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 * Skip inaccessible VMAs to avoid any confusion between
941 * PROT_NONE and NUMA hinting ptes
943 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
947 start = max(start, vma->vm_start);
948 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
949 end = min(end, vma->vm_end);
950 pages -= change_prot_numa(vma, start, end);
955 } while (end != vma->vm_end);
960 * It is possible to reach the end of the VMA list but the last few VMAs are
961 * not guaranteed to the vma_migratable. If they are not, we would find the
962 * !migratable VMA on the next scan but not reset the scanner to the start
966 mm->numa_scan_offset = start;
968 reset_ptenuma_scan(p);
969 up_read(&mm->mmap_sem);
973 * Drive the periodic memory faults..
975 void task_tick_numa(struct rq *rq, struct task_struct *curr)
977 struct callback_head *work = &curr->numa_work;
981 * We don't care about NUMA placement if we don't have memory.
983 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
987 * Using runtime rather than walltime has the dual advantage that
988 * we (mostly) drive the selection from busy threads and that the
989 * task needs to have done some actual work before we bother with
992 now = curr->se.sum_exec_runtime;
993 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
995 if (now - curr->node_stamp > period) {
996 if (!curr->node_stamp)
997 curr->numa_scan_period = sysctl_numa_balancing_scan_period_min;
998 curr->node_stamp = now;
1000 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1001 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1002 task_work_add(curr, work, true);
1007 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1010 #endif /* CONFIG_NUMA_BALANCING */
1013 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1015 update_load_add(&cfs_rq->load, se->load.weight);
1016 if (!parent_entity(se))
1017 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1019 if (entity_is_task(se))
1020 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1022 cfs_rq->nr_running++;
1026 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1028 update_load_sub(&cfs_rq->load, se->load.weight);
1029 if (!parent_entity(se))
1030 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1031 if (entity_is_task(se))
1032 list_del_init(&se->group_node);
1033 cfs_rq->nr_running--;
1036 #ifdef CONFIG_FAIR_GROUP_SCHED
1038 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1043 * Use this CPU's actual weight instead of the last load_contribution
1044 * to gain a more accurate current total weight. See
1045 * update_cfs_rq_load_contribution().
1047 tg_weight = atomic64_read(&tg->load_avg);
1048 tg_weight -= cfs_rq->tg_load_contrib;
1049 tg_weight += cfs_rq->load.weight;
1054 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1056 long tg_weight, load, shares;
1058 tg_weight = calc_tg_weight(tg, cfs_rq);
1059 load = cfs_rq->load.weight;
1061 shares = (tg->shares * load);
1063 shares /= tg_weight;
1065 if (shares < MIN_SHARES)
1066 shares = MIN_SHARES;
1067 if (shares > tg->shares)
1068 shares = tg->shares;
1072 # else /* CONFIG_SMP */
1073 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1077 # endif /* CONFIG_SMP */
1078 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1079 unsigned long weight)
1082 /* commit outstanding execution time */
1083 if (cfs_rq->curr == se)
1084 update_curr(cfs_rq);
1085 account_entity_dequeue(cfs_rq, se);
1088 update_load_set(&se->load, weight);
1091 account_entity_enqueue(cfs_rq, se);
1094 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1096 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1098 struct task_group *tg;
1099 struct sched_entity *se;
1103 se = tg->se[cpu_of(rq_of(cfs_rq))];
1104 if (!se || throttled_hierarchy(cfs_rq))
1107 if (likely(se->load.weight == tg->shares))
1110 shares = calc_cfs_shares(cfs_rq, tg);
1112 reweight_entity(cfs_rq_of(se), se, shares);
1114 #else /* CONFIG_FAIR_GROUP_SCHED */
1115 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1118 #endif /* CONFIG_FAIR_GROUP_SCHED */
1120 /* Only depends on SMP, FAIR_GROUP_SCHED may be removed when useful in lb */
1121 #if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)
1123 * We choose a half-life close to 1 scheduling period.
1124 * Note: The tables below are dependent on this value.
1126 #define LOAD_AVG_PERIOD 32
1127 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1128 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1130 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1131 static const u32 runnable_avg_yN_inv[] = {
1132 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1133 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1134 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1135 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1136 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1137 0x85aac367, 0x82cd8698,
1141 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1142 * over-estimates when re-combining.
1144 static const u32 runnable_avg_yN_sum[] = {
1145 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1146 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1147 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1152 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1154 static __always_inline u64 decay_load(u64 val, u64 n)
1156 unsigned int local_n;
1160 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1163 /* after bounds checking we can collapse to 32-bit */
1167 * As y^PERIOD = 1/2, we can combine
1168 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1169 * With a look-up table which covers k^n (n<PERIOD)
1171 * To achieve constant time decay_load.
1173 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1174 val >>= local_n / LOAD_AVG_PERIOD;
1175 local_n %= LOAD_AVG_PERIOD;
1178 val *= runnable_avg_yN_inv[local_n];
1179 /* We don't use SRR here since we always want to round down. */
1184 * For updates fully spanning n periods, the contribution to runnable
1185 * average will be: \Sum 1024*y^n
1187 * We can compute this reasonably efficiently by combining:
1188 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1190 static u32 __compute_runnable_contrib(u64 n)
1194 if (likely(n <= LOAD_AVG_PERIOD))
1195 return runnable_avg_yN_sum[n];
1196 else if (unlikely(n >= LOAD_AVG_MAX_N))
1197 return LOAD_AVG_MAX;
1199 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1201 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1202 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1204 n -= LOAD_AVG_PERIOD;
1205 } while (n > LOAD_AVG_PERIOD);
1207 contrib = decay_load(contrib, n);
1208 return contrib + runnable_avg_yN_sum[n];
1212 * We can represent the historical contribution to runnable average as the
1213 * coefficients of a geometric series. To do this we sub-divide our runnable
1214 * history into segments of approximately 1ms (1024us); label the segment that
1215 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1217 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1219 * (now) (~1ms ago) (~2ms ago)
1221 * Let u_i denote the fraction of p_i that the entity was runnable.
1223 * We then designate the fractions u_i as our co-efficients, yielding the
1224 * following representation of historical load:
1225 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1227 * We choose y based on the with of a reasonably scheduling period, fixing:
1230 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1231 * approximately half as much as the contribution to load within the last ms
1234 * When a period "rolls over" and we have new u_0`, multiplying the previous
1235 * sum again by y is sufficient to update:
1236 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1237 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1239 static __always_inline int __update_entity_runnable_avg(u64 now,
1240 struct sched_avg *sa,
1244 u32 runnable_contrib;
1245 int delta_w, decayed = 0;
1247 delta = now - sa->last_runnable_update;
1249 * This should only happen when time goes backwards, which it
1250 * unfortunately does during sched clock init when we swap over to TSC.
1252 if ((s64)delta < 0) {
1253 sa->last_runnable_update = now;
1258 * Use 1024ns as the unit of measurement since it's a reasonable
1259 * approximation of 1us and fast to compute.
1264 sa->last_runnable_update = now;
1266 /* delta_w is the amount already accumulated against our next period */
1267 delta_w = sa->runnable_avg_period % 1024;
1268 if (delta + delta_w >= 1024) {
1269 /* period roll-over */
1273 * Now that we know we're crossing a period boundary, figure
1274 * out how much from delta we need to complete the current
1275 * period and accrue it.
1277 delta_w = 1024 - delta_w;
1279 sa->runnable_avg_sum += delta_w;
1280 sa->runnable_avg_period += delta_w;
1284 /* Figure out how many additional periods this update spans */
1285 periods = delta / 1024;
1288 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1290 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1293 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1294 runnable_contrib = __compute_runnable_contrib(periods);
1296 sa->runnable_avg_sum += runnable_contrib;
1297 sa->runnable_avg_period += runnable_contrib;
1300 /* Remainder of delta accrued against u_0` */
1302 sa->runnable_avg_sum += delta;
1303 sa->runnable_avg_period += delta;
1308 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1309 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1311 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1312 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1314 decays -= se->avg.decay_count;
1318 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1319 se->avg.decay_count = 0;
1324 #ifdef CONFIG_FAIR_GROUP_SCHED
1325 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1328 struct task_group *tg = cfs_rq->tg;
1331 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1332 tg_contrib -= cfs_rq->tg_load_contrib;
1334 if (force_update || abs64(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1335 atomic64_add(tg_contrib, &tg->load_avg);
1336 cfs_rq->tg_load_contrib += tg_contrib;
1341 * Aggregate cfs_rq runnable averages into an equivalent task_group
1342 * representation for computing load contributions.
1344 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1345 struct cfs_rq *cfs_rq)
1347 struct task_group *tg = cfs_rq->tg;
1350 /* The fraction of a cpu used by this cfs_rq */
1351 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1352 sa->runnable_avg_period + 1);
1353 contrib -= cfs_rq->tg_runnable_contrib;
1355 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
1356 atomic_add(contrib, &tg->runnable_avg);
1357 cfs_rq->tg_runnable_contrib += contrib;
1361 static inline void __update_group_entity_contrib(struct sched_entity *se)
1363 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1364 struct task_group *tg = cfs_rq->tg;
1369 contrib = cfs_rq->tg_load_contrib * tg->shares;
1370 se->avg.load_avg_contrib = div64_u64(contrib,
1371 atomic64_read(&tg->load_avg) + 1);
1374 * For group entities we need to compute a correction term in the case
1375 * that they are consuming <1 cpu so that we would contribute the same
1376 * load as a task of equal weight.
1378 * Explicitly co-ordinating this measurement would be expensive, but
1379 * fortunately the sum of each cpus contribution forms a usable
1380 * lower-bound on the true value.
1382 * Consider the aggregate of 2 contributions. Either they are disjoint
1383 * (and the sum represents true value) or they are disjoint and we are
1384 * understating by the aggregate of their overlap.
1386 * Extending this to N cpus, for a given overlap, the maximum amount we
1387 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1388 * cpus that overlap for this interval and w_i is the interval width.
1390 * On a small machine; the first term is well-bounded which bounds the
1391 * total error since w_i is a subset of the period. Whereas on a
1392 * larger machine, while this first term can be larger, if w_i is the
1393 * of consequential size guaranteed to see n_i*w_i quickly converge to
1394 * our upper bound of 1-cpu.
1396 runnable_avg = atomic_read(&tg->runnable_avg);
1397 if (runnable_avg < NICE_0_LOAD) {
1398 se->avg.load_avg_contrib *= runnable_avg;
1399 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1403 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1404 int force_update) {}
1405 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1406 struct cfs_rq *cfs_rq) {}
1407 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1410 static inline void __update_task_entity_contrib(struct sched_entity *se)
1414 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1415 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1416 contrib /= (se->avg.runnable_avg_period + 1);
1417 se->avg.load_avg_contrib = scale_load(contrib);
1420 /* Compute the current contribution to load_avg by se, return any delta */
1421 static long __update_entity_load_avg_contrib(struct sched_entity *se)
1423 long old_contrib = se->avg.load_avg_contrib;
1425 if (entity_is_task(se)) {
1426 __update_task_entity_contrib(se);
1428 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1429 __update_group_entity_contrib(se);
1432 return se->avg.load_avg_contrib - old_contrib;
1435 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1438 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1439 cfs_rq->blocked_load_avg -= load_contrib;
1441 cfs_rq->blocked_load_avg = 0;
1444 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1446 /* Update a sched_entity's runnable average */
1447 static inline void update_entity_load_avg(struct sched_entity *se,
1450 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1455 * For a group entity we need to use their owned cfs_rq_clock_task() in
1456 * case they are the parent of a throttled hierarchy.
1458 if (entity_is_task(se))
1459 now = cfs_rq_clock_task(cfs_rq);
1461 now = cfs_rq_clock_task(group_cfs_rq(se));
1463 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
1466 contrib_delta = __update_entity_load_avg_contrib(se);
1472 cfs_rq->runnable_load_avg += contrib_delta;
1474 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1478 * Decay the load contributed by all blocked children and account this so that
1479 * their contribution may appropriately discounted when they wake up.
1481 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1483 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1486 decays = now - cfs_rq->last_decay;
1487 if (!decays && !force_update)
1490 if (atomic64_read(&cfs_rq->removed_load)) {
1491 u64 removed_load = atomic64_xchg(&cfs_rq->removed_load, 0);
1492 subtract_blocked_load_contrib(cfs_rq, removed_load);
1496 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1498 atomic64_add(decays, &cfs_rq->decay_counter);
1499 cfs_rq->last_decay = now;
1502 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1505 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1507 __update_entity_runnable_avg(rq->clock_task, &rq->avg, runnable);
1508 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1511 /* Add the load generated by se into cfs_rq's child load-average */
1512 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1513 struct sched_entity *se,
1517 * We track migrations using entity decay_count <= 0, on a wake-up
1518 * migration we use a negative decay count to track the remote decays
1519 * accumulated while sleeping.
1521 if (unlikely(se->avg.decay_count <= 0)) {
1522 se->avg.last_runnable_update = rq_of(cfs_rq)->clock_task;
1523 if (se->avg.decay_count) {
1525 * In a wake-up migration we have to approximate the
1526 * time sleeping. This is because we can't synchronize
1527 * clock_task between the two cpus, and it is not
1528 * guaranteed to be read-safe. Instead, we can
1529 * approximate this using our carried decays, which are
1530 * explicitly atomically readable.
1532 se->avg.last_runnable_update -= (-se->avg.decay_count)
1534 update_entity_load_avg(se, 0);
1535 /* Indicate that we're now synchronized and on-rq */
1536 se->avg.decay_count = 0;
1540 __synchronize_entity_decay(se);
1543 /* migrated tasks did not contribute to our blocked load */
1545 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1546 update_entity_load_avg(se, 0);
1549 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1550 /* we force update consideration on load-balancer moves */
1551 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1555 * Remove se's load from this cfs_rq child load-average, if the entity is
1556 * transitioning to a blocked state we track its projected decay using
1559 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1560 struct sched_entity *se,
1563 update_entity_load_avg(se, 1);
1564 /* we force update consideration on load-balancer moves */
1565 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1567 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1569 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1570 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1571 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1575 * Update the rq's load with the elapsed running time before entering
1576 * idle. if the last scheduled task is not a CFS task, idle_enter will
1577 * be the only way to update the runnable statistic.
1579 void idle_enter_fair(struct rq *this_rq)
1581 update_rq_runnable_avg(this_rq, 1);
1585 * Update the rq's load with the elapsed idle time before a task is
1586 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
1587 * be the only way to update the runnable statistic.
