr()->put() path */
2549 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2552 /* ensure the group is not already throttled */
2553 if (cfs_rq_throttled(cfs_rq))
2556 /* update runtime allocation */
2557 account_cfs_rq_runtime(cfs_rq, 0);
2558 if (cfs_rq->runtime_remaining <= 0)
2559 throttle_cfs_rq(cfs_rq);
2562 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2563 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2565 if (!cfs_bandwidth_used())
2568 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2572 * it's possible for a throttled entity to be forced into a running
2573 * state (e.g. set_curr_task), in this case we're finished.
2575 if (cfs_rq_throttled(cfs_rq))
2578 throttle_cfs_rq(cfs_rq);
2581 static inline u64 default_cfs_period(void);
2582 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun);
2583 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b);
2585 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2587 struct cfs_bandwidth *cfs_b =
2588 container_of(timer, struct cfs_bandwidth, slack_timer);
2589 do_sched_cfs_slack_timer(cfs_b);
2591 return HRTIMER_NORESTART;
2594 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2596 struct cfs_bandwidth *cfs_b =
2597 container_of(timer, struct cfs_bandwidth, period_timer);
2603 now = hrtimer_cb_get_time(timer);
2604 overrun = hrtimer_forward(timer, now, cfs_b->period);
2609 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2612 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2615 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2617 raw_spin_lock_init(&cfs_b->lock);
2619 cfs_b->quota = RUNTIME_INF;
2620 cfs_b->period = ns_to_ktime(default_cfs_period());
2622 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2623 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2624 cfs_b->period_timer.function = sched_cfs_period_timer;
2625 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2626 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2629 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2631 cfs_rq->runtime_enabled = 0;
2632 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2635 /* requires cfs_b->lock, may release to reprogram timer */
2636 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2639 * The timer may be active because we're trying to set a new bandwidth
2640 * period or because we're racing with the tear-down path
2641 * (timer_active==0 becomes visible before the hrtimer call-back
2642 * terminates). In either case we ensure that it's re-programmed
2644 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2645 raw_spin_unlock(&cfs_b->lock);
2646 /* ensure cfs_b->lock is available while we wait */
2647 hrtimer_cancel(&cfs_b->period_timer);
2649 raw_spin_lock(&cfs_b->lock);
2650 /* if someone else restarted the timer then we're done */
2651 if (cfs_b->timer_active)
2655 cfs_b->timer_active = 1;
2656 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2659 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2661 hrtimer_cancel(&cfs_b->period_timer);
2662 hrtimer_cancel(&cfs_b->slack_timer);
2665 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
2667 struct cfs_rq *cfs_rq;
2669 for_each_leaf_cfs_rq(rq, cfs_rq) {
2670 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2672 if (!cfs_rq->runtime_enabled)
2676 * clock_task is not advancing so we just need to make sure
2677 * there's some valid quota amount
2679 cfs_rq->runtime_remaining = cfs_b->quota;
2680 if (cfs_rq_throttled(cfs_rq))
2681 unthrottle_cfs_rq(cfs_rq);
2685 #else /* CONFIG_CFS_BANDWIDTH */
2686 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2688 return rq_of(cfs_rq)->clock_task;
2691 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2692 unsigned long delta_exec) {}
2693 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2694 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2695 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2697 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2702 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2707 static inline int throttled_lb_pair(struct task_group *tg,
2708 int src_cpu, int dest_cpu)
2713 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2715 #ifdef CONFIG_FAIR_GROUP_SCHED
2716 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2719 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2723 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2724 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2726 #endif /* CONFIG_CFS_BANDWIDTH */
2728 /**************************************************
2729 * CFS operations on tasks:
2732 #ifdef CONFIG_SCHED_HRTICK
2733 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2735 struct sched_entity *se = &p->se;
2736 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2738 WARN_ON(task_rq(p) != rq);
2740 if (cfs_rq->nr_running > 1) {
2741 u64 slice = sched_slice(cfs_rq, se);
2742 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2743 s64 delta = slice - ran;
2752 * Don't schedule slices shorter than 10000ns, that just
2753 * doesn't make sense. Rely on vruntime for fairness.
2756 delta = max_t(s64, 10000LL, delta);
2758 hrtick_start(rq, delta);
2763 * called from enqueue/dequeue and updates the hrtick when the
2764 * current task is from our class and nr_running is low enough
2767 static void hrtick_update(struct rq *rq)
2769 struct task_struct *curr = rq->curr;
2771 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2774 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2775 hrtick_start_fair(rq, curr);
2777 #else /* !CONFIG_SCHED_HRTICK */
2779 hrtick_start_fair(struct rq *rq, struct task_struct *p)
2783 static inline void hrtick_update(struct rq *rq)
2789 * The enqueue_task method is called before nr_running is
2790 * increased. Here we update the fair scheduling stats and
2791 * then put the task into the rbtree:
2794 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2796 struct cfs_rq *cfs_rq;
2797 struct sched_entity *se = &p->se;
2799 for_each_sched_entity(se) {
2802 cfs_rq = cfs_rq_of(se);
2803 enqueue_entity(cfs_rq, se, flags);
2806 * end evaluation on encountering a throttled cfs_rq
2808 * note: in the case of encountering a throttled cfs_rq we will
2809 * post the final h_nr_running increment below.
2811 if (cfs_rq_throttled(cfs_rq))
2813 cfs_rq->h_nr_running++;
2815 flags = ENQUEUE_WAKEUP;
2818 for_each_sched_entity(se) {
2819 cfs_rq = cfs_rq_of(se);
2820 cfs_rq->h_nr_running++;
2822 if (cfs_rq_throttled(cfs_rq))
2825 update_cfs_shares(cfs_rq);
2826 update_entity_load_avg(se, 1);
2830 update_rq_runnable_avg(rq, rq->nr_running);
2836 static void set_next_buddy(struct sched_entity *se);
2839 * The dequeue_task method is called before nr_running is
2840 * decreased. We remove the task from the rbtree and
2841 * update the fair scheduling stats:
2843 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2845 struct cfs_rq *cfs_rq;
2846 struct sched_entity *se = &p->se;
2847 int task_sleep = flags & DEQUEUE_SLEEP;
2849 for_each_sched_entity(se) {
2850 cfs_rq = cfs_rq_of(se);
2851 dequeue_entity(cfs_rq, se, flags);
2854 * end evaluation on encountering a throttled cfs_rq
2856 * note: in the case of encountering a throttled cfs_rq we will
2857 * post the final h_nr_running decrement below.
2859 if (cfs_rq_throttled(cfs_rq))
2861 cfs_rq->h_nr_running--;
2863 /* Don't dequeue parent if it has other entities besides us */
2864 if (cfs_rq->load.weight) {
2866 * Bias pick_next to pick a task from this cfs_rq, as
2867 * p is sleeping when it is within its sched_slice.
2869 if (task_sleep && parent_entity(se))
2870 set_next_buddy(parent_entity(se));
2872 /* avoid re-evaluating load for this entity */
2873 se = parent_entity(se);
2876 flags |= DEQUEUE_SLEEP;
2879 for_each_sched_entity(se) {
2880 cfs_rq = cfs_rq_of(se);
2881 cfs_rq->h_nr_running--;
2883 if (cfs_rq_throttled(cfs_rq))
2886 update_cfs_shares(cfs_rq);
2887 update_entity_load_avg(se, 1);
2892 update_rq_runnable_avg(rq, 1);
2898 /* Used instead of source_load when we know the type == 0 */
2899 static unsigned long weighted_cpuload(const int cpu)
2901 return cpu_rq(cpu)->load.weight;
2905 * Return a low guess at the load of a migration-source cpu weighted
2906 * according to the scheduling class and "nice" value.
2908 * We want to under-estimate the load of migration sources, to
2909 * balance conservatively.
2911 static unsigned long source_load(int cpu, int type)
2913 struct rq *rq = cpu_rq(cpu);
2914 unsigned long total = weighted_cpuload(cpu);
2916 if (type == 0 || !sched_feat(LB_BIAS))
2919 return min(rq->cpu_load[type-1], total);
2923 * Return a high guess at the load of a migration-target cpu weighted
2924 * according to the scheduling class and "nice" value.
2926 static unsigned long target_load(int cpu, int type)
2928 struct rq *rq = cpu_rq(cpu);
2929 unsigned long total = weighted_cpuload(cpu);
2931 if (type == 0 || !sched_feat(LB_BIAS))
2934 return max(rq->cpu_load[type-1], total);
2937 static unsigned long power_of(int cpu)
2939 return cpu_rq(cpu)->cpu_power;
2942 static unsigned long cpu_avg_load_per_task(int cpu)
2944 struct rq *rq = cpu_rq(cpu);
2945 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
2948 return rq->load.weight / nr_running;
2954 static void task_waking_fair(struct task_struct *p)
2956 struct sched_entity *se = &p->se;
2957 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2960 #ifndef CONFIG_64BIT
2961 u64 min_vruntime_copy;
2964 min_vruntime_copy = cfs_rq->min_vruntime_copy;
2966 min_vruntime = cfs_rq->min_vruntime;
2967 } while (min_vruntime != min_vruntime_copy);
2969 min_vruntime = cfs_rq->min_vruntime;
2972 se->vruntime -= min_vruntime;
2975 #ifdef CONFIG_FAIR_GROUP_SCHED
2977 * effective_load() calculates the load change as seen from the root_task_group
2979 * Adding load to a group doesn't make a group heavier, but can cause movement
2980 * of group shares between cpus. Assuming the shares were perfectly aligned one
2981 * can calculate the shift in shares.
2983 * Calculate the effective load difference if @wl is added (subtracted) to @tg
2984 * on this @cpu and results in a total addition (subtraction) of @wg to the
2985 * total group weight.
2987 * Given a runqueue weight distribution (rw_i) we can compute a shares
2988 * distribution (s_i) using:
2990 * s_i = rw_i / \Sum rw_j (1)
2992 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
2993 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
2994 * shares distribution (s_i):
2996 * rw_i = { 2, 4, 1, 0 }
2997 * s_i = { 2/7, 4/7, 1/7, 0 }
2999 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3000 * task used to run on and the CPU the waker is running on), we need to
3001 * compute the effect of waking a task on either CPU and, in case of a sync
3002 * wakeup, compute the effect of the current task going to sleep.
3004 * So for a change of @wl to the local @cpu with an overall group weight change
3005 * of @wl we can compute the new shares distribution (s'_i) using:
3007 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3009 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3010 * differences in waking a task to CPU 0. The additional task changes the
3011 * weight and shares distributions like:
3013 * rw'_i = { 3, 4, 1, 0 }
3014 * s'_i = { 3/8, 4/8, 1/8, 0 }
3016 * We can then compute the difference in effective weight by using:
3018 * dw_i = S * (s'_i - s_i) (3)
3020 * Where 'S' is the group weight as seen by its parent.
