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/task_work.h>
32 #include <trace/events/sched.h>
37 * Targeted preemption latency for CPU-bound tasks:
38 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
40 * NOTE: this latency value is not the same as the concept of
41 * 'timeslice length' - timeslices in CFS are of variable length
42 * and have no persistent notion like in traditional, time-slice
43 * based scheduling concepts.
45 * (to see the precise effective timeslice length of your workload,
46 * run vmstat and monitor the context-switches (cs) field)
48 unsigned int sysctl_sched_latency = 6000000ULL;
49 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
52 * The initial- and re-scaling of tunables is configurable
53 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
56 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
57 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
58 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
60 enum sched_tunable_scaling sysctl_sched_tunable_scaling
61 = SCHED_TUNABLESCALING_LOG;
64 * Minimal preemption granularity for CPU-bound tasks:
65 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
67 unsigned int sysctl_sched_min_granularity = 750000ULL;
68 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
71 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
73 static unsigned int sched_nr_latency = 8;
76 * After fork, child runs first. If set to 0 (default) then
77 * parent will (try to) run first.
79 unsigned int sysctl_sched_child_runs_first __read_mostly;
82 * SCHED_OTHER wake-up granularity.
83 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
85 * This option delays the preemption effects of decoupled workloads
86 * and reduces their over-scheduling. Synchronous workloads will still
87 * have immediate wakeup/sleep latencies.
89 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
90 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
92 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
95 * The exponential sliding window over which load is averaged for shares
99 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
101 #ifdef CONFIG_CFS_BANDWIDTH
103 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
104 * each time a cfs_rq requests quota.
106 * Note: in the case that the slice exceeds the runtime remaining (either due
107 * to consumption or the quota being specified to be smaller than the slice)
108 * we will always only issue the remaining available time.
110 * default: 5 msec, units: microseconds
112 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
116 * Increase the granularity value when there are more CPUs,
117 * because with more CPUs the 'effective latency' as visible
118 * to users decreases. But the relationship is not linear,
119 * so pick a second-best guess by going with the log2 of the
122 * This idea comes from the SD scheduler of Con Kolivas:
124 static int get_update_sysctl_factor(void)
126 unsigned int cpus = min_t(int, num_online_cpus(), 8);
129 switch (sysctl_sched_tunable_scaling) {
130 case SCHED_TUNABLESCALING_NONE:
133 case SCHED_TUNABLESCALING_LINEAR:
136 case SCHED_TUNABLESCALING_LOG:
138 factor = 1 + ilog2(cpus);
145 static void update_sysctl(void)
147 unsigned int factor = get_update_sysctl_factor();
149 #define SET_SYSCTL(name) \
150 (sysctl_##name = (factor) * normalized_sysctl_##name)
151 SET_SYSCTL(sched_min_granularity);
152 SET_SYSCTL(sched_latency);
153 SET_SYSCTL(sched_wakeup_granularity);
157 void sched_init_granularity(void)
162 #if BITS_PER_LONG == 32
163 # define WMULT_CONST (~0UL)
165 # define WMULT_CONST (1UL << 32)
168 #define WMULT_SHIFT 32
171 * Shift right and round:
173 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
176 * delta *= weight / lw
179 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
180 struct load_weight *lw)
185 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
186 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
187 * 2^SCHED_LOAD_RESOLUTION.
189 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
190 tmp = (u64)delta_exec * scale_load_down(weight);
192 tmp = (u64)delta_exec;
194 if (!lw->inv_weight) {
195 unsigned long w = scale_load_down(lw->weight);
197 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
199 else if (unlikely(!w))
200 lw->inv_weight = WMULT_CONST;
202 lw->inv_weight = WMULT_CONST / w;
206 * Check whether we'd overflow the 64-bit multiplication:
208 if (unlikely(tmp > WMULT_CONST))
209 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
212 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
214 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
218 const struct sched_class fair_sched_class;
220 /**************************************************************
221 * CFS operations on generic schedulable entities:
224 #ifdef CONFIG_FAIR_GROUP_SCHED
226 /* cpu runqueue to which this cfs_rq is attached */
227 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
232 /* An entity is a task if it doesn't "own" a runqueue */
233 #define entity_is_task(se) (!se->my_q)
235 static inline struct task_struct *task_of(struct sched_entity *se)
237 #ifdef CONFIG_SCHED_DEBUG
238 WARN_ON_ONCE(!entity_is_task(se));
240 return container_of(se, struct task_struct, se);
243 /* Walk up scheduling entities hierarchy */
244 #define for_each_sched_entity(se) \
245 for (; se; se = se->parent)
247 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
252 /* runqueue on which this entity is (to be) queued */
253 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
258 /* runqueue "owned" by this group */
259 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
264 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
266 if (!cfs_rq->on_list) {
268 * Ensure we either appear before our parent (if already
269 * enqueued) or force our parent to appear after us when it is
270 * enqueued. The fact that we always enqueue bottom-up
271 * reduces this to two cases.
273 if (cfs_rq->tg->parent &&
274 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
275 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
276 &rq_of(cfs_rq)->leaf_cfs_rq_list);
278 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
279 &rq_of(cfs_rq)->leaf_cfs_rq_list);
286 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
288 if (cfs_rq->on_list) {
289 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
294 /* Iterate thr' all leaf cfs_rq's on a runqueue */
295 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
296 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
298 /* Do the two (enqueued) entities belong to the same group ? */
300 is_same_group(struct sched_entity *se, struct sched_entity *pse)
302 if (se->cfs_rq == pse->cfs_rq)
308 static inline struct sched_entity *parent_entity(struct sched_entity *se)
313 /* return depth at which a sched entity is present in the hierarchy */
314 static inline int depth_se(struct sched_entity *se)
318 for_each_sched_entity(se)
325 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
327 int se_depth, pse_depth;
330 * preemption test can be made between sibling entities who are in the
331 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
332 * both tasks until we find their ancestors who are siblings of common
336 /* First walk up until both entities are at same depth */
337 se_depth = depth_se(*se);
338 pse_depth = depth_se(*pse);
340 while (se_depth > pse_depth) {
342 *se = parent_entity(*se);
345 while (pse_depth > se_depth) {
347 *pse = parent_entity(*pse);
350 while (!is_same_group(*se, *pse)) {
351 *se = parent_entity(*se);
352 *pse = parent_entity(*pse);
356 #else /* !CONFIG_FAIR_GROUP_SCHED */
358 static inline struct task_struct *task_of(struct sched_entity *se)
360 return container_of(se, struct task_struct, se);
363 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
365 return container_of(cfs_rq, struct rq, cfs);
368 #define entity_is_task(se) 1
370 #define for_each_sched_entity(se) \
371 for (; se; se = NULL)
373 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
375 return &task_rq(p)->cfs;
378 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
380 struct task_struct *p = task_of(se);
381 struct rq *rq = task_rq(p);
386 /* runqueue "owned" by this group */
387 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
392 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
396 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
400 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
401 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
404 is_same_group(struct sched_entity *se, struct sched_entity *pse)
409 static inline struct sched_entity *parent_entity(struct sched_entity *se)
415 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
419 #endif /* CONFIG_FAIR_GROUP_SCHED */
421 static __always_inline
422 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec);
424 /**************************************************************
425 * Scheduling class tree data structure manipulation methods:
428 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
430 s64 delta = (s64)(vruntime - min_vruntime);
432 min_vruntime = vruntime;
437 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
439 s64 delta = (s64)(vruntime - min_vruntime);
441 min_vruntime = vruntime;
446 static inline int entity_before(struct sched_entity *a,
447 struct sched_entity *b)
449 return (s64)(a->vruntime - b->vruntime) < 0;
452 static void update_min_vruntime(struct cfs_rq *cfs_rq)
454 u64 vruntime = cfs_rq->min_vruntime;
457 vruntime = cfs_rq->curr->vruntime;
459 if (cfs_rq->rb_leftmost) {
460 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
465 vruntime = se->vruntime;
467 vruntime = min_vruntime(vruntime, se->vruntime);
470 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
473 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
478 * Enqueue an entity into the rb-tree:
480 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
482 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
483 struct rb_node *parent = NULL;
484 struct sched_entity *entry;
488 * Find the right place in the rbtree:
492 entry = rb_entry(parent, struct sched_entity, run_node);
494 * We dont care about collisions. Nodes with
495 * the same key stay together.
497 if (entity_before(se, entry)) {
498 link = &parent->rb_left;
500 link = &parent->rb_right;
506 * Maintain a cache of leftmost tree entries (it is frequently
510 cfs_rq->rb_leftmost = &se->run_node;
512 rb_link_node(&se->run_node, parent, link);
513 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
516 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
518 if (cfs_rq->rb_leftmost == &se->run_node) {
519 struct rb_node *next_node;
521 next_node = rb_next(&se->run_node);
522 cfs_rq->rb_leftmost = next_node;
525 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
528 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
530 struct rb_node *left = cfs_rq->rb_leftmost;
535 return rb_entry(left, struct sched_entity, run_node);
538 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
540 struct rb_node *next = rb_next(&se->run_node);
545 return rb_entry(next, struct sched_entity, run_node);
548 #ifdef CONFIG_SCHED_DEBUG
549 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
551 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
556 return rb_entry(last, struct sched_entity, run_node);
559 /**************************************************************
560 * Scheduling class statistics methods:
563 int sched_proc_update_handler(struct ctl_table *table, int write,
564 void __user *buffer, size_t *lenp,
567 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
568 int factor = get_update_sysctl_factor();
573 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
574 sysctl_sched_min_granularity);
576 #define WRT_SYSCTL(name) \
577 (normalized_sysctl_##name = sysctl_##name / (factor))
578 WRT_SYSCTL(sched_min_granularity);
579 WRT_SYSCTL(sched_latency);
580 WRT_SYSCTL(sched_wakeup_granularity);
590 static inline unsigned long
591 calc_delta_fair(unsigned long delta, struct sched_entity *se)
593 if (unlikely(se->load.weight != NICE_0_LOAD))
594 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
600 * The idea is to set a period in which each task runs once.
602 * When there are too many tasks (sched_nr_latency) we have to stretch
603 * this period because otherwise the slices get too small.
605 * p = (nr <= nl) ? l : l*nr/nl
607 static u64 __sched_period(unsigned long nr_running)
609 u64 period = sysctl_sched_latency;
610 unsigned long nr_latency = sched_nr_latency;
612 if (unlikely(nr_running > nr_latency)) {
613 period = sysctl_sched_min_granularity;
614 period *= nr_running;
621 * We calculate the wall-time slice from the period by taking a part
622 * proportional to the weight.
626 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
628 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
630 for_each_sched_entity(se) {
631 struct load_weight *load;
632 struct load_weight lw;
634 cfs_rq = cfs_rq_of(se);
635 load = &cfs_rq->load;
637 if (unlikely(!se->on_rq)) {
640 update_load_add(&lw, se->load.weight);
643 slice = calc_delta_mine(slice, se->load.weight, load);
649 * We calculate the vruntime slice of a to be inserted task
653 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
655 return calc_delta_fair(sched_slice(cfs_rq, se), se);
658 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update);
659 static void update_cfs_shares(struct cfs_rq *cfs_rq);
662 * Update the current task's runtime statistics. Skip current tasks that
663 * are not in our scheduling class.
666 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
667 unsigned long delta_exec)
669 unsigned long delta_exec_weighted;
671 schedstat_set(curr->statistics.exec_max,
672 max((u64)delta_exec, curr->statistics.exec_max));
674 curr->sum_exec_runtime += delta_exec;
675 schedstat_add(cfs_rq, exec_clock, delta_exec);
676 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
678 curr->vruntime += delta_exec_weighted;
679 update_min_vruntime(cfs_rq);
681 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
682 cfs_rq->load_unacc_exec_time += delta_exec;
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: 5s
785 unsigned int sysctl_numa_balancing_scan_period_min = 5000;
786 unsigned int sysctl_numa_balancing_scan_period_max = 5000*16;
788 static void task_numa_placement(struct task_struct *p)
790 int seq = ACCESS_ONCE(p->mm->numa_scan_seq);
792 if (p->numa_scan_seq == seq)
794 p->numa_scan_seq = seq;
796 /* FIXME: Scheduling placement policy hints go here */
800 * Got a PROT_NONE fault for a page on @node.
802 void task_numa_fault(int node, int pages)
804 struct task_struct *p = current;
806 /* FIXME: Allocate task-specific structure for placement policy here */
808 task_numa_placement(p);
812 * The expensive part of numa migration is done from task_work context.
813 * Triggered from task_tick_numa().
815 void task_numa_work(struct callback_head *work)
817 unsigned long migrate, next_scan, now = jiffies;
818 struct task_struct *p = current;
819 struct mm_struct *mm = p->mm;
821 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
823 work->next = work; /* protect against double add */
825 * Who cares about NUMA placement when they're dying.
827 * NOTE: make sure not to dereference p->mm before this check,
828 * exit_task_work() happens _after_ exit_mm() so we could be called
829 * without p->mm even though we still had it when we enqueued this
832 if (p->flags & PF_EXITING)
836 * Enforce maximal scan/migration frequency..
838 migrate = mm->numa_next_scan;
839 if (time_before(now, migrate))
842 if (p->numa_scan_period == 0)
843 p->numa_scan_period = sysctl_numa_balancing_scan_period_min;
845 next_scan = now + 2*msecs_to_jiffies(p->numa_scan_period);
846 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
849 ACCESS_ONCE(mm->numa_scan_seq)++;
851 struct vm_area_struct *vma;
853 down_read(&mm->mmap_sem);
854 for (vma = mm->mmap; vma; vma = vma->vm_next) {
855 if (!vma_migratable(vma))
857 change_prot_numa(vma, vma->vm_start, vma->vm_end);
859 up_read(&mm->mmap_sem);
864 * Drive the periodic memory faults..
866 void task_tick_numa(struct rq *rq, struct task_struct *curr)
868 struct callback_head *work = &curr->numa_work;
872 * We don't care about NUMA placement if we don't have memory.