1589 void idle_exit_fair(struct rq *this_rq)
1591 update_rq_runnable_avg(this_rq, 0);
1595 static inline void update_entity_load_avg(struct sched_entity *se,
1596 int update_cfs_rq) {}
1597 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1598 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1599 struct sched_entity *se,
1601 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1602 struct sched_entity *se,
1604 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1605 int force_update) {}
1608 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1610 #ifdef CONFIG_SCHEDSTATS
1611 struct task_struct *tsk = NULL;
1613 if (entity_is_task(se))
1616 if (se->statistics.sleep_start) {
1617 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
1622 if (unlikely(delta > se->statistics.sleep_max))
1623 se->statistics.sleep_max = delta;
1625 se->statistics.sleep_start = 0;
1626 se->statistics.sum_sleep_runtime += delta;
1629 account_scheduler_latency(tsk, delta >> 10, 1);
1630 trace_sched_stat_sleep(tsk, delta);
1633 if (se->statistics.block_start) {
1634 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
1639 if (unlikely(delta > se->statistics.block_max))
1640 se->statistics.block_max = delta;
1642 se->statistics.block_start = 0;
1643 se->statistics.sum_sleep_runtime += delta;
1646 if (tsk->in_iowait) {
1647 se->statistics.iowait_sum += delta;
1648 se->statistics.iowait_count++;
1649 trace_sched_stat_iowait(tsk, delta);
1652 trace_sched_stat_blocked(tsk, delta);
1655 * Blocking time is in units of nanosecs, so shift by
1656 * 20 to get a milliseconds-range estimation of the
1657 * amount of time that the task spent sleeping:
1659 if (unlikely(prof_on == SLEEP_PROFILING)) {
1660 profile_hits(SLEEP_PROFILING,
1661 (void *)get_wchan(tsk),
1664 account_scheduler_latency(tsk, delta >> 10, 0);
1670 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1672 #ifdef CONFIG_SCHED_DEBUG
1673 s64 d = se->vruntime - cfs_rq->min_vruntime;
1678 if (d > 3*sysctl_sched_latency)
1679 schedstat_inc(cfs_rq, nr_spread_over);
1684 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1686 u64 vruntime = cfs_rq->min_vruntime;
1689 * The 'current' period is already promised to the current tasks,
1690 * however the extra weight of the new task will slow them down a
1691 * little, place the new task so that it fits in the slot that
1692 * stays open at the end.
1694 if (initial && sched_feat(START_DEBIT))
1695 vruntime += sched_vslice(cfs_rq, se);
1697 /* sleeps up to a single latency don't count. */
1699 unsigned long thresh = sysctl_sched_latency;
1702 * Halve their sleep time's effect, to allow
1703 * for a gentler effect of sleepers:
1705 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1711 /* ensure we never gain time by being placed backwards. */
1712 se->vruntime = max_vruntime(se->vruntime, vruntime);
1715 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1718 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1721 * Update the normalized vruntime before updating min_vruntime
1722 * through callig update_curr().
1724 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1725 se->vruntime += cfs_rq->min_vruntime;
1728 * Update run-time statistics of the 'current'.
1730 update_curr(cfs_rq);
1731 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
1732 account_entity_enqueue(cfs_rq, se);
1733 update_cfs_shares(cfs_rq);
1735 if (flags & ENQUEUE_WAKEUP) {
1736 place_entity(cfs_rq, se, 0);
1737 enqueue_sleeper(cfs_rq, se);
1740 update_stats_enqueue(cfs_rq, se);
1741 check_spread(cfs_rq, se);
1742 if (se != cfs_rq->curr)
1743 __enqueue_entity(cfs_rq, se);
1746 if (cfs_rq->nr_running == 1) {
1747 list_add_leaf_cfs_rq(cfs_rq);
1748 check_enqueue_throttle(cfs_rq);
1752 static void __clear_buddies_last(struct sched_entity *se)
1754 for_each_sched_entity(se) {
1755 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1756 if (cfs_rq->last == se)
1757 cfs_rq->last = NULL;
1763 static void __clear_buddies_next(struct sched_entity *se)
1765 for_each_sched_entity(se) {
1766 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1767 if (cfs_rq->next == se)
1768 cfs_rq->next = NULL;
1774 static void __clear_buddies_skip(struct sched_entity *se)
1776 for_each_sched_entity(se) {
1777 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1778 if (cfs_rq->skip == se)
1779 cfs_rq->skip = NULL;
1785 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1787 if (cfs_rq->last == se)
1788 __clear_buddies_last(se);
1790 if (cfs_rq->next == se)
1791 __clear_buddies_next(se);
1793 if (cfs_rq->skip == se)
1794 __clear_buddies_skip(se);
1797 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1800 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1803 * Update run-time statistics of the 'current'.
1805 update_curr(cfs_rq);
1806 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
1808 update_stats_dequeue(cfs_rq, se);
1809 if (flags & DEQUEUE_SLEEP) {
1810 #ifdef CONFIG_SCHEDSTATS
1811 if (entity_is_task(se)) {
1812 struct task_struct *tsk = task_of(se);
1814 if (tsk->state & TASK_INTERRUPTIBLE)
1815 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1816 if (tsk->state & TASK_UNINTERRUPTIBLE)
1817 se->statistics.block_start = rq_of(cfs_rq)->clock;
1822 clear_buddies(cfs_rq, se);
1824 if (se != cfs_rq->curr)
1825 __dequeue_entity(cfs_rq, se);
1827 account_entity_dequeue(cfs_rq, se);
1830 * Normalize the entity after updating the min_vruntime because the
1831 * update can refer to the ->curr item and we need to reflect this
1832 * movement in our normalized position.
1834 if (!(flags & DEQUEUE_SLEEP))
1835 se->vruntime -= cfs_rq->min_vruntime;
1837 /* return excess runtime on last dequeue */
1838 return_cfs_rq_runtime(cfs_rq);
1840 update_min_vruntime(cfs_rq);
1841 update_cfs_shares(cfs_rq);
1845 * Preempt the current task with a newly woken task if needed:
1848 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1850 unsigned long ideal_runtime, delta_exec;
1851 struct sched_entity *se;
1854 ideal_runtime = sched_slice(cfs_rq, curr);
1855 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1856 if (delta_exec > ideal_runtime) {
1857 resched_task(rq_of(cfs_rq)->curr);
1859 * The current task ran long enough, ensure it doesn't get
1860 * re-elected due to buddy favours.
1862 clear_buddies(cfs_rq, curr);
1867 * Ensure that a task that missed wakeup preemption by a
1868 * narrow margin doesn't have to wait for a full slice.
1869 * This also mitigates buddy induced latencies under load.
1871 if (delta_exec < sysctl_sched_min_granularity)
1874 se = __pick_first_entity(cfs_rq);
1875 delta = curr->vruntime - se->vruntime;
1880 if (delta > ideal_runtime)
1881 resched_task(rq_of(cfs_rq)->curr);
1885 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1887 /* 'current' is not kept within the tree. */
1890 * Any task has to be enqueued before it get to execute on
1891 * a CPU. So account for the time it spent waiting on the
1894 update_stats_wait_end(cfs_rq, se);
1895 __dequeue_entity(cfs_rq, se);
1898 update_stats_curr_start(cfs_rq, se);
1900 #ifdef CONFIG_SCHEDSTATS
1902 * Track our maximum slice length, if the CPU's load is at
1903 * least twice that of our own weight (i.e. dont track it
1904 * when there are only lesser-weight tasks around):
1906 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1907 se->statistics.slice_max = max(se->statistics.slice_max,
1908 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1911 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1915 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1918 * Pick the next process, keeping these things in mind, in this order:
1919 * 1) keep things fair between processes/task groups
1920 * 2) pick the "next" process, since someone really wants that to run
1921 * 3) pick the "last" process, for cache locality
1922 * 4) do not run the "skip" process, if something else is available
1924 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1926 struct sched_entity *se = __pick_first_entity(cfs_rq);
1927 struct sched_entity *left = se;
1930 * Avoid running the skip buddy, if running something else can
1931 * be done without getting too unfair.
1933 if (cfs_rq->skip == se) {
1934 struct sched_entity *second = __pick_next_entity(se);
1935 if (second && wakeup_preempt_entity(second, left) < 1)
1940 * Prefer last buddy, try to return the CPU to a preempted task.
1942 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1946 * Someone really wants this to run. If it's not unfair, run it.
1948 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1951 clear_buddies(cfs_rq, se);
1956 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1958 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1961 * If still on the runqueue then deactivate_task()
1962 * was not called and update_curr() has to be done:
1965 update_curr(cfs_rq);
1967 /* throttle cfs_rqs exceeding runtime */
1968 check_cfs_rq_runtime(cfs_rq);
1970 check_spread(cfs_rq, prev);
1972 update_stats_wait_start(cfs_rq, prev);
1973 /* Put 'current' back into the tree. */
1974 __enqueue_entity(cfs_rq, prev);
1975 /* in !on_rq case, update occurred at dequeue */
1976 update_entity_load_avg(prev, 1);
1978 cfs_rq->curr = NULL;
1982 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1985 * Update run-time statistics of the 'current'.
1987 update_curr(cfs_rq);
1990 * Ensure that runnable average is periodically updated.
1992 update_entity_load_avg(curr, 1);
1993 update_cfs_rq_blocked_load(cfs_rq, 1);
1994 update_cfs_shares(cfs_rq);
1996 #ifdef CONFIG_SCHED_HRTICK
1998 * queued ticks are scheduled to match the slice, so don't bother
1999 * validating it and just reschedule.
2002 resched_task(rq_of(cfs_rq)->curr);
2006 * don't let the period tick interfere with the hrtick preemption
2008 if (!sched_feat(DOUBLE_TICK) &&
2009 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2013 if (cfs_rq->nr_running > 1)
2014 check_preempt_tick(cfs_rq, curr);
2018 /**************************************************
2019 * CFS bandwidth control machinery
2022 #ifdef CONFIG_CFS_BANDWIDTH
2024 #ifdef HAVE_JUMP_LABEL
2025 static struct static_key __cfs_bandwidth_used;
2027 static inline bool cfs_bandwidth_used(void)
2029 return static_key_false(&__cfs_bandwidth_used);
2032 void cfs_bandwidth_usage_inc(void)
2034 static_key_slow_inc(&__cfs_bandwidth_used);
2037 void cfs_bandwidth_usage_dec(void)
2039 static_key_slow_dec(&__cfs_bandwidth_used);
2041 #else /* HAVE_JUMP_LABEL */
2042 static bool cfs_bandwidth_used(void)
2047 void cfs_bandwidth_usage_inc(void) {}
2048 void cfs_bandwidth_usage_dec(void) {}
2049 #endif /* HAVE_JUMP_LABEL */
2052 * default period for cfs group bandwidth.
2053 * default: 0.1s, units: nanoseconds
2055 static inline u64 default_cfs_period(void)
2057 return 100000000ULL;
2060 static inline u64 sched_cfs_bandwidth_slice(void)
2062 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2066 * Replenish runtime according to assigned quota and update expiration time.
2067 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2068 * additional synchronization around rq->lock.
2070 * requires cfs_b->lock
2072 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2076 if (cfs_b->quota == RUNTIME_INF)
2079 now = sched_clock_cpu(smp_processor_id());
2080 cfs_b->runtime = cfs_b->quota;
2081 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2084 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2086 return &tg->cfs_bandwidth;
2089 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2090 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2092 if (unlikely(cfs_rq->throttle_count))
2093 return cfs_rq->throttled_clock_task;
2095 return rq_of(cfs_rq)->clock_task - cfs_rq->throttled_clock_task_time;
2098 /* returns 0 on failure to allocate runtime */
2099 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2101 struct task_group *tg = cfs_rq->tg;
2102 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2103 u64 amount = 0, min_amount, expires;
2105 /* note: this is a positive sum as runtime_remaining <= 0 */
2106 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2108 raw_spin_lock(&cfs_b->lock);
2109 if (cfs_b->quota == RUNTIME_INF)
2110 amount = min_amount;
2113 * If the bandwidth pool has become inactive, then at least one
2114 * period must have elapsed since the last consumption.
2115 * Refresh the global state and ensure bandwidth timer becomes
2118 if (!cfs_b->timer_active) {
2119 __refill_cfs_bandwidth_runtime(cfs_b);
2120 __start_cfs_bandwidth(cfs_b);
2123 if (cfs_b->runtime > 0) {
2124 amount = min(cfs_b->runtime, min_amount);
2125 cfs_b->runtime -= amount;
2129 expires = cfs_b->runtime_expires;
2130 raw_spin_unlock(&cfs_b->lock);
2132 cfs_rq->runtime_remaining += amount;
2134 * we may have advanced our local expiration to account for allowed
2135 * spread between our sched_clock and the one on which runtime was
2138 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2139 cfs_rq->runtime_expires = expires;
2141 return cfs_rq->runtime_remaining > 0;
2145 * Note: This depends on the synchronization provided by sched_clock and the
2146 * fact that rq->clock snapshots this value.
2148 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2150 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2151 struct rq *rq = rq_of(cfs_rq);
2153 /* if the deadline is ahead of our clock, nothing to do */
2154 if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0))
2157 if (cfs_rq->runtime_remaining < 0)
2161 * If the local deadline has passed we have to consider the
2162 * possibility that our sched_clock is 'fast' and the global deadline
2163 * has not truly expired.
2165 * Fortunately we can check determine whether this the case by checking
2166 * whether the global deadline has advanced.
2169 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2170 /* extend local deadline, drift is bounded above by 2 ticks */
2171 cfs_rq->runtime_expires += TICK_NSEC;
2173 /* global deadline is ahead, expiration has passed */
2174 cfs_rq->runtime_remaining = 0;
2178 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2179 unsigned long delta_exec)
2181 /* dock delta_exec before expiring quota (as it could span periods) */
2182 cfs_rq->runtime_remaining -= delta_exec;
2183 expire_cfs_rq_runtime(cfs_rq);
2185 if (likely(cfs_rq->runtime_remaining > 0))
2189 * if we're unable to extend our runtime we resched so that the active
2190 * hierarchy can be throttled
2192 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2193 resched_task(rq_of(cfs_rq)->curr);
2196 static __always_inline
2197 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2199 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2202 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2205 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2207 return cfs_bandwidth_used() && cfs_rq->throttled;
2210 /* check whether cfs_rq, or any parent, is throttled */
2211 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2213 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2217 * Ensure that neither of the group entities corresponding to src_cpu or
2218 * dest_cpu are members of a throttled hierarchy when performing group
2219 * load-balance operations.
