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
785 unsigned int sysctl_numa_balancing_scan_period_min = 100;
786 unsigned int sysctl_numa_balancing_scan_period_max = 100*16;
788 /* Portion of address space to scan in MB */
789 unsigned int sysctl_numa_balancing_scan_size = 256;
791 static void task_numa_placement(struct task_struct *p)
793 int seq = ACCESS_ONCE(p->mm->numa_scan_seq);
795 if (p->numa_scan_seq == seq)
797 p->numa_scan_seq = seq;
799 /* FIXME: Scheduling placement policy hints go here */
803 * Got a PROT_NONE fault for a page on @node.
805 void task_numa_fault(int node, int pages)
807 struct task_struct *p = current;
809 /* FIXME: Allocate task-specific structure for placement policy here */
811 task_numa_placement(p);
814 static void reset_ptenuma_scan(struct task_struct *p)
816 ACCESS_ONCE(p->mm->numa_scan_seq)++;
817 p->mm->numa_scan_offset = 0;
821 * The expensive part of numa migration is done from task_work context.
822 * Triggered from task_tick_numa().
824 void task_numa_work(struct callback_head *work)
826 unsigned long migrate, next_scan, now = jiffies;
827 struct task_struct *p = current;
828 struct mm_struct *mm = p->mm;
829 struct vm_area_struct *vma;
830 unsigned long start, end;
833 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
835 work->next = work; /* protect against double add */
837 * Who cares about NUMA placement when they're dying.
839 * NOTE: make sure not to dereference p->mm before this check,
840 * exit_task_work() happens _after_ exit_mm() so we could be called
841 * without p->mm even though we still had it when we enqueued this
844 if (p->flags & PF_EXITING)
848 * Enforce maximal scan/migration frequency..
850 migrate = mm->numa_next_scan;
851 if (time_before(now, migrate))
854 if (p->numa_scan_period == 0)
855 p->numa_scan_period = sysctl_numa_balancing_scan_period_min;
857 next_scan = now + 2*msecs_to_jiffies(p->numa_scan_period);
858 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
861 start = mm->numa_scan_offset;
862 pages = sysctl_numa_balancing_scan_size;
863 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
867 down_read(&mm->mmap_sem);
868 vma = find_vma(mm, start);
870 reset_ptenuma_scan(p);
874 for (; vma; vma = vma->vm_next) {
875 if (!vma_migratable(vma))
878 /* Skip small VMAs. They are not likely to be of relevance */
879 if (((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) < HPAGE_PMD_NR)
883 start = max(start, vma->vm_start);
884 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
885 end = min(end, vma->vm_end);
886 pages -= change_prot_numa(vma, start, end);
891 } while (end != vma->vm_end);
896 * It is possible to reach the end of the VMA list but the last few VMAs are
897 * not guaranteed to the vma_migratable. If they are not, we would find the
898 * !migratable VMA on the next scan but not reset the scanner to the start
902 mm->numa_scan_offset = start;
904 reset_ptenuma_scan(p);
905 up_read(&mm->mmap_sem);
909 * Drive the periodic memory faults..
911 void task_tick_numa(struct rq *rq, struct task_struct *curr)
913 struct callback_head *work = &curr->numa_work;
917 * We don't care about NUMA placement if we don't have memory.