874 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
878 * Using runtime rather than walltime has the dual advantage that
879 * we (mostly) drive the selection from busy threads and that the
880 * task needs to have done some actual work before we bother with
883 now = curr->se.sum_exec_runtime;
884 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
886 if (now - curr->node_stamp > period) {
887 curr->node_stamp = now;
889 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
890 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
891 task_work_add(curr, work, true);
896 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
899 #endif /* CONFIG_NUMA_BALANCING */
902 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
904 update_load_add(&cfs_rq->load, se->load.weight);
905 if (!parent_entity(se))
906 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
908 if (entity_is_task(se))
909 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
911 cfs_rq->nr_running++;
915 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
917 update_load_sub(&cfs_rq->load, se->load.weight);
918 if (!parent_entity(se))
919 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
920 if (entity_is_task(se))
921 list_del_init(&se->group_node);
922 cfs_rq->nr_running--;
925 #ifdef CONFIG_FAIR_GROUP_SCHED
926 /* we need this in update_cfs_load and load-balance functions below */
927 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
929 static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq,
932 struct task_group *tg = cfs_rq->tg;
935 load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1);
936 load_avg -= cfs_rq->load_contribution;
938 if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) {
939 atomic_add(load_avg, &tg->load_weight);
940 cfs_rq->load_contribution += load_avg;
944 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
946 u64 period = sysctl_sched_shares_window;
948 unsigned long load = cfs_rq->load.weight;
950 if (cfs_rq->tg == &root_task_group || throttled_hierarchy(cfs_rq))
953 now = rq_of(cfs_rq)->clock_task;
954 delta = now - cfs_rq->load_stamp;
956 /* truncate load history at 4 idle periods */
957 if (cfs_rq->load_stamp > cfs_rq->load_last &&
958 now - cfs_rq->load_last > 4 * period) {
959 cfs_rq->load_period = 0;
960 cfs_rq->load_avg = 0;
964 cfs_rq->load_stamp = now;
965 cfs_rq->load_unacc_exec_time = 0;
966 cfs_rq->load_period += delta;
968 cfs_rq->load_last = now;
969 cfs_rq->load_avg += delta * load;
972 /* consider updating load contribution on each fold or truncate */
973 if (global_update || cfs_rq->load_period > period
974 || !cfs_rq->load_period)
975 update_cfs_rq_load_contribution(cfs_rq, global_update);
977 while (cfs_rq->load_period > period) {
979 * Inline assembly required to prevent the compiler
980 * optimising this loop into a divmod call.
981 * See __iter_div_u64_rem() for another example of this.
983 asm("" : "+rm" (cfs_rq->load_period));
984 cfs_rq->load_period /= 2;
985 cfs_rq->load_avg /= 2;
988 if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg)
989 list_del_leaf_cfs_rq(cfs_rq);
992 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
997 * Use this CPU's actual weight instead of the last load_contribution
998 * to gain a more accurate current total weight. See
999 * update_cfs_rq_load_contribution().
1001 tg_weight = atomic_read(&tg->load_weight);
1002 tg_weight -= cfs_rq->load_contribution;
1003 tg_weight += cfs_rq->load.weight;
1008 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1010 long tg_weight, load, shares;
1012 tg_weight = calc_tg_weight(tg, cfs_rq);
1013 load = cfs_rq->load.weight;
1015 shares = (tg->shares * load);
1017 shares /= tg_weight;
1019 if (shares < MIN_SHARES)
1020 shares = MIN_SHARES;
1021 if (shares > tg->shares)
1022 shares = tg->shares;
1027 static void update_entity_shares_tick(struct cfs_rq *cfs_rq)
1029 if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) {
1030 update_cfs_load(cfs_rq, 0);
1031 update_cfs_shares(cfs_rq);
1034 # else /* CONFIG_SMP */
1035 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
1039 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1044 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
1047 # endif /* CONFIG_SMP */
1048 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1049 unsigned long weight)
1052 /* commit outstanding execution time */
1053 if (cfs_rq->curr == se)
1054 update_curr(cfs_rq);
1055 account_entity_dequeue(cfs_rq, se);
1058 update_load_set(&se->load, weight);
1061 account_entity_enqueue(cfs_rq, se);
1064 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1066 struct task_group *tg;
1067 struct sched_entity *se;
1071 se = tg->se[cpu_of(rq_of(cfs_rq))];
1072 if (!se || throttled_hierarchy(cfs_rq))
1075 if (likely(se->load.weight == tg->shares))
1078 shares = calc_cfs_shares(cfs_rq, tg);
1080 reweight_entity(cfs_rq_of(se), se, shares);
1082 #else /* CONFIG_FAIR_GROUP_SCHED */
1083 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
1087 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1091 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
1094 #endif /* CONFIG_FAIR_GROUP_SCHED */
1096 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1098 #ifdef CONFIG_SCHEDSTATS
1099 struct task_struct *tsk = NULL;
1101 if (entity_is_task(se))
1104 if (se->statistics.sleep_start) {
1105 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
1110 if (unlikely(delta > se->statistics.sleep_max))
1111 se->statistics.sleep_max = delta;
1113 se->statistics.sleep_start = 0;
1114 se->statistics.sum_sleep_runtime += delta;
1117 account_scheduler_latency(tsk, delta >> 10, 1);
1118 trace_sched_stat_sleep(tsk, delta);
1121 if (se->statistics.block_start) {
1122 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
1127 if (unlikely(delta > se->statistics.block_max))
1128 se->statistics.block_max = delta;
1130 se->statistics.block_start = 0;
1131 se->statistics.sum_sleep_runtime += delta;
1134 if (tsk->in_iowait) {
1135 se->statistics.iowait_sum += delta;
1136 se->statistics.iowait_count++;
1137 trace_sched_stat_iowait(tsk, delta);
1140 trace_sched_stat_blocked(tsk, delta);
1143 * Blocking time is in units of nanosecs, so shift by
1144 * 20 to get a milliseconds-range estimation of the
1145 * amount of time that the task spent sleeping:
1147 if (unlikely(prof_on == SLEEP_PROFILING)) {
1148 profile_hits(SLEEP_PROFILING,
1149 (void *)get_wchan(tsk),
1152 account_scheduler_latency(tsk, delta >> 10, 0);
1158 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1160 #ifdef CONFIG_SCHED_DEBUG
1161 s64 d = se->vruntime - cfs_rq->min_vruntime;
1166 if (d > 3*sysctl_sched_latency)
1167 schedstat_inc(cfs_rq, nr_spread_over);
1172 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1174 u64 vruntime = cfs_rq->min_vruntime;
1177 * The 'current' period is already promised to the current tasks,
1178 * however the extra weight of the new task will slow them down a
1179 * little, place the new task so that it fits in the slot that
1180 * stays open at the end.
1182 if (initial && sched_feat(START_DEBIT))
1183 vruntime += sched_vslice(cfs_rq, se);
1185 /* sleeps up to a single latency don't count. */
1187 unsigned long thresh = sysctl_sched_latency;
1190 * Halve their sleep time's effect, to allow
1191 * for a gentler effect of sleepers:
1193 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1199 /* ensure we never gain time by being placed backwards. */
1200 vruntime = max_vruntime(se->vruntime, vruntime);
1202 se->vruntime = vruntime;
1205 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1208 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1211 * Update the normalized vruntime before updating min_vruntime
1212 * through callig update_curr().
1214 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1215 se->vruntime += cfs_rq->min_vruntime;
1218 * Update run-time statistics of the 'current'.
1220 update_curr(cfs_rq);
1221 update_cfs_load(cfs_rq, 0);
1222 account_entity_enqueue(cfs_rq, se);
1223 update_cfs_shares(cfs_rq);
1225 if (flags & ENQUEUE_WAKEUP) {
1226 place_entity(cfs_rq, se, 0);
1227 enqueue_sleeper(cfs_rq, se);
1230 update_stats_enqueue(cfs_rq, se);
1231 check_spread(cfs_rq, se);
1232 if (se != cfs_rq->curr)
1233 __enqueue_entity(cfs_rq, se);
1236 if (cfs_rq->nr_running == 1) {
1237 list_add_leaf_cfs_rq(cfs_rq);
1238 check_enqueue_throttle(cfs_rq);
1242 static void __clear_buddies_last(struct sched_entity *se)
1244 for_each_sched_entity(se) {
1245 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1246 if (cfs_rq->last == se)
1247 cfs_rq->last = NULL;
1253 static void __clear_buddies_next(struct sched_entity *se)
1255 for_each_sched_entity(se) {
1256 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1257 if (cfs_rq->next == se)
1258 cfs_rq->next = NULL;
1264 static void __clear_buddies_skip(struct sched_entity *se)
1266 for_each_sched_entity(se) {
1267 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1268 if (cfs_rq->skip == se)
1269 cfs_rq->skip = NULL;
1275 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1277 if (cfs_rq->last == se)
1278 __clear_buddies_last(se);
1280 if (cfs_rq->next == se)
1281 __clear_buddies_next(se);
1283 if (cfs_rq->skip == se)
1284 __clear_buddies_skip(se);
1287 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1290 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1293 * Update run-time statistics of the 'current'.
1295 update_curr(cfs_rq);
1297 update_stats_dequeue(cfs_rq, se);
1298 if (flags & DEQUEUE_SLEEP) {
1299 #ifdef CONFIG_SCHEDSTATS
1300 if (entity_is_task(se)) {
1301 struct task_struct *tsk = task_of(se);
1303 if (tsk->state & TASK_INTERRUPTIBLE)
1304 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1305 if (tsk->state & TASK_UNINTERRUPTIBLE)
1306 se->statistics.block_start = rq_of(cfs_rq)->clock;
1311 clear_buddies(cfs_rq, se);
1313 if (se != cfs_rq->curr)
1314 __dequeue_entity(cfs_rq, se);
1316 update_cfs_load(cfs_rq, 0);
1317 account_entity_dequeue(cfs_rq, se);
1320 * Normalize the entity after updating the min_vruntime because the
1321 * update can refer to the ->curr item and we need to reflect this
1322 * movement in our normalized position.
1324 if (!(flags & DEQUEUE_SLEEP))
1325 se->vruntime -= cfs_rq->min_vruntime;
1327 /* return excess runtime on last dequeue */
1328 return_cfs_rq_runtime(cfs_rq);
1330 update_min_vruntime(cfs_rq);
1331 update_cfs_shares(cfs_rq);
1335 * Preempt the current task with a newly woken task if needed:
1338 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1340 unsigned long ideal_runtime, delta_exec;
1341 struct sched_entity *se;
1344 ideal_runtime = sched_slice(cfs_rq, curr);
1345 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1346 if (delta_exec > ideal_runtime) {
1347 resched_task(rq_of(cfs_rq)->curr);
1349 * The current task ran long enough, ensure it doesn't get
1350 * re-elected due to buddy favours.
1352 clear_buddies(cfs_rq, curr);
1357 * Ensure that a task that missed wakeup preemption by a
1358 * narrow margin doesn't have to wait for a full slice.
1359 * This also mitigates buddy induced latencies under load.
1361 if (delta_exec < sysctl_sched_min_granularity)
1364 se = __pick_first_entity(cfs_rq);
1365 delta = curr->vruntime - se->vruntime;
1370 if (delta > ideal_runtime)
1371 resched_task(rq_of(cfs_rq)->curr);
1375 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1377 /* 'current' is not kept within the tree. */
1380 * Any task has to be enqueued before it get to execute on
1381 * a CPU. So account for the time it spent waiting on the
1384 update_stats_wait_end(cfs_rq, se);
1385 __dequeue_entity(cfs_rq, se);
1388 update_stats_curr_start(cfs_rq, se);
1390 #ifdef CONFIG_SCHEDSTATS
1392 * Track our maximum slice length, if the CPU's load is at
1393 * least twice that of our own weight (i.e. dont track it
1394 * when there are only lesser-weight tasks around):
1396 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1397 se->statistics.slice_max = max(se->statistics.slice_max,
1398 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1401 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1405 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1408 * Pick the next process, keeping these things in mind, in this order:
1409 * 1) keep things fair between processes/task groups
1410 * 2) pick the "next" process, since someone really wants that to run
1411 * 3) pick the "last" process, for cache locality
1412 * 4) do not run the "skip" process, if something else is available
1414 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1416 struct sched_entity *se = __pick_first_entity(cfs_rq);
1417 struct sched_entity *left = se;
1420 * Avoid running the skip buddy, if running something else can
1421 * be done without getting too unfair.
1423 if (cfs_rq->skip == se) {
1424 struct sched_entity *second = __pick_next_entity(se);
1425 if (second && wakeup_preempt_entity(second, left) < 1)
1430 * Prefer last buddy, try to return the CPU to a preempted task.
1432 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1436 * Someone really wants this to run. If it's not unfair, run it.
1438 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1441 clear_buddies(cfs_rq, se);
1446 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1448 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1451 * If still on the runqueue then deactivate_task()
1452 * was not called and update_curr() has to be done:
1455 update_curr(cfs_rq);
1457 /* throttle cfs_rqs exceeding runtime */
1458 check_cfs_rq_runtime(cfs_rq);
1460 check_spread(cfs_rq, prev);
1462 update_stats_wait_start(cfs_rq, prev);
1463 /* Put 'current' back into the tree. */
1464 __enqueue_entity(cfs_rq, prev);
1466 cfs_rq->curr = NULL;
1470 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1473 * Update run-time statistics of the 'current'.
1475 update_curr(cfs_rq);
1478 * Update share accounting for long-running entities.
1480 update_entity_shares_tick(cfs_rq);
1482 #ifdef CONFIG_SCHED_HRTICK
1484 * queued ticks are scheduled to match the slice, so don't bother
1485 * validating it and just reschedule.
1488 resched_task(rq_of(cfs_rq)->curr);
1492 * don't let the period tick interfere with the hrtick preemption
1494 if (!sched_feat(DOUBLE_TICK) &&
1495 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1499 if (cfs_rq->nr_running > 1)
1500 check_preempt_tick(cfs_rq, curr);
1504 /**************************************************
1505 * CFS bandwidth control machinery
1508 #ifdef CONFIG_CFS_BANDWIDTH
1510 #ifdef HAVE_JUMP_LABEL
1511 static struct static_key __cfs_bandwidth_used;
1513 static inline bool cfs_bandwidth_used(void)
1515 return static_key_false(&__cfs_bandwidth_used);
1518 void account_cfs_bandwidth_used(int enabled, int was_enabled)
1520 /* only need to count groups transitioning between enabled/!enabled */
1521 if (enabled && !was_enabled)
1522 static_key_slow_inc(&__cfs_bandwidth_used);
1523 else if (!enabled && was_enabled)
1524 static_key_slow_dec(&__cfs_bandwidth_used);
1526 #else /* HAVE_JUMP_LABEL */
1527 static bool cfs_bandwidth_used(void)
1532 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
1533 #endif /* HAVE_JUMP_LABEL */
1536 * default period for cfs group bandwidth.
1537 * default: 0.1s, units: nanoseconds
1539 static inline u64 default_cfs_period(void)
1541 return 100000000ULL;
1544 static inline u64 sched_cfs_bandwidth_slice(void)
1546 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
1550 * Replenish runtime according to assigned quota and update expiration time.
1551 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
1552 * additional synchronization around rq->lock.
1554 * requires cfs_b->lock
1556 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
1560 if (cfs_b->quota == RUNTIME_INF)
1563 now = sched_clock_cpu(smp_processor_id());
1564 cfs_b->runtime = cfs_b->quota;
1565 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
1568 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
1570 return &tg->cfs_bandwidth;
1573 /* returns 0 on failure to allocate runtime */
1574 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1576 struct task_group *tg = cfs_rq->tg;
1577 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
1578 u64 amount = 0, min_amount, expires;
1580 /* note: this is a positive sum as runtime_remaining <= 0 */
1581 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
1583 raw_spin_lock(&cfs_b->lock);
1584 if (cfs_b->quota == RUNTIME_INF)
1585 amount = min_amount;
1588 * If the bandwidth pool has become inactive, then at least one
1589 * period must have elapsed since the last consumption.