2221 static inline int throttled_lb_pair(struct task_group *tg,
2222 int src_cpu, int dest_cpu)
2224 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2226 src_cfs_rq = tg->cfs_rq[src_cpu];
2227 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2229 return throttled_hierarchy(src_cfs_rq) ||
2230 throttled_hierarchy(dest_cfs_rq);
2233 /* updated child weight may affect parent so we have to do this bottom up */
2234 static int tg_unthrottle_up(struct task_group *tg, void *data)
2236 struct rq *rq = data;
2237 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2239 cfs_rq->throttle_count--;
2241 if (!cfs_rq->throttle_count) {
2242 /* adjust cfs_rq_clock_task() */
2243 cfs_rq->throttled_clock_task_time += rq->clock_task -
2244 cfs_rq->throttled_clock_task;
2251 static int tg_throttle_down(struct task_group *tg, void *data)
2253 struct rq *rq = data;
2254 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2256 /* group is entering throttled state, stop time */
2257 if (!cfs_rq->throttle_count)
2258 cfs_rq->throttled_clock_task = rq->clock_task;
2259 cfs_rq->throttle_count++;
2264 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2266 struct rq *rq = rq_of(cfs_rq);
2267 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2268 struct sched_entity *se;
2269 long task_delta, dequeue = 1;
2271 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2273 /* freeze hierarchy runnable averages while throttled */
2275 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2278 task_delta = cfs_rq->h_nr_running;
2279 for_each_sched_entity(se) {
2280 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2281 /* throttled entity or throttle-on-deactivate */
2286 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2287 qcfs_rq->h_nr_running -= task_delta;
2289 if (qcfs_rq->load.weight)
2294 rq->nr_running -= task_delta;
2296 cfs_rq->throttled = 1;
2297 cfs_rq->throttled_clock = rq->clock;
2298 raw_spin_lock(&cfs_b->lock);
2299 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2300 if (!cfs_b->timer_active)
2301 __start_cfs_bandwidth(cfs_b);
2302 raw_spin_unlock(&cfs_b->lock);
2305 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2307 struct rq *rq = rq_of(cfs_rq);
2308 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2309 struct sched_entity *se;
2313 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2315 cfs_rq->throttled = 0;
2316 raw_spin_lock(&cfs_b->lock);
2317 cfs_b->throttled_time += rq->clock - cfs_rq->throttled_clock;
2318 list_del_rcu(&cfs_rq->throttled_list);
2319 raw_spin_unlock(&cfs_b->lock);
2321 update_rq_clock(rq);
2322 /* update hierarchical throttle state */
2323 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2325 if (!cfs_rq->load.weight)
2328 task_delta = cfs_rq->h_nr_running;
2329 for_each_sched_entity(se) {
2333 cfs_rq = cfs_rq_of(se);
2335 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2336 cfs_rq->h_nr_running += task_delta;
2338 if (cfs_rq_throttled(cfs_rq))
2343 rq->nr_running += task_delta;
2345 /* determine whether we need to wake up potentially idle cpu */
2346 if (rq->curr == rq->idle && rq->cfs.nr_running)
2347 resched_task(rq->curr);
2350 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2351 u64 remaining, u64 expires)
2353 struct cfs_rq *cfs_rq;
2354 u64 runtime = remaining;
2357 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2359 struct rq *rq = rq_of(cfs_rq);
2361 raw_spin_lock(&rq->lock);
2362 if (!cfs_rq_throttled(cfs_rq))
2365 runtime = -cfs_rq->runtime_remaining + 1;
2366 if (runtime > remaining)
2367 runtime = remaining;
2368 remaining -= runtime;
2370 cfs_rq->runtime_remaining += runtime;
2371 cfs_rq->runtime_expires = expires;
2373 /* we check whether we're throttled above */
2374 if (cfs_rq->runtime_remaining > 0)
2375 unthrottle_cfs_rq(cfs_rq);
2378 raw_spin_unlock(&rq->lock);
2389 * Responsible for refilling a task_group's bandwidth and unthrottling its
2390 * cfs_rqs as appropriate. If there has been no activity within the last
2391 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2392 * used to track this state.
2394 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2396 u64 runtime, runtime_expires;
2397 int idle = 1, throttled;
2399 raw_spin_lock(&cfs_b->lock);
2400 /* no need to continue the timer with no bandwidth constraint */
2401 if (cfs_b->quota == RUNTIME_INF)
2404 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2405 /* idle depends on !throttled (for the case of a large deficit) */
2406 idle = cfs_b->idle && !throttled;
2407 cfs_b->nr_periods += overrun;
2409 /* if we're going inactive then everything else can be deferred */
2414 * if we have relooped after returning idle once, we need to update our
2415 * status as actually running, so that other cpus doing
2416 * __start_cfs_bandwidth will stop trying to cancel us.
2418 cfs_b->timer_active = 1;
2420 __refill_cfs_bandwidth_runtime(cfs_b);
2423 /* mark as potentially idle for the upcoming period */
2428 /* account preceding periods in which throttling occurred */
2429 cfs_b->nr_throttled += overrun;
2432 * There are throttled entities so we must first use the new bandwidth
2433 * to unthrottle them before making it generally available. This
2434 * ensures that all existing debts will be paid before a new cfs_rq is
2437 runtime = cfs_b->runtime;
2438 runtime_expires = cfs_b->runtime_expires;
2442 * This check is repeated as we are holding onto the new bandwidth
2443 * while we unthrottle. This can potentially race with an unthrottled
2444 * group trying to acquire new bandwidth from the global pool.
2446 while (throttled && runtime > 0) {
2447 raw_spin_unlock(&cfs_b->lock);
2448 /* we can't nest cfs_b->lock while distributing bandwidth */
2449 runtime = distribute_cfs_runtime(cfs_b, runtime,
2451 raw_spin_lock(&cfs_b->lock);
2453 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2456 /* return (any) remaining runtime */
2457 cfs_b->runtime = runtime;
2459 * While we are ensured activity in the period following an
2460 * unthrottle, this also covers the case in which the new bandwidth is
2461 * insufficient to cover the existing bandwidth deficit. (Forcing the
2462 * timer to remain active while there are any throttled entities.)
2467 cfs_b->timer_active = 0;
2468 raw_spin_unlock(&cfs_b->lock);
2473 /* a cfs_rq won't donate quota below this amount */
2474 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2475 /* minimum remaining period time to redistribute slack quota */
2476 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2477 /* how long we wait to gather additional slack before distributing */
2478 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2481 * Are we near the end of the current quota period?
2483 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
2484 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
2485 * migrate_hrtimers, base is never cleared, so we are fine.
2487 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2489 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2492 /* if the call-back is running a quota refresh is already occurring */
2493 if (hrtimer_callback_running(refresh_timer))
2496 /* is a quota refresh about to occur? */
2497 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2498 if (remaining < min_expire)
2504 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2506 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2508 /* if there's a quota refresh soon don't bother with slack */
2509 if (runtime_refresh_within(cfs_b, min_left))
2512 start_bandwidth_timer(&cfs_b->slack_timer,
2513 ns_to_ktime(cfs_bandwidth_slack_period));
2516 /* we know any runtime found here is valid as update_curr() precedes return */
2517 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2519 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2520 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2522 if (slack_runtime <= 0)
2525 raw_spin_lock(&cfs_b->lock);
2526 if (cfs_b->quota != RUNTIME_INF &&
2527 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2528 cfs_b->runtime += slack_runtime;
2530 /* we are under rq->lock, defer unthrottling using a timer */
2531 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2532 !list_empty(&cfs_b->throttled_cfs_rq))
2533 start_cfs_slack_bandwidth(cfs_b);
2535 raw_spin_unlock(&cfs_b->lock);
2537 /* even if it's not valid for return we don't want to try again */
2538 cfs_rq->runtime_remaining -= slack_runtime;
2541 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2543 if (!cfs_bandwidth_used())
2546 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2549 __return_cfs_rq_runtime(cfs_rq);
2553 * This is done with a timer (instead of inline with bandwidth return) since
2554 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2556 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2558 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2561 /* confirm we're still not at a refresh boundary */
2562 raw_spin_lock(&cfs_b->lock);
2563 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
2564 raw_spin_unlock(&cfs_b->lock);
2568 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2569 runtime = cfs_b->runtime;
2572 expires = cfs_b->runtime_expires;
2573 raw_spin_unlock(&cfs_b->lock);
2578 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2580 raw_spin_lock(&cfs_b->lock);
2581 if (expires == cfs_b->runtime_expires)
2582 cfs_b->runtime = runtime;
2583 raw_spin_unlock(&cfs_b->lock);
2587 * When a group wakes up we want to make sure that its quota is not already
2588 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2589 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2591 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2593 if (!cfs_bandwidth_used())
2596 /* an active group must be handled by the update_curr()->put() path */
2597 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2600 /* ensure the group is not already throttled */
2601 if (cfs_rq_throttled(cfs_rq))
2604 /* update runtime allocation */
2605 account_cfs_rq_runtime(cfs_rq, 0);
2606 if (cfs_rq->runtime_remaining <= 0)
2607 throttle_cfs_rq(cfs_rq);
2610 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2611 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2613 if (!cfs_bandwidth_used())
2616 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2620 * it's possible for a throttled entity to be forced into a running
2621 * state (e.g. set_curr_task), in this case we're finished.
2623 if (cfs_rq_throttled(cfs_rq))
2626 throttle_cfs_rq(cfs_rq);
2629 static inline u64 default_cfs_period(void);
2630 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun);
2631 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b);
2633 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2635 struct cfs_bandwidth *cfs_b =
2636 container_of(timer, struct cfs_bandwidth, slack_timer);
2637 do_sched_cfs_slack_timer(cfs_b);
2639 return HRTIMER_NORESTART;
2642 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2644 struct cfs_bandwidth *cfs_b =
2645 container_of(timer, struct cfs_bandwidth, period_timer);
2651 now = hrtimer_cb_get_time(timer);
2652 overrun = hrtimer_forward(timer, now, cfs_b->period);
2657 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2660 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2663 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2665 raw_spin_lock_init(&cfs_b->lock);
2667 cfs_b->quota = RUNTIME_INF;
2668 cfs_b->period = ns_to_ktime(default_cfs_period());
2670 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2671 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2672 cfs_b->period_timer.function = sched_cfs_period_timer;
2673 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2674 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2677 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2679 cfs_rq->runtime_enabled = 0;
2680 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2683 /* requires cfs_b->lock, may release to reprogram timer */
2684 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2687 * The timer may be active because we're trying to set a new bandwidth
2688 * period or because we're racing with the tear-down path
2689 * (timer_active==0 becomes visible before the hrtimer call-back
2690 * terminates). In either case we ensure that it's re-programmed
2692 while (unlikely(hrtimer_active(&cfs_b->period_timer)) &&
2693 hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) {
2694 /* bounce the lock to allow do_sched_cfs_period_timer to run */
2695 raw_spin_unlock(&cfs_b->lock);
2697 raw_spin_lock(&cfs_b->lock);
2698 /* if someone else restarted the timer then we're done */
2699 if (cfs_b->timer_active)
2703 cfs_b->timer_active = 1;
2704 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2707 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2709 hrtimer_cancel(&cfs_b->period_timer);
2710 hrtimer_cancel(&cfs_b->slack_timer);
2713 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
2715 struct cfs_rq *cfs_rq;
2717 for_each_leaf_cfs_rq(rq, cfs_rq) {
2718 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2720 if (!cfs_rq->runtime_enabled)
2724 * clock_task is not advancing so we just need to make sure
2725 * there's some valid quota amount
2727 cfs_rq->runtime_remaining = cfs_b->quota;
2728 if (cfs_rq_throttled(cfs_rq))
2729 unthrottle_cfs_rq(cfs_rq);
2733 #else /* CONFIG_CFS_BANDWIDTH */
2734 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2736 return rq_of(cfs_rq)->clock_task;
2739 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2740 unsigned long delta_exec) {}
2741 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2742 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2743 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2745 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2750 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2755 static inline int throttled_lb_pair(struct task_group *tg,
2756 int src_cpu, int dest_cpu)
2761 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2763 #ifdef CONFIG_FAIR_GROUP_SCHED
2764 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2767 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2771 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2772 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2774 #endif /* CONFIG_CFS_BANDWIDTH */
2776 /**************************************************
2777 * CFS operations on tasks:
2780 #ifdef CONFIG_SCHED_HRTICK
2781 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2783 struct sched_entity *se = &p->se;
2784 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2786 WARN_ON(task_rq(p) != rq);
2788 if (cfs_rq->nr_running > 1) {
2789 u64 slice = sched_slice(cfs_rq, se);
2790 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2791 s64 delta = slice - ran;
2800 * Don't schedule slices shorter than 10000ns, that just
2801 * doesn't make sense. Rely on vruntime for fairness.
2804 delta = max_t(s64, 10000LL, delta);
2806 hrtick_start(rq, delta);
2811 * called from enqueue/dequeue and updates the hrtick when the
2812 * current task is from our class and nr_running is low enough
2815 static void hrtick_update(struct rq *rq)
2817 struct task_struct *curr = rq->curr;
2819 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2822 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2823 hrtick_start_fair(rq, curr);
2825 #else /* !CONFIG_SCHED_HRTICK */
2827 hrtick_start_fair(struct rq *rq, struct task_struct *p)
2831 static inline void hrtick_update(struct rq *rq)
2837 * The enqueue_task method is called before nr_running is
2838 * increased. Here we update the fair scheduling stats and
2839 * then put the task into the rbtree:
2842 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2844 struct cfs_rq *cfs_rq;
2845 struct sched_entity *se = &p->se;
2847 for_each_sched_entity(se) {
2850 cfs_rq = cfs_rq_of(se);
2851 enqueue_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 increment below.
2859 if (cfs_rq_throttled(cfs_rq))
2861 cfs_rq->h_nr_running++;
2863 flags = ENQUEUE_WAKEUP;
2866 for_each_sched_entity(se) {
2867 cfs_rq = cfs_rq_of(se);
2868 cfs_rq->h_nr_running++;
2870 if (cfs_rq_throttled(cfs_rq))
2873 update_cfs_shares(cfs_rq);
2874 update_entity_load_avg(se, 1);
2878 update_rq_runnable_avg(rq, rq->nr_running);
2884 static void set_next_buddy(struct sched_entity *se);
2887 * The dequeue_task method is called before nr_running is
2888 * decreased. We remove the task from the rbtree and
2889 * update the fair scheduling stats:
2891 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2893 struct cfs_rq *cfs_rq;
2894 struct sched_entity *se = &p->se;
2895 int task_sleep = flags & DEQUEUE_SLEEP;
2897 for_each_sched_entity(se) {
2898 cfs_rq = cfs_rq_of(se);
2899 dequeue_entity(cfs_rq, se, flags);
2902 * end evaluation on encountering a throttled cfs_rq
2904 * note: in the case of encountering a throttled cfs_rq we will
2905 * post the final h_nr_running decrement below.
2907 if (cfs_rq_throttled(cfs_rq))
2909 cfs_rq->h_nr_running--;
2911 /* Don't dequeue parent if it has other entities besides us */
2912 if (cfs_rq->load.weight) {
2914 * Bias pick_next to pick a task from this cfs_rq, as
2915 * p is sleeping when it is within its sched_slice.
2917 if (task_sleep && parent_entity(se))
2918 set_next_buddy(parent_entity(se));
2920 /* avoid re-evaluating load for this entity */
2921 se = parent_entity(se);
2924 flags |= DEQUEUE_SLEEP;
2927 for_each_sched_entity(se) {
2928 cfs_rq = cfs_rq_of(se);
2929 cfs_rq->h_nr_running--;
2931 if (cfs_rq_throttled(cfs_rq))
2934 update_cfs_shares(cfs_rq);
2935 update_entity_load_avg(se, 1);
2940 update_rq_runnable_avg(rq, 1);
2946 /* Used instead of source_load when we know the type == 0 */
2947 static unsigned long weighted_cpuload(const int cpu)
2949 return cpu_rq(cpu)->load.weight;
2953 * Return a low guess at the load of a migration-source cpu weighted
2954 * according to the scheduling class and "nice" value.
2956 * We want to under-estimate the load of migration sources, to
2957 * balance conservatively.
2959 static unsigned long source_load(int cpu, int type)
2961 struct rq *rq = cpu_rq(cpu);
2962 unsigned long total = weighted_cpuload(cpu);
2964 if (type == 0 || !sched_feat(LB_BIAS))
2967 return min(rq->cpu_load[type-1], total);
2971 * Return a high guess at the load of a migration-target cpu weighted
2972 * according to the scheduling class and "nice" value.