3022 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3023 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3024 * 4/7) times the weight of the group.
3026 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3028 struct sched_entity *se = tg->se[cpu];
3030 if (!tg->parent) /* the trivial, non-cgroup case */
3033 for_each_sched_entity(se) {
3039 * W = @wg + \Sum rw_j
3041 W = wg + calc_tg_weight(tg, se->my_q);
3046 w = se->my_q->load.weight + wl;
3049 * wl = S * s'_i; see (2)
3052 wl = (w * tg->shares) / W;
3057 * Per the above, wl is the new se->load.weight value; since
3058 * those are clipped to [MIN_SHARES, ...) do so now. See
3059 * calc_cfs_shares().
3061 if (wl < MIN_SHARES)
3065 * wl = dw_i = S * (s'_i - s_i); see (3)
3067 wl -= se->load.weight;
3070 * Recursively apply this logic to all parent groups to compute
3071 * the final effective load change on the root group. Since
3072 * only the @tg group gets extra weight, all parent groups can
3073 * only redistribute existing shares. @wl is the shift in shares
3074 * resulting from this level per the above.
3083 static inline unsigned long effective_load(struct task_group *tg, int cpu,
3084 unsigned long wl, unsigned long wg)
3091 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3093 s64 this_load, load;
3094 int idx, this_cpu, prev_cpu;
3095 unsigned long tl_per_task;
3096 struct task_group *tg;
3097 unsigned long weight;
3101 this_cpu = smp_processor_id();
3102 prev_cpu = task_cpu(p);
3103 load = source_load(prev_cpu, idx);
3104 this_load = target_load(this_cpu, idx);
3107 * If sync wakeup then subtract the (maximum possible)
3108 * effect of the currently running task from the load
3109 * of the current CPU:
3112 tg = task_group(current);
3113 weight = current->se.load.weight;
3115 this_load += effective_load(tg, this_cpu, -weight, -weight);
3116 load += effective_load(tg, prev_cpu, 0, -weight);
3120 weight = p->se.load.weight;
3123 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3124 * due to the sync cause above having dropped this_load to 0, we'll
3125 * always have an imbalance, but there's really nothing you can do
3126 * about that, so that's good too.
3128 * Otherwise check if either cpus are near enough in load to allow this
3129 * task to be woken on this_cpu.
3131 if (this_load > 0) {
3132 s64 this_eff_load, prev_eff_load;
3134 this_eff_load = 100;
3135 this_eff_load *= power_of(prev_cpu);
3136 this_eff_load *= this_load +
3137 effective_load(tg, this_cpu, weight, weight);
3139 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3140 prev_eff_load *= power_of(this_cpu);
3141 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3143 balanced = this_eff_load <= prev_eff_load;
3148 * If the currently running task will sleep within
3149 * a reasonable amount of time then attract this newly
3152 if (sync && balanced)
3155 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3156 tl_per_task = cpu_avg_load_per_task(this_cpu);
3159 (this_load <= load &&
3160 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3162 * This domain has SD_WAKE_AFFINE and
3163 * p is cache cold in this domain, and
3164 * there is no bad imbalance.
3166 schedstat_inc(sd, ttwu_move_affine);
3167 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3175 * find_idlest_group finds and returns the least busy CPU group within the
3178 static struct sched_group *
3179 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3180 int this_cpu, int load_idx)
3182 struct sched_group *idlest = NULL, *group = sd->groups;
3183 unsigned long min_load = ULONG_MAX, this_load = 0;
3184 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3187 unsigned long load, avg_load;
3191 /* Skip over this group if it has no CPUs allowed */
3192 if (!cpumask_intersects(sched_group_cpus(group),
3193 tsk_cpus_allowed(p)))
3196 local_group = cpumask_test_cpu(this_cpu,
3197 sched_group_cpus(group));
3199 /* Tally up the load of all CPUs in the group */
3202 for_each_cpu(i, sched_group_cpus(group)) {
3203 /* Bias balancing toward cpus of our domain */
3205 load = source_load(i, load_idx);
3207 load = target_load(i, load_idx);
3212 /* Adjust by relative CPU power of the group */
3213 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3216 this_load = avg_load;
3217 } else if (avg_load < min_load) {
3218 min_load = avg_load;
3221 } while (group = group->next, group != sd->groups);
3223 if (!idlest || 100*this_load < imbalance*min_load)
3229 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3232 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3234 unsigned long load, min_load = ULONG_MAX;
3238 /* Traverse only the allowed CPUs */
3239 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3240 load = weighted_cpuload(i);
3242 if (load < min_load || (load == min_load && i == this_cpu)) {
3252 * Try and locate an idle CPU in the sched_domain.
3254 static int select_idle_sibling(struct task_struct *p, int target)
3256 struct sched_domain *sd;
3257 struct sched_group *sg;
3258 int i = task_cpu(p);
3260 if (idle_cpu(target))
3264 * If the prevous cpu is cache affine and idle, don't be stupid.
3266 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3270 * Otherwise, iterate the domains and find an elegible idle cpu.
3272 sd = rcu_dereference(per_cpu(sd_llc, target));
3273 for_each_lower_domain(sd) {
3276 if (!cpumask_intersects(sched_group_cpus(sg),
3277 tsk_cpus_allowed(p)))
3280 for_each_cpu(i, sched_group_cpus(sg)) {
3281 if (i == target || !idle_cpu(i))
3285 target = cpumask_first_and(sched_group_cpus(sg),
3286 tsk_cpus_allowed(p));
3290 } while (sg != sd->groups);
3297 * sched_balance_self: balance the current task (running on cpu) in domains
3298 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3301 * Balance, ie. select the least loaded group.
3303 * Returns the target CPU number, or the same CPU if no balancing is needed.
3305 * preempt must be disabled.
3308 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
3310 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3311 int cpu = smp_processor_id();
3312 int prev_cpu = task_cpu(p);
3314 int want_affine = 0;
3315 int sync = wake_flags & WF_SYNC;
3317 if (p->nr_cpus_allowed == 1)
3320 if (sd_flag & SD_BALANCE_WAKE) {
3321 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
3327 for_each_domain(cpu, tmp) {
3328 if (!(tmp->flags & SD_LOAD_BALANCE))
3332 * If both cpu and prev_cpu are part of this domain,
3333 * cpu is a valid SD_WAKE_AFFINE target.
3335 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3336 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3341 if (tmp->flags & sd_flag)
3346 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3349 new_cpu = select_idle_sibling(p, prev_cpu);
3354 int load_idx = sd->forkexec_idx;
3355 struct sched_group *group;
3358 if (!(sd->flags & sd_flag)) {
3363 if (sd_flag & SD_BALANCE_WAKE)
3364 load_idx = sd->wake_idx;
3366 group = find_idlest_group(sd, p, cpu, load_idx);
3372 new_cpu = find_idlest_cpu(group, p, cpu);
3373 if (new_cpu == -1 || new_cpu == cpu) {
3374 /* Now try balancing at a lower domain level of cpu */
3379 /* Now try balancing at a lower domain level of new_cpu */
3381 weight = sd->span_weight;
3383 for_each_domain(cpu, tmp) {
3384 if (weight <= tmp->span_weight)
3386 if (tmp->flags & sd_flag)
3389 /* while loop will break here if sd == NULL */
3398 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
3399 * removed when useful for applications beyond shares distribution (e.g.
3402 #ifdef CONFIG_FAIR_GROUP_SCHED
3404 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3405 * cfs_rq_of(p) references at time of call are still valid and identify the
3406 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3407 * other assumptions, including the state of rq->lock, should be made.
3410 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
3412 struct sched_entity *se = &p->se;
3413 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3416 * Load tracking: accumulate removed load so that it can be processed
3417 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3418 * to blocked load iff they have a positive decay-count. It can never
3419 * be negative here since on-rq tasks have decay-count == 0.
3421 if (se->avg.decay_count) {
3422 se->avg.decay_count = -__synchronize_entity_decay(se);
3423 atomic64_add(se->avg.load_avg_contrib, &cfs_rq->removed_load);
3427 #endif /* CONFIG_SMP */
3429 static unsigned long
3430 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3432 unsigned long gran = sysctl_sched_wakeup_granularity;
3435 * Since its curr running now, convert the gran from real-time
3436 * to virtual-time in his units.
3438 * By using 'se' instead of 'curr' we penalize light tasks, so
3439 * they get preempted easier. That is, if 'se' < 'curr' then
3440 * the resulting gran will be larger, therefore penalizing the
3441 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3442 * be smaller, again penalizing the lighter task.
3444 * This is especially important for buddies when the leftmost
3445 * task is higher priority than the buddy.
3447 return calc_delta_fair(gran, se);
3451 * Should 'se' preempt 'curr'.
3465 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3467 s64 gran, vdiff = curr->vruntime - se->vruntime;
3472 gran = wakeup_gran(curr, se);
3479 static void set_last_buddy(struct sched_entity *se)
3481 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3484 for_each_sched_entity(se)
3485 cfs_rq_of(se)->last = se;
3488 static void set_next_buddy(struct sched_entity *se)
3490 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3493 for_each_sched_entity(se)
3494 cfs_rq_of(se)->next = se;
3497 static void set_skip_buddy(struct sched_entity *se)
3499 for_each_sched_entity(se)
3500 cfs_rq_of(se)->skip = se;
3504 * Preempt the current task with a newly woken task if needed:
3506 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3508 struct task_struct *curr = rq->curr;
3509 struct sched_entity *se = &curr->se, *pse = &p->se;
3510 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3511 int scale = cfs_rq->nr_running >= sched_nr_latency;
3512 int next_buddy_marked = 0;
3514 if (unlikely(se == pse))
3518 * This is possible from callers such as move_task(), in which we
3519 * unconditionally check_prempt_curr() after an enqueue (which may have
3520 * lead to a throttle). This both saves work and prevents false
3521 * next-buddy nomination below.
3523 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3526 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3527 set_next_buddy(pse);
3528 next_buddy_marked = 1;
3532 * We can come here with TIF_NEED_RESCHED already set from new task
3535 * Note: this also catches the edge-case of curr being in a throttled
3536 * group (e.g. via set_curr_task), since update_curr() (in the
3537 * enqueue of curr) will have resulted in resched being set. This
3538 * prevents us from potentially nominating it as a false LAST_BUDDY
3541 if (test_tsk_need_resched(curr))
3544 /* Idle tasks are by definition preempted by non-idle tasks. */
3545 if (unlikely(curr->policy == SCHED_IDLE) &&
3546 likely(p->policy != SCHED_IDLE))
3550 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3551 * is driven by the tick):
3553 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
3556 find_matching_se(&se, &pse);
3557 update_curr(cfs_rq_of(se));
3559 if (wakeup_preempt_entity(se, pse) == 1) {
3561 * Bias pick_next to pick the sched entity that is
3562 * triggering this preemption.