1590 * Refresh the global state and ensure bandwidth timer becomes
1593 if (!cfs_b->timer_active) {
1594 __refill_cfs_bandwidth_runtime(cfs_b);
1595 __start_cfs_bandwidth(cfs_b);
1598 if (cfs_b->runtime > 0) {
1599 amount = min(cfs_b->runtime, min_amount);
1600 cfs_b->runtime -= amount;
1604 expires = cfs_b->runtime_expires;
1605 raw_spin_unlock(&cfs_b->lock);
1607 cfs_rq->runtime_remaining += amount;
1609 * we may have advanced our local expiration to account for allowed
1610 * spread between our sched_clock and the one on which runtime was
1613 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
1614 cfs_rq->runtime_expires = expires;
1616 return cfs_rq->runtime_remaining > 0;
1620 * Note: This depends on the synchronization provided by sched_clock and the
1621 * fact that rq->clock snapshots this value.
1623 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1625 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1626 struct rq *rq = rq_of(cfs_rq);
1628 /* if the deadline is ahead of our clock, nothing to do */
1629 if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0))
1632 if (cfs_rq->runtime_remaining < 0)
1636 * If the local deadline has passed we have to consider the
1637 * possibility that our sched_clock is 'fast' and the global deadline
1638 * has not truly expired.
1640 * Fortunately we can check determine whether this the case by checking
1641 * whether the global deadline has advanced.
1644 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
1645 /* extend local deadline, drift is bounded above by 2 ticks */
1646 cfs_rq->runtime_expires += TICK_NSEC;
1648 /* global deadline is ahead, expiration has passed */
1649 cfs_rq->runtime_remaining = 0;
1653 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1654 unsigned long delta_exec)
1656 /* dock delta_exec before expiring quota (as it could span periods) */
1657 cfs_rq->runtime_remaining -= delta_exec;
1658 expire_cfs_rq_runtime(cfs_rq);
1660 if (likely(cfs_rq->runtime_remaining > 0))
1664 * if we're unable to extend our runtime we resched so that the active
1665 * hierarchy can be throttled
1667 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
1668 resched_task(rq_of(cfs_rq)->curr);
1671 static __always_inline
1672 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
1674 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
1677 __account_cfs_rq_runtime(cfs_rq, delta_exec);
1680 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
1682 return cfs_bandwidth_used() && cfs_rq->throttled;
1685 /* check whether cfs_rq, or any parent, is throttled */
1686 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
1688 return cfs_bandwidth_used() && cfs_rq->throttle_count;
1692 * Ensure that neither of the group entities corresponding to src_cpu or
1693 * dest_cpu are members of a throttled hierarchy when performing group
1694 * load-balance operations.
1696 static inline int throttled_lb_pair(struct task_group *tg,
1697 int src_cpu, int dest_cpu)
1699 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
1701 src_cfs_rq = tg->cfs_rq[src_cpu];
1702 dest_cfs_rq = tg->cfs_rq[dest_cpu];
1704 return throttled_hierarchy(src_cfs_rq) ||
1705 throttled_hierarchy(dest_cfs_rq);
1708 /* updated child weight may affect parent so we have to do this bottom up */
1709 static int tg_unthrottle_up(struct task_group *tg, void *data)
1711 struct rq *rq = data;
1712 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1714 cfs_rq->throttle_count--;
1716 if (!cfs_rq->throttle_count) {
1717 u64 delta = rq->clock_task - cfs_rq->load_stamp;
1719 /* leaving throttled state, advance shares averaging windows */
1720 cfs_rq->load_stamp += delta;
1721 cfs_rq->load_last += delta;
1723 /* update entity weight now that we are on_rq again */
1724 update_cfs_shares(cfs_rq);
1731 static int tg_throttle_down(struct task_group *tg, void *data)
1733 struct rq *rq = data;
1734 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1736 /* group is entering throttled state, record last load */
1737 if (!cfs_rq->throttle_count)
1738 update_cfs_load(cfs_rq, 0);
1739 cfs_rq->throttle_count++;
1744 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
1746 struct rq *rq = rq_of(cfs_rq);
1747 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1748 struct sched_entity *se;
1749 long task_delta, dequeue = 1;
1751 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1753 /* account load preceding throttle */
1755 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
1758 task_delta = cfs_rq->h_nr_running;
1759 for_each_sched_entity(se) {
1760 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
1761 /* throttled entity or throttle-on-deactivate */
1766 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
1767 qcfs_rq->h_nr_running -= task_delta;
1769 if (qcfs_rq->load.weight)
1774 rq->nr_running -= task_delta;
1776 cfs_rq->throttled = 1;
1777 cfs_rq->throttled_timestamp = rq->clock;
1778 raw_spin_lock(&cfs_b->lock);
1779 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
1780 raw_spin_unlock(&cfs_b->lock);
1783 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
1785 struct rq *rq = rq_of(cfs_rq);
1786 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1787 struct sched_entity *se;
1791 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1793 cfs_rq->throttled = 0;
1794 raw_spin_lock(&cfs_b->lock);
1795 cfs_b->throttled_time += rq->clock - cfs_rq->throttled_timestamp;
1796 list_del_rcu(&cfs_rq->throttled_list);
1797 raw_spin_unlock(&cfs_b->lock);
1798 cfs_rq->throttled_timestamp = 0;
1800 update_rq_clock(rq);
1801 /* update hierarchical throttle state */
1802 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
1804 if (!cfs_rq->load.weight)
1807 task_delta = cfs_rq->h_nr_running;
1808 for_each_sched_entity(se) {
1812 cfs_rq = cfs_rq_of(se);
1814 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
1815 cfs_rq->h_nr_running += task_delta;
1817 if (cfs_rq_throttled(cfs_rq))
1822 rq->nr_running += task_delta;
1824 /* determine whether we need to wake up potentially idle cpu */
1825 if (rq->curr == rq->idle && rq->cfs.nr_running)
1826 resched_task(rq->curr);
1829 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
1830 u64 remaining, u64 expires)
1832 struct cfs_rq *cfs_rq;
1833 u64 runtime = remaining;
1836 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
1838 struct rq *rq = rq_of(cfs_rq);
1840 raw_spin_lock(&rq->lock);
1841 if (!cfs_rq_throttled(cfs_rq))
1844 runtime = -cfs_rq->runtime_remaining + 1;
1845 if (runtime > remaining)
1846 runtime = remaining;
1847 remaining -= runtime;
1849 cfs_rq->runtime_remaining += runtime;
1850 cfs_rq->runtime_expires = expires;
1852 /* we check whether we're throttled above */
1853 if (cfs_rq->runtime_remaining > 0)
1854 unthrottle_cfs_rq(cfs_rq);
1857 raw_spin_unlock(&rq->lock);
1868 * Responsible for refilling a task_group's bandwidth and unthrottling its
1869 * cfs_rqs as appropriate. If there has been no activity within the last
1870 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
1871 * used to track this state.
1873 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
1875 u64 runtime, runtime_expires;
1876 int idle = 1, throttled;
1878 raw_spin_lock(&cfs_b->lock);
1879 /* no need to continue the timer with no bandwidth constraint */
1880 if (cfs_b->quota == RUNTIME_INF)
1883 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
1884 /* idle depends on !throttled (for the case of a large deficit) */
1885 idle = cfs_b->idle && !throttled;
1886 cfs_b->nr_periods += overrun;
1888 /* if we're going inactive then everything else can be deferred */
1892 __refill_cfs_bandwidth_runtime(cfs_b);
1895 /* mark as potentially idle for the upcoming period */
1900 /* account preceding periods in which throttling occurred */
1901 cfs_b->nr_throttled += overrun;
1904 * There are throttled entities so we must first use the new bandwidth
1905 * to unthrottle them before making it generally available. This
1906 * ensures that all existing debts will be paid before a new cfs_rq is
1909 runtime = cfs_b->runtime;
1910 runtime_expires = cfs_b->runtime_expires;
1914 * This check is repeated as we are holding onto the new bandwidth
1915 * while we unthrottle. This can potentially race with an unthrottled
1916 * group trying to acquire new bandwidth from the global pool.
1918 while (throttled && runtime > 0) {
1919 raw_spin_unlock(&cfs_b->lock);
1920 /* we can't nest cfs_b->lock while distributing bandwidth */
1921 runtime = distribute_cfs_runtime(cfs_b, runtime,
1923 raw_spin_lock(&cfs_b->lock);
1925 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
1928 /* return (any) remaining runtime */
1929 cfs_b->runtime = runtime;
1931 * While we are ensured activity in the period following an
1932 * unthrottle, this also covers the case in which the new bandwidth is
1933 * insufficient to cover the existing bandwidth deficit. (Forcing the
1934 * timer to remain active while there are any throttled entities.)
1939 cfs_b->timer_active = 0;
1940 raw_spin_unlock(&cfs_b->lock);
1945 /* a cfs_rq won't donate quota below this amount */
1946 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
1947 /* minimum remaining period time to redistribute slack quota */
1948 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
1949 /* how long we wait to gather additional slack before distributing */
1950 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
1952 /* are we near the end of the current quota period? */
1953 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
1955 struct hrtimer *refresh_timer = &cfs_b->period_timer;
1958 /* if the call-back is running a quota refresh is already occurring */
1959 if (hrtimer_callback_running(refresh_timer))
1962 /* is a quota refresh about to occur? */
1963 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
1964 if (remaining < min_expire)
1970 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
1972 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
1974 /* if there's a quota refresh soon don't bother with slack */
1975 if (runtime_refresh_within(cfs_b, min_left))
1978 start_bandwidth_timer(&cfs_b->slack_timer,
1979 ns_to_ktime(cfs_bandwidth_slack_period));
1982 /* we know any runtime found here is valid as update_curr() precedes return */
1983 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1985 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1986 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
1988 if (slack_runtime <= 0)
1991 raw_spin_lock(&cfs_b->lock);
1992 if (cfs_b->quota != RUNTIME_INF &&
1993 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
1994 cfs_b->runtime += slack_runtime;
1996 /* we are under rq->lock, defer unthrottling using a timer */
1997 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
1998 !list_empty(&cfs_b->throttled_cfs_rq))
1999 start_cfs_slack_bandwidth(cfs_b);
2001 raw_spin_unlock(&cfs_b->lock);
2003 /* even if it's not valid for return we don't want to try again */
2004 cfs_rq->runtime_remaining -= slack_runtime;
2007 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2009 if (!cfs_bandwidth_used())
2012 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2015 __return_cfs_rq_runtime(cfs_rq);
2019 * This is done with a timer (instead of inline with bandwidth return) since
2020 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2022 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2024 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2027 /* confirm we're still not at a refresh boundary */
2028 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2031 raw_spin_lock(&cfs_b->lock);
2032 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2033 runtime = cfs_b->runtime;
2036 expires = cfs_b->runtime_expires;
2037 raw_spin_unlock(&cfs_b->lock);
2042 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2044 raw_spin_lock(&cfs_b->lock);
2045 if (expires == cfs_b->runtime_expires)
2046 cfs_b->runtime = runtime;
2047 raw_spin_unlock(&cfs_b->lock);
2051 * When a group wakes up we want to make sure that its quota is not already
2052 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2053 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2055 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2057 if (!cfs_bandwidth_used())
2060 /* an active group must be handled by the update_curr()->put() path */
2061 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2064 /* ensure the group is not already throttled */
2065 if (cfs_rq_throttled(cfs_rq))
2068 /* update runtime allocation */
2069 account_cfs_rq_runtime(cfs_rq, 0);
2070 if (cfs_rq->runtime_remaining <= 0)
2071 throttle_cfs_rq(cfs_rq);
2074 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2075 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2077 if (!cfs_bandwidth_used())
2080 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2084 * it's possible for a throttled entity to be forced into a running
2085 * state (e.g. set_curr_task), in this case we're finished.
2087 if (cfs_rq_throttled(cfs_rq))
2090 throttle_cfs_rq(cfs_rq);
2093 static inline u64 default_cfs_period(void);
2094 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun);
2095 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b);
2097 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2099 struct cfs_bandwidth *cfs_b =
2100 container_of(timer, struct cfs_bandwidth, slack_timer);
2101 do_sched_cfs_slack_timer(cfs_b);
2103 return HRTIMER_NORESTART;
2106 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2108 struct cfs_bandwidth *cfs_b =
2109 container_of(timer, struct cfs_bandwidth, period_timer);
2115 now = hrtimer_cb_get_time(timer);
2116 overrun = hrtimer_forward(timer, now, cfs_b->period);
2121 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2124 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2127 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2129 raw_spin_lock_init(&cfs_b->lock);
2131 cfs_b->quota = RUNTIME_INF;
2132 cfs_b->period = ns_to_ktime(default_cfs_period());
2134 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2135 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2136 cfs_b->period_timer.function = sched_cfs_period_timer;
2137 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2138 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2141 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2143 cfs_rq->runtime_enabled = 0;
2144 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2147 /* requires cfs_b->lock, may release to reprogram timer */
2148 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2151 * The timer may be active because we're trying to set a new bandwidth
2152 * period or because we're racing with the tear-down path
2153 * (timer_active==0 becomes visible before the hrtimer call-back
2154 * terminates). In either case we ensure that it's re-programmed
2156 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2157 raw_spin_unlock(&cfs_b->lock);
2158 /* ensure cfs_b->lock is available while we wait */
2159 hrtimer_cancel(&cfs_b->period_timer);
2161 raw_spin_lock(&cfs_b->lock);
2162 /* if someone else restarted the timer then we're done */
2163 if (cfs_b->timer_active)
2167 cfs_b->timer_active = 1;
2168 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2171 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2173 hrtimer_cancel(&cfs_b->period_timer);
2174 hrtimer_cancel(&cfs_b->slack_timer);
2177 static void unthrottle_offline_cfs_rqs(struct rq *rq)
2179 struct cfs_rq *cfs_rq;
2181 for_each_leaf_cfs_rq(rq, cfs_rq) {
2182 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2184 if (!cfs_rq->runtime_enabled)
2188 * clock_task is not advancing so we just need to make sure
2189 * there's some valid quota amount
2191 cfs_rq->runtime_remaining = cfs_b->quota;
2192 if (cfs_rq_throttled(cfs_rq))
2193 unthrottle_cfs_rq(cfs_rq);
2197 #else /* CONFIG_CFS_BANDWIDTH */
2198 static __always_inline
2199 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec) {}
2200 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2201 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2202 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2204 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2209 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2214 static inline int throttled_lb_pair(struct task_group *tg,
2215 int src_cpu, int dest_cpu)
2220 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2222 #ifdef CONFIG_FAIR_GROUP_SCHED
2223 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2226 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2230 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2231 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2233 #endif /* CONFIG_CFS_BANDWIDTH */
2235 /**************************************************
2236 * CFS operations on tasks:
2239 #ifdef CONFIG_SCHED_HRTICK
2240 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2242 struct sched_entity *se = &p->se;
2243 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2245 WARN_ON(task_rq(p) != rq);
2247 if (cfs_rq->nr_running > 1) {
2248 u64 slice = sched_slice(cfs_rq, se);
2249 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2250 s64 delta = slice - ran;
2259 * Don't schedule slices shorter than 10000ns, that just
2260 * doesn't make sense. Rely on vruntime for fairness.