2974 static unsigned long target_load(int cpu, int type)
2976 struct rq *rq = cpu_rq(cpu);
2977 unsigned long total = weighted_cpuload(cpu);
2979 if (type == 0 || !sched_feat(LB_BIAS))
2982 return max(rq->cpu_load[type-1], total);
2985 static unsigned long power_of(int cpu)
2987 return cpu_rq(cpu)->cpu_power;
2990 static unsigned long cpu_avg_load_per_task(int cpu)
2992 struct rq *rq = cpu_rq(cpu);
2993 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
2996 return rq->load.weight / nr_running;
3002 static void task_waking_fair(struct task_struct *p)
3004 struct sched_entity *se = &p->se;
3005 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3008 #ifndef CONFIG_64BIT
3009 u64 min_vruntime_copy;
3012 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3014 min_vruntime = cfs_rq->min_vruntime;
3015 } while (min_vruntime != min_vruntime_copy);
3017 min_vruntime = cfs_rq->min_vruntime;
3020 se->vruntime -= min_vruntime;
3023 #ifdef CONFIG_FAIR_GROUP_SCHED
3025 * effective_load() calculates the load change as seen from the root_task_group
3027 * Adding load to a group doesn't make a group heavier, but can cause movement
3028 * of group shares between cpus. Assuming the shares were perfectly aligned one
3029 * can calculate the shift in shares.
3031 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3032 * on this @cpu and results in a total addition (subtraction) of @wg to the
3033 * total group weight.
3035 * Given a runqueue weight distribution (rw_i) we can compute a shares
3036 * distribution (s_i) using:
3038 * s_i = rw_i / \Sum rw_j (1)
3040 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3041 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3042 * shares distribution (s_i):
3044 * rw_i = { 2, 4, 1, 0 }
3045 * s_i = { 2/7, 4/7, 1/7, 0 }
3047 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3048 * task used to run on and the CPU the waker is running on), we need to
3049 * compute the effect of waking a task on either CPU and, in case of a sync
3050 * wakeup, compute the effect of the current task going to sleep.
3052 * So for a change of @wl to the local @cpu with an overall group weight change
3053 * of @wl we can compute the new shares distribution (s'_i) using:
3055 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3057 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3058 * differences in waking a task to CPU 0. The additional task changes the
3059 * weight and shares distributions like:
3061 * rw'_i = { 3, 4, 1, 0 }
3062 * s'_i = { 3/8, 4/8, 1/8, 0 }
3064 * We can then compute the difference in effective weight by using:
3066 * dw_i = S * (s'_i - s_i) (3)
3068 * Where 'S' is the group weight as seen by its parent.
3070 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3071 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3072 * 4/7) times the weight of the group.
3074 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3076 struct sched_entity *se = tg->se[cpu];
3078 if (!tg->parent) /* the trivial, non-cgroup case */
3081 for_each_sched_entity(se) {
3087 * W = @wg + \Sum rw_j
3089 W = wg + calc_tg_weight(tg, se->my_q);
3094 w = se->my_q->load.weight + wl;
3097 * wl = S * s'_i; see (2)
3100 wl = (w * tg->shares) / W;
3105 * Per the above, wl is the new se->load.weight value; since
3106 * those are clipped to [MIN_SHARES, ...) do so now. See
3107 * calc_cfs_shares().
3109 if (wl < MIN_SHARES)
3113 * wl = dw_i = S * (s'_i - s_i); see (3)
3115 wl -= se->load.weight;
3118 * Recursively apply this logic to all parent groups to compute
3119 * the final effective load change on the root group. Since
3120 * only the @tg group gets extra weight, all parent groups can
3121 * only redistribute existing shares. @wl is the shift in shares
3122 * resulting from this level per the above.
3131 static inline unsigned long effective_load(struct task_group *tg, int cpu,
3132 unsigned long wl, unsigned long wg)
3139 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3141 s64 this_load, load;
3142 int idx, this_cpu, prev_cpu;
3143 unsigned long tl_per_task;
3144 struct task_group *tg;
3145 unsigned long weight;
3149 this_cpu = smp_processor_id();
3150 prev_cpu = task_cpu(p);
3151 load = source_load(prev_cpu, idx);
3152 this_load = target_load(this_cpu, idx);
3155 * If sync wakeup then subtract the (maximum possible)
3156 * effect of the currently running task from the load
3157 * of the current CPU:
3160 tg = task_group(current);
3161 weight = current->se.load.weight;
3163 this_load += effective_load(tg, this_cpu, -weight, -weight);
3164 load += effective_load(tg, prev_cpu, 0, -weight);
3168 weight = p->se.load.weight;
3171 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3172 * due to the sync cause above having dropped this_load to 0, we'll
3173 * always have an imbalance, but there's really nothing you can do
3174 * about that, so that's good too.
3176 * Otherwise check if either cpus are near enough in load to allow this
3177 * task to be woken on this_cpu.
3179 if (this_load > 0) {
3180 s64 this_eff_load, prev_eff_load;
3182 this_eff_load = 100;
3183 this_eff_load *= power_of(prev_cpu);
3184 this_eff_load *= this_load +
3185 effective_load(tg, this_cpu, weight, weight);
3187 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3188 prev_eff_load *= power_of(this_cpu);
3189 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3191 balanced = this_eff_load <= prev_eff_load;
3196 * If the currently running task will sleep within
3197 * a reasonable amount of time then attract this newly
3200 if (sync && balanced)
3203 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3204 tl_per_task = cpu_avg_load_per_task(this_cpu);
3207 (this_load <= load &&
3208 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3210 * This domain has SD_WAKE_AFFINE and
3211 * p is cache cold in this domain, and
3212 * there is no bad imbalance.
3214 schedstat_inc(sd, ttwu_move_affine);
3215 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3223 * find_idlest_group finds and returns the least busy CPU group within the
3226 static struct sched_group *
3227 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3228 int this_cpu, int load_idx)
3230 struct sched_group *idlest = NULL, *group = sd->groups;
3231 unsigned long min_load = ULONG_MAX, this_load = 0;
3232 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3235 unsigned long load, avg_load;
3239 /* Skip over this group if it has no CPUs allowed */
3240 if (!cpumask_intersects(sched_group_cpus(group),
3241 tsk_cpus_allowed(p)))
3244 local_group = cpumask_test_cpu(this_cpu,
3245 sched_group_cpus(group));
3247 /* Tally up the load of all CPUs in the group */
3250 for_each_cpu(i, sched_group_cpus(group)) {
3251 /* Bias balancing toward cpus of our domain */
3253 load = source_load(i, load_idx);
3255 load = target_load(i, load_idx);
3260 /* Adjust by relative CPU power of the group */
3261 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3264 this_load = avg_load;
3265 } else if (avg_load < min_load) {
3266 min_load = avg_load;
3269 } while (group = group->next, group != sd->groups);
3271 if (!idlest || 100*this_load < imbalance*min_load)
3277 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3280 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3282 unsigned long load, min_load = ULONG_MAX;
3286 /* Traverse only the allowed CPUs */
3287 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3288 load = weighted_cpuload(i);
3290 if (load < min_load || (load == min_load && i == this_cpu)) {
3300 * Try and locate an idle CPU in the sched_domain.
3302 static int select_idle_sibling(struct task_struct *p, int target)
3304 struct sched_domain *sd;
3305 struct sched_group *sg;
3306 int i = task_cpu(p);
3308 if (idle_cpu(target))
3312 * If the prevous cpu is cache affine and idle, don't be stupid.
3314 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3318 * Otherwise, iterate the domains and find an elegible idle cpu.
3320 sd = rcu_dereference(per_cpu(sd_llc, target));
3321 for_each_lower_domain(sd) {
3324 if (!cpumask_intersects(sched_group_cpus(sg),
3325 tsk_cpus_allowed(p)))
3328 for_each_cpu(i, sched_group_cpus(sg)) {
3329 if (i == target || !idle_cpu(i))
3333 target = cpumask_first_and(sched_group_cpus(sg),
3334 tsk_cpus_allowed(p));
3338 } while (sg != sd->groups);
3345 * sched_balance_self: balance the current task (running on cpu) in domains
3346 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3349 * Balance, ie. select the least loaded group.
3351 * Returns the target CPU number, or the same CPU if no balancing is needed.
3353 * preempt must be disabled.
3356 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
3358 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3359 int cpu = smp_processor_id();
3360 int prev_cpu = task_cpu(p);
3362 int want_affine = 0;
3363 int sync = wake_flags & WF_SYNC;
3365 if (p->nr_cpus_allowed == 1)
3368 if (sd_flag & SD_BALANCE_WAKE) {
3369 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
3375 for_each_domain(cpu, tmp) {
3376 if (!(tmp->flags & SD_LOAD_BALANCE))
3380 * If both cpu and prev_cpu are part of this domain,
3381 * cpu is a valid SD_WAKE_AFFINE target.
3383 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3384 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3389 if (tmp->flags & sd_flag)
3394 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3397 new_cpu = select_idle_sibling(p, prev_cpu);
3402 int load_idx = sd->forkexec_idx;
3403 struct sched_group *group;
3406 if (!(sd->flags & sd_flag)) {
3411 if (sd_flag & SD_BALANCE_WAKE)
3412 load_idx = sd->wake_idx;
3414 group = find_idlest_group(sd, p, cpu, load_idx);
3420 new_cpu = find_idlest_cpu(group, p, cpu);
3421 if (new_cpu == -1 || new_cpu == cpu) {
3422 /* Now try balancing at a lower domain level of cpu */
3427 /* Now try balancing at a lower domain level of new_cpu */
3429 weight = sd->span_weight;
3431 for_each_domain(cpu, tmp) {
3432 if (weight <= tmp->span_weight)
3434 if (tmp->flags & sd_flag)
3437 /* while loop will break here if sd == NULL */
3446 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
3447 * removed when useful for applications beyond shares distribution (e.g.
3450 #ifdef CONFIG_FAIR_GROUP_SCHED
3452 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3453 * cfs_rq_of(p) references at time of call are still valid and identify the
3454 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3455 * other assumptions, including the state of rq->lock, should be made.
3458 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
3460 struct sched_entity *se = &p->se;
3461 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3464 * Load tracking: accumulate removed load so that it can be processed
3465 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3466 * to blocked load iff they have a positive decay-count. It can never
3467 * be negative here since on-rq tasks have decay-count == 0.
3469 if (se->avg.decay_count) {
3470 se->avg.decay_count = -__synchronize_entity_decay(se);
3471 atomic64_add(se->avg.load_avg_contrib, &cfs_rq->removed_load);
3475 #endif /* CONFIG_SMP */
3477 static unsigned long
3478 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3480 unsigned long gran = sysctl_sched_wakeup_granularity;
3483 * Since its curr running now, convert the gran from real-time
3484 * to virtual-time in his units.
3486 * By using 'se' instead of 'curr' we penalize light tasks, so
3487 * they get preempted easier. That is, if 'se' < 'curr' then
3488 * the resulting gran will be larger, therefore penalizing the
3489 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3490 * be smaller, again penalizing the lighter task.
3492 * This is especially important for buddies when the leftmost
3493 * task is higher priority than the buddy.
3495 return calc_delta_fair(gran, se);
3499 * Should 'se' preempt 'curr'.
3513 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3515 s64 gran, vdiff = curr->vruntime - se->vruntime;
3520 gran = wakeup_gran(curr, se);
3527 static void set_last_buddy(struct sched_entity *se)
3529 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3532 for_each_sched_entity(se)
3533 cfs_rq_of(se)->last = se;
3536 static void set_next_buddy(struct sched_entity *se)
3538 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3541 for_each_sched_entity(se)
3542 cfs_rq_of(se)->next = se;
3545 static void set_skip_buddy(struct sched_entity *se)
3547 for_each_sched_entity(se)
3548 cfs_rq_of(se)->skip = se;
3552 * Preempt the current task with a newly woken task if needed:
3554 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3556 struct task_struct *curr = rq->curr;
3557 struct sched_entity *se = &curr->se, *pse = &p->se;
3558 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3559 int scale = cfs_rq->nr_running >= sched_nr_latency;
3560 int next_buddy_marked = 0;
3562 if (unlikely(se == pse))
3566 * This is possible from callers such as move_task(), in which we
3567 * unconditionally check_prempt_curr() after an enqueue (which may have
3568 * lead to a throttle). This both saves work and prevents false
3569 * next-buddy nomination below.
3571 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3574 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3575 set_next_buddy(pse);
3576 next_buddy_marked = 1;
3580 * We can come here with TIF_NEED_RESCHED already set from new task
3583 * Note: this also catches the edge-case of curr being in a throttled
3584 * group (e.g. via set_curr_task), since update_curr() (in the
3585 * enqueue of curr) will have resulted in resched being set. This
3586 * prevents us from potentially nominating it as a false LAST_BUDDY
3589 if (test_tsk_need_resched(curr))
3592 /* Idle tasks are by definition preempted by non-idle tasks. */
3593 if (unlikely(curr->policy == SCHED_IDLE) &&
3594 likely(p->policy != SCHED_IDLE))
3598 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3599 * is driven by the tick):
3601 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
3604 find_matching_se(&se, &pse);
3605 update_curr(cfs_rq_of(se));
3607 if (wakeup_preempt_entity(se, pse) == 1) {
3609 * Bias pick_next to pick the sched entity that is
3610 * triggering this preemption.
3612 if (!next_buddy_marked)
3613 set_next_buddy(pse);
3622 * Only set the backward buddy when the current task is still
3623 * on the rq. This can happen when a wakeup gets interleaved
3624 * with schedule on the ->pre_schedule() or idle_balance()
3625 * point, either of which can * drop the rq lock.
3627 * Also, during early boot the idle thread is in the fair class,
3628 * for obvious reasons its a bad idea to schedule back to it.
3630 if (unlikely(!se->on_rq || curr == rq->idle))
3633 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3637 static struct task_struct *pick_next_task_fair(struct rq *rq)
3639 struct task_struct *p;
3640 struct cfs_rq *cfs_rq = &rq->cfs;
3641 struct sched_entity *se;
3643 if (!cfs_rq->nr_running)
3647 se = pick_next_entity(cfs_rq);
3648 set_next_entity(cfs_rq, se);
3649 cfs_rq = group_cfs_rq(se);
3653 if (hrtick_enabled(rq))
3654 hrtick_start_fair(rq, p);
3660 * Account for a descheduled task:
3662 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3664 struct sched_entity *se = &prev->se;
3665 struct cfs_rq *cfs_rq;
3667 for_each_sched_entity(se) {
3668 cfs_rq = cfs_rq_of(se);
3669 put_prev_entity(cfs_rq, se);
3674 * sched_yield() is very simple
3676 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3678 static void yield_task_fair(struct rq *rq)
3680 struct task_struct *curr = rq->curr;
3681 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3682 struct sched_entity *se = &curr->se;
3685 * Are we the only task in the tree?
3687 if (unlikely(rq->nr_running == 1))
3690 clear_buddies(cfs_rq, se);
3692 if (curr->policy != SCHED_BATCH) {
3693 update_rq_clock(rq);
3695 * Update run-time statistics of the 'current'.
3697 update_curr(cfs_rq);
3699 * Tell update_rq_clock() that we've just updated,
3700 * so we don't do microscopic update in schedule()
3701 * and double the fastpath cost.
3703 rq->skip_clock_update = 1;
3709 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3711 struct sched_entity *se = &p->se;
3713 /* throttled hierarchies are not runnable */
3714 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3717 /* Tell the scheduler that we'd really like pse to run next. */
3720 yield_task_fair(rq);
3726 /**************************************************
3727 * Fair scheduling class load-balancing methods.