919 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
923 * Using runtime rather than walltime has the dual advantage that
924 * we (mostly) drive the selection from busy threads and that the
925 * task needs to have done some actual work before we bother with
928 now = curr->se.sum_exec_runtime;
929 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
931 if (now - curr->node_stamp > period) {
932 curr->node_stamp = now;
934 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
935 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
936 task_work_add(curr, work, true);
941 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
944 #endif /* CONFIG_NUMA_BALANCING */
947 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
949 update_load_add(&cfs_rq->load, se->load.weight);
950 if (!parent_entity(se))
951 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
953 if (entity_is_task(se))
954 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
956 cfs_rq->nr_running++;
960 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
962 update_load_sub(&cfs_rq->load, se->load.weight);
963 if (!parent_entity(se))
964 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
965 if (entity_is_task(se))
966 list_del_init(&se->group_node);
967 cfs_rq->nr_running--;
970 #ifdef CONFIG_FAIR_GROUP_SCHED
971 /* we need this in update_cfs_load and load-balance functions below */
972 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
974 static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq,
977 struct task_group *tg = cfs_rq->tg;
980 load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1);
981 load_avg -= cfs_rq->load_contribution;
983 if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) {
984 atomic_add(load_avg, &tg->load_weight);
985 cfs_rq->load_contribution += load_avg;
989 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
991 u64 period = sysctl_sched_shares_window;
993 unsigned long load = cfs_rq->load.weight;
995 if (cfs_rq->tg == &root_task_group || throttled_hierarchy(cfs_rq))
998 now = rq_of(cfs_rq)->clock_task;
999 delta = now - cfs_rq->load_stamp;
1001 /* truncate load history at 4 idle periods */
1002 if (cfs_rq->load_stamp > cfs_rq->load_last &&
1003 now - cfs_rq->load_last > 4 * period) {
1004 cfs_rq->load_period = 0;
1005 cfs_rq->load_avg = 0;
1009 cfs_rq->load_stamp = now;
1010 cfs_rq->load_unacc_exec_time = 0;
1011 cfs_rq->load_period += delta;
1013 cfs_rq->load_last = now;
1014 cfs_rq->load_avg += delta * load;
1017 /* consider updating load contribution on each fold or truncate */
1018 if (global_update || cfs_rq->load_period > period
1019 || !cfs_rq->load_period)
1020 update_cfs_rq_load_contribution(cfs_rq, global_update);
1022 while (cfs_rq->load_period > period) {
1024 * Inline assembly required to prevent the compiler
1025 * optimising this loop into a divmod call.
1026 * See __iter_div_u64_rem() for another example of this.
1028 asm("" : "+rm" (cfs_rq->load_period));
1029 cfs_rq->load_period /= 2;
1030 cfs_rq->load_avg /= 2;
1033 if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg)
1034 list_del_leaf_cfs_rq(cfs_rq);
1037 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1042 * Use this CPU's actual weight instead of the last load_contribution
1043 * to gain a more accurate current total weight. See
1044 * update_cfs_rq_load_contribution().
1046 tg_weight = atomic_read(&tg->load_weight);
1047 tg_weight -= cfs_rq->load_contribution;
1048 tg_weight += cfs_rq->load.weight;
1053 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1055 long tg_weight, load, shares;
1057 tg_weight = calc_tg_weight(tg, cfs_rq);
1058 load = cfs_rq->load.weight;
1060 shares = (tg->shares * load);
1062 shares /= tg_weight;
1064 if (shares < MIN_SHARES)
1065 shares = MIN_SHARES;
1066 if (shares > tg->shares)
1067 shares = tg->shares;
1072 static void update_entity_shares_tick(struct cfs_rq *cfs_rq)
1074 if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) {
1075 update_cfs_load(cfs_rq, 0);
1076 update_cfs_shares(cfs_rq);
1079 # else /* CONFIG_SMP */
1080 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
1084 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1089 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
1092 # endif /* CONFIG_SMP */
1093 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1094 unsigned long weight)
1097 /* commit outstanding execution time */
1098 if (cfs_rq->curr == se)
1099 update_curr(cfs_rq);
1100 account_entity_dequeue(cfs_rq, se);
1103 update_load_set(&se->load, weight);
1106 account_entity_enqueue(cfs_rq, se);
1109 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1111 struct task_group *tg;
1112 struct sched_entity *se;
1116 se = tg->se[cpu_of(rq_of(cfs_rq))];