2263 delta = max_t(s64, 10000LL, delta);
2265 hrtick_start(rq, delta);
2270 * called from enqueue/dequeue and updates the hrtick when the
2271 * current task is from our class and nr_running is low enough
2274 static void hrtick_update(struct rq *rq)
2276 struct task_struct *curr = rq->curr;
2278 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2281 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2282 hrtick_start_fair(rq, curr);
2284 #else /* !CONFIG_SCHED_HRTICK */
2286 hrtick_start_fair(struct rq *rq, struct task_struct *p)
2290 static inline void hrtick_update(struct rq *rq)
2296 * The enqueue_task method is called before nr_running is
2297 * increased. Here we update the fair scheduling stats and
2298 * then put the task into the rbtree:
2301 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2303 struct cfs_rq *cfs_rq;
2304 struct sched_entity *se = &p->se;
2306 for_each_sched_entity(se) {
2309 cfs_rq = cfs_rq_of(se);
2310 enqueue_entity(cfs_rq, se, flags);
2313 * end evaluation on encountering a throttled cfs_rq
2315 * note: in the case of encountering a throttled cfs_rq we will
2316 * post the final h_nr_running increment below.
2318 if (cfs_rq_throttled(cfs_rq))
2320 cfs_rq->h_nr_running++;
2322 flags = ENQUEUE_WAKEUP;
2325 for_each_sched_entity(se) {
2326 cfs_rq = cfs_rq_of(se);
2327 cfs_rq->h_nr_running++;
2329 if (cfs_rq_throttled(cfs_rq))
2332 update_cfs_load(cfs_rq, 0);
2333 update_cfs_shares(cfs_rq);
2341 static void set_next_buddy(struct sched_entity *se);
2344 * The dequeue_task method is called before nr_running is
2345 * decreased. We remove the task from the rbtree and
2346 * update the fair scheduling stats:
2348 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2350 struct cfs_rq *cfs_rq;
2351 struct sched_entity *se = &p->se;
2352 int task_sleep = flags & DEQUEUE_SLEEP;
2354 for_each_sched_entity(se) {
2355 cfs_rq = cfs_rq_of(se);
2356 dequeue_entity(cfs_rq, se, flags);
2359 * end evaluation on encountering a throttled cfs_rq
2361 * note: in the case of encountering a throttled cfs_rq we will
2362 * post the final h_nr_running decrement below.
2364 if (cfs_rq_throttled(cfs_rq))
2366 cfs_rq->h_nr_running--;
2368 /* Don't dequeue parent if it has other entities besides us */
2369 if (cfs_rq->load.weight) {
2371 * Bias pick_next to pick a task from this cfs_rq, as
2372 * p is sleeping when it is within its sched_slice.
2374 if (task_sleep && parent_entity(se))
2375 set_next_buddy(parent_entity(se));
2377 /* avoid re-evaluating load for this entity */
2378 se = parent_entity(se);
2381 flags |= DEQUEUE_SLEEP;
2384 for_each_sched_entity(se) {
2385 cfs_rq = cfs_rq_of(se);
2386 cfs_rq->h_nr_running--;
2388 if (cfs_rq_throttled(cfs_rq))
2391 update_cfs_load(cfs_rq, 0);
2392 update_cfs_shares(cfs_rq);
2401 /* Used instead of source_load when we know the type == 0 */
2402 static unsigned long weighted_cpuload(const int cpu)
2404 return cpu_rq(cpu)->load.weight;
2408 * Return a low guess at the load of a migration-source cpu weighted
2409 * according to the scheduling class and "nice" value.
2411 * We want to under-estimate the load of migration sources, to
2412 * balance conservatively.
2414 static unsigned long source_load(int cpu, int type)
2416 struct rq *rq = cpu_rq(cpu);
2417 unsigned long total = weighted_cpuload(cpu);
2419 if (type == 0 || !sched_feat(LB_BIAS))
2422 return min(rq->cpu_load[type-1], total);
2426 * Return a high guess at the load of a migration-target cpu weighted
2427 * according to the scheduling class and "nice" value.
2429 static unsigned long target_load(int cpu, int type)
2431 struct rq *rq = cpu_rq(cpu);
2432 unsigned long total = weighted_cpuload(cpu);
2434 if (type == 0 || !sched_feat(LB_BIAS))
2437 return max(rq->cpu_load[type-1], total);
2440 static unsigned long power_of(int cpu)
2442 return cpu_rq(cpu)->cpu_power;
2445 static unsigned long cpu_avg_load_per_task(int cpu)
2447 struct rq *rq = cpu_rq(cpu);
2448 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
2451 return rq->load.weight / nr_running;
2457 static void task_waking_fair(struct task_struct *p)
2459 struct sched_entity *se = &p->se;
2460 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2463 #ifndef CONFIG_64BIT
2464 u64 min_vruntime_copy;
2467 min_vruntime_copy = cfs_rq->min_vruntime_copy;
2469 min_vruntime = cfs_rq->min_vruntime;
2470 } while (min_vruntime != min_vruntime_copy);
2472 min_vruntime = cfs_rq->min_vruntime;
2475 se->vruntime -= min_vruntime;
2478 #ifdef CONFIG_FAIR_GROUP_SCHED
2480 * effective_load() calculates the load change as seen from the root_task_group
2482 * Adding load to a group doesn't make a group heavier, but can cause movement
2483 * of group shares between cpus. Assuming the shares were perfectly aligned one
2484 * can calculate the shift in shares.
2486 * Calculate the effective load difference if @wl is added (subtracted) to @tg
2487 * on this @cpu and results in a total addition (subtraction) of @wg to the
2488 * total group weight.
2490 * Given a runqueue weight distribution (rw_i) we can compute a shares
2491 * distribution (s_i) using:
2493 * s_i = rw_i / \Sum rw_j (1)
2495 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
2496 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
2497 * shares distribution (s_i):
2499 * rw_i = { 2, 4, 1, 0 }
2500 * s_i = { 2/7, 4/7, 1/7, 0 }
2502 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
2503 * task used to run on and the CPU the waker is running on), we need to
2504 * compute the effect of waking a task on either CPU and, in case of a sync
2505 * wakeup, compute the effect of the current task going to sleep.
2507 * So for a change of @wl to the local @cpu with an overall group weight change
2508 * of @wl we can compute the new shares distribution (s'_i) using:
2510 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
2512 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
2513 * differences in waking a task to CPU 0. The additional task changes the
2514 * weight and shares distributions like:
2516 * rw'_i = { 3, 4, 1, 0 }
2517 * s'_i = { 3/8, 4/8, 1/8, 0 }
2519 * We can then compute the difference in effective weight by using:
2521 * dw_i = S * (s'_i - s_i) (3)
2523 * Where 'S' is the group weight as seen by its parent.
2525 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
2526 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
2527 * 4/7) times the weight of the group.
2529 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
2531 struct sched_entity *se = tg->se[cpu];
2533 if (!tg->parent) /* the trivial, non-cgroup case */
2536 for_each_sched_entity(se) {
2542 * W = @wg + \Sum rw_j
2544 W = wg + calc_tg_weight(tg, se->my_q);
2549 w = se->my_q->load.weight + wl;
2552 * wl = S * s'_i; see (2)
2555 wl = (w * tg->shares) / W;
2560 * Per the above, wl is the new se->load.weight value; since
2561 * those are clipped to [MIN_SHARES, ...) do so now. See
2562 * calc_cfs_shares().
2564 if (wl < MIN_SHARES)
2568 * wl = dw_i = S * (s'_i - s_i); see (3)
2570 wl -= se->load.weight;
2573 * Recursively apply this logic to all parent groups to compute
2574 * the final effective load change on the root group. Since
2575 * only the @tg group gets extra weight, all parent groups can
2576 * only redistribute existing shares. @wl is the shift in shares
2577 * resulting from this level per the above.
2586 static inline unsigned long effective_load(struct task_group *tg, int cpu,
2587 unsigned long wl, unsigned long wg)
2594 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
2596 s64 this_load, load;
2597 int idx, this_cpu, prev_cpu;
2598 unsigned long tl_per_task;
2599 struct task_group *tg;
2600 unsigned long weight;
2604 this_cpu = smp_processor_id();
2605 prev_cpu = task_cpu(p);
2606 load = source_load(prev_cpu, idx);
2607 this_load = target_load(this_cpu, idx);
2610 * If sync wakeup then subtract the (maximum possible)
2611 * effect of the currently running task from the load
2612 * of the current CPU:
2615 tg = task_group(current);
2616 weight = current->se.load.weight;
2618 this_load += effective_load(tg, this_cpu, -weight, -weight);
2619 load += effective_load(tg, prev_cpu, 0, -weight);
2623 weight = p->se.load.weight;
2626 * In low-load situations, where prev_cpu is idle and this_cpu is idle
2627 * due to the sync cause above having dropped this_load to 0, we'll
2628 * always have an imbalance, but there's really nothing you can do
2629 * about that, so that's good too.
2631 * Otherwise check if either cpus are near enough in load to allow this
2632 * task to be woken on this_cpu.
2634 if (this_load > 0) {
2635 s64 this_eff_load, prev_eff_load;
2637 this_eff_load = 100;
2638 this_eff_load *= power_of(prev_cpu);
2639 this_eff_load *= this_load +
2640 effective_load(tg, this_cpu, weight, weight);
2642 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
2643 prev_eff_load *= power_of(this_cpu);
2644 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
2646 balanced = this_eff_load <= prev_eff_load;
2651 * If the currently running task will sleep within
2652 * a reasonable amount of time then attract this newly
2655 if (sync && balanced)
2658 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
2659 tl_per_task = cpu_avg_load_per_task(this_cpu);
2662 (this_load <= load &&
2663 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
2665 * This domain has SD_WAKE_AFFINE and
2666 * p is cache cold in this domain, and
2667 * there is no bad imbalance.
2669 schedstat_inc(sd, ttwu_move_affine);
2670 schedstat_inc(p, se.statistics.nr_wakeups_affine);
2678 * find_idlest_group finds and returns the least busy CPU group within the
2681 static struct sched_group *
2682 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
2683 int this_cpu, int load_idx)
2685 struct sched_group *idlest = NULL, *group = sd->groups;
2686 unsigned long min_load = ULONG_MAX, this_load = 0;
2687 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2690 unsigned long load, avg_load;
2694 /* Skip over this group if it has no CPUs allowed */
2695 if (!cpumask_intersects(sched_group_cpus(group),
2696 tsk_cpus_allowed(p)))
2699 local_group = cpumask_test_cpu(this_cpu,
2700 sched_group_cpus(group));
2702 /* Tally up the load of all CPUs in the group */
2705 for_each_cpu(i, sched_group_cpus(group)) {
2706 /* Bias balancing toward cpus of our domain */
2708 load = source_load(i, load_idx);
2710 load = target_load(i, load_idx);
2715 /* Adjust by relative CPU power of the group */
2716 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
2719 this_load = avg_load;
2720 } else if (avg_load < min_load) {
2721 min_load = avg_load;
2724 } while (group = group->next, group != sd->groups);
2726 if (!idlest || 100*this_load < imbalance*min_load)
2732 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2735 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2737 unsigned long load, min_load = ULONG_MAX;
2741 /* Traverse only the allowed CPUs */
2742 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
2743 load = weighted_cpuload(i);
2745 if (load < min_load || (load == min_load && i == this_cpu)) {
2755 * Try and locate an idle CPU in the sched_domain.
2757 static int select_idle_sibling(struct task_struct *p, int target)
2759 int cpu = smp_processor_id();
2760 int prev_cpu = task_cpu(p);
2761 struct sched_domain *sd;
2762 struct sched_group *sg;
2766 * If the task is going to be woken-up on this cpu and if it is
2767 * already idle, then it is the right target.
2769 if (target == cpu && idle_cpu(cpu))
2773 * If the task is going to be woken-up on the cpu where it previously
2774 * ran and if it is currently idle, then it the right target.
2776 if (target == prev_cpu && idle_cpu(prev_cpu))
2780 * Otherwise, iterate the domains and find an elegible idle cpu.
2782 sd = rcu_dereference(per_cpu(sd_llc, target));
2783 for_each_lower_domain(sd) {
2786 if (!cpumask_intersects(sched_group_cpus(sg),
2787 tsk_cpus_allowed(p)))
2790 for_each_cpu(i, sched_group_cpus(sg)) {
2795 target = cpumask_first_and(sched_group_cpus(sg),
2796 tsk_cpus_allowed(p));
2800 } while (sg != sd->groups);
2807 * sched_balance_self: balance the current task (running on cpu) in domains
2808 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2811 * Balance, ie. select the least loaded group.
2813 * Returns the target CPU number, or the same CPU if no balancing is needed.
2815 * preempt must be disabled.
2818 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
2820 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
2821 int cpu = smp_processor_id();
2822 int prev_cpu = task_cpu(p);
2824 int want_affine = 0;
2825 int sync = wake_flags & WF_SYNC;
2827 if (p->nr_cpus_allowed == 1)
2830 if (sd_flag & SD_BALANCE_WAKE) {
2831 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
2837 for_each_domain(cpu, tmp) {
2838 if (!(tmp->flags & SD_LOAD_BALANCE))
2842 * If both cpu and prev_cpu are part of this domain,
2843 * cpu is a valid SD_WAKE_AFFINE target.
2845 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
2846 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
2851 if (tmp->flags & sd_flag)
2856 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
2859 new_cpu = select_idle_sibling(p, prev_cpu);
2864 int load_idx = sd->forkexec_idx;
2865 struct sched_group *group;
2868 if (!(sd->flags & sd_flag)) {
2873 if (sd_flag & SD_BALANCE_WAKE)
2874 load_idx = sd->wake_idx;
2876 group = find_idlest_group(sd, p, cpu, load_idx);
2882 new_cpu = find_idlest_cpu(group, p, cpu);
2883 if (new_cpu == -1 || new_cpu == cpu) {
2884 /* Now try balancing at a lower domain level of cpu */
2889 /* Now try balancing at a lower domain level of new_cpu */
2891 weight = sd->span_weight;
2893 for_each_domain(cpu, tmp) {
2894 if (weight <= tmp->span_weight)
2896 if (tmp->flags & sd_flag)
2899 /* while loop will break here if sd == NULL */
2906 #endif /* CONFIG_SMP */
2908 static unsigned long
2909 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
2911 unsigned long gran = sysctl_sched_wakeup_granularity;
2914 * Since its curr running now, convert the gran from real-time
2915 * to virtual-time in his units.
2917 * By using 'se' instead of 'curr' we penalize light tasks, so
2918 * they get preempted easier. That is, if 'se' < 'curr' then
2919 * the resulting gran will be larger, therefore penalizing the
2920 * lighter, if otoh 'se' > 'curr' then the resulting gran will
2921 * be smaller, again penalizing the lighter task.
2923 * This is especially important for buddies when the leftmost
2924 * task is higher priority than the buddy.
2926 return calc_delta_fair(gran, se);
2930 * Should 'se' preempt 'curr'.