3731 * The purpose of load-balancing is to achieve the same basic fairness the
3732 * per-cpu scheduler provides, namely provide a proportional amount of compute
3733 * time to each task. This is expressed in the following equation:
3735 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
3737 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
3738 * W_i,0 is defined as:
3740 * W_i,0 = \Sum_j w_i,j (2)
3742 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
3743 * is derived from the nice value as per prio_to_weight[].
3745 * The weight average is an exponential decay average of the instantaneous
3748 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
3750 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
3751 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
3752 * can also include other factors [XXX].
3754 * To achieve this balance we define a measure of imbalance which follows
3755 * directly from (1):
3757 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
3759 * We them move tasks around to minimize the imbalance. In the continuous
3760 * function space it is obvious this converges, in the discrete case we get
3761 * a few fun cases generally called infeasible weight scenarios.
3764 * - infeasible weights;
3765 * - local vs global optima in the discrete case. ]
3770 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
3771 * for all i,j solution, we create a tree of cpus that follows the hardware
3772 * topology where each level pairs two lower groups (or better). This results
3773 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
3774 * tree to only the first of the previous level and we decrease the frequency
3775 * of load-balance at each level inv. proportional to the number of cpus in
3781 * \Sum { --- * --- * 2^i } = O(n) (5)
3783 * `- size of each group
3784 * | | `- number of cpus doing load-balance
3786 * `- sum over all levels
3788 * Coupled with a limit on how many tasks we can migrate every balance pass,
3789 * this makes (5) the runtime complexity of the balancer.
3791 * An important property here is that each CPU is still (indirectly) connected
3792 * to every other cpu in at most O(log n) steps:
3794 * The adjacency matrix of the resulting graph is given by:
3797 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
3800 * And you'll find that:
3802 * A^(log_2 n)_i,j != 0 for all i,j (7)
3804 * Showing there's indeed a path between every cpu in at most O(log n) steps.
3805 * The task movement gives a factor of O(m), giving a convergence complexity
3808 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
3813 * In order to avoid CPUs going idle while there's still work to do, new idle
3814 * balancing is more aggressive and has the newly idle cpu iterate up the domain
3815 * tree itself instead of relying on other CPUs to bring it work.
3817 * This adds some complexity to both (5) and (8) but it reduces the total idle
3825 * Cgroups make a horror show out of (2), instead of a simple sum we get:
3828 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
3833 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
3835 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
3837 * The big problem is S_k, its a global sum needed to compute a local (W_i)
3840 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
3841 * rewrite all of this once again.]
3844 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
3846 #define LBF_ALL_PINNED 0x01
3847 #define LBF_NEED_BREAK 0x02
3848 #define LBF_SOME_PINNED 0x04
3851 struct sched_domain *sd;
3859 struct cpumask *dst_grpmask;
3861 enum cpu_idle_type idle;
3863 /* The set of CPUs under consideration for load-balancing */
3864 struct cpumask *cpus;
3869 unsigned int loop_break;
3870 unsigned int loop_max;
3874 * move_task - move a task from one runqueue to another runqueue.
3875 * Both runqueues must be locked.
3877 static void move_task(struct task_struct *p, struct lb_env *env)
3879 deactivate_task(env->src_rq, p, 0);
3880 set_task_cpu(p, env->dst_cpu);
3881 activate_task(env->dst_rq, p, 0);
3882 check_preempt_curr(env->dst_rq, p, 0);
3886 * Is this task likely cache-hot:
3889 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
3893 if (p->sched_class != &fair_sched_class)
3896 if (unlikely(p->policy == SCHED_IDLE))
3900 * Buddy candidates are cache hot:
3902 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
3903 (&p->se == cfs_rq_of(&p->se)->next ||
3904 &p->se == cfs_rq_of(&p->se)->last))
3907 if (sysctl_sched_migration_cost == -1)
3909 if (sysctl_sched_migration_cost == 0)
3912 delta = now - p->se.exec_start;
3914 return delta < (s64)sysctl_sched_migration_cost;
3918 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3921 int can_migrate_task(struct task_struct *p, struct lb_env *env)
3923 int tsk_cache_hot = 0;
3925 * We do not migrate tasks that are:
3926 * 1) throttled_lb_pair, or
3927 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3928 * 3) running (obviously), or
3929 * 4) are cache-hot on their current CPU.
3931 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
3934 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
3937 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
3940 * Remember if this task can be migrated to any other cpu in
3941 * our sched_group. We may want to revisit it if we couldn't
3942 * meet load balance goals by pulling other tasks on src_cpu.
3944 * Also avoid computing new_dst_cpu if we have already computed
3945 * one in current iteration.
3947 if (!env->dst_grpmask || (env->flags & LBF_SOME_PINNED))
3950 /* Prevent to re-select dst_cpu via env's cpus */
3951 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
3952 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
3953 env->flags |= LBF_SOME_PINNED;
3954 env->new_dst_cpu = cpu;
3962 /* Record that we found atleast one task that could run on dst_cpu */
3963 env->flags &= ~LBF_ALL_PINNED;
3965 if (task_running(env->src_rq, p)) {
3966 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
3971 * Aggressive migration if:
3972 * 1) task is cache cold, or
3973 * 2) too many balance attempts have failed.
3976 tsk_cache_hot = task_hot(p, env->src_rq->clock_task, env->sd);
3977 if (!tsk_cache_hot ||
3978 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
3980 if (tsk_cache_hot) {
3981 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
3982 schedstat_inc(p, se.statistics.nr_forced_migrations);
3988 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
3993 * move_one_task tries to move exactly one task from busiest to this_rq, as
3994 * part of active balancing operations within "domain".
3995 * Returns 1 if successful and 0 otherwise.
3997 * Called with both runqueues locked.
3999 static int move_one_task(struct lb_env *env)
4001 struct task_struct *p, *n;
4003 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4004 if (!can_migrate_task(p, env))
4009 * Right now, this is only the second place move_task()
4010 * is called, so we can safely collect move_task()
4011 * stats here rather than inside move_task().
4013 schedstat_inc(env->sd, lb_gained[env->idle]);
4019 static unsigned long task_h_load(struct task_struct *p);
4021 static const unsigned int sched_nr_migrate_break = 32;
4024 * move_tasks tries to move up to imbalance weighted load from busiest to
4025 * this_rq, as part of a balancing operation within domain "sd".
4026 * Returns 1 if successful and 0 otherwise.
4028 * Called with both runqueues locked.
4030 static int move_tasks(struct lb_env *env)
4032 struct list_head *tasks = &env->src_rq->cfs_tasks;
4033 struct task_struct *p;
4037 if (env->imbalance <= 0)
4040 while (!list_empty(tasks)) {
4041 p = list_first_entry(tasks, struct task_struct, se.group_node);
4044 /* We've more or less seen every task there is, call it quits */
4045 if (env->loop > env->loop_max)
4048 /* take a breather every nr_migrate tasks */
4049 if (env->loop > env->loop_break) {
4050 env->loop_break += sched_nr_migrate_break;
4051 env->flags |= LBF_NEED_BREAK;
4055 if (!can_migrate_task(p, env))
4058 load = task_h_load(p);
4060 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4063 if ((load / 2) > env->imbalance)
4068 env->imbalance -= load;
4070 #ifdef CONFIG_PREEMPT
4072 * NEWIDLE balancing is a source of latency, so preemptible
4073 * kernels will stop after the first task is pulled to minimize
4074 * the critical section.
4076 if (env->idle == CPU_NEWLY_IDLE)
4081 * We only want to steal up to the prescribed amount of
4084 if (env->imbalance <= 0)
4089 list_move_tail(&p->se.group_node, tasks);
4093 * Right now, this is one of only two places move_task() is called,
4094 * so we can safely collect move_task() stats here rather than
4095 * inside move_task().
4097 schedstat_add(env->sd, lb_gained[env->idle], pulled);
4102 #ifdef CONFIG_FAIR_GROUP_SCHED
4104 * update tg->load_weight by folding this cpu's load_avg
4106 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
4108 struct sched_entity *se = tg->se[cpu];
4109 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
4111 /* throttled entities do not contribute to load */
4112 if (throttled_hierarchy(cfs_rq))
4115 update_cfs_rq_blocked_load(cfs_rq, 1);
4118 update_entity_load_avg(se, 1);
4120 * We pivot on our runnable average having decayed to zero for
4121 * list removal. This generally implies that all our children
4122 * have also been removed (modulo rounding error or bandwidth
4123 * control); however, such cases are rare and we can fix these
4126 * TODO: fix up out-of-order children on enqueue.
4128 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
4129 list_del_leaf_cfs_rq(cfs_rq);
4131 struct rq *rq = rq_of(cfs_rq);
4132 update_rq_runnable_avg(rq, rq->nr_running);
4136 static void update_blocked_averages(int cpu)
4138 struct rq *rq = cpu_rq(cpu);
4139 struct cfs_rq *cfs_rq;
4140 unsigned long flags;
4142 raw_spin_lock_irqsave(&rq->lock, flags);
4143 update_rq_clock(rq);
4145 * Iterates the task_group tree in a bottom up fashion, see
4146 * list_add_leaf_cfs_rq() for details.
4148 for_each_leaf_cfs_rq(rq, cfs_rq) {
4150 * Note: We may want to consider periodically releasing
4151 * rq->lock about these updates so that creating many task
4152 * groups does not result in continually extending hold time.
4154 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
4157 raw_spin_unlock_irqrestore(&rq->lock, flags);
4161 * Compute the cpu's hierarchical load factor for each task group.
4162 * This needs to be done in a top-down fashion because the load of a child
4163 * group is a fraction of its parents load.
4165 static int tg_load_down(struct task_group *tg, void *data)
4168 long cpu = (long)data;
4171 load = cpu_rq(cpu)->load.weight;
4173 load = tg->parent->cfs_rq[cpu]->h_load;
4174 load *= tg->se[cpu]->load.weight;
4175 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
4178 tg->cfs_rq[cpu]->h_load = load;
4183 static void update_h_load(long cpu)
4185 struct rq *rq = cpu_rq(cpu);
4186 unsigned long now = jiffies;
4188 if (rq->h_load_throttle == now)
4191 rq->h_load_throttle = now;
4194 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
4198 static unsigned long task_h_load(struct task_struct *p)
4200 struct cfs_rq *cfs_rq = task_cfs_rq(p);
4203 load = p->se.load.weight;
4204 load = div_u64(load * cfs_rq->h_load, cfs_rq->load.weight + 1);
4209 static inline void update_blocked_averages(int cpu)
4213 static inline void update_h_load(long cpu)
4217 static unsigned long task_h_load(struct task_struct *p)
4219 return p->se.load.weight;
4223 /********** Helpers for find_busiest_group ************************/
4225 * sd_lb_stats - Structure to store the statistics of a sched_domain
4226 * during load balancing.
4228 struct sd_lb_stats {
4229 struct sched_group *busiest; /* Busiest group in this sd */
4230 struct sched_group *this; /* Local group in this sd */
4231 unsigned long total_load; /* Total load of all groups in sd */
4232 unsigned long total_pwr; /* Total power of all groups in sd */
4233 unsigned long avg_load; /* Average load across all groups in sd */
4235 /** Statistics of this group */
4236 unsigned long this_load;
4237 unsigned long this_load_per_task;
4238 unsigned long this_nr_running;
4239 unsigned long this_has_capacity;
4240 unsigned int this_idle_cpus;
4242 /* Statistics of the busiest group */
4243 unsigned int busiest_idle_cpus;
4244 unsigned long max_load;
4245 unsigned long busiest_load_per_task;
4246 unsigned long busiest_nr_running;
4247 unsigned long busiest_group_capacity;
4248 unsigned long busiest_has_capacity;
4249 unsigned int busiest_group_weight;
4251 int group_imb; /* Is there imbalance in this sd */
4255 * sg_lb_stats - stats of a sched_group required for load_balancing
4257 struct sg_lb_stats {
4258 unsigned long avg_load; /*Avg load across the CPUs of the group */
4259 unsigned long group_load; /* Total load over the CPUs of the group */
4260 unsigned long sum_nr_running; /* Nr tasks running in the group */
4261 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
4262 unsigned long group_capacity;
4263 unsigned long idle_cpus;
4264 unsigned long group_weight;
4265 int group_imb; /* Is there an imbalance in the group ? */
4266 int group_has_capacity; /* Is there extra capacity in the group? */
4270 * get_sd_load_idx - Obtain the load index for a given sched domain.
4271 * @sd: The sched_domain whose load_idx is to be obtained.
4272 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4274 static inline int get_sd_load_idx(struct sched_domain *sd,
4275 enum cpu_idle_type idle)
4281 load_idx = sd->busy_idx;
4284 case CPU_NEWLY_IDLE:
4285 load_idx = sd->newidle_idx;
4288 load_idx = sd->idle_idx;
4295 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
4297 return SCHED_POWER_SCALE;
4300 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
4302 return default_scale_freq_power(sd, cpu);
4305 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
4307 unsigned long weight = sd->span_weight;
4308 unsigned long smt_gain = sd->smt_gain;
4315 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
4317 return default_scale_smt_power(sd, cpu);
4320 static unsigned long scale_rt_power(int cpu)
4322 struct rq *rq = cpu_rq(cpu);
4323 u64 total, available, age_stamp, avg;
4326 * Since we're reading these variables without serialization make sure
4327 * we read them once before doing sanity checks on them.
4329 age_stamp = ACCESS_ONCE(rq->age_stamp);
4330 avg = ACCESS_ONCE(rq->rt_avg);
4332 total = sched_avg_period() + (rq->clock - age_stamp);
4334 if (unlikely(total < avg)) {
4335 /* Ensures that power won't end up being negative */
4338 available = total - avg;
4341 if (unlikely((s64)total < SCHED_POWER_SCALE))
4342 total = SCHED_POWER_SCALE;
4344 total >>= SCHED_POWER_SHIFT;
4346 return div_u64(available, total);
4349 static void update_cpu_power(struct sched_domain *sd, int cpu)
4351 unsigned long weight = sd->span_weight;
4352 unsigned long power = SCHED_POWER_SCALE;
4353 struct sched_group *sdg = sd->groups;
4355 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
4356 if (sched_feat(ARCH_POWER))
4357 power *= arch_scale_smt_power(sd, cpu);
4359 power *= default_scale_smt_power(sd, cpu);
4361 power >>= SCHED_POWER_SHIFT;
4364 sdg->sgp->power_orig = power;
4366 if (sched_feat(ARCH_POWER))
4367 power *= arch_scale_freq_power(sd, cpu);
4369 power *= default_scale_freq_power(sd, cpu);
4371 power >>= SCHED_POWER_SHIFT;
4373 power *= scale_rt_power(cpu);
4374 power >>= SCHED_POWER_SHIFT;
4379 cpu_rq(cpu)->cpu_power = power;
4380 sdg->sgp->power = power;
4383 void update_group_power(struct sched_domain *sd, int cpu)
4385 struct sched_domain *child = sd->child;
4386 struct sched_group *group, *sdg = sd->groups;
4387 unsigned long power;
4388 unsigned long interval;
4390 interval = msecs_to_jiffies(sd->balance_interval);
4391 interval = clamp(interval, 1UL, max_load_balance_interval);
4392 sdg->sgp->next_update = jiffies + interval;
4395 update_cpu_power(sd, cpu);
4401 if (child->flags & SD_OVERLAP) {
4403 * SD_OVERLAP domains cannot assume that child groups
4404 * span the current group.