3564 if (!next_buddy_marked)
3565 set_next_buddy(pse);
3574 * Only set the backward buddy when the current task is still
3575 * on the rq. This can happen when a wakeup gets interleaved
3576 * with schedule on the ->pre_schedule() or idle_balance()
3577 * point, either of which can * drop the rq lock.
3579 * Also, during early boot the idle thread is in the fair class,
3580 * for obvious reasons its a bad idea to schedule back to it.
3582 if (unlikely(!se->on_rq || curr == rq->idle))
3585 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3589 static struct task_struct *pick_next_task_fair(struct rq *rq)
3591 struct task_struct *p;
3592 struct cfs_rq *cfs_rq = &rq->cfs;
3593 struct sched_entity *se;
3595 if (!cfs_rq->nr_running)
3599 se = pick_next_entity(cfs_rq);
3600 set_next_entity(cfs_rq, se);
3601 cfs_rq = group_cfs_rq(se);
3605 if (hrtick_enabled(rq))
3606 hrtick_start_fair(rq, p);
3612 * Account for a descheduled task:
3614 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3616 struct sched_entity *se = &prev->se;
3617 struct cfs_rq *cfs_rq;
3619 for_each_sched_entity(se) {
3620 cfs_rq = cfs_rq_of(se);
3621 put_prev_entity(cfs_rq, se);
3626 * sched_yield() is very simple
3628 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3630 static void yield_task_fair(struct rq *rq)
3632 struct task_struct *curr = rq->curr;
3633 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3634 struct sched_entity *se = &curr->se;
3637 * Are we the only task in the tree?
3639 if (unlikely(rq->nr_running == 1))
3642 clear_buddies(cfs_rq, se);
3644 if (curr->policy != SCHED_BATCH) {
3645 update_rq_clock(rq);
3647 * Update run-time statistics of the 'current'.
3649 update_curr(cfs_rq);
3651 * Tell update_rq_clock() that we've just updated,
3652 * so we don't do microscopic update in schedule()
3653 * and double the fastpath cost.
3655 rq->skip_clock_update = 1;
3661 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3663 struct sched_entity *se = &p->se;
3665 /* throttled hierarchies are not runnable */
3666 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3669 /* Tell the scheduler that we'd really like pse to run next. */
3672 yield_task_fair(rq);
3678 /**************************************************
3679 * Fair scheduling class load-balancing methods.
3683 * The purpose of load-balancing is to achieve the same basic fairness the
3684 * per-cpu scheduler provides, namely provide a proportional amount of compute
3685 * time to each task. This is expressed in the following equation:
3687 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
3689 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
3690 * W_i,0 is defined as:
3692 * W_i,0 = \Sum_j w_i,j (2)
3694 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
3695 * is derived from the nice value as per prio_to_weight[].
3697 * The weight average is an exponential decay average of the instantaneous
3700 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
3702 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
3703 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
3704 * can also include other factors [XXX].
3706 * To achieve this balance we define a measure of imbalance which follows
3707 * directly from (1):
3709 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
3711 * We them move tasks around to minimize the imbalance. In the continuous
3712 * function space it is obvious this converges, in the discrete case we get
3713 * a few fun cases generally called infeasible weight scenarios.
3716 * - infeasible weights;
3717 * - local vs global optima in the discrete case. ]
3722 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
3723 * for all i,j solution, we create a tree of cpus that follows the hardware
3724 * topology where each level pairs two lower groups (or better). This results
3725 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
3726 * tree to only the first of the previous level and we decrease the frequency
3727 * of load-balance at each level inv. proportional to the number of cpus in
3733 * \Sum { --- * --- * 2^i } = O(n) (5)
3735 * `- size of each group
3736 * | | `- number of cpus doing load-balance
3738 * `- sum over all levels
3740 * Coupled with a limit on how many tasks we can migrate every balance pass,
3741 * this makes (5) the runtime complexity of the balancer.
3743 * An important property here is that each CPU is still (indirectly) connected
3744 * to every other cpu in at most O(log n) steps:
3746 * The adjacency matrix of the resulting graph is given by:
3749 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
3752 * And you'll find that:
3754 * A^(log_2 n)_i,j != 0 for all i,j (7)
3756 * Showing there's indeed a path between every cpu in at most O(log n) steps.
3757 * The task movement gives a factor of O(m), giving a convergence complexity
3760 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
3765 * In order to avoid CPUs going idle while there's still work to do, new idle
3766 * balancing is more aggressive and has the newly idle cpu iterate up the domain
3767 * tree itself instead of relying on other CPUs to bring it work.
3769 * This adds some complexity to both (5) and (8) but it reduces the total idle
3777 * Cgroups make a horror show out of (2), instead of a simple sum we get:
3780 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
3785 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
3787 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
3789 * The big problem is S_k, its a global sum needed to compute a local (W_i)
3792 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
3793 * rewrite all of this once again.]
3796 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
3798 #define LBF_ALL_PINNED 0x01
3799 #define LBF_NEED_BREAK 0x02
3800 #define LBF_SOME_PINNED 0x04
3803 struct sched_domain *sd;
3811 struct cpumask *dst_grpmask;
3813 enum cpu_idle_type idle;
3815 /* The set of CPUs under consideration for load-balancing */
3816 struct cpumask *cpus;
3821 unsigned int loop_break;
3822 unsigned int loop_max;
3826 * move_task - move a task from one runqueue to another runqueue.
3827 * Both runqueues must be locked.
3829 static void move_task(struct task_struct *p, struct lb_env *env)
3831 deactivate_task(env->src_rq, p, 0);
3832 set_task_cpu(p, env->dst_cpu);
3833 activate_task(env->dst_rq, p, 0);
3834 check_preempt_curr(env->dst_rq, p, 0);
3838 * Is this task likely cache-hot:
3841 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
3845 if (p->sched_class != &fair_sched_class)
3848 if (unlikely(p->policy == SCHED_IDLE))
3852 * Buddy candidates are cache hot:
3854 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
3855 (&p->se == cfs_rq_of(&p->se)->next ||
3856 &p->se == cfs_rq_of(&p->se)->last))
3859 if (sysctl_sched_migration_cost == -1)
3861 if (sysctl_sched_migration_cost == 0)
3864 delta = now - p->se.exec_start;
3866 return delta < (s64)sysctl_sched_migration_cost;
3870 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3873 int can_migrate_task(struct task_struct *p, struct lb_env *env)
3875 int tsk_cache_hot = 0;
3877 * We do not migrate tasks that are:
3878 * 1) running (obviously), or
3879 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3880 * 3) are cache-hot on their current CPU.
3882 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
3885 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
3888 * Remember if this task can be migrated to any other cpu in
3889 * our sched_group. We may want to revisit it if we couldn't
3890 * meet load balance goals by pulling other tasks on src_cpu.
3892 * Also avoid computing new_dst_cpu if we have already computed
3893 * one in current iteration.
3895 if (!env->dst_grpmask || (env->flags & LBF_SOME_PINNED))
3898 new_dst_cpu = cpumask_first_and(env->dst_grpmask,
3899 tsk_cpus_allowed(p));
3900 if (new_dst_cpu < nr_cpu_ids) {
3901 env->flags |= LBF_SOME_PINNED;
3902 env->new_dst_cpu = new_dst_cpu;
3907 /* Record that we found atleast one task that could run on dst_cpu */
3908 env->flags &= ~LBF_ALL_PINNED;
3910 if (task_running(env->src_rq, p)) {
3911 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
3916 * Aggressive migration if:
3917 * 1) task is cache cold, or
3918 * 2) too many balance attempts have failed.
3921 tsk_cache_hot = task_hot(p, env->src_rq->clock_task, env->sd);
3922 if (!tsk_cache_hot ||
3923 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
3925 if (tsk_cache_hot) {
3926 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
3927 schedstat_inc(p, se.statistics.nr_forced_migrations);
3933 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
3938 * move_one_task tries to move exactly one task from busiest to this_rq, as
3939 * part of active balancing operations within "domain".
3940 * Returns 1 if successful and 0 otherwise.
3942 * Called with both runqueues locked.
3944 static int move_one_task(struct lb_env *env)
3946 struct task_struct *p, *n;
3948 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
3949 if (throttled_lb_pair(task_group(p), env->src_rq->cpu, env->dst_cpu))
3952 if (!can_migrate_task(p, env))
3957 * Right now, this is only the second place move_task()
3958 * is called, so we can safely collect move_task()
3959 * stats here rather than inside move_task().
3961 schedstat_inc(env->sd, lb_gained[env->idle]);
3967 static unsigned long task_h_load(struct task_struct *p);
3969 static const unsigned int sched_nr_migrate_break = 32;
3972 * move_tasks tries to move up to imbalance weighted load from busiest to
3973 * this_rq, as part of a balancing operation within domain "sd".
3974 * Returns 1 if successful and 0 otherwise.
3976 * Called with both runqueues locked.
3978 static int move_tasks(struct lb_env *env)
3980 struct list_head *tasks = &env->src_rq->cfs_tasks;
3981 struct task_struct *p;
3985 if (env->imbalance <= 0)
3988 while (!list_empty(tasks)) {
3989 p = list_first_entry(tasks, struct task_struct, se.group_node);
3992 /* We've more or less seen every task there is, call it quits */
3993 if (env->loop > env->loop_max)
3996 /* take a breather every nr_migrate tasks */
3997 if (env->loop > env->loop_break) {
3998 env->loop_break += sched_nr_migrate_break;
3999 env->flags |= LBF_NEED_BREAK;
4003 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4006 load = task_h_load(p);
4008 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4011 if ((load / 2) > env->imbalance)
4014 if (!can_migrate_task(p, env))
4019 env->imbalance -= load;
4021 #ifdef CONFIG_PREEMPT
4023 * NEWIDLE balancing is a source of latency, so preemptible
4024 * kernels will stop after the first task is pulled to minimize
4025 * the critical section.
4027 if (env->idle == CPU_NEWLY_IDLE)
4032 * We only want to steal up to the prescribed amount of
4035 if (env->imbalance <= 0)
4040 list_move_tail(&p->se.group_node, tasks);
4044 * Right now, this is one of only two places move_task() is called,
4045 * so we can safely collect move_task() stats here rather than
4046 * inside move_task().