2944 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
2946 s64 gran, vdiff = curr->vruntime - se->vruntime;
2951 gran = wakeup_gran(curr, se);
2958 static void set_last_buddy(struct sched_entity *se)
2960 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
2963 for_each_sched_entity(se)
2964 cfs_rq_of(se)->last = se;
2967 static void set_next_buddy(struct sched_entity *se)
2969 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
2972 for_each_sched_entity(se)
2973 cfs_rq_of(se)->next = se;
2976 static void set_skip_buddy(struct sched_entity *se)
2978 for_each_sched_entity(se)
2979 cfs_rq_of(se)->skip = se;
2983 * Preempt the current task with a newly woken task if needed:
2985 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2987 struct task_struct *curr = rq->curr;
2988 struct sched_entity *se = &curr->se, *pse = &p->se;
2989 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
2990 int scale = cfs_rq->nr_running >= sched_nr_latency;
2991 int next_buddy_marked = 0;
2993 if (unlikely(se == pse))
2997 * This is possible from callers such as move_task(), in which we
2998 * unconditionally check_prempt_curr() after an enqueue (which may have
2999 * lead to a throttle). This both saves work and prevents false
3000 * next-buddy nomination below.
3002 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3005 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3006 set_next_buddy(pse);
3007 next_buddy_marked = 1;
3011 * We can come here with TIF_NEED_RESCHED already set from new task
3014 * Note: this also catches the edge-case of curr being in a throttled
3015 * group (e.g. via set_curr_task), since update_curr() (in the
3016 * enqueue of curr) will have resulted in resched being set. This
3017 * prevents us from potentially nominating it as a false LAST_BUDDY
3020 if (test_tsk_need_resched(curr))
3023 /* Idle tasks are by definition preempted by non-idle tasks. */
3024 if (unlikely(curr->policy == SCHED_IDLE) &&
3025 likely(p->policy != SCHED_IDLE))
3029 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3030 * is driven by the tick):
3032 if (unlikely(p->policy != SCHED_NORMAL))
3035 find_matching_se(&se, &pse);
3036 update_curr(cfs_rq_of(se));
3038 if (wakeup_preempt_entity(se, pse) == 1) {
3040 * Bias pick_next to pick the sched entity that is
3041 * triggering this preemption.
3043 if (!next_buddy_marked)
3044 set_next_buddy(pse);
3053 * Only set the backward buddy when the current task is still
3054 * on the rq. This can happen when a wakeup gets interleaved
3055 * with schedule on the ->pre_schedule() or idle_balance()
3056 * point, either of which can * drop the rq lock.
3058 * Also, during early boot the idle thread is in the fair class,
3059 * for obvious reasons its a bad idea to schedule back to it.
3061 if (unlikely(!se->on_rq || curr == rq->idle))
3064 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3068 static struct task_struct *pick_next_task_fair(struct rq *rq)
3070 struct task_struct *p;
3071 struct cfs_rq *cfs_rq = &rq->cfs;
3072 struct sched_entity *se;
3074 if (!cfs_rq->nr_running)
3078 se = pick_next_entity(cfs_rq);
3079 set_next_entity(cfs_rq, se);
3080 cfs_rq = group_cfs_rq(se);
3084 if (hrtick_enabled(rq))
3085 hrtick_start_fair(rq, p);
3091 * Account for a descheduled task:
3093 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3095 struct sched_entity *se = &prev->se;
3096 struct cfs_rq *cfs_rq;
3098 for_each_sched_entity(se) {
3099 cfs_rq = cfs_rq_of(se);
3100 put_prev_entity(cfs_rq, se);
3105 * sched_yield() is very simple
3107 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3109 static void yield_task_fair(struct rq *rq)
3111 struct task_struct *curr = rq->curr;
3112 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3113 struct sched_entity *se = &curr->se;
3116 * Are we the only task in the tree?
3118 if (unlikely(rq->nr_running == 1))
3121 clear_buddies(cfs_rq, se);
3123 if (curr->policy != SCHED_BATCH) {
3124 update_rq_clock(rq);
3126 * Update run-time statistics of the 'current'.
3128 update_curr(cfs_rq);
3130 * Tell update_rq_clock() that we've just updated,
3131 * so we don't do microscopic update in schedule()
3132 * and double the fastpath cost.
3134 rq->skip_clock_update = 1;
3140 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3142 struct sched_entity *se = &p->se;
3144 /* throttled hierarchies are not runnable */
3145 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3148 /* Tell the scheduler that we'd really like pse to run next. */
3151 yield_task_fair(rq);
3157 /**************************************************
3158 * Fair scheduling class load-balancing methods:
3161 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
3163 #define LBF_ALL_PINNED 0x01
3164 #define LBF_NEED_BREAK 0x02
3165 #define LBF_SOME_PINNED 0x04
3168 struct sched_domain *sd;
3176 struct cpumask *dst_grpmask;
3178 enum cpu_idle_type idle;
3180 /* The set of CPUs under consideration for load-balancing */
3181 struct cpumask *cpus;
3186 unsigned int loop_break;
3187 unsigned int loop_max;
3191 * move_task - move a task from one runqueue to another runqueue.
3192 * Both runqueues must be locked.
3194 static void move_task(struct task_struct *p, struct lb_env *env)
3196 deactivate_task(env->src_rq, p, 0);
3197 set_task_cpu(p, env->dst_cpu);
3198 activate_task(env->dst_rq, p, 0);
3199 check_preempt_curr(env->dst_rq, p, 0);
3203 * Is this task likely cache-hot:
3206 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
3210 if (p->sched_class != &fair_sched_class)
3213 if (unlikely(p->policy == SCHED_IDLE))
3217 * Buddy candidates are cache hot:
3219 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
3220 (&p->se == cfs_rq_of(&p->se)->next ||
3221 &p->se == cfs_rq_of(&p->se)->last))
3224 if (sysctl_sched_migration_cost == -1)
3226 if (sysctl_sched_migration_cost == 0)
3229 delta = now - p->se.exec_start;
3231 return delta < (s64)sysctl_sched_migration_cost;
3235 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3238 int can_migrate_task(struct task_struct *p, struct lb_env *env)
3240 int tsk_cache_hot = 0;
3242 * We do not migrate tasks that are:
3243 * 1) running (obviously), or
3244 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3245 * 3) are cache-hot on their current CPU.
3247 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
3250 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
3253 * Remember if this task can be migrated to any other cpu in
3254 * our sched_group. We may want to revisit it if we couldn't
3255 * meet load balance goals by pulling other tasks on src_cpu.
3257 * Also avoid computing new_dst_cpu if we have already computed
3258 * one in current iteration.
3260 if (!env->dst_grpmask || (env->flags & LBF_SOME_PINNED))
3263 new_dst_cpu = cpumask_first_and(env->dst_grpmask,
3264 tsk_cpus_allowed(p));
3265 if (new_dst_cpu < nr_cpu_ids) {
3266 env->flags |= LBF_SOME_PINNED;
3267 env->new_dst_cpu = new_dst_cpu;
3272 /* Record that we found atleast one task that could run on dst_cpu */
3273 env->flags &= ~LBF_ALL_PINNED;
3275 if (task_running(env->src_rq, p)) {
3276 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
3281 * Aggressive migration if:
3282 * 1) task is cache cold, or
3283 * 2) too many balance attempts have failed.
3286 tsk_cache_hot = task_hot(p, env->src_rq->clock_task, env->sd);
3287 if (!tsk_cache_hot ||
3288 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
3289 #ifdef CONFIG_SCHEDSTATS
3290 if (tsk_cache_hot) {
3291 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
3292 schedstat_inc(p, se.statistics.nr_forced_migrations);
3298 if (tsk_cache_hot) {
3299 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
3306 * move_one_task tries to move exactly one task from busiest to this_rq, as
3307 * part of active balancing operations within "domain".
3308 * Returns 1 if successful and 0 otherwise.
3310 * Called with both runqueues locked.
3312 static int move_one_task(struct lb_env *env)
3314 struct task_struct *p, *n;
3316 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
3317 if (throttled_lb_pair(task_group(p), env->src_rq->cpu, env->dst_cpu))
3320 if (!can_migrate_task(p, env))
3325 * Right now, this is only the second place move_task()
3326 * is called, so we can safely collect move_task()
3327 * stats here rather than inside move_task().
3329 schedstat_inc(env->sd, lb_gained[env->idle]);
3335 static unsigned long task_h_load(struct task_struct *p);
3337 static const unsigned int sched_nr_migrate_break = 32;
3340 * move_tasks tries to move up to imbalance weighted load from busiest to
3341 * this_rq, as part of a balancing operation within domain "sd".
3342 * Returns 1 if successful and 0 otherwise.
3344 * Called with both runqueues locked.
3346 static int move_tasks(struct lb_env *env)
3348 struct list_head *tasks = &env->src_rq->cfs_tasks;
3349 struct task_struct *p;
3353 if (env->imbalance <= 0)
3356 while (!list_empty(tasks)) {
3357 p = list_first_entry(tasks, struct task_struct, se.group_node);
3360 /* We've more or less seen every task there is, call it quits */
3361 if (env->loop > env->loop_max)
3364 /* take a breather every nr_migrate tasks */
3365 if (env->loop > env->loop_break) {
3366 env->loop_break += sched_nr_migrate_break;
3367 env->flags |= LBF_NEED_BREAK;
3371 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
3374 load = task_h_load(p);
3376 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
3379 if ((load / 2) > env->imbalance)
3382 if (!can_migrate_task(p, env))
3387 env->imbalance -= load;
3389 #ifdef CONFIG_PREEMPT
3391 * NEWIDLE balancing is a source of latency, so preemptible
3392 * kernels will stop after the first task is pulled to minimize
3393 * the critical section.
3395 if (env->idle == CPU_NEWLY_IDLE)
3400 * We only want to steal up to the prescribed amount of
3403 if (env->imbalance <= 0)
3408 list_move_tail(&p->se.group_node, tasks);
3412 * Right now, this is one of only two places move_task() is called,
3413 * so we can safely collect move_task() stats here rather than
3414 * inside move_task().
3416 schedstat_add(env->sd, lb_gained[env->idle], pulled);
3421 #ifdef CONFIG_FAIR_GROUP_SCHED
3423 * update tg->load_weight by folding this cpu's load_avg
3425 static int update_shares_cpu(struct task_group *tg, int cpu)
3427 struct cfs_rq *cfs_rq;
3428 unsigned long flags;
3435 cfs_rq = tg->cfs_rq[cpu];
3437 raw_spin_lock_irqsave(&rq->lock, flags);
3439 update_rq_clock(rq);
3440 update_cfs_load(cfs_rq, 1);
3443 * We need to update shares after updating tg->load_weight in
3444 * order to adjust the weight of groups with long running tasks.
3446 update_cfs_shares(cfs_rq);
3448 raw_spin_unlock_irqrestore(&rq->lock, flags);
3453 static void update_shares(int cpu)
3455 struct cfs_rq *cfs_rq;
3456 struct rq *rq = cpu_rq(cpu);
3460 * Iterates the task_group tree in a bottom up fashion, see
3461 * list_add_leaf_cfs_rq() for details.
3463 for_each_leaf_cfs_rq(rq, cfs_rq) {
3464 /* throttled entities do not contribute to load */
3465 if (throttled_hierarchy(cfs_rq))
3468 update_shares_cpu(cfs_rq->tg, cpu);
3474 * Compute the cpu's hierarchical load factor for each task group.
3475 * This needs to be done in a top-down fashion because the load of a child
3476 * group is a fraction of its parents load.
3478 static int tg_load_down(struct task_group *tg, void *data)
3481 long cpu = (long)data;
3484 load = cpu_rq(cpu)->load.weight;
3486 load = tg->parent->cfs_rq[cpu]->h_load;
3487 load *= tg->se[cpu]->load.weight;
3488 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
3491 tg->cfs_rq[cpu]->h_load = load;
3496 static void update_h_load(long cpu)
3498 struct rq *rq = cpu_rq(cpu);
3499 unsigned long now = jiffies;
3501 if (rq->h_load_throttle == now)
3504 rq->h_load_throttle = now;
3507 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
3511 static unsigned long task_h_load(struct task_struct *p)
3513 struct cfs_rq *cfs_rq = task_cfs_rq(p);
3516 load = p->se.load.weight;
3517 load = div_u64(load * cfs_rq->h_load, cfs_rq->load.weight + 1);
3522 static inline void update_shares(int cpu)
3526 static inline void update_h_load(long cpu)
3530 static unsigned long task_h_load(struct task_struct *p)
3532 return p->se.load.weight;
3536 /********** Helpers for find_busiest_group ************************/
3538 * sd_lb_stats - Structure to store the statistics of a sched_domain
3539 * during load balancing.
3541 struct sd_lb_stats {
3542 struct sched_group *busiest; /* Busiest group in this sd */
3543 struct sched_group *this; /* Local group in this sd */
3544 unsigned long total_load; /* Total load of all groups in sd */
3545 unsigned long total_pwr; /* Total power of all groups in sd */
3546 unsigned long avg_load; /* Average load across all groups in sd */
3548 /** Statistics of this group */
3549 unsigned long this_load;
3550 unsigned long this_load_per_task;
3551 unsigned long this_nr_running;
3552 unsigned long this_has_capacity;
3553 unsigned int this_idle_cpus;
3555 /* Statistics of the busiest group */
3556 unsigned int busiest_idle_cpus;
3557 unsigned long max_load;
3558 unsigned long busiest_load_per_task;
3559 unsigned long busiest_nr_running;
3560 unsigned long busiest_group_capacity;
3561 unsigned long busiest_has_capacity;
3562 unsigned int busiest_group_weight;
3564 int group_imb; /* Is there imbalance in this sd */
3568 * sg_lb_stats - stats of a sched_group required for load_balancing
3570 struct sg_lb_stats {
3571 unsigned long avg_load; /*Avg load across the CPUs of the group */
3572 unsigned long group_load; /* Total load over the CPUs of the group */
3573 unsigned long sum_nr_running; /* Nr tasks running in the group */
3574 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3575 unsigned long group_capacity;
3576 unsigned long idle_cpus;
3577 unsigned long group_weight;
3578 int group_imb; /* Is there an imbalance in the group ? */
3579 int group_has_capacity; /* Is there extra capacity in the group? */
3583 * get_sd_load_idx - Obtain the load index for a given sched domain.
3584 * @sd: The sched_domain whose load_idx is to be obtained.
3585 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3587 static inline int get_sd_load_idx(struct sched_domain *sd,
3588 enum cpu_idle_type idle)
3594 load_idx = sd->busy_idx;
3597 case CPU_NEWLY_IDLE:
3598 load_idx = sd->newidle_idx;
3601 load_idx = sd->idle_idx;
3608 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3610 return SCHED_POWER_SCALE;
3613 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3615 return default_scale_freq_power(sd, cpu);
3618 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3620 unsigned long weight = sd->span_weight;
3621 unsigned long smt_gain = sd->smt_gain;
3628 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3630 return default_scale_smt_power(sd, cpu);
3633 unsigned long scale_rt_power(int cpu)
3635 struct rq *rq = cpu_rq(cpu);
3636 u64 total, available, age_stamp, avg;
3639 * Since we're reading these variables without serialization make sure
3640 * we read them once before doing sanity checks on them.