4407 for_each_cpu(cpu, sched_group_cpus(sdg))
4408 power += power_of(cpu);
4411 * !SD_OVERLAP domains can assume that child groups
4412 * span the current group.
4415 group = child->groups;
4417 power += group->sgp->power;
4418 group = group->next;
4419 } while (group != child->groups);
4422 sdg->sgp->power_orig = sdg->sgp->power = power;
4426 * Try and fix up capacity for tiny siblings, this is needed when
4427 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4428 * which on its own isn't powerful enough.
4430 * See update_sd_pick_busiest() and check_asym_packing().
4433 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
4436 * Only siblings can have significantly less than SCHED_POWER_SCALE
4438 if (!(sd->flags & SD_SHARE_CPUPOWER))
4442 * If ~90% of the cpu_power is still there, we're good.
4444 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
4451 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4452 * @env: The load balancing environment.
4453 * @group: sched_group whose statistics are to be updated.
4454 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4455 * @local_group: Does group contain this_cpu.
4456 * @balance: Should we balance.
4457 * @sgs: variable to hold the statistics for this group.
4459 static inline void update_sg_lb_stats(struct lb_env *env,
4460 struct sched_group *group, int load_idx,
4461 int local_group, int *balance, struct sg_lb_stats *sgs)
4463 unsigned long nr_running, max_nr_running, min_nr_running;
4464 unsigned long load, max_cpu_load, min_cpu_load;
4465 unsigned int balance_cpu = -1, first_idle_cpu = 0;
4466 unsigned long avg_load_per_task = 0;
4470 balance_cpu = group_balance_cpu(group);
4472 /* Tally up the load of all CPUs in the group */
4474 min_cpu_load = ~0UL;
4476 min_nr_running = ~0UL;
4478 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
4479 struct rq *rq = cpu_rq(i);
4481 nr_running = rq->nr_running;
4483 /* Bias balancing toward cpus of our domain */
4485 if (idle_cpu(i) && !first_idle_cpu &&
4486 cpumask_test_cpu(i, sched_group_mask(group))) {
4491 load = target_load(i, load_idx);
4493 load = source_load(i, load_idx);
4494 if (load > max_cpu_load)
4495 max_cpu_load = load;
4496 if (min_cpu_load > load)
4497 min_cpu_load = load;
4499 if (nr_running > max_nr_running)
4500 max_nr_running = nr_running;
4501 if (min_nr_running > nr_running)
4502 min_nr_running = nr_running;
4505 sgs->group_load += load;
4506 sgs->sum_nr_running += nr_running;
4507 sgs->sum_weighted_load += weighted_cpuload(i);
4513 * First idle cpu or the first cpu(busiest) in this sched group
4514 * is eligible for doing load balancing at this and above
4515 * domains. In the newly idle case, we will allow all the cpu's
4516 * to do the newly idle load balance.
4519 if (env->idle != CPU_NEWLY_IDLE) {
4520 if (balance_cpu != env->dst_cpu) {
4524 update_group_power(env->sd, env->dst_cpu);
4525 } else if (time_after_eq(jiffies, group->sgp->next_update))
4526 update_group_power(env->sd, env->dst_cpu);
4529 /* Adjust by relative CPU power of the group */
4530 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
4533 * Consider the group unbalanced when the imbalance is larger
4534 * than the average weight of a task.
4536 * APZ: with cgroup the avg task weight can vary wildly and
4537 * might not be a suitable number - should we keep a
4538 * normalized nr_running number somewhere that negates
4541 if (sgs->sum_nr_running)
4542 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
4544 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task &&
4545 (max_nr_running - min_nr_running) > 1)
4548 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
4550 if (!sgs->group_capacity)
4551 sgs->group_capacity = fix_small_capacity(env->sd, group);
4552 sgs->group_weight = group->group_weight;
4554 if (sgs->group_capacity > sgs->sum_nr_running)
4555 sgs->group_has_capacity = 1;
4559 * update_sd_pick_busiest - return 1 on busiest group
4560 * @env: The load balancing environment.
4561 * @sds: sched_domain statistics
4562 * @sg: sched_group candidate to be checked for being the busiest
4563 * @sgs: sched_group statistics
4565 * Determine if @sg is a busier group than the previously selected
4568 static bool update_sd_pick_busiest(struct lb_env *env,
4569 struct sd_lb_stats *sds,
4570 struct sched_group *sg,
4571 struct sg_lb_stats *sgs)
4573 if (sgs->avg_load <= sds->max_load)
4576 if (sgs->sum_nr_running > sgs->group_capacity)
4583 * ASYM_PACKING needs to move all the work to the lowest
4584 * numbered CPUs in the group, therefore mark all groups
4585 * higher than ourself as busy.
4587 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
4588 env->dst_cpu < group_first_cpu(sg)) {
4592 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
4600 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
4601 * @env: The load balancing environment.
4602 * @balance: Should we balance.
4603 * @sds: variable to hold the statistics for this sched_domain.
4605 static inline void update_sd_lb_stats(struct lb_env *env,
4606 int *balance, struct sd_lb_stats *sds)
4608 struct sched_domain *child = env->sd->child;
4609 struct sched_group *sg = env->sd->groups;
4610 struct sg_lb_stats sgs;
4611 int load_idx, prefer_sibling = 0;
4613 if (child && child->flags & SD_PREFER_SIBLING)
4616 load_idx = get_sd_load_idx(env->sd, env->idle);
4621 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
4622 memset(&sgs, 0, sizeof(sgs));
4623 update_sg_lb_stats(env, sg, load_idx, local_group, balance, &sgs);
4625 if (local_group && !(*balance))
4628 sds->total_load += sgs.group_load;
4629 sds->total_pwr += sg->sgp->power;
4632 * In case the child domain prefers tasks go to siblings
4633 * first, lower the sg capacity to one so that we'll try
4634 * and move all the excess tasks away. We lower the capacity
4635 * of a group only if the local group has the capacity to fit
4636 * these excess tasks, i.e. nr_running < group_capacity. The
4637 * extra check prevents the case where you always pull from the
4638 * heaviest group when it is already under-utilized (possible
4639 * with a large weight task outweighs the tasks on the system).
4641 if (prefer_sibling && !local_group && sds->this_has_capacity)
4642 sgs.group_capacity = min(sgs.group_capacity, 1UL);
4645 sds->this_load = sgs.avg_load;
4647 sds->this_nr_running = sgs.sum_nr_running;
4648 sds->this_load_per_task = sgs.sum_weighted_load;
4649 sds->this_has_capacity = sgs.group_has_capacity;
4650 sds->this_idle_cpus = sgs.idle_cpus;
4651 } else if (update_sd_pick_busiest(env, sds, sg, &sgs)) {
4652 sds->max_load = sgs.avg_load;
4654 sds->busiest_nr_running = sgs.sum_nr_running;
4655 sds->busiest_idle_cpus = sgs.idle_cpus;
4656 sds->busiest_group_capacity = sgs.group_capacity;
4657 sds->busiest_load_per_task = sgs.sum_weighted_load;
4658 sds->busiest_has_capacity = sgs.group_has_capacity;
4659 sds->busiest_group_weight = sgs.group_weight;
4660 sds->group_imb = sgs.group_imb;
4664 } while (sg != env->sd->groups);
4668 * check_asym_packing - Check to see if the group is packed into the
4671 * This is primarily intended to used at the sibling level. Some
4672 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4673 * case of POWER7, it can move to lower SMT modes only when higher
4674 * threads are idle. When in lower SMT modes, the threads will
4675 * perform better since they share less core resources. Hence when we
4676 * have idle threads, we want them to be the higher ones.
4678 * This packing function is run on idle threads. It checks to see if
4679 * the busiest CPU in this domain (core in the P7 case) has a higher
4680 * CPU number than the packing function is being run on. Here we are
4681 * assuming lower CPU number will be equivalent to lower a SMT thread
4684 * Returns 1 when packing is required and a task should be moved to
4685 * this CPU. The amount of the imbalance is returned in *imbalance.
4687 * @env: The load balancing environment.
4688 * @sds: Statistics of the sched_domain which is to be packed
4690 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
4694 if (!(env->sd->flags & SD_ASYM_PACKING))
4700 busiest_cpu = group_first_cpu(sds->busiest);
4701 if (env->dst_cpu > busiest_cpu)
4704 env->imbalance = DIV_ROUND_CLOSEST(
4705 sds->max_load * sds->busiest->sgp->power, SCHED_POWER_SCALE);
4711 * fix_small_imbalance - Calculate the minor imbalance that exists
4712 * amongst the groups of a sched_domain, during
4714 * @env: The load balancing environment.
4715 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4718 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4720 unsigned long tmp, pwr_now = 0, pwr_move = 0;
4721 unsigned int imbn = 2;
4722 unsigned long scaled_busy_load_per_task;
4724 if (sds->this_nr_running) {
4725 sds->this_load_per_task /= sds->this_nr_running;
4726 if (sds->busiest_load_per_task >
4727 sds->this_load_per_task)
4730 sds->this_load_per_task =
4731 cpu_avg_load_per_task(env->dst_cpu);
4734 scaled_busy_load_per_task = sds->busiest_load_per_task
4735 * SCHED_POWER_SCALE;
4736 scaled_busy_load_per_task /= sds->busiest->sgp->power;
4738 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
4739 (scaled_busy_load_per_task * imbn)) {
4740 env->imbalance = sds->busiest_load_per_task;
4745 * OK, we don't have enough imbalance to justify moving tasks,
4746 * however we may be able to increase total CPU power used by
4750 pwr_now += sds->busiest->sgp->power *
4751 min(sds->busiest_load_per_task, sds->max_load);
4752 pwr_now += sds->this->sgp->power *
4753 min(sds->this_load_per_task, sds->this_load);
4754 pwr_now /= SCHED_POWER_SCALE;
4756 /* Amount of load we'd subtract */
4757 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4758 sds->busiest->sgp->power;
4759 if (sds->max_load > tmp)
4760 pwr_move += sds->busiest->sgp->power *
4761 min(sds->busiest_load_per_task, sds->max_load - tmp);
4763 /* Amount of load we'd add */
4764 if (sds->max_load * sds->busiest->sgp->power <
4765 sds->busiest_load_per_task * SCHED_POWER_SCALE)
4766 tmp = (sds->max_load * sds->busiest->sgp->power) /
4767 sds->this->sgp->power;
4769 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4770 sds->this->sgp->power;
4771 pwr_move += sds->this->sgp->power *
4772 min(sds->this_load_per_task, sds->this_load + tmp);
4773 pwr_move /= SCHED_POWER_SCALE;
4775 /* Move if we gain throughput */
4776 if (pwr_move > pwr_now)
4777 env->imbalance = sds->busiest_load_per_task;
4781 * calculate_imbalance - Calculate the amount of imbalance present within the
4782 * groups of a given sched_domain during load balance.
4783 * @env: load balance environment
4784 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4786 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4788 unsigned long max_pull, load_above_capacity = ~0UL;
4790 sds->busiest_load_per_task /= sds->busiest_nr_running;
4791 if (sds->group_imb) {
4792 sds->busiest_load_per_task =
4793 min(sds->busiest_load_per_task, sds->avg_load);
4797 * In the presence of smp nice balancing, certain scenarios can have
4798 * max load less than avg load(as we skip the groups at or below
4799 * its cpu_power, while calculating max_load..)
4801 if (sds->max_load < sds->avg_load) {
4803 return fix_small_imbalance(env, sds);
4806 if (!sds->group_imb) {
4808 * Don't want to pull so many tasks that a group would go idle.
4810 load_above_capacity = (sds->busiest_nr_running -
4811 sds->busiest_group_capacity);
4813 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
4815 load_above_capacity /= sds->busiest->sgp->power;
4819 * We're trying to get all the cpus to the average_load, so we don't
4820 * want to push ourselves above the average load, nor do we wish to
4821 * reduce the max loaded cpu below the average load. At the same time,
4822 * we also don't want to reduce the group load below the group capacity
4823 * (so that we can implement power-savings policies etc). Thus we look
4824 * for the minimum possible imbalance.
4825 * Be careful of negative numbers as they'll appear as very large values
4826 * with unsigned longs.
4828 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
4830 /* How much load to actually move to equalise the imbalance */
4831 env->imbalance = min(max_pull * sds->busiest->sgp->power,
4832 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
4833 / SCHED_POWER_SCALE;
4836 * if *imbalance is less than the average load per runnable task
4837 * there is no guarantee that any tasks will be moved so we'll have
4838 * a think about bumping its value to force at least one task to be
4841 if (env->imbalance < sds->busiest_load_per_task)
4842 return fix_small_imbalance(env, sds);
4846 /******* find_busiest_group() helpers end here *********************/
4849 * find_busiest_group - Returns the busiest group within the sched_domain
4850 * if there is an imbalance. If there isn't an imbalance, and
4851 * the user has opted for power-savings, it returns a group whose
4852 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4853 * such a group exists.
4855 * Also calculates the amount of weighted load which should be moved
4856 * to restore balance.
4858 * @env: The load balancing environment.
4859 * @balance: Pointer to a variable indicating if this_cpu
4860 * is the appropriate cpu to perform load balancing at this_level.
4862 * Returns: - the busiest group if imbalance exists.
4863 * - If no imbalance and user has opted for power-savings balance,
4864 * return the least loaded group whose CPUs can be
4865 * put to idle by rebalancing its tasks onto our group.
4867 static struct sched_group *
4868 find_busiest_group(struct lb_env *env, int *balance)
4870 struct sd_lb_stats sds;
4872 memset(&sds, 0, sizeof(sds));
4875 * Compute the various statistics relavent for load balancing at
4878 update_sd_lb_stats(env, balance, &sds);
4881 * this_cpu is not the appropriate cpu to perform load balancing at
4887 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
4888 check_asym_packing(env, &sds))
4891 /* There is no busy sibling group to pull tasks from */
4892 if (!sds.busiest || sds.busiest_nr_running == 0)
4895 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
4898 * If the busiest group is imbalanced the below checks don't
4899 * work because they assumes all things are equal, which typically
4900 * isn't true due to cpus_allowed constraints and the like.
4905 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
4906 if (env->idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
4907 !sds.busiest_has_capacity)
4911 * If the local group is more busy than the selected busiest group
4912 * don't try and pull any tasks.
4914 if (sds.this_load >= sds.max_load)
4918 * Don't pull any tasks if this group is already above the domain
4921 if (sds.this_load >= sds.avg_load)
4924 if (env->idle == CPU_IDLE) {
4926 * This cpu is idle. If the busiest group load doesn't
4927 * have more tasks than the number of available cpu's and
4928 * there is no imbalance between this and busiest group
4929 * wrt to idle cpu's, it is balanced.
4931 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
4932 sds.busiest_nr_running <= sds.busiest_group_weight)
4936 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
4937 * imbalance_pct to be conservative.