1117 if (!se || throttled_hierarchy(cfs_rq))
1120 if (likely(se->load.weight == tg->shares))
1123 shares = calc_cfs_shares(cfs_rq, tg);
1125 reweight_entity(cfs_rq_of(se), se, shares);
1127 #else /* CONFIG_FAIR_GROUP_SCHED */
1128 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
1132 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1136 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
1139 #endif /* CONFIG_FAIR_GROUP_SCHED */
1141 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1143 #ifdef CONFIG_SCHEDSTATS
1144 struct task_struct *tsk = NULL;
1146 if (entity_is_task(se))
1149 if (se->statistics.sleep_start) {
1150 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
1155 if (unlikely(delta > se->statistics.sleep_max))
1156 se->statistics.sleep_max = delta;
1158 se->statistics.sleep_start = 0;
1159 se->statistics.sum_sleep_runtime += delta;
1162 account_scheduler_latency(tsk, delta >> 10, 1);
1163 trace_sched_stat_sleep(tsk, delta);
1166 if (se->statistics.block_start) {
1167 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
1172 if (unlikely(delta > se->statistics.block_max))
1173 se->statistics.block_max = delta;
1175 se->statistics.block_start = 0;
1176 se->statistics.sum_sleep_runtime += delta;
1179 if (tsk->in_iowait) {
1180 se->statistics.iowait_sum += delta;
1181 se->statistics.iowait_count++;
1182 trace_sched_stat_iowait(tsk, delta);
1185 trace_sched_stat_blocked(tsk, delta);
1188 * Blocking time is in units of nanosecs, so shift by
1189 * 20 to get a milliseconds-range estimation of the
1190 * amount of time that the task spent sleeping:
1192 if (unlikely(prof_on == SLEEP_PROFILING)) {
1193 profile_hits(SLEEP_PROFILING,
1194 (void *)get_wchan(tsk),
1197 account_scheduler_latency(tsk, delta >> 10, 0);
1203 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1205 #ifdef CONFIG_SCHED_DEBUG
1206 s64 d = se->vruntime - cfs_rq->min_vruntime;
1211 if (d > 3*sysctl_sched_latency)
1212 schedstat_inc(cfs_rq, nr_spread_over);
1217 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1219 u64 vruntime = cfs_rq->min_vruntime;
1222 * The 'current' period is already promised to the current tasks,
1223 * however the extra weight of the new task will slow them down a
1224 * little, place the new task so that it fits in the slot that
1225 * stays open at the end.
1227 if (initial && sched_feat(START_DEBIT))
1228 vruntime += sched_vslice(cfs_rq, se);
1230 /* sleeps up to a single latency don't count. */
1232 unsigned long thresh = sysctl_sched_latency;
1235 * Halve their sleep time's effect, to allow
1236 * for a gentler effect of sleepers:
1238 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1244 /* ensure we never gain time by being placed backwards. */
1245 vruntime = max_vruntime(se->vruntime, vruntime);
1247 se->vruntime = vruntime;
1250 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1253 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1256 * Update the normalized vruntime before updating min_vruntime
1257 * through callig update_curr().
1259 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1260 se->vruntime += cfs_rq->min_vruntime;
1263 * Update run-time statistics of the 'current'.
1265 update_curr(cfs_rq);
1266 update_cfs_load(cfs_rq, 0);
1267 account_entity_enqueue(cfs_rq, se);
1268 update_cfs_shares(cfs_rq);
1270 if (flags & ENQUEUE_WAKEUP) {
1271 place_entity(cfs_rq, se, 0);
1272 enqueue_sleeper(cfs_rq, se);
1275 update_stats_enqueue(cfs_rq, se);
1276 check_spread(cfs_rq, se);
1277 if (se != cfs_rq->curr)
1278 __enqueue_entity(cfs_rq, se);
1281 if (cfs_rq->nr_running == 1) {
1282 list_add_leaf_cfs_rq(cfs_rq);
1283 check_enqueue_throttle(cfs_rq);
1287 static void __clear_buddies_last(struct sched_entity *se)
1289 for_each_sched_entity(se) {
1290 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1291 if (cfs_rq->last == se)
1292 cfs_rq->last = NULL;
1298 static void __clear_buddies_next(struct sched_entity *se)
1300 for_each_sched_entity(se) {
1301 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1302 if (cfs_rq->next == se)
1303 cfs_rq->next = NULL;
1309 static void __clear_buddies_skip(struct sched_entity *se)
1311 for_each_sched_entity(se) {
1312 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1313 if (cfs_rq->skip == se)
1314 cfs_rq->skip = NULL;
1320 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1322 if (cfs_rq->last == se)
1323 __clear_buddies_last(se);
1325 if (cfs_rq->next == se)
1326 __clear_buddies_next(se);
1328 if (cfs_rq->skip == se)
1329 __clear_buddies_skip(se);
1332 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1335 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1338 * Update run-time statistics of the 'current'.