3642 age_stamp = ACCESS_ONCE(rq->age_stamp);
3643 avg = ACCESS_ONCE(rq->rt_avg);
3645 total = sched_avg_period() + (rq->clock - age_stamp);
3647 if (unlikely(total < avg)) {
3648 /* Ensures that power won't end up being negative */
3651 available = total - avg;
3654 if (unlikely((s64)total < SCHED_POWER_SCALE))
3655 total = SCHED_POWER_SCALE;
3657 total >>= SCHED_POWER_SHIFT;
3659 return div_u64(available, total);
3662 static void update_cpu_power(struct sched_domain *sd, int cpu)
3664 unsigned long weight = sd->span_weight;
3665 unsigned long power = SCHED_POWER_SCALE;
3666 struct sched_group *sdg = sd->groups;
3668 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3669 if (sched_feat(ARCH_POWER))
3670 power *= arch_scale_smt_power(sd, cpu);
3672 power *= default_scale_smt_power(sd, cpu);
3674 power >>= SCHED_POWER_SHIFT;
3677 sdg->sgp->power_orig = power;
3679 if (sched_feat(ARCH_POWER))
3680 power *= arch_scale_freq_power(sd, cpu);
3682 power *= default_scale_freq_power(sd, cpu);
3684 power >>= SCHED_POWER_SHIFT;
3686 power *= scale_rt_power(cpu);
3687 power >>= SCHED_POWER_SHIFT;
3692 cpu_rq(cpu)->cpu_power = power;
3693 sdg->sgp->power = power;
3696 void update_group_power(struct sched_domain *sd, int cpu)
3698 struct sched_domain *child = sd->child;
3699 struct sched_group *group, *sdg = sd->groups;
3700 unsigned long power;
3701 unsigned long interval;
3703 interval = msecs_to_jiffies(sd->balance_interval);
3704 interval = clamp(interval, 1UL, max_load_balance_interval);
3705 sdg->sgp->next_update = jiffies + interval;
3708 update_cpu_power(sd, cpu);
3714 if (child->flags & SD_OVERLAP) {
3716 * SD_OVERLAP domains cannot assume that child groups
3717 * span the current group.
3720 for_each_cpu(cpu, sched_group_cpus(sdg))
3721 power += power_of(cpu);
3724 * !SD_OVERLAP domains can assume that child groups
3725 * span the current group.
3728 group = child->groups;
3730 power += group->sgp->power;
3731 group = group->next;
3732 } while (group != child->groups);
3735 sdg->sgp->power_orig = sdg->sgp->power = power;
3739 * Try and fix up capacity for tiny siblings, this is needed when
3740 * things like SD_ASYM_PACKING need f_b_g to select another sibling
3741 * which on its own isn't powerful enough.
3743 * See update_sd_pick_busiest() and check_asym_packing().
3746 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
3749 * Only siblings can have significantly less than SCHED_POWER_SCALE
3751 if (!(sd->flags & SD_SHARE_CPUPOWER))
3755 * If ~90% of the cpu_power is still there, we're good.
3757 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
3764 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3765 * @env: The load balancing environment.
3766 * @group: sched_group whose statistics are to be updated.
3767 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3768 * @local_group: Does group contain this_cpu.
3769 * @balance: Should we balance.
3770 * @sgs: variable to hold the statistics for this group.
3772 static inline void update_sg_lb_stats(struct lb_env *env,
3773 struct sched_group *group, int load_idx,
3774 int local_group, int *balance, struct sg_lb_stats *sgs)
3776 unsigned long nr_running, max_nr_running, min_nr_running;
3777 unsigned long load, max_cpu_load, min_cpu_load;
3778 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3779 unsigned long avg_load_per_task = 0;
3783 balance_cpu = group_balance_cpu(group);
3785 /* Tally up the load of all CPUs in the group */
3787 min_cpu_load = ~0UL;
3789 min_nr_running = ~0UL;
3791 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
3792 struct rq *rq = cpu_rq(i);
3794 nr_running = rq->nr_running;
3796 /* Bias balancing toward cpus of our domain */
3798 if (idle_cpu(i) && !first_idle_cpu &&
3799 cpumask_test_cpu(i, sched_group_mask(group))) {
3804 load = target_load(i, load_idx);
3806 load = source_load(i, load_idx);
3807 if (load > max_cpu_load)
3808 max_cpu_load = load;
3809 if (min_cpu_load > load)
3810 min_cpu_load = load;
3812 if (nr_running > max_nr_running)
3813 max_nr_running = nr_running;
3814 if (min_nr_running > nr_running)
3815 min_nr_running = nr_running;
3818 sgs->group_load += load;
3819 sgs->sum_nr_running += nr_running;
3820 sgs->sum_weighted_load += weighted_cpuload(i);
3826 * First idle cpu or the first cpu(busiest) in this sched group
3827 * is eligible for doing load balancing at this and above
3828 * domains. In the newly idle case, we will allow all the cpu's
3829 * to do the newly idle load balance.
3832 if (env->idle != CPU_NEWLY_IDLE) {
3833 if (balance_cpu != env->dst_cpu) {
3837 update_group_power(env->sd, env->dst_cpu);
3838 } else if (time_after_eq(jiffies, group->sgp->next_update))
3839 update_group_power(env->sd, env->dst_cpu);
3842 /* Adjust by relative CPU power of the group */
3843 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
3846 * Consider the group unbalanced when the imbalance is larger
3847 * than the average weight of a task.
3849 * APZ: with cgroup the avg task weight can vary wildly and
3850 * might not be a suitable number - should we keep a
3851 * normalized nr_running number somewhere that negates
3854 if (sgs->sum_nr_running)
3855 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
3857 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task &&
3858 (max_nr_running - min_nr_running) > 1)
3861 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
3863 if (!sgs->group_capacity)
3864 sgs->group_capacity = fix_small_capacity(env->sd, group);
3865 sgs->group_weight = group->group_weight;
3867 if (sgs->group_capacity > sgs->sum_nr_running)
3868 sgs->group_has_capacity = 1;
3872 * update_sd_pick_busiest - return 1 on busiest group
3873 * @env: The load balancing environment.
3874 * @sds: sched_domain statistics
3875 * @sg: sched_group candidate to be checked for being the busiest
3876 * @sgs: sched_group statistics
3878 * Determine if @sg is a busier group than the previously selected
3881 static bool update_sd_pick_busiest(struct lb_env *env,
3882 struct sd_lb_stats *sds,
3883 struct sched_group *sg,
3884 struct sg_lb_stats *sgs)
3886 if (sgs->avg_load <= sds->max_load)
3889 if (sgs->sum_nr_running > sgs->group_capacity)
3896 * ASYM_PACKING needs to move all the work to the lowest
3897 * numbered CPUs in the group, therefore mark all groups
3898 * higher than ourself as busy.
3900 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
3901 env->dst_cpu < group_first_cpu(sg)) {
3905 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
3913 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
3914 * @env: The load balancing environment.
3915 * @balance: Should we balance.
3916 * @sds: variable to hold the statistics for this sched_domain.
3918 static inline void update_sd_lb_stats(struct lb_env *env,
3919 int *balance, struct sd_lb_stats *sds)
3921 struct sched_domain *child = env->sd->child;
3922 struct sched_group *sg = env->sd->groups;
3923 struct sg_lb_stats sgs;
3924 int load_idx, prefer_sibling = 0;
3926 if (child && child->flags & SD_PREFER_SIBLING)
3929 load_idx = get_sd_load_idx(env->sd, env->idle);
3934 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
3935 memset(&sgs, 0, sizeof(sgs));
3936 update_sg_lb_stats(env, sg, load_idx, local_group, balance, &sgs);
3938 if (local_group && !(*balance))
3941 sds->total_load += sgs.group_load;
3942 sds->total_pwr += sg->sgp->power;
3945 * In case the child domain prefers tasks go to siblings
3946 * first, lower the sg capacity to one so that we'll try
3947 * and move all the excess tasks away. We lower the capacity
3948 * of a group only if the local group has the capacity to fit
3949 * these excess tasks, i.e. nr_running < group_capacity. The
3950 * extra check prevents the case where you always pull from the
3951 * heaviest group when it is already under-utilized (possible
3952 * with a large weight task outweighs the tasks on the system).
3954 if (prefer_sibling && !local_group && sds->this_has_capacity)
3955 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3958 sds->this_load = sgs.avg_load;
3960 sds->this_nr_running = sgs.sum_nr_running;
3961 sds->this_load_per_task = sgs.sum_weighted_load;
3962 sds->this_has_capacity = sgs.group_has_capacity;
3963 sds->this_idle_cpus = sgs.idle_cpus;
3964 } else if (update_sd_pick_busiest(env, sds, sg, &sgs)) {
3965 sds->max_load = sgs.avg_load;
3967 sds->busiest_nr_running = sgs.sum_nr_running;
3968 sds->busiest_idle_cpus = sgs.idle_cpus;
3969 sds->busiest_group_capacity = sgs.group_capacity;
3970 sds->busiest_load_per_task = sgs.sum_weighted_load;
3971 sds->busiest_has_capacity = sgs.group_has_capacity;
3972 sds->busiest_group_weight = sgs.group_weight;
3973 sds->group_imb = sgs.group_imb;
3977 } while (sg != env->sd->groups);
3981 * check_asym_packing - Check to see if the group is packed into the
3984 * This is primarily intended to used at the sibling level. Some
3985 * cores like POWER7 prefer to use lower numbered SMT threads. In the
3986 * case of POWER7, it can move to lower SMT modes only when higher
3987 * threads are idle. When in lower SMT modes, the threads will
3988 * perform better since they share less core resources. Hence when we
3989 * have idle threads, we want them to be the higher ones.
3991 * This packing function is run on idle threads. It checks to see if
3992 * the busiest CPU in this domain (core in the P7 case) has a higher
3993 * CPU number than the packing function is being run on. Here we are
3994 * assuming lower CPU number will be equivalent to lower a SMT thread
3997 * Returns 1 when packing is required and a task should be moved to
3998 * this CPU. The amount of the imbalance is returned in *imbalance.
4000 * @env: The load balancing environment.
4001 * @sds: Statistics of the sched_domain which is to be packed
4003 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
4007 if (!(env->sd->flags & SD_ASYM_PACKING))
4013 busiest_cpu = group_first_cpu(sds->busiest);
4014 if (env->dst_cpu > busiest_cpu)
4017 env->imbalance = DIV_ROUND_CLOSEST(
4018 sds->max_load * sds->busiest->sgp->power, SCHED_POWER_SCALE);
4024 * fix_small_imbalance - Calculate the minor imbalance that exists
4025 * amongst the groups of a sched_domain, during
4027 * @env: The load balancing environment.
4028 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4031 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4033 unsigned long tmp, pwr_now = 0, pwr_move = 0;
4034 unsigned int imbn = 2;
4035 unsigned long scaled_busy_load_per_task;
4037 if (sds->this_nr_running) {
4038 sds->this_load_per_task /= sds->this_nr_running;
4039 if (sds->busiest_load_per_task >
4040 sds->this_load_per_task)
4043 sds->this_load_per_task =
4044 cpu_avg_load_per_task(env->dst_cpu);
4047 scaled_busy_load_per_task = sds->busiest_load_per_task
4048 * SCHED_POWER_SCALE;
4049 scaled_busy_load_per_task /= sds->busiest->sgp->power;
4051 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
4052 (scaled_busy_load_per_task * imbn)) {
4053 env->imbalance = sds->busiest_load_per_task;
4058 * OK, we don't have enough imbalance to justify moving tasks,
4059 * however we may be able to increase total CPU power used by
4063 pwr_now += sds->busiest->sgp->power *
4064 min(sds->busiest_load_per_task, sds->max_load);
4065 pwr_now += sds->this->sgp->power *
4066 min(sds->this_load_per_task, sds->this_load);
4067 pwr_now /= SCHED_POWER_SCALE;
4069 /* Amount of load we'd subtract */
4070 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4071 sds->busiest->sgp->power;
4072 if (sds->max_load > tmp)
4073 pwr_move += sds->busiest->sgp->power *
4074 min(sds->busiest_load_per_task, sds->max_load - tmp);
4076 /* Amount of load we'd add */
4077 if (sds->max_load * sds->busiest->sgp->power <
4078 sds->busiest_load_per_task * SCHED_POWER_SCALE)
4079 tmp = (sds->max_load * sds->busiest->sgp->power) /
4080 sds->this->sgp->power;
4082 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4083 sds->this->sgp->power;
4084 pwr_move += sds->this->sgp->power *
4085 min(sds->this_load_per_task, sds->this_load + tmp);
4086 pwr_move /= SCHED_POWER_SCALE;
4088 /* Move if we gain throughput */
4089 if (pwr_move > pwr_now)
4090 env->imbalance = sds->busiest_load_per_task;
4094 * calculate_imbalance - Calculate the amount of imbalance present within the
4095 * groups of a given sched_domain during load balance.
4096 * @env: load balance environment
4097 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4099 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4101 unsigned long max_pull, load_above_capacity = ~0UL;
4103 sds->busiest_load_per_task /= sds->busiest_nr_running;
4104 if (sds->group_imb) {
4105 sds->busiest_load_per_task =
4106 min(sds->busiest_load_per_task, sds->avg_load);
4110 * In the presence of smp nice balancing, certain scenarios can have
4111 * max load less than avg load(as we skip the groups at or below
4112 * its cpu_power, while calculating max_load..)
4114 if (sds->max_load < sds->avg_load) {
4116 return fix_small_imbalance(env, sds);
4119 if (!sds->group_imb) {
4121 * Don't want to pull so many tasks that a group would go idle.
4123 load_above_capacity = (sds->busiest_nr_running -
4124 sds->busiest_group_capacity);
4126 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
4128 load_above_capacity /= sds->busiest->sgp->power;
4132 * We're trying to get all the cpus to the average_load, so we don't
4133 * want to push ourselves above the average load, nor do we wish to
4134 * reduce the max loaded cpu below the average load. At the same time,
4135 * we also don't want to reduce the group load below the group capacity
4136 * (so that we can implement power-savings policies etc). Thus we look
4137 * for the minimum possible imbalance.
4138 * Be careful of negative numbers as they'll appear as very large values
4139 * with unsigned longs.
4141 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
4143 /* How much load to actually move to equalise the imbalance */
4144 env->imbalance = min(max_pull * sds->busiest->sgp->power,
4145 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
4146 / SCHED_POWER_SCALE;
4149 * if *imbalance is less than the average load per runnable task
4150 * there is no guarantee that any tasks will be moved so we'll have
4151 * a think about bumping its value to force at least one task to be
4154 if (env->imbalance < sds->busiest_load_per_task)
4155 return fix_small_imbalance(env, sds);
4159 /******* find_busiest_group() helpers end here *********************/
4162 * find_busiest_group - Returns the busiest group within the sched_domain
4163 * if there is an imbalance. If there isn't an imbalance, and
4164 * the user has opted for power-savings, it returns a group whose
4165 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4166 * such a group exists.
4168 * Also calculates the amount of weighted load which should be moved
4169 * to restore balance.
4171 * @env: The load balancing environment.
4172 * @balance: Pointer to a variable indicating if this_cpu
4173 * is the appropriate cpu to perform load balancing at this_level.
4175 * Returns: - the busiest group if imbalance exists.
4176 * - If no imbalance and user has opted for power-savings balance,
4177 * return the least loaded group whose CPUs can be
4178 * put to idle by rebalancing its tasks onto our group.