4048 schedstat_add(env->sd, lb_gained[env->idle], pulled);
4053 #ifdef CONFIG_FAIR_GROUP_SCHED
4055 * update tg->load_weight by folding this cpu's load_avg
4057 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
4059 struct sched_entity *se = tg->se[cpu];
4060 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
4062 /* throttled entities do not contribute to load */
4063 if (throttled_hierarchy(cfs_rq))
4066 update_cfs_rq_blocked_load(cfs_rq, 1);
4069 update_entity_load_avg(se, 1);
4071 * We pivot on our runnable average having decayed to zero for
4072 * list removal. This generally implies that all our children
4073 * have also been removed (modulo rounding error or bandwidth
4074 * control); however, such cases are rare and we can fix these
4077 * TODO: fix up out-of-order children on enqueue.
4079 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
4080 list_del_leaf_cfs_rq(cfs_rq);
4082 struct rq *rq = rq_of(cfs_rq);
4083 update_rq_runnable_avg(rq, rq->nr_running);
4087 static void update_blocked_averages(int cpu)
4089 struct rq *rq = cpu_rq(cpu);
4090 struct cfs_rq *cfs_rq;
4091 unsigned long flags;
4093 raw_spin_lock_irqsave(&rq->lock, flags);
4094 update_rq_clock(rq);
4096 * Iterates the task_group tree in a bottom up fashion, see
4097 * list_add_leaf_cfs_rq() for details.
4099 for_each_leaf_cfs_rq(rq, cfs_rq) {
4101 * Note: We may want to consider periodically releasing
4102 * rq->lock about these updates so that creating many task
4103 * groups does not result in continually extending hold time.
4105 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
4108 raw_spin_unlock_irqrestore(&rq->lock, flags);
4112 * Compute the cpu's hierarchical load factor for each task group.
4113 * This needs to be done in a top-down fashion because the load of a child
4114 * group is a fraction of its parents load.
4116 static int tg_load_down(struct task_group *tg, void *data)
4119 long cpu = (long)data;
4122 load = cpu_rq(cpu)->load.weight;
4124 load = tg->parent->cfs_rq[cpu]->h_load;
4125 load *= tg->se[cpu]->load.weight;
4126 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
4129 tg->cfs_rq[cpu]->h_load = load;
4134 static void update_h_load(long cpu)
4136 struct rq *rq = cpu_rq(cpu);
4137 unsigned long now = jiffies;
4139 if (rq->h_load_throttle == now)
4142 rq->h_load_throttle = now;
4145 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
4149 static unsigned long task_h_load(struct task_struct *p)
4151 struct cfs_rq *cfs_rq = task_cfs_rq(p);
4154 load = p->se.load.weight;
4155 load = div_u64(load * cfs_rq->h_load, cfs_rq->load.weight + 1);
4160 static inline void update_blocked_averages(int cpu)
4164 static inline void update_h_load(long cpu)
4168 static unsigned long task_h_load(struct task_struct *p)
4170 return p->se.load.weight;
4174 /********** Helpers for find_busiest_group ************************/
4176 * sd_lb_stats - Structure to store the statistics of a sched_domain
4177 * during load balancing.
4179 struct sd_lb_stats {
4180 struct sched_group *busiest; /* Busiest group in this sd */
4181 struct sched_group *this; /* Local group in this sd */
4182 unsigned long total_load; /* Total load of all groups in sd */
4183 unsigned long total_pwr; /* Total power of all groups in sd */
4184 unsigned long avg_load; /* Average load across all groups in sd */
4186 /** Statistics of this group */
4187 unsigned long this_load;
4188 unsigned long this_load_per_task;
4189 unsigned long this_nr_running;
4190 unsigned long this_has_capacity;
4191 unsigned int this_idle_cpus;
4193 /* Statistics of the busiest group */
4194 unsigned int busiest_idle_cpus;
4195 unsigned long max_load;
4196 unsigned long busiest_load_per_task;
4197 unsigned long busiest_nr_running;
4198 unsigned long busiest_group_capacity;
4199 unsigned long busiest_has_capacity;
4200 unsigned int busiest_group_weight;
4202 int group_imb; /* Is there imbalance in this sd */
4206 * sg_lb_stats - stats of a sched_group required for load_balancing
4208 struct sg_lb_stats {
4209 unsigned long avg_load; /*Avg load across the CPUs of the group */
4210 unsigned long group_load; /* Total load over the CPUs of the group */
4211 unsigned long sum_nr_running; /* Nr tasks running in the group */
4212 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
4213 unsigned long group_capacity;
4214 unsigned long idle_cpus;
4215 unsigned long group_weight;
4216 int group_imb; /* Is there an imbalance in the group ? */
4217 int group_has_capacity; /* Is there extra capacity in the group? */
4221 * get_sd_load_idx - Obtain the load index for a given sched domain.
4222 * @sd: The sched_domain whose load_idx is to be obtained.
4223 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4225 static inline int get_sd_load_idx(struct sched_domain *sd,
4226 enum cpu_idle_type idle)
4232 load_idx = sd->busy_idx;
4235 case CPU_NEWLY_IDLE:
4236 load_idx = sd->newidle_idx;
4239 load_idx = sd->idle_idx;
4246 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
4248 return SCHED_POWER_SCALE;
4251 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
4253 return default_scale_freq_power(sd, cpu);
4256 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
4258 unsigned long weight = sd->span_weight;
4259 unsigned long smt_gain = sd->smt_gain;
4266 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
4268 return default_scale_smt_power(sd, cpu);
4271 static unsigned long scale_rt_power(int cpu)
4273 struct rq *rq = cpu_rq(cpu);
4274 u64 total, available, age_stamp, avg;
4277 * Since we're reading these variables without serialization make sure
4278 * we read them once before doing sanity checks on them.
4280 age_stamp = ACCESS_ONCE(rq->age_stamp);
4281 avg = ACCESS_ONCE(rq->rt_avg);
4283 total = sched_avg_period() + (rq->clock - age_stamp);
4285 if (unlikely(total < avg)) {
4286 /* Ensures that power won't end up being negative */
4289 available = total - avg;
4292 if (unlikely((s64)total < SCHED_POWER_SCALE))
4293 total = SCHED_POWER_SCALE;
4295 total >>= SCHED_POWER_SHIFT;
4297 return div_u64(available, total);
4300 static void update_cpu_power(struct sched_domain *sd, int cpu)
4302 unsigned long weight = sd->span_weight;
4303 unsigned long power = SCHED_POWER_SCALE;
4304 struct sched_group *sdg = sd->groups;
4306 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
4307 if (sched_feat(ARCH_POWER))
4308 power *= arch_scale_smt_power(sd, cpu);
4310 power *= default_scale_smt_power(sd, cpu);
4312 power >>= SCHED_POWER_SHIFT;
4315 sdg->sgp->power_orig = power;
4317 if (sched_feat(ARCH_POWER))
4318 power *= arch_scale_freq_power(sd, cpu);
4320 power *= default_scale_freq_power(sd, cpu);
4322 power >>= SCHED_POWER_SHIFT;
4324 power *= scale_rt_power(cpu);
4325 power >>= SCHED_POWER_SHIFT;
4330 cpu_rq(cpu)->cpu_power = power;
4331 sdg->sgp->power = power;
4334 void update_group_power(struct sched_domain *sd, int cpu)
4336 struct sched_domain *child = sd->child;
4337 struct sched_group *group, *sdg = sd->groups;
4338 unsigned long power;
4339 unsigned long interval;
4341 interval = msecs_to_jiffies(sd->balance_interval);
4342 interval = clamp(interval, 1UL, max_load_balance_interval);
4343 sdg->sgp->next_update = jiffies + interval;
4346 update_cpu_power(sd, cpu);
4352 if (child->flags & SD_OVERLAP) {
4354 * SD_OVERLAP domains cannot assume that child groups
4355 * span the current group.
4358 for_each_cpu(cpu, sched_group_cpus(sdg))
4359 power += power_of(cpu);
4362 * !SD_OVERLAP domains can assume that child groups
4363 * span the current group.
4366 group = child->groups;
4368 power += group->sgp->power;
4369 group = group->next;
4370 } while (group != child->groups);
4373 sdg->sgp->power_orig = sdg->sgp->power = power;
4377 * Try and fix up capacity for tiny siblings, this is needed when
4378 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4379 * which on its own isn't powerful enough.
4381 * See update_sd_pick_busiest() and check_asym_packing().
4384 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
4387 * Only siblings can have significantly less than SCHED_POWER_SCALE
4389 if (!(sd->flags & SD_SHARE_CPUPOWER))
4393 * If ~90% of the cpu_power is still there, we're good.
4395 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
4402 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4403 * @env: The load balancing environment.
4404 * @group: sched_group whose statistics are to be updated.
4405 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4406 * @local_group: Does group contain this_cpu.
4407 * @balance: Should we balance.
4408 * @sgs: variable to hold the statistics for this group.
4410 static inline void update_sg_lb_stats(struct lb_env *env,
4411 struct sched_group *group, int load_idx,
4412 int local_group, int *balance, struct sg_lb_stats *sgs)
4414 unsigned long nr_running, max_nr_running, min_nr_running;
4415 unsigned long load, max_cpu_load, min_cpu_load;
4416 unsigned int balance_cpu = -1, first_idle_cpu = 0;
4417 unsigned long avg_load_per_task = 0;
4421 balance_cpu = group_balance_cpu(group);
4423 /* Tally up the load of all CPUs in the group */
4425 min_cpu_load = ~0UL;
4427 min_nr_running = ~0UL;
4429 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
4430 struct rq *rq = cpu_rq(i);
4432 nr_running = rq->nr_running;
4434 /* Bias balancing toward cpus of our domain */
4436 if (idle_cpu(i) && !first_idle_cpu &&
4437 cpumask_test_cpu(i, sched_group_mask(group))) {
4442 load = target_load(i, load_idx);
4444 load = source_load(i, load_idx);
4445 if (load > max_cpu_load)
4446 max_cpu_load = load;
4447 if (min_cpu_load > load)
4448 min_cpu_load = load;
4450 if (nr_running > max_nr_running)
4451 max_nr_running = nr_running;
4452 if (min_nr_running > nr_running)
4453 min_nr_running = nr_running;
4456 sgs->group_load += load;
4457 sgs->sum_nr_running += nr_running;
4458 sgs->sum_weighted_load += weighted_cpuload(i);
4464 * First idle cpu or the first cpu(busiest) in this sched group
4465 * is eligible for doing load balancing at this and above
4466 * domains. In the newly idle case, we will allow all the cpu's
4467 * to do the newly idle load balance.