4939 if (100 * sds.max_load <= env->sd->imbalance_pct * sds.this_load)
4944 /* Looks like there is an imbalance. Compute it */
4945 calculate_imbalance(env, &sds);
4955 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4957 static struct rq *find_busiest_queue(struct lb_env *env,
4958 struct sched_group *group)
4960 struct rq *busiest = NULL, *rq;
4961 unsigned long max_load = 0;
4964 for_each_cpu(i, sched_group_cpus(group)) {
4965 unsigned long power = power_of(i);
4966 unsigned long capacity = DIV_ROUND_CLOSEST(power,
4971 capacity = fix_small_capacity(env->sd, group);
4973 if (!cpumask_test_cpu(i, env->cpus))
4977 wl = weighted_cpuload(i);
4980 * When comparing with imbalance, use weighted_cpuload()
4981 * which is not scaled with the cpu power.
4983 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
4987 * For the load comparisons with the other cpu's, consider
4988 * the weighted_cpuload() scaled with the cpu power, so that
4989 * the load can be moved away from the cpu that is potentially
4990 * running at a lower capacity.
4992 wl = (wl * SCHED_POWER_SCALE) / power;
4994 if (wl > max_load) {
5004 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5005 * so long as it is large enough.
5007 #define MAX_PINNED_INTERVAL 512
5009 /* Working cpumask for load_balance and load_balance_newidle. */
5010 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5012 static int need_active_balance(struct lb_env *env)
5014 struct sched_domain *sd = env->sd;
5016 if (env->idle == CPU_NEWLY_IDLE) {
5019 * ASYM_PACKING needs to force migrate tasks from busy but
5020 * higher numbered CPUs in order to pack all tasks in the
5021 * lowest numbered CPUs.
5023 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
5027 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
5030 static int active_load_balance_cpu_stop(void *data);
5033 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5034 * tasks if there is an imbalance.
5036 static int load_balance(int this_cpu, struct rq *this_rq,
5037 struct sched_domain *sd, enum cpu_idle_type idle,
5040 int ld_moved, cur_ld_moved, active_balance = 0;
5041 struct sched_group *group;
5043 unsigned long flags;
5044 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
5046 struct lb_env env = {
5048 .dst_cpu = this_cpu,
5050 .dst_grpmask = sched_group_cpus(sd->groups),
5052 .loop_break = sched_nr_migrate_break,
5057 * For NEWLY_IDLE load_balancing, we don't need to consider
5058 * other cpus in our group
5060 if (idle == CPU_NEWLY_IDLE)
5061 env.dst_grpmask = NULL;
5063 cpumask_copy(cpus, cpu_active_mask);
5065 schedstat_inc(sd, lb_count[idle]);
5068 group = find_busiest_group(&env, balance);
5074 schedstat_inc(sd, lb_nobusyg[idle]);
5078 busiest = find_busiest_queue(&env, group);
5080 schedstat_inc(sd, lb_nobusyq[idle]);
5084 BUG_ON(busiest == env.dst_rq);
5086 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
5089 if (busiest->nr_running > 1) {
5091 * Attempt to move tasks. If find_busiest_group has found
5092 * an imbalance but busiest->nr_running <= 1, the group is
5093 * still unbalanced. ld_moved simply stays zero, so it is
5094 * correctly treated as an imbalance.
5096 env.flags |= LBF_ALL_PINNED;
5097 env.src_cpu = busiest->cpu;
5098 env.src_rq = busiest;
5099 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
5101 update_h_load(env.src_cpu);
5103 local_irq_save(flags);
5104 double_rq_lock(env.dst_rq, busiest);
5107 * cur_ld_moved - load moved in current iteration
5108 * ld_moved - cumulative load moved across iterations
5110 cur_ld_moved = move_tasks(&env);
5111 ld_moved += cur_ld_moved;
5112 double_rq_unlock(env.dst_rq, busiest);
5113 local_irq_restore(flags);
5116 * some other cpu did the load balance for us.
5118 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
5119 resched_cpu(env.dst_cpu);
5121 if (env.flags & LBF_NEED_BREAK) {
5122 env.flags &= ~LBF_NEED_BREAK;
5127 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5128 * us and move them to an alternate dst_cpu in our sched_group
5129 * where they can run. The upper limit on how many times we
5130 * iterate on same src_cpu is dependent on number of cpus in our
5133 * This changes load balance semantics a bit on who can move
5134 * load to a given_cpu. In addition to the given_cpu itself
5135 * (or a ilb_cpu acting on its behalf where given_cpu is
5136 * nohz-idle), we now have balance_cpu in a position to move
5137 * load to given_cpu. In rare situations, this may cause
5138 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5139 * _independently_ and at _same_ time to move some load to
5140 * given_cpu) causing exceess load to be moved to given_cpu.
5141 * This however should not happen so much in practice and
5142 * moreover subsequent load balance cycles should correct the
5143 * excess load moved.
5145 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
5147 env.dst_rq = cpu_rq(env.new_dst_cpu);
5148 env.dst_cpu = env.new_dst_cpu;
5149 env.flags &= ~LBF_SOME_PINNED;
5151 env.loop_break = sched_nr_migrate_break;
5153 /* Prevent to re-select dst_cpu via env's cpus */
5154 cpumask_clear_cpu(env.dst_cpu, env.cpus);
5157 * Go back to "more_balance" rather than "redo" since we
5158 * need to continue with same src_cpu.
5163 /* All tasks on this runqueue were pinned by CPU affinity */
5164 if (unlikely(env.flags & LBF_ALL_PINNED)) {
5165 cpumask_clear_cpu(cpu_of(busiest), cpus);
5166 if (!cpumask_empty(cpus)) {
5168 env.loop_break = sched_nr_migrate_break;
5176 schedstat_inc(sd, lb_failed[idle]);
5178 * Increment the failure counter only on periodic balance.
5179 * We do not want newidle balance, which can be very
5180 * frequent, pollute the failure counter causing
5181 * excessive cache_hot migrations and active balances.
5183 if (idle != CPU_NEWLY_IDLE)
5184 sd->nr_balance_failed++;
5186 if (need_active_balance(&env)) {
5187 raw_spin_lock_irqsave(&busiest->lock, flags);
5189 /* don't kick the active_load_balance_cpu_stop,
5190 * if the curr task on busiest cpu can't be
5193 if (!cpumask_test_cpu(this_cpu,
5194 tsk_cpus_allowed(busiest->curr))) {
5195 raw_spin_unlock_irqrestore(&busiest->lock,
5197 env.flags |= LBF_ALL_PINNED;
5198 goto out_one_pinned;
5202 * ->active_balance synchronizes accesses to
5203 * ->active_balance_work. Once set, it's cleared
5204 * only after active load balance is finished.
5206 if (!busiest->active_balance) {
5207 busiest->active_balance = 1;
5208 busiest->push_cpu = this_cpu;
5211 raw_spin_unlock_irqrestore(&busiest->lock, flags);
5213 if (active_balance) {
5214 stop_one_cpu_nowait(cpu_of(busiest),
5215 active_load_balance_cpu_stop, busiest,
5216 &busiest->active_balance_work);
5220 * We've kicked active balancing, reset the failure
5223 sd->nr_balance_failed = sd->cache_nice_tries+1;
5226 sd->nr_balance_failed = 0;
5228 if (likely(!active_balance)) {
5229 /* We were unbalanced, so reset the balancing interval */
5230 sd->balance_interval = sd->min_interval;
5233 * If we've begun active balancing, start to back off. This
5234 * case may not be covered by the all_pinned logic if there
5235 * is only 1 task on the busy runqueue (because we don't call
5238 if (sd->balance_interval < sd->max_interval)
5239 sd->balance_interval *= 2;
5245 schedstat_inc(sd, lb_balanced[idle]);
5247 sd->nr_balance_failed = 0;
5250 /* tune up the balancing interval */
5251 if (((env.flags & LBF_ALL_PINNED) &&
5252 sd->balance_interval < MAX_PINNED_INTERVAL) ||
5253 (sd->balance_interval < sd->max_interval))
5254 sd->balance_interval *= 2;
5262 * idle_balance is called by schedule() if this_cpu is about to become
5263 * idle. Attempts to pull tasks from other CPUs.
5265 void idle_balance(int this_cpu, struct rq *this_rq)
5267 struct sched_domain *sd;
5268 int pulled_task = 0;
5269 unsigned long next_balance = jiffies + HZ;
5271 this_rq->idle_stamp = this_rq->clock;
5273 if (this_rq->avg_idle < sysctl_sched_migration_cost)
5277 * Drop the rq->lock, but keep IRQ/preempt disabled.
5279 raw_spin_unlock(&this_rq->lock);
5281 update_blocked_averages(this_cpu);
5283 for_each_domain(this_cpu, sd) {
5284 unsigned long interval;
5287 if (!(sd->flags & SD_LOAD_BALANCE))
5290 if (sd->flags & SD_BALANCE_NEWIDLE) {
5291 /* If we've pulled tasks over stop searching: */
5292 pulled_task = load_balance(this_cpu, this_rq,
5293 sd, CPU_NEWLY_IDLE, &balance);
5296 interval = msecs_to_jiffies(sd->balance_interval);
5297 if (time_after(next_balance, sd->last_balance + interval))
5298 next_balance = sd->last_balance + interval;
5300 this_rq->idle_stamp = 0;
5306 raw_spin_lock(&this_rq->lock);
5308 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
5310 * We are going idle. next_balance may be set based on
5311 * a busy processor. So reset next_balance.
5313 this_rq->next_balance = next_balance;
5318 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5319 * running tasks off the busiest CPU onto idle CPUs. It requires at
5320 * least 1 task to be running on each physical CPU where possible, and
5321 * avoids physical / logical imbalances.
5323 static int active_load_balance_cpu_stop(void *data)
5325 struct rq *busiest_rq = data;
5326 int busiest_cpu = cpu_of(busiest_rq);
5327 int target_cpu = busiest_rq->push_cpu;
5328 struct rq *target_rq = cpu_rq(target_cpu);
5329 struct sched_domain *sd;
5331 raw_spin_lock_irq(&busiest_rq->lock);
5333 /* make sure the requested cpu hasn't gone down in the meantime */
5334 if (unlikely(busiest_cpu != smp_processor_id() ||
5335 !busiest_rq->active_balance))
5338 /* Is there any task to move? */
5339 if (busiest_rq->nr_running <= 1)
5343 * This condition is "impossible", if it occurs
5344 * we need to fix it. Originally reported by
5345 * Bjorn Helgaas on a 128-cpu setup.
5347 BUG_ON(busiest_rq == target_rq);
5349 /* move a task from busiest_rq to target_rq */
5350 double_lock_balance(busiest_rq, target_rq);
5352 /* Search for an sd spanning us and the target CPU. */
5354 for_each_domain(target_cpu, sd) {
5355 if ((sd->flags & SD_LOAD_BALANCE) &&
5356 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
5361 struct lb_env env = {
5363 .dst_cpu = target_cpu,
5364 .dst_rq = target_rq,
5365 .src_cpu = busiest_rq->cpu,
5366 .src_rq = busiest_rq,
5370 schedstat_inc(sd, alb_count);
5372 if (move_one_task(&env))
5373 schedstat_inc(sd, alb_pushed);
5375 schedstat_inc(sd, alb_failed);
5378 double_unlock_balance(busiest_rq, target_rq);
5380 busiest_rq->active_balance = 0;
5381 raw_spin_unlock_irq(&busiest_rq->lock);
5385 #ifdef CONFIG_NO_HZ_COMMON
5387 * idle load balancing details
5388 * - When one of the busy CPUs notice that there may be an idle rebalancing
5389 * needed, they will kick the idle load balancer, which then does idle
5390 * load balancing for all the idle CPUs.
5393 cpumask_var_t idle_cpus_mask;
5395 unsigned long next_balance; /* in jiffy units */
5396 } nohz ____cacheline_aligned;
5398 static inline int find_new_ilb(int call_cpu)
5400 int ilb = cpumask_first(nohz.idle_cpus_mask);
5402 if (ilb < nr_cpu_ids && idle_cpu(ilb))
5409 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5410 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5411 * CPU (if there is one).
5413 static void nohz_balancer_kick(int cpu)
5417 nohz.next_balance++;
5419 ilb_cpu = find_new_ilb(cpu);
5421 if (ilb_cpu >= nr_cpu_ids)
5424 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
5427 * Use smp_send_reschedule() instead of resched_cpu().
5428 * This way we generate a sched IPI on the target cpu which
5429 * is idle. And the softirq performing nohz idle load balance
5430 * will be run before returning from the IPI.
5432 smp_send_reschedule(ilb_cpu);
5436 static inline void nohz_balance_exit_idle(int cpu)
5438 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5439 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5440 atomic_dec(&nohz.nr_cpus);
5441 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5445 static inline void set_cpu_sd_state_busy(void)
5447 struct sched_domain *sd;
5448 int cpu = smp_processor_id();
5451 sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd);
5453 if (!sd || !sd->nohz_idle)
5457 for (; sd; sd = sd->parent)
5458 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
5463 void set_cpu_sd_state_idle(void)
5465 struct sched_domain *sd;
5466 int cpu = smp_processor_id();
5469 sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd);
5471 if (!sd || sd->nohz_idle)
5475 for (; sd; sd = sd->parent)
5476 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
5482 * This routine will record that the cpu is going idle with tick stopped.
5483 * This info will be used in performing idle load balancing in the future.
5485 void nohz_balance_enter_idle(int cpu)
5488 * If this cpu is going down, then nothing needs to be done.
5490 if (!cpu_active(cpu))
5493 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
5496 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
5497 atomic_inc(&nohz.nr_cpus);
5498 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5501 static int __cpuinit sched_ilb_notifier(struct notifier_block *nfb,
5502 unsigned long action, void *hcpu)
5504 switch (action & ~CPU_TASKS_FROZEN) {
5506 nohz_balance_exit_idle(smp_processor_id());
5514 static DEFINE_SPINLOCK(balancing);
5517 * Scale the max load_balance interval with the number of CPUs in the system.
5518 * This trades load-balance latency on larger machines for less cross talk.
5520 void update_max_interval(void)
5522 max_load_balance_interval = HZ*num_online_cpus()/10;
5526 * It checks each scheduling domain to see if it is due to be balanced,
5527 * and initiates a balancing operation if so.
5529 * Balancing parameters are set up in init_sched_domains.
5531 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
5534 struct rq *rq = cpu_rq(cpu);
5535 unsigned long interval;
5536 struct sched_domain *sd;
5537 /* Earliest time when we have to do rebalance again */
5538 unsigned long next_balance = jiffies + 60*HZ;
5539 int update_next_balance = 0;
5542 update_blocked_averages(cpu);
5545 for_each_domain(cpu, sd) {
5546 if (!(sd->flags & SD_LOAD_BALANCE))
5549 interval = sd->balance_interval;
5550 if (idle != CPU_IDLE)
5551 interval *= sd->busy_factor;
5553 /* scale ms to jiffies */
5554 interval = msecs_to_jiffies(interval);
5555 interval = clamp(interval, 1UL, max_load_balance_interval);
5557 need_serialize = sd->flags & SD_SERIALIZE;
5559 if (need_serialize) {
5560 if (!spin_trylock(&balancing))
5564 if (time_after_eq(jiffies, sd->last_balance + interval)) {
5565 if (load_balance(cpu, rq, sd, idle, &balance)) {
5567 * The LBF_SOME_PINNED logic could have changed
5568 * env->dst_cpu, so we can't know our idle
5569 * state even if we migrated tasks. Update it.