4180 static struct sched_group *
4181 find_busiest_group(struct lb_env *env, int *balance)
4183 struct sd_lb_stats sds;
4185 memset(&sds, 0, sizeof(sds));
4188 * Compute the various statistics relavent for load balancing at
4191 update_sd_lb_stats(env, balance, &sds);
4194 * this_cpu is not the appropriate cpu to perform load balancing at
4200 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
4201 check_asym_packing(env, &sds))
4204 /* There is no busy sibling group to pull tasks from */
4205 if (!sds.busiest || sds.busiest_nr_running == 0)
4208 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
4211 * If the busiest group is imbalanced the below checks don't
4212 * work because they assumes all things are equal, which typically
4213 * isn't true due to cpus_allowed constraints and the like.
4218 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
4219 if (env->idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
4220 !sds.busiest_has_capacity)
4224 * If the local group is more busy than the selected busiest group
4225 * don't try and pull any tasks.
4227 if (sds.this_load >= sds.max_load)
4231 * Don't pull any tasks if this group is already above the domain
4234 if (sds.this_load >= sds.avg_load)
4237 if (env->idle == CPU_IDLE) {
4239 * This cpu is idle. If the busiest group load doesn't
4240 * have more tasks than the number of available cpu's and
4241 * there is no imbalance between this and busiest group
4242 * wrt to idle cpu's, it is balanced.
4244 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
4245 sds.busiest_nr_running <= sds.busiest_group_weight)
4249 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
4250 * imbalance_pct to be conservative.
4252 if (100 * sds.max_load <= env->sd->imbalance_pct * sds.this_load)
4257 /* Looks like there is an imbalance. Compute it */
4258 calculate_imbalance(env, &sds);
4268 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4270 static struct rq *find_busiest_queue(struct lb_env *env,
4271 struct sched_group *group)
4273 struct rq *busiest = NULL, *rq;
4274 unsigned long max_load = 0;
4277 for_each_cpu(i, sched_group_cpus(group)) {
4278 unsigned long power = power_of(i);
4279 unsigned long capacity = DIV_ROUND_CLOSEST(power,
4284 capacity = fix_small_capacity(env->sd, group);
4286 if (!cpumask_test_cpu(i, env->cpus))
4290 wl = weighted_cpuload(i);
4293 * When comparing with imbalance, use weighted_cpuload()
4294 * which is not scaled with the cpu power.
4296 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
4300 * For the load comparisons with the other cpu's, consider
4301 * the weighted_cpuload() scaled with the cpu power, so that
4302 * the load can be moved away from the cpu that is potentially
4303 * running at a lower capacity.
4305 wl = (wl * SCHED_POWER_SCALE) / power;
4307 if (wl > max_load) {
4317 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4318 * so long as it is large enough.
4320 #define MAX_PINNED_INTERVAL 512
4322 /* Working cpumask for load_balance and load_balance_newidle. */
4323 DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4325 static int need_active_balance(struct lb_env *env)
4327 struct sched_domain *sd = env->sd;
4329 if (env->idle == CPU_NEWLY_IDLE) {
4332 * ASYM_PACKING needs to force migrate tasks from busy but
4333 * higher numbered CPUs in order to pack all tasks in the
4334 * lowest numbered CPUs.
4336 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
4340 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
4343 static int active_load_balance_cpu_stop(void *data);
4346 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4347 * tasks if there is an imbalance.
4349 static int load_balance(int this_cpu, struct rq *this_rq,
4350 struct sched_domain *sd, enum cpu_idle_type idle,
4353 int ld_moved, cur_ld_moved, active_balance = 0;
4354 int lb_iterations, max_lb_iterations;
4355 struct sched_group *group;
4357 unsigned long flags;
4358 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4360 struct lb_env env = {
4362 .dst_cpu = this_cpu,
4364 .dst_grpmask = sched_group_cpus(sd->groups),
4366 .loop_break = sched_nr_migrate_break,
4370 cpumask_copy(cpus, cpu_active_mask);
4371 max_lb_iterations = cpumask_weight(env.dst_grpmask);
4373 schedstat_inc(sd, lb_count[idle]);
4376 group = find_busiest_group(&env, balance);
4382 schedstat_inc(sd, lb_nobusyg[idle]);
4386 busiest = find_busiest_queue(&env, group);
4388 schedstat_inc(sd, lb_nobusyq[idle]);
4392 BUG_ON(busiest == env.dst_rq);
4394 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
4398 if (busiest->nr_running > 1) {
4400 * Attempt to move tasks. If find_busiest_group has found
4401 * an imbalance but busiest->nr_running <= 1, the group is
4402 * still unbalanced. ld_moved simply stays zero, so it is
4403 * correctly treated as an imbalance.
4405 env.flags |= LBF_ALL_PINNED;
4406 env.src_cpu = busiest->cpu;
4407 env.src_rq = busiest;
4408 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
4410 update_h_load(env.src_cpu);
4412 local_irq_save(flags);
4413 double_rq_lock(env.dst_rq, busiest);
4416 * cur_ld_moved - load moved in current iteration
4417 * ld_moved - cumulative load moved across iterations
4419 cur_ld_moved = move_tasks(&env);
4420 ld_moved += cur_ld_moved;
4421 double_rq_unlock(env.dst_rq, busiest);
4422 local_irq_restore(flags);
4424 if (env.flags & LBF_NEED_BREAK) {
4425 env.flags &= ~LBF_NEED_BREAK;
4430 * some other cpu did the load balance for us.
4432 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
4433 resched_cpu(env.dst_cpu);
4436 * Revisit (affine) tasks on src_cpu that couldn't be moved to
4437 * us and move them to an alternate dst_cpu in our sched_group
4438 * where they can run. The upper limit on how many times we
4439 * iterate on same src_cpu is dependent on number of cpus in our
4442 * This changes load balance semantics a bit on who can move
4443 * load to a given_cpu. In addition to the given_cpu itself
4444 * (or a ilb_cpu acting on its behalf where given_cpu is
4445 * nohz-idle), we now have balance_cpu in a position to move
4446 * load to given_cpu. In rare situations, this may cause
4447 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
4448 * _independently_ and at _same_ time to move some load to
4449 * given_cpu) causing exceess load to be moved to given_cpu.
4450 * This however should not happen so much in practice and
4451 * moreover subsequent load balance cycles should correct the
4452 * excess load moved.
4454 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0 &&
4455 lb_iterations++ < max_lb_iterations) {
4457 env.dst_rq = cpu_rq(env.new_dst_cpu);
4458 env.dst_cpu = env.new_dst_cpu;
4459 env.flags &= ~LBF_SOME_PINNED;
4461 env.loop_break = sched_nr_migrate_break;
4463 * Go back to "more_balance" rather than "redo" since we
4464 * need to continue with same src_cpu.
4469 /* All tasks on this runqueue were pinned by CPU affinity */
4470 if (unlikely(env.flags & LBF_ALL_PINNED)) {
4471 cpumask_clear_cpu(cpu_of(busiest), cpus);
4472 if (!cpumask_empty(cpus)) {
4474 env.loop_break = sched_nr_migrate_break;
4482 schedstat_inc(sd, lb_failed[idle]);
4484 * Increment the failure counter only on periodic balance.
4485 * We do not want newidle balance, which can be very
4486 * frequent, pollute the failure counter causing
4487 * excessive cache_hot migrations and active balances.
4489 if (idle != CPU_NEWLY_IDLE)
4490 sd->nr_balance_failed++;
4492 if (need_active_balance(&env)) {
4493 raw_spin_lock_irqsave(&busiest->lock, flags);
4495 /* don't kick the active_load_balance_cpu_stop,
4496 * if the curr task on busiest cpu can't be
4499 if (!cpumask_test_cpu(this_cpu,
4500 tsk_cpus_allowed(busiest->curr))) {
4501 raw_spin_unlock_irqrestore(&busiest->lock,
4503 env.flags |= LBF_ALL_PINNED;
4504 goto out_one_pinned;
4508 * ->active_balance synchronizes accesses to
4509 * ->active_balance_work. Once set, it's cleared
4510 * only after active load balance is finished.
4512 if (!busiest->active_balance) {
4513 busiest->active_balance = 1;
4514 busiest->push_cpu = this_cpu;
4517 raw_spin_unlock_irqrestore(&busiest->lock, flags);
4519 if (active_balance) {
4520 stop_one_cpu_nowait(cpu_of(busiest),
4521 active_load_balance_cpu_stop, busiest,
4522 &busiest->active_balance_work);
4526 * We've kicked active balancing, reset the failure
4529 sd->nr_balance_failed = sd->cache_nice_tries+1;
4532 sd->nr_balance_failed = 0;
4534 if (likely(!active_balance)) {
4535 /* We were unbalanced, so reset the balancing interval */
4536 sd->balance_interval = sd->min_interval;
4539 * If we've begun active balancing, start to back off. This
4540 * case may not be covered by the all_pinned logic if there
4541 * is only 1 task on the busy runqueue (because we don't call
4544 if (sd->balance_interval < sd->max_interval)
4545 sd->balance_interval *= 2;
4551 schedstat_inc(sd, lb_balanced[idle]);
4553 sd->nr_balance_failed = 0;
4556 /* tune up the balancing interval */
4557 if (((env.flags & LBF_ALL_PINNED) &&
4558 sd->balance_interval < MAX_PINNED_INTERVAL) ||
4559 (sd->balance_interval < sd->max_interval))
4560 sd->balance_interval *= 2;
4568 * idle_balance is called by schedule() if this_cpu is about to become
4569 * idle. Attempts to pull tasks from other CPUs.
4571 void idle_balance(int this_cpu, struct rq *this_rq)
4573 struct sched_domain *sd;
4574 int pulled_task = 0;
4575 unsigned long next_balance = jiffies + HZ;
4577 this_rq->idle_stamp = this_rq->clock;
4579 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4583 * Drop the rq->lock, but keep IRQ/preempt disabled.
4585 raw_spin_unlock(&this_rq->lock);
4587 update_shares(this_cpu);
4589 for_each_domain(this_cpu, sd) {
4590 unsigned long interval;
4593 if (!(sd->flags & SD_LOAD_BALANCE))
4596 if (sd->flags & SD_BALANCE_NEWIDLE) {
4597 /* If we've pulled tasks over stop searching: */
4598 pulled_task = load_balance(this_cpu, this_rq,
4599 sd, CPU_NEWLY_IDLE, &balance);
4602 interval = msecs_to_jiffies(sd->balance_interval);
4603 if (time_after(next_balance, sd->last_balance + interval))
4604 next_balance = sd->last_balance + interval;
4606 this_rq->idle_stamp = 0;
4612 raw_spin_lock(&this_rq->lock);
4614 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4616 * We are going idle. next_balance may be set based on
4617 * a busy processor. So reset next_balance.
4619 this_rq->next_balance = next_balance;
4624 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
4625 * running tasks off the busiest CPU onto idle CPUs. It requires at
4626 * least 1 task to be running on each physical CPU where possible, and
4627 * avoids physical / logical imbalances.
4629 static int active_load_balance_cpu_stop(void *data)
4631 struct rq *busiest_rq = data;
4632 int busiest_cpu = cpu_of(busiest_rq);
4633 int target_cpu = busiest_rq->push_cpu;
4634 struct rq *target_rq = cpu_rq(target_cpu);
4635 struct sched_domain *sd;
4637 raw_spin_lock_irq(&busiest_rq->lock);
4639 /* make sure the requested cpu hasn't gone down in the meantime */
4640 if (unlikely(busiest_cpu != smp_processor_id() ||
4641 !busiest_rq->active_balance))
4644 /* Is there any task to move? */
4645 if (busiest_rq->nr_running <= 1)
4649 * This condition is "impossible", if it occurs
4650 * we need to fix it. Originally reported by
4651 * Bjorn Helgaas on a 128-cpu setup.
4653 BUG_ON(busiest_rq == target_rq);
4655 /* move a task from busiest_rq to target_rq */
4656 double_lock_balance(busiest_rq, target_rq);
4658 /* Search for an sd spanning us and the target CPU. */
4660 for_each_domain(target_cpu, sd) {
4661 if ((sd->flags & SD_LOAD_BALANCE) &&
4662 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4667 struct lb_env env = {
4669 .dst_cpu = target_cpu,
4670 .dst_rq = target_rq,
4671 .src_cpu = busiest_rq->cpu,
4672 .src_rq = busiest_rq,
4676 schedstat_inc(sd, alb_count);
4678 if (move_one_task(&env))
4679 schedstat_inc(sd, alb_pushed);
4681 schedstat_inc(sd, alb_failed);
4684 double_unlock_balance(busiest_rq, target_rq);
4686 busiest_rq->active_balance = 0;
4687 raw_spin_unlock_irq(&busiest_rq->lock);
4693 * idle load balancing details
4694 * - When one of the busy CPUs notice that there may be an idle rebalancing
4695 * needed, they will kick the idle load balancer, which then does idle
4696 * load balancing for all the idle CPUs.
4699 cpumask_var_t idle_cpus_mask;
4701 unsigned long next_balance; /* in jiffy units */
4702 } nohz ____cacheline_aligned;
4704 static inline int find_new_ilb(int call_cpu)
4706 int ilb = cpumask_first(nohz.idle_cpus_mask);
4708 if (ilb < nr_cpu_ids && idle_cpu(ilb))
4715 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
4716 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
4717 * CPU (if there is one).
4719 static void nohz_balancer_kick(int cpu)
4723 nohz.next_balance++;
4725 ilb_cpu = find_new_ilb(cpu);
4727 if (ilb_cpu >= nr_cpu_ids)
4730 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
4733 * Use smp_send_reschedule() instead of resched_cpu().
4734 * This way we generate a sched IPI on the target cpu which
4735 * is idle. And the softirq performing nohz idle load balance
4736 * will be run before returning from the IPI.
4738 smp_send_reschedule(ilb_cpu);
4742 static inline void nohz_balance_exit_idle(int cpu)
4744 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
4745 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
4746 atomic_dec(&nohz.nr_cpus);
4747 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
4751 static inline void set_cpu_sd_state_busy(void)
4753 struct sched_domain *sd;
4754 int cpu = smp_processor_id();
4756 if (!test_bit(NOHZ_IDLE, nohz_flags(cpu)))
4758 clear_bit(NOHZ_IDLE, nohz_flags(cpu));
4761 for_each_domain(cpu, sd)
4762 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
4766 void set_cpu_sd_state_idle(void)
4768 struct sched_domain *sd;
4769 int cpu = smp_processor_id();
4771 if (test_bit(NOHZ_IDLE, nohz_flags(cpu)))
4773 set_bit(NOHZ_IDLE, nohz_flags(cpu));
4776 for_each_domain(cpu, sd)
4777 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
4782 * This routine will record that the cpu is going idle with tick stopped.
4783 * This info will be used in performing idle load balancing in the future.
4785 void nohz_balance_enter_idle(int cpu)
4788 * If this cpu is going down, then nothing needs to be done.
4790 if (!cpu_active(cpu))
4793 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
4796 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
4797 atomic_inc(&nohz.nr_cpus);
4798 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
4801 static int __cpuinit sched_ilb_notifier(struct notifier_block *nfb,
4802 unsigned long action, void *hcpu)
4804 switch (action & ~CPU_TASKS_FROZEN) {
4806 nohz_balance_exit_idle(smp_processor_id());
4814 static DEFINE_SPINLOCK(balancing);
4817 * Scale the max load_balance interval with the number of CPUs in the system.