4470 if (env->idle != CPU_NEWLY_IDLE) {
4471 if (balance_cpu != env->dst_cpu) {
4475 update_group_power(env->sd, env->dst_cpu);
4476 } else if (time_after_eq(jiffies, group->sgp->next_update))
4477 update_group_power(env->sd, env->dst_cpu);
4480 /* Adjust by relative CPU power of the group */
4481 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
4484 * Consider the group unbalanced when the imbalance is larger
4485 * than the average weight of a task.
4487 * APZ: with cgroup the avg task weight can vary wildly and
4488 * might not be a suitable number - should we keep a
4489 * normalized nr_running number somewhere that negates
4492 if (sgs->sum_nr_running)
4493 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
4495 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task &&
4496 (max_nr_running - min_nr_running) > 1)
4499 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
4501 if (!sgs->group_capacity)
4502 sgs->group_capacity = fix_small_capacity(env->sd, group);
4503 sgs->group_weight = group->group_weight;
4505 if (sgs->group_capacity > sgs->sum_nr_running)
4506 sgs->group_has_capacity = 1;
4510 * update_sd_pick_busiest - return 1 on busiest group
4511 * @env: The load balancing environment.
4512 * @sds: sched_domain statistics
4513 * @sg: sched_group candidate to be checked for being the busiest
4514 * @sgs: sched_group statistics
4516 * Determine if @sg is a busier group than the previously selected
4519 static bool update_sd_pick_busiest(struct lb_env *env,
4520 struct sd_lb_stats *sds,
4521 struct sched_group *sg,
4522 struct sg_lb_stats *sgs)
4524 if (sgs->avg_load <= sds->max_load)
4527 if (sgs->sum_nr_running > sgs->group_capacity)
4534 * ASYM_PACKING needs to move all the work to the lowest
4535 * numbered CPUs in the group, therefore mark all groups
4536 * higher than ourself as busy.
4538 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
4539 env->dst_cpu < group_first_cpu(sg)) {
4543 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
4551 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
4552 * @env: The load balancing environment.
4553 * @balance: Should we balance.
4554 * @sds: variable to hold the statistics for this sched_domain.
4556 static inline void update_sd_lb_stats(struct lb_env *env,
4557 int *balance, struct sd_lb_stats *sds)
4559 struct sched_domain *child = env->sd->child;
4560 struct sched_group *sg = env->sd->groups;
4561 struct sg_lb_stats sgs;
4562 int load_idx, prefer_sibling = 0;
4564 if (child && child->flags & SD_PREFER_SIBLING)
4567 load_idx = get_sd_load_idx(env->sd, env->idle);
4572 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
4573 memset(&sgs, 0, sizeof(sgs));
4574 update_sg_lb_stats(env, sg, load_idx, local_group, balance, &sgs);
4576 if (local_group && !(*balance))
4579 sds->total_load += sgs.group_load;
4580 sds->total_pwr += sg->sgp->power;
4583 * In case the child domain prefers tasks go to siblings
4584 * first, lower the sg capacity to one so that we'll try
4585 * and move all the excess tasks away. We lower the capacity
4586 * of a group only if the local group has the capacity to fit
4587 * these excess tasks, i.e. nr_running < group_capacity. The
4588 * extra check prevents the case where you always pull from the
4589 * heaviest group when it is already under-utilized (possible
4590 * with a large weight task outweighs the tasks on the system).
4592 if (prefer_sibling && !local_group && sds->this_has_capacity)
4593 sgs.group_capacity = min(sgs.group_capacity, 1UL);
4596 sds->this_load = sgs.avg_load;
4598 sds->this_nr_running = sgs.sum_nr_running;
4599 sds->this_load_per_task = sgs.sum_weighted_load;
4600 sds->this_has_capacity = sgs.group_has_capacity;
4601 sds->this_idle_cpus = sgs.idle_cpus;
4602 } else if (update_sd_pick_busiest(env, sds, sg, &sgs)) {
4603 sds->max_load = sgs.avg_load;
4605 sds->busiest_nr_running = sgs.sum_nr_running;
4606 sds->busiest_idle_cpus = sgs.idle_cpus;
4607 sds->busiest_group_capacity = sgs.group_capacity;
4608 sds->busiest_load_per_task = sgs.sum_weighted_load;
4609 sds->busiest_has_capacity = sgs.group_has_capacity;
4610 sds->busiest_group_weight = sgs.group_weight;
4611 sds->group_imb = sgs.group_imb;
4615 } while (sg != env->sd->groups);
4619 * check_asym_packing - Check to see if the group is packed into the
4622 * This is primarily intended to used at the sibling level. Some
4623 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4624 * case of POWER7, it can move to lower SMT modes only when higher
4625 * threads are idle. When in lower SMT modes, the threads will
4626 * perform better since they share less core resources. Hence when we
4627 * have idle threads, we want them to be the higher ones.
4629 * This packing function is run on idle threads. It checks to see if
4630 * the busiest CPU in this domain (core in the P7 case) has a higher
4631 * CPU number than the packing function is being run on. Here we are
4632 * assuming lower CPU number will be equivalent to lower a SMT thread
4635 * Returns 1 when packing is required and a task should be moved to
4636 * this CPU. The amount of the imbalance is returned in *imbalance.
4638 * @env: The load balancing environment.
4639 * @sds: Statistics of the sched_domain which is to be packed
4641 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
4645 if (!(env->sd->flags & SD_ASYM_PACKING))
4651 busiest_cpu = group_first_cpu(sds->busiest);
4652 if (env->dst_cpu > busiest_cpu)
4655 env->imbalance = DIV_ROUND_CLOSEST(
4656 sds->max_load * sds->busiest->sgp->power, SCHED_POWER_SCALE);
4662 * fix_small_imbalance - Calculate the minor imbalance that exists
4663 * amongst the groups of a sched_domain, during
4665 * @env: The load balancing environment.
4666 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4669 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4671 unsigned long tmp, pwr_now = 0, pwr_move = 0;
4672 unsigned int imbn = 2;
4673 unsigned long scaled_busy_load_per_task;
4675 if (sds->this_nr_running) {
4676 sds->this_load_per_task /= sds->this_nr_running;
4677 if (sds->busiest_load_per_task >
4678 sds->this_load_per_task)
4681 sds->this_load_per_task =
4682 cpu_avg_load_per_task(env->dst_cpu);
4685 scaled_busy_load_per_task = sds->busiest_load_per_task
4686 * SCHED_POWER_SCALE;
4687 scaled_busy_load_per_task /= sds->busiest->sgp->power;
4689 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
4690 (scaled_busy_load_per_task * imbn)) {
4691 env->imbalance = sds->busiest_load_per_task;
4696 * OK, we don't have enough imbalance to justify moving tasks,
4697 * however we may be able to increase total CPU power used by
4701 pwr_now += sds->busiest->sgp->power *
4702 min(sds->busiest_load_per_task, sds->max_load);
4703 pwr_now += sds->this->sgp->power *
4704 min(sds->this_load_per_task, sds->this_load);
4705 pwr_now /= SCHED_POWER_SCALE;
4707 /* Amount of load we'd subtract */
4708 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4709 sds->busiest->sgp->power;
4710 if (sds->max_load > tmp)
4711 pwr_move += sds->busiest->sgp->power *
4712 min(sds->busiest_load_per_task, sds->max_load - tmp);
4714 /* Amount of load we'd add */
4715 if (sds->max_load * sds->busiest->sgp->power <
4716 sds->busiest_load_per_task * SCHED_POWER_SCALE)
4717 tmp = (sds->max_load * sds->busiest->sgp->power) /
4718 sds->this->sgp->power;
4720 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4721 sds->this->sgp->power;
4722 pwr_move += sds->this->sgp->power *
4723 min(sds->this_load_per_task, sds->this_load + tmp);
4724 pwr_move /= SCHED_POWER_SCALE;
4726 /* Move if we gain throughput */
4727 if (pwr_move > pwr_now)
4728 env->imbalance = sds->busiest_load_per_task;
4732 * calculate_imbalance - Calculate the amount of imbalance present within the
4733 * groups of a given sched_domain during load balance.
4734 * @env: load balance environment
4735 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4737 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4739 unsigned long max_pull, load_above_capacity = ~0UL;
4741 sds->busiest_load_per_task /= sds->busiest_nr_running;
4742 if (sds->group_imb) {
4743 sds->busiest_load_per_task =
4744 min(sds->busiest_load_per_task, sds->avg_load);
4748 * In the presence of smp nice balancing, certain scenarios can have
4749 * max load less than avg load(as we skip the groups at or below
4750 * its cpu_power, while calculating max_load..)
4752 if (sds->max_load < sds->avg_load) {
4754 return fix_small_imbalance(env, sds);
4757 if (!sds->group_imb) {
4759 * Don't want to pull so many tasks that a group would go idle.
4761 load_above_capacity = (sds->busiest_nr_running -
4762 sds->busiest_group_capacity);
4764 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
4766 load_above_capacity /= sds->busiest->sgp->power;
4770 * We're trying to get all the cpus to the average_load, so we don't
4771 * want to push ourselves above the average load, nor do we wish to
4772 * reduce the max loaded cpu below the average load. At the same time,
4773 * we also don't want to reduce the group load below the group capacity
4774 * (so that we can implement power-savings policies etc). Thus we look
4775 * for the minimum possible imbalance.
4776 * Be careful of negative numbers as they'll appear as very large values
4777 * with unsigned longs.
4779 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
4781 /* How much load to actually move to equalise the imbalance */
4782 env->imbalance = min(max_pull * sds->busiest->sgp->power,
4783 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
4784 / SCHED_POWER_SCALE;
4787 * if *imbalance is less than the average load per runnable task
4788 * there is no guarantee that any tasks will be moved so we'll have
4789 * a think about bumping its value to force at least one task to be
4792 if (env->imbalance < sds->busiest_load_per_task)
4793 return fix_small_imbalance(env, sds);
4797 /******* find_busiest_group() helpers end here *********************/
4800 * find_busiest_group - Returns the busiest group within the sched_domain
4801 * if there is an imbalance. If there isn't an imbalance, and
4802 * the user has opted for power-savings, it returns a group whose
4803 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4804 * such a group exists.
4806 * Also calculates the amount of weighted load which should be moved
4807 * to restore balance.
4809 * @env: The load balancing environment.
4810 * @balance: Pointer to a variable indicating if this_cpu
4811 * is the appropriate cpu to perform load balancing at this_level.
4813 * Returns: - the busiest group if imbalance exists.
4814 * - If no imbalance and user has opted for power-savings balance,
4815 * return the least loaded group whose CPUs can be
4816 * put to idle by rebalancing its tasks onto our group.