1340 update_curr(cfs_rq);
1342 update_stats_dequeue(cfs_rq, se);
1343 if (flags & DEQUEUE_SLEEP) {
1344 #ifdef CONFIG_SCHEDSTATS
1345 if (entity_is_task(se)) {
1346 struct task_struct *tsk = task_of(se);
1348 if (tsk->state & TASK_INTERRUPTIBLE)
1349 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1350 if (tsk->state & TASK_UNINTERRUPTIBLE)
1351 se->statistics.block_start = rq_of(cfs_rq)->clock;
1356 clear_buddies(cfs_rq, se);
1358 if (se != cfs_rq->curr)
1359 __dequeue_entity(cfs_rq, se);
1361 update_cfs_load(cfs_rq, 0);
1362 account_entity_dequeue(cfs_rq, se);
1365 * Normalize the entity after updating the min_vruntime because the
1366 * update can refer to the ->curr item and we need to reflect this
1367 * movement in our normalized position.
1369 if (!(flags & DEQUEUE_SLEEP))
1370 se->vruntime -= cfs_rq->min_vruntime;
1372 /* return excess runtime on last dequeue */
1373 return_cfs_rq_runtime(cfs_rq);
1375 update_min_vruntime(cfs_rq);
1376 update_cfs_shares(cfs_rq);
1380 * Preempt the current task with a newly woken task if needed:
1383 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1385 unsigned long ideal_runtime, delta_exec;
1386 struct sched_entity *se;
1389 ideal_runtime = sched_slice(cfs_rq, curr);
1390 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1391 if (delta_exec > ideal_runtime) {
1392 resched_task(rq_of(cfs_rq)->curr);
1394 * The current task ran long enough, ensure it doesn't get
1395 * re-elected due to buddy favours.
1397 clear_buddies(cfs_rq, curr);
1402 * Ensure that a task that missed wakeup preemption by a
1403 * narrow margin doesn't have to wait for a full slice.
1404 * This also mitigates buddy induced latencies under load.
1406 if (delta_exec < sysctl_sched_min_granularity)
1409 se = __pick_first_entity(cfs_rq);
1410 delta = curr->vruntime - se->vruntime;
1415 if (delta > ideal_runtime)
1416 resched_task(rq_of(cfs_rq)->curr);
1420 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1422 /* 'current' is not kept within the tree. */
1425 * Any task has to be enqueued before it get to execute on
1426 * a CPU. So account for the time it spent waiting on the
1429 update_stats_wait_end(cfs_rq, se);
1430 __dequeue_entity(cfs_rq, se);
1433 update_stats_curr_start(cfs_rq, se);
1435 #ifdef CONFIG_SCHEDSTATS
1437 * Track our maximum slice length, if the CPU's load is at
1438 * least twice that of our own weight (i.e. dont track it
1439 * when there are only lesser-weight tasks around):
1441 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1442 se->statistics.slice_max = max(se->statistics.slice_max,
1443 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1446 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1450 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1453 * Pick the next process, keeping these things in mind, in this order:
1454 * 1) keep things fair between processes/task groups
1455 * 2) pick the "next" process, since someone really wants that to run
1456 * 3) pick the "last" process, for cache locality
1457 * 4) do not run the "skip" process, if something else is available
1459 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1461 struct sched_entity *se = __pick_first_entity(cfs_rq);
1462 struct sched_entity *left = se;
1465 * Avoid running the skip buddy, if running something else can
1466 * be done without getting too unfair.
1468 if (cfs_rq->skip == se) {
1469 struct sched_entity *second = __pick_next_entity(se);
1470 if (second && wakeup_preempt_entity(second, left) < 1)
1475 * Prefer last buddy, try to return the CPU to a preempted task.
1477 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1481 * Someone really wants this to run. If it's not unfair, run it.
1483 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1486 clear_buddies(cfs_rq, se);
1491 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1493 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1496 * If still on the runqueue then deactivate_task()
1497 * was not called and update_curr() has to be done:
1500 update_curr(cfs_rq);
1502 /* throttle cfs_rqs exceeding runtime */
1503 check_cfs_rq_runtime(cfs_rq);
1505 check_spread(cfs_rq, prev);
1507 update_stats_wait_start(cfs_rq, prev);
1508 /* Put 'current' back into the tree. */
1509 __enqueue_entity(cfs_rq, prev);
1511 cfs_rq->curr = NULL;
1515 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1518 * Update run-time statistics of the 'current'.