4818 * This trades load-balance latency on larger machines for less cross talk.
4820 void update_max_interval(void)
4822 max_load_balance_interval = HZ*num_online_cpus()/10;
4826 * It checks each scheduling domain to see if it is due to be balanced,
4827 * and initiates a balancing operation if so.
4829 * Balancing parameters are set up in arch_init_sched_domains.
4831 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4834 struct rq *rq = cpu_rq(cpu);
4835 unsigned long interval;
4836 struct sched_domain *sd;
4837 /* Earliest time when we have to do rebalance again */
4838 unsigned long next_balance = jiffies + 60*HZ;
4839 int update_next_balance = 0;
4845 for_each_domain(cpu, sd) {
4846 if (!(sd->flags & SD_LOAD_BALANCE))
4849 interval = sd->balance_interval;
4850 if (idle != CPU_IDLE)
4851 interval *= sd->busy_factor;
4853 /* scale ms to jiffies */
4854 interval = msecs_to_jiffies(interval);
4855 interval = clamp(interval, 1UL, max_load_balance_interval);
4857 need_serialize = sd->flags & SD_SERIALIZE;
4859 if (need_serialize) {
4860 if (!spin_trylock(&balancing))
4864 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4865 if (load_balance(cpu, rq, sd, idle, &balance)) {
4867 * We've pulled tasks over so either we're no
4870 idle = CPU_NOT_IDLE;
4872 sd->last_balance = jiffies;
4875 spin_unlock(&balancing);
4877 if (time_after(next_balance, sd->last_balance + interval)) {
4878 next_balance = sd->last_balance + interval;
4879 update_next_balance = 1;
4883 * Stop the load balance at this level. There is another
4884 * CPU in our sched group which is doing load balancing more
4893 * next_balance will be updated only when there is a need.
4894 * When the cpu is attached to null domain for ex, it will not be
4897 if (likely(update_next_balance))
4898 rq->next_balance = next_balance;
4903 * In CONFIG_NO_HZ case, the idle balance kickee will do the
4904 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4906 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
4908 struct rq *this_rq = cpu_rq(this_cpu);
4912 if (idle != CPU_IDLE ||
4913 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
4916 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
4917 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
4921 * If this cpu gets work to do, stop the load balancing
4922 * work being done for other cpus. Next load
4923 * balancing owner will pick it up.
4928 rq = cpu_rq(balance_cpu);
4930 raw_spin_lock_irq(&rq->lock);
4931 update_rq_clock(rq);
4932 update_idle_cpu_load(rq);
4933 raw_spin_unlock_irq(&rq->lock);
4935 rebalance_domains(balance_cpu, CPU_IDLE);
4937 if (time_after(this_rq->next_balance, rq->next_balance))
4938 this_rq->next_balance = rq->next_balance;
4940 nohz.next_balance = this_rq->next_balance;
4942 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
4946 * Current heuristic for kicking the idle load balancer in the presence
4947 * of an idle cpu is the system.
4948 * - This rq has more than one task.
4949 * - At any scheduler domain level, this cpu's scheduler group has multiple
4950 * busy cpu's exceeding the group's power.
4951 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
4952 * domain span are idle.
4954 static inline int nohz_kick_needed(struct rq *rq, int cpu)
4956 unsigned long now = jiffies;
4957 struct sched_domain *sd;
4959 if (unlikely(idle_cpu(cpu)))
4963 * We may be recently in ticked or tickless idle mode. At the first
4964 * busy tick after returning from idle, we will update the busy stats.
4966 set_cpu_sd_state_busy();
4967 nohz_balance_exit_idle(cpu);
4970 * None are in tickless mode and hence no need for NOHZ idle load
4973 if (likely(!atomic_read(&nohz.nr_cpus)))
4976 if (time_before(now, nohz.next_balance))
4979 if (rq->nr_running >= 2)
4983 for_each_domain(cpu, sd) {
4984 struct sched_group *sg = sd->groups;
4985 struct sched_group_power *sgp = sg->sgp;
4986 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
4988 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
4989 goto need_kick_unlock;
4991 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
4992 && (cpumask_first_and(nohz.idle_cpus_mask,
4993 sched_domain_span(sd)) < cpu))
4994 goto need_kick_unlock;
4996 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
5008 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
5012 * run_rebalance_domains is triggered when needed from the scheduler tick.
5013 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
5015 static void run_rebalance_domains(struct softirq_action *h)
5017 int this_cpu = smp_processor_id();
5018 struct rq *this_rq = cpu_rq(this_cpu);
5019 enum cpu_idle_type idle = this_rq->idle_balance ?
5020 CPU_IDLE : CPU_NOT_IDLE;
5022 rebalance_domains(this_cpu, idle);
5025 * If this cpu has a pending nohz_balance_kick, then do the
5026 * balancing on behalf of the other idle cpus whose ticks are
5029 nohz_idle_balance(this_cpu, idle);
5032 static inline int on_null_domain(int cpu)
5034 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
5038 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
5040 void trigger_load_balance(struct rq *rq, int cpu)
5042 /* Don't need to rebalance while attached to NULL domain */
5043 if (time_after_eq(jiffies, rq->next_balance) &&
5044 likely(!on_null_domain(cpu)))
5045 raise_softirq(SCHED_SOFTIRQ);
5047 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
5048 nohz_balancer_kick(cpu);
5052 static void rq_online_fair(struct rq *rq)
5057 static void rq_offline_fair(struct rq *rq)
5061 /* Ensure any throttled groups are reachable by pick_next_task */
5062 unthrottle_offline_cfs_rqs(rq);
5065 #endif /* CONFIG_SMP */
5068 * scheduler tick hitting a task of our scheduling class:
5070 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
5072 struct cfs_rq *cfs_rq;
5073 struct sched_entity *se = &curr->se;
5075 for_each_sched_entity(se) {
5076 cfs_rq = cfs_rq_of(se);
5077 entity_tick(cfs_rq, se, queued);
5080 if (sched_feat_numa(NUMA))
5081 task_tick_numa(rq, curr);
5085 * called on fork with the child task as argument from the parent's context
5086 * - child not yet on the tasklist
5087 * - preemption disabled
5089 static void task_fork_fair(struct task_struct *p)
5091 struct cfs_rq *cfs_rq;
5092 struct sched_entity *se = &p->se, *curr;
5093 int this_cpu = smp_processor_id();
5094 struct rq *rq = this_rq();
5095 unsigned long flags;
5097 raw_spin_lock_irqsave(&rq->lock, flags);
5099 update_rq_clock(rq);
5101 cfs_rq = task_cfs_rq(current);
5102 curr = cfs_rq->curr;
5104 if (unlikely(task_cpu(p) != this_cpu)) {
5106 __set_task_cpu(p, this_cpu);
5110 update_curr(cfs_rq);
5113 se->vruntime = curr->vruntime;
5114 place_entity(cfs_rq, se, 1);
5116 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
5118 * Upon rescheduling, sched_class::put_prev_task() will place
5119 * 'current' within the tree based on its new key value.
5121 swap(curr->vruntime, se->vruntime);
5122 resched_task(rq->curr);
5125 se->vruntime -= cfs_rq->min_vruntime;
5127 raw_spin_unlock_irqrestore(&rq->lock, flags);
5131 * Priority of the task has changed. Check to see if we preempt
5135 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
5141 * Reschedule if we are currently running on this runqueue and
5142 * our priority decreased, or if we are not currently running on
5143 * this runqueue and our priority is higher than the current's
5145 if (rq->curr == p) {
5146 if (p->prio > oldprio)
5147 resched_task(rq->curr);
5149 check_preempt_curr(rq, p, 0);
5152 static void switched_from_fair(struct rq *rq, struct task_struct *p)
5154 struct sched_entity *se = &p->se;
5155 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5158 * Ensure the task's vruntime is normalized, so that when its
5159 * switched back to the fair class the enqueue_entity(.flags=0) will
5160 * do the right thing.
5162 * If it was on_rq, then the dequeue_entity(.flags=0) will already
5163 * have normalized the vruntime, if it was !on_rq, then only when
5164 * the task is sleeping will it still have non-normalized vruntime.
5166 if (!se->on_rq && p->state != TASK_RUNNING) {
5168 * Fix up our vruntime so that the current sleep doesn't
5169 * cause 'unlimited' sleep bonus.
5171 place_entity(cfs_rq, se, 0);
5172 se->vruntime -= cfs_rq->min_vruntime;
5177 * We switched to the sched_fair class.
5179 static void switched_to_fair(struct rq *rq, struct task_struct *p)
5185 * We were most likely switched from sched_rt, so
5186 * kick off the schedule if running, otherwise just see
5187 * if we can still preempt the current task.
5190 resched_task(rq->curr);
5192 check_preempt_curr(rq, p, 0);
5195 /* Account for a task changing its policy or group.
5197 * This routine is mostly called to set cfs_rq->curr field when a task
5198 * migrates between groups/classes.
5200 static void set_curr_task_fair(struct rq *rq)
5202 struct sched_entity *se = &rq->curr->se;
5204 for_each_sched_entity(se) {
5205 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5207 set_next_entity(cfs_rq, se);
5208 /* ensure bandwidth has been allocated on our new cfs_rq */
5209 account_cfs_rq_runtime(cfs_rq, 0);
5213 void init_cfs_rq(struct cfs_rq *cfs_rq)
5215 cfs_rq->tasks_timeline = RB_ROOT;
5216 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
5217 #ifndef CONFIG_64BIT
5218 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
5222 #ifdef CONFIG_FAIR_GROUP_SCHED
5223 static void task_move_group_fair(struct task_struct *p, int on_rq)
5226 * If the task was not on the rq at the time of this cgroup movement
5227 * it must have been asleep, sleeping tasks keep their ->vruntime
5228 * absolute on their old rq until wakeup (needed for the fair sleeper
5229 * bonus in place_entity()).
5231 * If it was on the rq, we've just 'preempted' it, which does convert
5232 * ->vruntime to a relative base.
5234 * Make sure both cases convert their relative position when migrating
5235 * to another cgroup's rq. This does somewhat interfere with the
5236 * fair sleeper stuff for the first placement, but who cares.
5239 * When !on_rq, vruntime of the task has usually NOT been normalized.
5240 * But there are some cases where it has already been normalized:
5242 * - Moving a forked child which is waiting for being woken up by
5243 * wake_up_new_task().
5244 * - Moving a task which has been woken up by try_to_wake_up() and
5245 * waiting for actually being woken up by sched_ttwu_pending().
5247 * To prevent boost or penalty in the new cfs_rq caused by delta
5248 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
5250 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
5254 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
5255 set_task_rq(p, task_cpu(p));
5257 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
5260 void free_fair_sched_group(struct task_group *tg)
5264 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
5266 for_each_possible_cpu(i) {
5268 kfree(tg->cfs_rq[i]);
5277 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
5279 struct cfs_rq *cfs_rq;
5280 struct sched_entity *se;
5283 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
5286 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
5290 tg->shares = NICE_0_LOAD;
5292 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
5294 for_each_possible_cpu(i) {
5295 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
5296 GFP_KERNEL, cpu_to_node(i));
5300 se = kzalloc_node(sizeof(struct sched_entity),
5301 GFP_KERNEL, cpu_to_node(i));
5305 init_cfs_rq(cfs_rq);
5306 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
5317 void unregister_fair_sched_group(struct task_group *tg, int cpu)
5319 struct rq *rq = cpu_rq(cpu);
5320 unsigned long flags;
5323 * Only empty task groups can be destroyed; so we can speculatively
5324 * check on_list without danger of it being re-added.
5326 if (!tg->cfs_rq[cpu]->on_list)
5329 raw_spin_lock_irqsave(&rq->lock, flags);
5330 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
5331 raw_spin_unlock_irqrestore(&rq->lock, flags);
5334 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
5335 struct sched_entity *se, int cpu,
5336 struct sched_entity *parent)
5338 struct rq *rq = cpu_rq(cpu);
5343 /* allow initial update_cfs_load() to truncate */
5344 cfs_rq->load_stamp = 1;
5346 init_cfs_rq_runtime(cfs_rq);
5348 tg->cfs_rq[cpu] = cfs_rq;
5351 /* se could be NULL for root_task_group */
5356 se->cfs_rq = &rq->cfs;
5358 se->cfs_rq = parent->my_q;
5361 update_load_set(&se->load, 0);
5362 se->parent = parent;
5365 static DEFINE_MUTEX(shares_mutex);
5367 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
5370 unsigned long flags;
5373 * We can't change the weight of the root cgroup.
5378 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
5380 mutex_lock(&shares_mutex);
5381 if (tg->shares == shares)
5384 tg->shares = shares;
5385 for_each_possible_cpu(i) {
5386 struct rq *rq = cpu_rq(i);
5387 struct sched_entity *se;
5390 /* Propagate contribution to hierarchy */
5391 raw_spin_lock_irqsave(&rq->lock, flags);
5392 for_each_sched_entity(se)
5393 update_cfs_shares(group_cfs_rq(se));
5394 raw_spin_unlock_irqrestore(&rq->lock, flags);
5398 mutex_unlock(&shares_mutex);
5401 #else /* CONFIG_FAIR_GROUP_SCHED */
5403 void free_fair_sched_group(struct task_group *tg) { }
5405 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
5410 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
5412 #endif /* CONFIG_FAIR_GROUP_SCHED */
5415 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
5417 struct sched_entity *se = &task->se;
5418 unsigned int rr_interval = 0;
5421 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
5424 if (rq->cfs.load.weight)
5425 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5431 * All the scheduling class methods:
5433 const struct sched_class fair_sched_class = {
5434 .next = &idle_sched_class,
5435 .enqueue_task = enqueue_task_fair,
5436 .dequeue_task = dequeue_task_fair,
5437 .yield_task = yield_task_fair,
5438 .yield_to_task = yield_to_task_fair,
5440 .check_preempt_curr = check_preempt_wakeup,
5442 .pick_next_task = pick_next_task_fair,
5443 .put_prev_task = put_prev_task_fair,
5446 .select_task_rq = select_task_rq_fair,
5448 .rq_online = rq_online_fair,
5449 .rq_offline = rq_offline_fair,
5451 .task_waking = task_waking_fair,
5454 .set_curr_task = set_curr_task_fair,
5455 .task_tick = task_tick_fair,
5456 .task_fork = task_fork_fair,
5458 .prio_changed = prio_changed_fair,
5459 .switched_from = switched_from_fair,
5460 .switched_to = switched_to_fair,
5462 .get_rr_interval = get_rr_interval_fair,
5464 #ifdef CONFIG_FAIR_GROUP_SCHED
5465 .task_move_group = task_move_group_fair,
5469 #ifdef CONFIG_SCHED_DEBUG
5470 void print_cfs_stats(struct seq_file *m, int cpu)
5472 struct cfs_rq *cfs_rq;
5475 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
5476 print_cfs_rq(m, cpu, cfs_rq);
5481 __init void init_sched_fair_class(void)
5484 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
5487 nohz.next_balance = jiffies;
5488 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
5489 cpu_notifier(sched_ilb_notifier, 0);