4818 static struct sched_group *
4819 find_busiest_group(struct lb_env *env, int *balance)
4821 struct sd_lb_stats sds;
4823 memset(&sds, 0, sizeof(sds));
4826 * Compute the various statistics relavent for load balancing at
4829 update_sd_lb_stats(env, balance, &sds);
4832 * this_cpu is not the appropriate cpu to perform load balancing at
4838 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
4839 check_asym_packing(env, &sds))
4842 /* There is no busy sibling group to pull tasks from */
4843 if (!sds.busiest || sds.busiest_nr_running == 0)
4846 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
4849 * If the busiest group is imbalanced the below checks don't
4850 * work because they assumes all things are equal, which typically
4851 * isn't true due to cpus_allowed constraints and the like.
4856 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
4857 if (env->idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
4858 !sds.busiest_has_capacity)
4862 * If the local group is more busy than the selected busiest group
4863 * don't try and pull any tasks.
4865 if (sds.this_load >= sds.max_load)
4869 * Don't pull any tasks if this group is already above the domain
4872 if (sds.this_load >= sds.avg_load)
4875 if (env->idle == CPU_IDLE) {
4877 * This cpu is idle. If the busiest group load doesn't
4878 * have more tasks than the number of available cpu's and
4879 * there is no imbalance between this and busiest group
4880 * wrt to idle cpu's, it is balanced.
4882 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
4883 sds.busiest_nr_running <= sds.busiest_group_weight)
4887 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
4888 * imbalance_pct to be conservative.
4890 if (100 * sds.max_load <= env->sd->imbalance_pct * sds.this_load)
4895 /* Looks like there is an imbalance. Compute it */
4896 calculate_imbalance(env, &sds);
4906 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4908 static struct rq *find_busiest_queue(struct lb_env *env,
4909 struct sched_group *group)
4911 struct rq *busiest = NULL, *rq;
4912 unsigned long max_load = 0;
4915 for_each_cpu(i, sched_group_cpus(group)) {
4916 unsigned long power = power_of(i);
4917 unsigned long capacity = DIV_ROUND_CLOSEST(power,
4922 capacity = fix_small_capacity(env->sd, group);
4924 if (!cpumask_test_cpu(i, env->cpus))
4928 wl = weighted_cpuload(i);
4931 * When comparing with imbalance, use weighted_cpuload()
4932 * which is not scaled with the cpu power.
4934 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
4938 * For the load comparisons with the other cpu's, consider
4939 * the weighted_cpuload() scaled with the cpu power, so that
4940 * the load can be moved away from the cpu that is potentially
4941 * running at a lower capacity.
4943 wl = (wl * SCHED_POWER_SCALE) / power;
4945 if (wl > max_load) {
4955 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4956 * so long as it is large enough.
4958 #define MAX_PINNED_INTERVAL 512
4960 /* Working cpumask for load_balance and load_balance_newidle. */
4961 DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4963 static int need_active_balance(struct lb_env *env)
4965 struct sched_domain *sd = env->sd;
4967 if (env->idle == CPU_NEWLY_IDLE) {
4970 * ASYM_PACKING needs to force migrate tasks from busy but
4971 * higher numbered CPUs in order to pack all tasks in the
4972 * lowest numbered CPUs.
4974 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
4978 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
4981 static int active_load_balance_cpu_stop(void *data);
4984 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4985 * tasks if there is an imbalance.
4987 static int load_balance(int this_cpu, struct rq *this_rq,
4988 struct sched_domain *sd, enum cpu_idle_type idle,
4991 int ld_moved, cur_ld_moved, active_balance = 0;
4992 int lb_iterations, max_lb_iterations;
4993 struct sched_group *group;
4995 unsigned long flags;
4996 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4998 struct lb_env env = {
5000 .dst_cpu = this_cpu,
5002 .dst_grpmask = sched_group_cpus(sd->groups),
5004 .loop_break = sched_nr_migrate_break,
5008 cpumask_copy(cpus, cpu_active_mask);
5009 max_lb_iterations = cpumask_weight(env.dst_grpmask);
5011 schedstat_inc(sd, lb_count[idle]);
5014 group = find_busiest_group(&env, balance);
5020 schedstat_inc(sd, lb_nobusyg[idle]);
5024 busiest = find_busiest_queue(&env, group);
5026 schedstat_inc(sd, lb_nobusyq[idle]);
5030 BUG_ON(busiest == env.dst_rq);
5032 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
5036 if (busiest->nr_running > 1) {
5038 * Attempt to move tasks. If find_busiest_group has found
5039 * an imbalance but busiest->nr_running <= 1, the group is
5040 * still unbalanced. ld_moved simply stays zero, so it is
5041 * correctly treated as an imbalance.
5043 env.flags |= LBF_ALL_PINNED;
5044 env.src_cpu = busiest->cpu;
5045 env.src_rq = busiest;
5046 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
5048 update_h_load(env.src_cpu);
5050 local_irq_save(flags);
5051 double_rq_lock(env.dst_rq, busiest);
5054 * cur_ld_moved - load moved in current iteration
5055 * ld_moved - cumulative load moved across iterations
5057 cur_ld_moved = move_tasks(&env);
5058 ld_moved += cur_ld_moved;
5059 double_rq_unlock(env.dst_rq, busiest);
5060 local_irq_restore(flags);
5062 if (env.flags & LBF_NEED_BREAK) {
5063 env.flags &= ~LBF_NEED_BREAK;
5068 * some other cpu did the load balance for us.
5070 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
5071 resched_cpu(env.dst_cpu);
5074 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5075 * us and move them to an alternate dst_cpu in our sched_group
5076 * where they can run. The upper limit on how many times we
5077 * iterate on same src_cpu is dependent on number of cpus in our
5080 * This changes load balance semantics a bit on who can move
5081 * load to a given_cpu. In addition to the given_cpu itself
5082 * (or a ilb_cpu acting on its behalf where given_cpu is
5083 * nohz-idle), we now have balance_cpu in a position to move
5084 * load to given_cpu. In rare situations, this may cause
5085 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5086 * _independently_ and at _same_ time to move some load to
5087 * given_cpu) causing exceess load to be moved to given_cpu.
5088 * This however should not happen so much in practice and
5089 * moreover subsequent load balance cycles should correct the
5090 * excess load moved.
5092 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0 &&
5093 lb_iterations++ < max_lb_iterations) {
5095 env.dst_rq = cpu_rq(env.new_dst_cpu);
5096 env.dst_cpu = env.new_dst_cpu;
5097 env.flags &= ~LBF_SOME_PINNED;
5099 env.loop_break = sched_nr_migrate_break;
5101 * Go back to "more_balance" rather than "redo" since we
5102 * need to continue with same src_cpu.
5107 /* All tasks on this runqueue were pinned by CPU affinity */
5108 if (unlikely(env.flags & LBF_ALL_PINNED)) {
5109 cpumask_clear_cpu(cpu_of(busiest), cpus);
5110 if (!cpumask_empty(cpus)) {
5112 env.loop_break = sched_nr_migrate_break;
5120 schedstat_inc(sd, lb_failed[idle]);
5122 * Increment the failure counter only on periodic balance.
5123 * We do not want newidle balance, which can be very
5124 * frequent, pollute the failure counter causing
5125 * excessive cache_hot migrations and active balances.
5127 if (idle != CPU_NEWLY_IDLE)
5128 sd->nr_balance_failed++;
5130 if (need_active_balance(&env)) {
5131 raw_spin_lock_irqsave(&busiest->lock, flags);
5133 /* don't kick the active_load_balance_cpu_stop,
5134 * if the curr task on busiest cpu can't be
5137 if (!cpumask_test_cpu(this_cpu,
5138 tsk_cpus_allowed(busiest->curr))) {
5139 raw_spin_unlock_irqrestore(&busiest->lock,
5141 env.flags |= LBF_ALL_PINNED;
5142 goto out_one_pinned;
5146 * ->active_balance synchronizes accesses to
5147 * ->active_balance_work. Once set, it's cleared
5148 * only after active load balance is finished.
5150 if (!busiest->active_balance) {
5151 busiest->active_balance = 1;
5152 busiest->push_cpu = this_cpu;
5155 raw_spin_unlock_irqrestore(&busiest->lock, flags);
5157 if (active_balance) {
5158 stop_one_cpu_nowait(cpu_of(busiest),
5159 active_load_balance_cpu_stop, busiest,
5160 &busiest->active_balance_work);
5164 * We've kicked active balancing, reset the failure
5167 sd->nr_balance_failed = sd->cache_nice_tries+1;
5170 sd->nr_balance_failed = 0;
5172 if (likely(!active_balance)) {
5173 /* We were unbalanced, so reset the balancing interval */
5174 sd->balance_interval = sd->min_interval;
5177 * If we've begun active balancing, start to back off. This
5178 * case may not be covered by the all_pinned logic if there
5179 * is only 1 task on the busy runqueue (because we don't call
5182 if (sd->balance_interval < sd->max_interval)
5183 sd->balance_interval *= 2;
5189 schedstat_inc(sd, lb_balanced[idle]);
5191 sd->nr_balance_failed = 0;
5194 /* tune up the balancing interval */
5195 if (((env.flags & LBF_ALL_PINNED) &&
5196 sd->balance_interval < MAX_PINNED_INTERVAL) ||
5197 (sd->balance_interval < sd->max_interval))
5198 sd->balance_interval *= 2;
5206 * idle_balance is called by schedule() if this_cpu is about to become
5207 * idle. Attempts to pull tasks from other CPUs.
5209 void idle_balance(int this_cpu, struct rq *this_rq)
5211 struct sched_domain *sd;
5212 int pulled_task = 0;
5213 unsigned long next_balance = jiffies + HZ;
5215 this_rq->idle_stamp = this_rq->clock;
5217 if (this_rq->avg_idle < sysctl_sched_migration_cost)
5220 update_rq_runnable_avg(this_rq, 1);
5223 * Drop the rq->lock, but keep IRQ/preempt disabled.
5225 raw_spin_unlock(&this_rq->lock);
5227 update_blocked_averages(this_cpu);
5229 for_each_domain(this_cpu, sd) {
5230 unsigned long interval;
5233 if (!(sd->flags & SD_LOAD_BALANCE))
5236 if (sd->flags & SD_BALANCE_NEWIDLE) {
5237 /* If we've pulled tasks over stop searching: */
5238 pulled_task = load_balance(this_cpu, this_rq,
5239 sd, CPU_NEWLY_IDLE, &balance);
5242 interval = msecs_to_jiffies(sd->balance_interval);
5243 if (time_after(next_balance, sd->last_balance + interval))
5244 next_balance = sd->last_balance + interval;
5246 this_rq->idle_stamp = 0;
5252 raw_spin_lock(&this_rq->lock);
5254 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
5256 * We are going idle. next_balance may be set based on
5257 * a busy processor. So reset next_balance.