5571 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
5573 sd->last_balance = jiffies;
5576 spin_unlock(&balancing);
5578 if (time_after(next_balance, sd->last_balance + interval)) {
5579 next_balance = sd->last_balance + interval;
5580 update_next_balance = 1;
5584 * Stop the load balance at this level. There is another
5585 * CPU in our sched group which is doing load balancing more
5594 * next_balance will be updated only when there is a need.
5595 * When the cpu is attached to null domain for ex, it will not be
5598 if (likely(update_next_balance))
5599 rq->next_balance = next_balance;
5602 #ifdef CONFIG_NO_HZ_COMMON
5604 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
5605 * rebalancing for all the cpus for whom scheduler ticks are stopped.
5607 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
5609 struct rq *this_rq = cpu_rq(this_cpu);
5613 if (idle != CPU_IDLE ||
5614 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
5617 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
5618 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
5622 * If this cpu gets work to do, stop the load balancing
5623 * work being done for other cpus. Next load
5624 * balancing owner will pick it up.
5629 rq = cpu_rq(balance_cpu);
5631 raw_spin_lock_irq(&rq->lock);
5632 update_rq_clock(rq);
5633 update_idle_cpu_load(rq);
5634 raw_spin_unlock_irq(&rq->lock);
5636 rebalance_domains(balance_cpu, CPU_IDLE);
5638 if (time_after(this_rq->next_balance, rq->next_balance))
5639 this_rq->next_balance = rq->next_balance;
5641 nohz.next_balance = this_rq->next_balance;
5643 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
5647 * Current heuristic for kicking the idle load balancer in the presence
5648 * of an idle cpu is the system.
5649 * - This rq has more than one task.
5650 * - At any scheduler domain level, this cpu's scheduler group has multiple
5651 * busy cpu's exceeding the group's power.
5652 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
5653 * domain span are idle.
5655 static inline int nohz_kick_needed(struct rq *rq, int cpu)
5657 unsigned long now = jiffies;
5658 struct sched_domain *sd;
5660 if (unlikely(idle_cpu(cpu)))
5664 * We may be recently in ticked or tickless idle mode. At the first
5665 * busy tick after returning from idle, we will update the busy stats.
5667 set_cpu_sd_state_busy();
5668 nohz_balance_exit_idle(cpu);
5671 * None are in tickless mode and hence no need for NOHZ idle load
5674 if (likely(!atomic_read(&nohz.nr_cpus)))
5677 if (time_before(now, nohz.next_balance))
5680 if (rq->nr_running >= 2)
5684 for_each_domain(cpu, sd) {
5685 struct sched_group *sg = sd->groups;
5686 struct sched_group_power *sgp = sg->sgp;
5687 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
5689 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
5690 goto need_kick_unlock;
5692 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
5693 && (cpumask_first_and(nohz.idle_cpus_mask,
5694 sched_domain_span(sd)) < cpu))
5695 goto need_kick_unlock;
5697 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
5709 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
5713 * run_rebalance_domains is triggered when needed from the scheduler tick.
5714 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
5716 static void run_rebalance_domains(struct softirq_action *h)
5718 int this_cpu = smp_processor_id();
5719 struct rq *this_rq = cpu_rq(this_cpu);
5720 enum cpu_idle_type idle = this_rq->idle_balance ?
5721 CPU_IDLE : CPU_NOT_IDLE;
5723 rebalance_domains(this_cpu, idle);
5726 * If this cpu has a pending nohz_balance_kick, then do the
5727 * balancing on behalf of the other idle cpus whose ticks are
5730 nohz_idle_balance(this_cpu, idle);
5733 static inline int on_null_domain(int cpu)
5735 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
5739 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
5741 void trigger_load_balance(struct rq *rq, int cpu)
5743 /* Don't need to rebalance while attached to NULL domain */
5744 if (time_after_eq(jiffies, rq->next_balance) &&
5745 likely(!on_null_domain(cpu)))
5746 raise_softirq(SCHED_SOFTIRQ);
5747 #ifdef CONFIG_NO_HZ_COMMON
5748 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
5749 nohz_balancer_kick(cpu);
5753 static void rq_online_fair(struct rq *rq)
5758 static void rq_offline_fair(struct rq *rq)
5762 /* Ensure any throttled groups are reachable by pick_next_task */
5763 unthrottle_offline_cfs_rqs(rq);
5766 #endif /* CONFIG_SMP */
5769 * scheduler tick hitting a task of our scheduling class:
5771 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
5773 struct cfs_rq *cfs_rq;
5774 struct sched_entity *se = &curr->se;
5776 for_each_sched_entity(se) {
5777 cfs_rq = cfs_rq_of(se);
5778 entity_tick(cfs_rq, se, queued);
5781 if (sched_feat_numa(NUMA))
5782 task_tick_numa(rq, curr);
5784 update_rq_runnable_avg(rq, 1);
5788 * called on fork with the child task as argument from the parent's context
5789 * - child not yet on the tasklist
5790 * - preemption disabled
5792 static void task_fork_fair(struct task_struct *p)
5794 struct cfs_rq *cfs_rq;
5795 struct sched_entity *se = &p->se, *curr;
5796 int this_cpu = smp_processor_id();
5797 struct rq *rq = this_rq();
5798 unsigned long flags;
5800 raw_spin_lock_irqsave(&rq->lock, flags);
5802 update_rq_clock(rq);
5804 cfs_rq = task_cfs_rq(current);
5805 curr = cfs_rq->curr;
5808 * Not only the cpu but also the task_group of the parent might have
5809 * been changed after parent->se.parent,cfs_rq were copied to
5810 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
5811 * of child point to valid ones.
5814 __set_task_cpu(p, this_cpu);
5817 update_curr(cfs_rq);
5820 se->vruntime = curr->vruntime;
5821 place_entity(cfs_rq, se, 1);
5823 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
5825 * Upon rescheduling, sched_class::put_prev_task() will place
5826 * 'current' within the tree based on its new key value.
5828 swap(curr->vruntime, se->vruntime);
5829 resched_task(rq->curr);
5832 se->vruntime -= cfs_rq->min_vruntime;
5834 raw_spin_unlock_irqrestore(&rq->lock, flags);
5838 * Priority of the task has changed. Check to see if we preempt
5842 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
5848 * Reschedule if we are currently running on this runqueue and
5849 * our priority decreased, or if we are not currently running on
5850 * this runqueue and our priority is higher than the current's
5852 if (rq->curr == p) {
5853 if (p->prio > oldprio)
5854 resched_task(rq->curr);
5856 check_preempt_curr(rq, p, 0);
5859 static void switched_from_fair(struct rq *rq, struct task_struct *p)
5861 struct sched_entity *se = &p->se;
5862 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5865 * Ensure the task's vruntime is normalized, so that when it's
5866 * switched back to the fair class the enqueue_entity(.flags=0) will
5867 * do the right thing.
5869 * If it's on_rq, then the dequeue_entity(.flags=0) will already
5870 * have normalized the vruntime, if it's !on_rq, then only when
5871 * the task is sleeping will it still have non-normalized vruntime.
5873 if (!p->on_rq && p->state != TASK_RUNNING) {
5875 * Fix up our vruntime so that the current sleep doesn't
5876 * cause 'unlimited' sleep bonus.
5878 place_entity(cfs_rq, se, 0);
5879 se->vruntime -= cfs_rq->min_vruntime;
5882 #if defined(CONFIG_FAIR_GROUP_SCHED) && defined(CONFIG_SMP)
5884 * Remove our load from contribution when we leave sched_fair
5885 * and ensure we don't carry in an old decay_count if we
5888 if (p->se.avg.decay_count) {
5889 struct cfs_rq *cfs_rq = cfs_rq_of(&p->se);
5890 __synchronize_entity_decay(&p->se);
5891 subtract_blocked_load_contrib(cfs_rq,
5892 p->se.avg.load_avg_contrib);
5898 * We switched to the sched_fair class.
5900 static void switched_to_fair(struct rq *rq, struct task_struct *p)
5906 * We were most likely switched from sched_rt, so
5907 * kick off the schedule if running, otherwise just see
5908 * if we can still preempt the current task.
5911 resched_task(rq->curr);
5913 check_preempt_curr(rq, p, 0);
5916 /* Account for a task changing its policy or group.
5918 * This routine is mostly called to set cfs_rq->curr field when a task
5919 * migrates between groups/classes.
5921 static void set_curr_task_fair(struct rq *rq)
5923 struct sched_entity *se = &rq->curr->se;
5925 for_each_sched_entity(se) {
5926 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5928 set_next_entity(cfs_rq, se);
5929 /* ensure bandwidth has been allocated on our new cfs_rq */
5930 account_cfs_rq_runtime(cfs_rq, 0);
5934 void init_cfs_rq(struct cfs_rq *cfs_rq)
5936 cfs_rq->tasks_timeline = RB_ROOT;
5937 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
5938 #ifndef CONFIG_64BIT
5939 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
5941 #if defined(CONFIG_FAIR_GROUP_SCHED) && defined(CONFIG_SMP)
5942 atomic64_set(&cfs_rq->decay_counter, 1);
5943 atomic64_set(&cfs_rq->removed_load, 0);
5947 #ifdef CONFIG_FAIR_GROUP_SCHED
5948 static void task_move_group_fair(struct task_struct *p, int on_rq)
5950 struct cfs_rq *cfs_rq;
5952 * If the task was not on the rq at the time of this cgroup movement
5953 * it must have been asleep, sleeping tasks keep their ->vruntime
5954 * absolute on their old rq until wakeup (needed for the fair sleeper
5955 * bonus in place_entity()).
5957 * If it was on the rq, we've just 'preempted' it, which does convert
5958 * ->vruntime to a relative base.
5960 * Make sure both cases convert their relative position when migrating
5961 * to another cgroup's rq. This does somewhat interfere with the
5962 * fair sleeper stuff for the first placement, but who cares.
5965 * When !on_rq, vruntime of the task has usually NOT been normalized.
5966 * But there are some cases where it has already been normalized:
5968 * - Moving a forked child which is waiting for being woken up by
5969 * wake_up_new_task().
5970 * - Moving a task which has been woken up by try_to_wake_up() and
5971 * waiting for actually being woken up by sched_ttwu_pending().
5973 * To prevent boost or penalty in the new cfs_rq caused by delta
5974 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
5976 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
5980 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
5981 set_task_rq(p, task_cpu(p));
5983 cfs_rq = cfs_rq_of(&p->se);
5984 p->se.vruntime += cfs_rq->min_vruntime;
5987 * migrate_task_rq_fair() will have removed our previous
5988 * contribution, but we must synchronize for ongoing future
5991 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
5992 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
5997 void free_fair_sched_group(struct task_group *tg)
6001 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
6003 for_each_possible_cpu(i) {
6005 kfree(tg->cfs_rq[i]);
6014 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6016 struct cfs_rq *cfs_rq;
6017 struct sched_entity *se;
6020 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
6023 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
6027 tg->shares = NICE_0_LOAD;
6029 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
6031 for_each_possible_cpu(i) {
6032 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
6033 GFP_KERNEL, cpu_to_node(i));
6037 se = kzalloc_node(sizeof(struct sched_entity),
6038 GFP_KERNEL, cpu_to_node(i));
6042 init_cfs_rq(cfs_rq);
6043 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
6054 void unregister_fair_sched_group(struct task_group *tg, int cpu)
6056 struct rq *rq = cpu_rq(cpu);
6057 unsigned long flags;
6060 * Only empty task groups can be destroyed; so we can speculatively
6061 * check on_list without danger of it being re-added.
6063 if (!tg->cfs_rq[cpu]->on_list)
6066 raw_spin_lock_irqsave(&rq->lock, flags);
6067 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
6068 raw_spin_unlock_irqrestore(&rq->lock, flags);
6071 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
6072 struct sched_entity *se, int cpu,
6073 struct sched_entity *parent)
6075 struct rq *rq = cpu_rq(cpu);
6079 init_cfs_rq_runtime(cfs_rq);
6081 tg->cfs_rq[cpu] = cfs_rq;
6084 /* se could be NULL for root_task_group */
6089 se->cfs_rq = &rq->cfs;
6091 se->cfs_rq = parent->my_q;
6094 /* guarantee group entities always have weight */
6095 update_load_set(&se->load, NICE_0_LOAD);
6096 se->parent = parent;
6099 static DEFINE_MUTEX(shares_mutex);
6101 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6104 unsigned long flags;
6107 * We can't change the weight of the root cgroup.
6112 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
6114 mutex_lock(&shares_mutex);
6115 if (tg->shares == shares)
6118 tg->shares = shares;
6119 for_each_possible_cpu(i) {
6120 struct rq *rq = cpu_rq(i);
6121 struct sched_entity *se;
6124 /* Propagate contribution to hierarchy */
6125 raw_spin_lock_irqsave(&rq->lock, flags);
6126 for_each_sched_entity(se)
6127 update_cfs_shares(group_cfs_rq(se));
6128 raw_spin_unlock_irqrestore(&rq->lock, flags);
6132 mutex_unlock(&shares_mutex);
6135 #else /* CONFIG_FAIR_GROUP_SCHED */
6137 void free_fair_sched_group(struct task_group *tg) { }
6139 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6144 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
6146 #endif /* CONFIG_FAIR_GROUP_SCHED */
6149 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
6151 struct sched_entity *se = &task->se;
6152 unsigned int rr_interval = 0;
6155 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6158 if (rq->cfs.load.weight)
6159 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
6165 * All the scheduling class methods:
6167 const struct sched_class fair_sched_class = {
6168 .next = &idle_sched_class,
6169 .enqueue_task = enqueue_task_fair,
6170 .dequeue_task = dequeue_task_fair,
6171 .yield_task = yield_task_fair,
6172 .yield_to_task = yield_to_task_fair,
6174 .check_preempt_curr = check_preempt_wakeup,
6176 .pick_next_task = pick_next_task_fair,
6177 .put_prev_task = put_prev_task_fair,
6180 .select_task_rq = select_task_rq_fair,
6181 #ifdef CONFIG_FAIR_GROUP_SCHED
6182 .migrate_task_rq = migrate_task_rq_fair,
6184 .rq_online = rq_online_fair,
6185 .rq_offline = rq_offline_fair,
6187 .task_waking = task_waking_fair,
6190 .set_curr_task = set_curr_task_fair,
6191 .task_tick = task_tick_fair,
6192 .task_fork = task_fork_fair,
6194 .prio_changed = prio_changed_fair,
6195 .switched_from = switched_from_fair,
6196 .switched_to = switched_to_fair,
6198 .get_rr_interval = get_rr_interval_fair,
6200 #ifdef CONFIG_FAIR_GROUP_SCHED
6201 .task_move_group = task_move_group_fair,
6205 #ifdef CONFIG_SCHED_DEBUG
6206 void print_cfs_stats(struct seq_file *m, int cpu)
6208 struct cfs_rq *cfs_rq;
6211 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
6212 print_cfs_rq(m, cpu, cfs_rq);
6217 __init void init_sched_fair_class(void)
6220 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
6222 #ifdef CONFIG_NO_HZ_COMMON
6223 nohz.next_balance = jiffies;
6224 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
6225 cpu_notifier(sched_ilb_notifier, 0);