5259 this_rq->next_balance = next_balance;
5264 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5265 * running tasks off the busiest CPU onto idle CPUs. It requires at
5266 * least 1 task to be running on each physical CPU where possible, and
5267 * avoids physical / logical imbalances.
5269 static int active_load_balance_cpu_stop(void *data)
5271 struct rq *busiest_rq = data;
5272 int busiest_cpu = cpu_of(busiest_rq);
5273 int target_cpu = busiest_rq->push_cpu;
5274 struct rq *target_rq = cpu_rq(target_cpu);
5275 struct sched_domain *sd;
5277 raw_spin_lock_irq(&busiest_rq->lock);
5279 /* make sure the requested cpu hasn't gone down in the meantime */
5280 if (unlikely(busiest_cpu != smp_processor_id() ||
5281 !busiest_rq->active_balance))
5284 /* Is there any task to move? */
5285 if (busiest_rq->nr_running <= 1)
5289 * This condition is "impossible", if it occurs
5290 * we need to fix it. Originally reported by
5291 * Bjorn Helgaas on a 128-cpu setup.
5293 BUG_ON(busiest_rq == target_rq);
5295 /* move a task from busiest_rq to target_rq */
5296 double_lock_balance(busiest_rq, target_rq);
5298 /* Search for an sd spanning us and the target CPU. */
5300 for_each_domain(target_cpu, sd) {
5301 if ((sd->flags & SD_LOAD_BALANCE) &&
5302 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
5307 struct lb_env env = {
5309 .dst_cpu = target_cpu,
5310 .dst_rq = target_rq,
5311 .src_cpu = busiest_rq->cpu,
5312 .src_rq = busiest_rq,
5316 schedstat_inc(sd, alb_count);
5318 if (move_one_task(&env))
5319 schedstat_inc(sd, alb_pushed);
5321 schedstat_inc(sd, alb_failed);
5324 double_unlock_balance(busiest_rq, target_rq);
5326 busiest_rq->active_balance = 0;
5327 raw_spin_unlock_irq(&busiest_rq->lock);
5333 * idle load balancing details
5334 * - When one of the busy CPUs notice that there may be an idle rebalancing
5335 * needed, they will kick the idle load balancer, which then does idle
5336 * load balancing for all the idle CPUs.
5339 cpumask_var_t idle_cpus_mask;
5341 unsigned long next_balance; /* in jiffy units */
5342 } nohz ____cacheline_aligned;
5344 static inline int find_new_ilb(int call_cpu)
5346 int ilb = cpumask_first(nohz.idle_cpus_mask);
5348 if (ilb < nr_cpu_ids && idle_cpu(ilb))
5355 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5356 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5357 * CPU (if there is one).
5359 static void nohz_balancer_kick(int cpu)
5363 nohz.next_balance++;
5365 ilb_cpu = find_new_ilb(cpu);
5367 if (ilb_cpu >= nr_cpu_ids)
5370 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
5373 * Use smp_send_reschedule() instead of resched_cpu().
5374 * This way we generate a sched IPI on the target cpu which
5375 * is idle. And the softirq performing nohz idle load balance
5376 * will be run before returning from the IPI.
5378 smp_send_reschedule(ilb_cpu);
5382 static inline void nohz_balance_exit_idle(int cpu)
5384 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5385 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5386 atomic_dec(&nohz.nr_cpus);
5387 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5391 static inline void set_cpu_sd_state_busy(void)
5393 struct sched_domain *sd;
5394 int cpu = smp_processor_id();
5396 if (!test_bit(NOHZ_IDLE, nohz_flags(cpu)))
5398 clear_bit(NOHZ_IDLE, nohz_flags(cpu));
5401 for_each_domain(cpu, sd)
5402 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
5406 void set_cpu_sd_state_idle(void)
5408 struct sched_domain *sd;
5409 int cpu = smp_processor_id();
5411 if (test_bit(NOHZ_IDLE, nohz_flags(cpu)))
5413 set_bit(NOHZ_IDLE, nohz_flags(cpu));
5416 for_each_domain(cpu, sd)
5417 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
5422 * This routine will record that the cpu is going idle with tick stopped.
5423 * This info will be used in performing idle load balancing in the future.
5425 void nohz_balance_enter_idle(int cpu)
5428 * If this cpu is going down, then nothing needs to be done.
5430 if (!cpu_active(cpu))
5433 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
5436 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
5437 atomic_inc(&nohz.nr_cpus);
5438 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5441 static int __cpuinit sched_ilb_notifier(struct notifier_block *nfb,
5442 unsigned long action, void *hcpu)
5444 switch (action & ~CPU_TASKS_FROZEN) {
5446 nohz_balance_exit_idle(smp_processor_id());
5454 static DEFINE_SPINLOCK(balancing);
5457 * Scale the max load_balance interval with the number of CPUs in the system.
5458 * This trades load-balance latency on larger machines for less cross talk.
5460 void update_max_interval(void)
5462 max_load_balance_interval = HZ*num_online_cpus()/10;
5466 * It checks each scheduling domain to see if it is due to be balanced,
5467 * and initiates a balancing operation if so.
5469 * Balancing parameters are set up in arch_init_sched_domains.
5471 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
5474 struct rq *rq = cpu_rq(cpu);
5475 unsigned long interval;
5476 struct sched_domain *sd;
5477 /* Earliest time when we have to do rebalance again */
5478 unsigned long next_balance = jiffies + 60*HZ;
5479 int update_next_balance = 0;
5482 update_blocked_averages(cpu);
5485 for_each_domain(cpu, sd) {
5486 if (!(sd->flags & SD_LOAD_BALANCE))
5489 interval = sd->balance_interval;
5490 if (idle != CPU_IDLE)
5491 interval *= sd->busy_factor;
5493 /* scale ms to jiffies */
5494 interval = msecs_to_jiffies(interval);
5495 interval = clamp(interval, 1UL, max_load_balance_interval);
5497 need_serialize = sd->flags & SD_SERIALIZE;
5499 if (need_serialize) {
5500 if (!spin_trylock(&balancing))
5504 if (time_after_eq(jiffies, sd->last_balance + interval)) {
5505 if (load_balance(cpu, rq, sd, idle, &balance)) {
5507 * We've pulled tasks over so either we're no
5510 idle = CPU_NOT_IDLE;
5512 sd->last_balance = jiffies;
5515 spin_unlock(&balancing);
5517 if (time_after(next_balance, sd->last_balance + interval)) {
5518 next_balance = sd->last_balance + interval;
5519 update_next_balance = 1;
5523 * Stop the load balance at this level. There is another
5524 * CPU in our sched group which is doing load balancing more
5533 * next_balance will be updated only when there is a need.
5534 * When the cpu is attached to null domain for ex, it will not be
5537 if (likely(update_next_balance))
5538 rq->next_balance = next_balance;
5543 * In CONFIG_NO_HZ case, the idle balance kickee will do the
5544 * rebalancing for all the cpus for whom scheduler ticks are stopped.
5546 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
5548 struct rq *this_rq = cpu_rq(this_cpu);
5552 if (idle != CPU_IDLE ||
5553 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
5556 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
5557 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
5561 * If this cpu gets work to do, stop the load balancing
5562 * work being done for other cpus. Next load
5563 * balancing owner will pick it up.
5568 rq = cpu_rq(balance_cpu);
5570 raw_spin_lock_irq(&rq->lock);
5571 update_rq_clock(rq);
5572 update_idle_cpu_load(rq);
5573 raw_spin_unlock_irq(&rq->lock);
5575 rebalance_domains(balance_cpu, CPU_IDLE);
5577 if (time_after(this_rq->next_balance, rq->next_balance))
5578 this_rq->next_balance = rq->next_balance;
5580 nohz.next_balance = this_rq->next_balance;
5582 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
5586 * Current heuristic for kicking the idle load balancer in the presence
5587 * of an idle cpu is the system.
5588 * - This rq has more than one task.
5589 * - At any scheduler domain level, this cpu's scheduler group has multiple
5590 * busy cpu's exceeding the group's power.
5591 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
5592 * domain span are idle.
5594 static inline int nohz_kick_needed(struct rq *rq, int cpu)
5596 unsigned long now = jiffies;
5597 struct sched_domain *sd;
5599 if (unlikely(idle_cpu(cpu)))
5603 * We may be recently in ticked or tickless idle mode. At the first
5604 * busy tick after returning from idle, we will update the busy stats.
5606 set_cpu_sd_state_busy();
5607 nohz_balance_exit_idle(cpu);
5610 * None are in tickless mode and hence no need for NOHZ idle load
5613 if (likely(!atomic_read(&nohz.nr_cpus)))
5616 if (time_before(now, nohz.next_balance))
5619 if (rq->nr_running >= 2)
5623 for_each_domain(cpu, sd) {
5624 struct sched_group *sg = sd->groups;
5625 struct sched_group_power *sgp = sg->sgp;
5626 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
5628 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
5629 goto need_kick_unlock;
5631 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
5632 && (cpumask_first_and(nohz.idle_cpus_mask,
5633 sched_domain_span(sd)) < cpu))
5634 goto need_kick_unlock;
5636 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
5648 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
5652 * run_rebalance_domains is triggered when needed from the scheduler tick.
5653 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
5655 static void run_rebalance_domains(struct softirq_action *h)
5657 int this_cpu = smp_processor_id();
5658 struct rq *this_rq = cpu_rq(this_cpu);
5659 enum cpu_idle_type idle = this_rq->idle_balance ?
5660 CPU_IDLE : CPU_NOT_IDLE;
5662 rebalance_domains(this_cpu, idle);
5665 * If this cpu has a pending nohz_balance_kick, then do the
5666 * balancing on behalf of the other idle cpus whose ticks are
5669 nohz_idle_balance(this_cpu, idle);
5672 static inline int on_null_domain(int cpu)
5674 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
5678 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
5680 void trigger_load_balance(struct rq *rq, int cpu)
5682 /* Don't need to rebalance while attached to NULL domain */
5683 if (time_after_eq(jiffies, rq->next_balance) &&
5684 likely(!on_null_domain(cpu)))
5685 raise_softirq(SCHED_SOFTIRQ);
5687 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
5688 nohz_balancer_kick(cpu);
5692 static void rq_online_fair(struct rq *rq)
5697 static void rq_offline_fair(struct rq *rq)
5701 /* Ensure any throttled groups are reachable by pick_next_task */
5702 unthrottle_offline_cfs_rqs(rq);
5705 #endif /* CONFIG_SMP */
5708 * scheduler tick hitting a task of our scheduling class:
5710 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
5712 struct cfs_rq *cfs_rq;
5713 struct sched_entity *se = &curr->se;
5715 for_each_sched_entity(se) {
5716 cfs_rq = cfs_rq_of(se);
5717 entity_tick(cfs_rq, se, queued);
5720 if (sched_feat_numa(NUMA))
5721 task_tick_numa(rq, curr);
5723 update_rq_runnable_avg(rq, 1);
5727 * called on fork with the child task as argument from the parent's context
5728 * - child not yet on the tasklist
5729 * - preemption disabled
5731 static void task_fork_fair(struct task_struct *p)
5733 struct cfs_rq *cfs_rq;
5734 struct sched_entity *se = &p->se, *curr;
5735 int this_cpu = smp_processor_id();
5736 struct rq *rq = this_rq();
5737 unsigned long flags;
5739 raw_spin_lock_irqsave(&rq->lock, flags);
5741 update_rq_clock(rq);
5743 cfs_rq = task_cfs_rq(current);
5744 curr = cfs_rq->curr;
5746 if (unlikely(task_cpu(p) != this_cpu)) {
5748 __set_task_cpu(p, this_cpu);
5752 update_curr(cfs_rq);
5755 se->vruntime = curr->vruntime;
5756 place_entity(cfs_rq, se, 1);
5758 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
5760 * Upon rescheduling, sched_class::put_prev_task() will place
5761 * 'current' within the tree based on its new key value.
5763 swap(curr->vruntime, se->vruntime);
5764 resched_task(rq->curr);
5767 se->vruntime -= cfs_rq->min_vruntime;
5769 raw_spin_unlock_irqrestore(&rq->lock, flags);
5773 * Priority of the task has changed. Check to see if we preempt
5777 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
5783 * Reschedule if we are currently running on this runqueue and
5784 * our priority decreased, or if we are not currently running on
5785 * this runqueue and our priority is higher than the current's
5787 if (rq->curr == p) {
5788 if (p->prio > oldprio)
5789 resched_task(rq->curr);
5791 check_preempt_curr(rq, p, 0);
5794 static void switched_from_fair(struct rq *rq, struct task_struct *p)
5796 struct sched_entity *se = &p->se;
5797 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5800 * Ensure the task's vruntime is normalized, so that when its
5801 * switched back to the fair class the enqueue_entity(.flags=0) will
5802 * do the right thing.
5804 * If it was on_rq, then the dequeue_entity(.flags=0) will already
5805 * have normalized the vruntime, if it was !on_rq, then only when
5806 * the task is sleeping will it still have non-normalized vruntime.
5808 if (!se->on_rq && p->state != TASK_RUNNING) {
5810 * Fix up our vruntime so that the current sleep doesn't
5811 * cause 'unlimited' sleep bonus.
5813 place_entity(cfs_rq, se, 0);
5814 se->vruntime -= cfs_rq->min_vruntime;
5817 #if defined(CONFIG_FAIR_GROUP_SCHED) && defined(CONFIG_SMP)
5819 * Remove our load from contribution when we leave sched_fair
5820 * and ensure we don't carry in an old decay_count if we
5823 if (p->se.avg.decay_count) {
5824 struct cfs_rq *cfs_rq = cfs_rq_of(&p->se);
5825 __synchronize_entity_decay(&p->se);
5826 subtract_blocked_load_contrib(cfs_rq,
5827 p->se.avg.load_avg_contrib);
5833 * We switched to the sched_fair class.
5835 static void switched_to_fair(struct rq *rq, struct task_struct *p)
5841 * We were most likely switched from sched_rt, so
5842 * kick off the schedule if running, otherwise just see
5843 * if we can still preempt the current task.
5846 resched_task(rq->curr);
5848 check_preempt_curr(rq, p, 0);
5851 /* Account for a task changing its policy or group.
5853 * This routine is mostly called to set cfs_rq->curr field when a task
5854 * migrates between groups/classes.
5856 static void set_curr_task_fair(struct rq *rq)
5858 struct sched_entity *se = &rq->curr->se;
5860 for_each_sched_entity(se) {
5861 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5863 set_next_entity(cfs_rq, se);
5864 /* ensure bandwidth has been allocated on our new cfs_rq */
5865 account_cfs_rq_runtime(cfs_rq, 0);
5869 void init_cfs_rq(struct cfs_rq *cfs_rq)
5871 cfs_rq->tasks_timeline = RB_ROOT;
5872 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
5873 #ifndef CONFIG_64BIT
5874 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
5876 #if defined(CONFIG_FAIR_GROUP_SCHED) && defined(CONFIG_SMP)
5877 atomic64_set(&cfs_rq->decay_counter, 1);
5878 atomic64_set(&cfs_rq->removed_load, 0);
5882 #ifdef CONFIG_FAIR_GROUP_SCHED
5883 static void task_move_group_fair(struct task_struct *p, int on_rq)
5885 struct cfs_rq *cfs_rq;
5887 * If the task was not on the rq at the time of this cgroup movement
5888 * it must have been asleep, sleeping tasks keep their ->vruntime
5889 * absolute on their old rq until wakeup (needed for the fair sleeper
5890 * bonus in place_entity()).
5892 * If it was on the rq, we've just 'preempted' it, which does convert
5893 * ->vruntime to a relative base.
5895 * Make sure both cases convert their relative position when migrating
5896 * to another cgroup's rq. This does somewhat interfere with the
5897 * fair sleeper stuff for the first placement, but who cares.
5900 * When !on_rq, vruntime of the task has usually NOT been normalized.
5901 * But there are some cases where it has already been normalized:
5903 * - Moving a forked child which is waiting for being woken up by
5904 * wake_up_new_task().
5905 * - Moving a task which has been woken up by try_to_wake_up() and
5906 * waiting for actually being woken up by sched_ttwu_pending().
5908 * To prevent boost or penalty in the new cfs_rq caused by delta
5909 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
5911 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
5915 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
5916 set_task_rq(p, task_cpu(p));
5918 cfs_rq = cfs_rq_of(&p->se);
5919 p->se.vruntime += cfs_rq->min_vruntime;
5922 * migrate_task_rq_fair() will have removed our previous
5923 * contribution, but we must synchronize for ongoing future
5926 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
5927 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
5932 void free_fair_sched_group(struct task_group *tg)
5936 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
5938 for_each_possible_cpu(i) {
5940 kfree(tg->cfs_rq[i]);
5949 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
5951 struct cfs_rq *cfs_rq;
5952 struct sched_entity *se;
5955 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
5958 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
5962 tg->shares = NICE_0_LOAD;
5964 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
5966 for_each_possible_cpu(i) {
5967 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
5968 GFP_KERNEL, cpu_to_node(i));
5972 se = kzalloc_node(sizeof(struct sched_entity),
5973 GFP_KERNEL, cpu_to_node(i));
5977 init_cfs_rq(cfs_rq);
5978 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
5989 void unregister_fair_sched_group(struct task_group *tg, int cpu)
5991 struct rq *rq = cpu_rq(cpu);
5992 unsigned long flags;
5995 * Only empty task groups can be destroyed; so we can speculatively
5996 * check on_list without danger of it being re-added.
5998 if (!tg->cfs_rq[cpu]->on_list)
6001 raw_spin_lock_irqsave(&rq->lock, flags);
6002 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
6003 raw_spin_unlock_irqrestore(&rq->lock, flags);
6006 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
6007 struct sched_entity *se, int cpu,
6008 struct sched_entity *parent)
6010 struct rq *rq = cpu_rq(cpu);
6014 init_cfs_rq_runtime(cfs_rq);
6016 tg->cfs_rq[cpu] = cfs_rq;
6019 /* se could be NULL for root_task_group */
6024 se->cfs_rq = &rq->cfs;
6026 se->cfs_rq = parent->my_q;
6029 update_load_set(&se->load, 0);
6030 se->parent = parent;
6033 static DEFINE_MUTEX(shares_mutex);
6035 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6038 unsigned long flags;
6041 * We can't change the weight of the root cgroup.
6046 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
6048 mutex_lock(&shares_mutex);
6049 if (tg->shares == shares)
6052 tg->shares = shares;
6053 for_each_possible_cpu(i) {
6054 struct rq *rq = cpu_rq(i);
6055 struct sched_entity *se;
6058 /* Propagate contribution to hierarchy */
6059 raw_spin_lock_irqsave(&rq->lock, flags);
6060 for_each_sched_entity(se)
6061 update_cfs_shares(group_cfs_rq(se));
6062 raw_spin_unlock_irqrestore(&rq->lock, flags);
6066 mutex_unlock(&shares_mutex);
6069 #else /* CONFIG_FAIR_GROUP_SCHED */
6071 void free_fair_sched_group(struct task_group *tg) { }
6073 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6078 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
6080 #endif /* CONFIG_FAIR_GROUP_SCHED */
6083 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
6085 struct sched_entity *se = &task->se;
6086 unsigned int rr_interval = 0;
6089 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6092 if (rq->cfs.load.weight)
6093 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
6099 * All the scheduling class methods:
6101 const struct sched_class fair_sched_class = {
6102 .next = &idle_sched_class,
6103 .enqueue_task = enqueue_task_fair,
6104 .dequeue_task = dequeue_task_fair,
6105 .yield_task = yield_task_fair,
6106 .yield_to_task = yield_to_task_fair,
6108 .check_preempt_curr = check_preempt_wakeup,
6110 .pick_next_task = pick_next_task_fair,
6111 .put_prev_task = put_prev_task_fair,
6114 .select_task_rq = select_task_rq_fair,
6115 #ifdef CONFIG_FAIR_GROUP_SCHED
6116 .migrate_task_rq = migrate_task_rq_fair,
6118 .rq_online = rq_online_fair,
6119 .rq_offline = rq_offline_fair,
6121 .task_waking = task_waking_fair,
6124 .set_curr_task = set_curr_task_fair,
6125 .task_tick = task_tick_fair,
6126 .task_fork = task_fork_fair,
6128 .prio_changed = prio_changed_fair,
6129 .switched_from = switched_from_fair,
6130 .switched_to = switched_to_fair,
6132 .get_rr_interval = get_rr_interval_fair,
6134 #ifdef CONFIG_FAIR_GROUP_SCHED
6135 .task_move_group = task_move_group_fair,
6139 #ifdef CONFIG_SCHED_DEBUG
6140 void print_cfs_stats(struct seq_file *m, int cpu)
6142 struct cfs_rq *cfs_rq;
6145 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
6146 print_cfs_rq(m, cpu, cfs_rq);
6151 __init void init_sched_fair_class(void)
6154 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
6157 nohz.next_balance = jiffies;
6158 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
6159 cpu_notifier(sched_ilb_notifier, 0);