1520 update_curr(cfs_rq);
1523 * Update share accounting for long-running entities.
1525 update_entity_shares_tick(cfs_rq);
1527 #ifdef CONFIG_SCHED_HRTICK
1529 * queued ticks are scheduled to match the slice, so don't bother
1530 * validating it and just reschedule.
1533 resched_task(rq_of(cfs_rq)->curr);
1537 * don't let the period tick interfere with the hrtick preemption
1539 if (!sched_feat(DOUBLE_TICK) &&
1540 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1544 if (cfs_rq->nr_running > 1)
1545 check_preempt_tick(cfs_rq, curr);
1549 /**************************************************
1550 * CFS bandwidth control machinery
1553 #ifdef CONFIG_CFS_BANDWIDTH
1555 #ifdef HAVE_JUMP_LABEL
1556 static struct static_key __cfs_bandwidth_used;
1558 static inline bool cfs_bandwidth_used(void)
1560 return static_key_false(&__cfs_bandwidth_used);
1563 void account_cfs_bandwidth_used(int enabled, int was_enabled)
1565 /* only need to count groups transitioning between enabled/!enabled */
1566 if (enabled && !was_enabled)
1567 static_key_slow_inc(&__cfs_bandwidth_used);
1568 else if (!enabled && was_enabled)
1569 static_key_slow_dec(&__cfs_bandwidth_used);
1571 #else /* HAVE_JUMP_LABEL */
1572 static bool cfs_bandwidth_used(void)
1577 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
1578 #endif /* HAVE_JUMP_LABEL */
1581 * default period for cfs group bandwidth.
1582 * default: 0.1s, units: nanoseconds
1584 static inline u64 default_cfs_period(void)
1586 return 100000000ULL;
1589 static inline u64 sched_cfs_bandwidth_slice(void)
1591 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
1595 * Replenish runtime according to assigned quota and update expiration time.
1596 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
1597 * additional synchronization around rq->lock.
1599 * requires cfs_b->lock
1601 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
1605 if (cfs_b->quota == RUNTIME_INF)
1608 now = sched_clock_cpu(smp_processor_id());
1609 cfs_b->runtime = cfs_b->quota;
1610 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
1613 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
1615 return &tg->cfs_bandwidth;
1618 /* returns 0 on failure to allocate runtime */
1619 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1621 struct task_group *tg = cfs_rq->tg;
1622 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
1623 u64 amount = 0, min_amount, expires;
1625 /* note: this is a positive sum as runtime_remaining <= 0 */
1626 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
1628 raw_spin_lock(&cfs_b->lock);
1629 if (cfs_b->quota == RUNTIME_INF)
1630 amount = min_amount;
1633 * If the bandwidth pool has become inactive, then at least one
1634 * period must have elapsed since the last consumption.
1635 * Refresh the global state and ensure bandwidth timer becomes
1638 if (!cfs_b->timer_active) {
1639 __refill_cfs_bandwidth_runtime(cfs_b);
1640 __start_cfs_bandwidth(cfs_b);
1643 if (cfs_b->runtime > 0) {
1644 amount = min(cfs_b->runtime, min_amount);
1645 cfs_b->runtime -= amount;
1649 expires = cfs_b->runtime_expires;
1650 raw_spin_unlock(&cfs_b->lock);
1652 cfs_rq->runtime_remaining += amount;
1654 * we may have advanced our local expiration to account for allowed
1655 * spread between our sched_clock and the one on which runtime was
1658 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
1659 cfs_rq->runtime_expires = expires;
1661 return cfs_rq->runtime_remaining > 0;
1665 * Note: This depends on the synchronization provided by sched_clock and the
1666 * fact that rq->clock snapshots this value.
1668 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1670 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1671 struct rq *rq = rq_of(cfs_rq);
1673 /* if the deadline is ahead of our clock, nothing to do */
1674 if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0))
1677 if (cfs_rq->runtime_remaining < 0)
1681 * If the local deadline has passed we have to consider the
1682 * possibility that our sched_clock is 'fast' and the global deadline
1683 * has not truly expired.
1685 * Fortunately we can check determine whether this the case by checking
1686 * whether the global deadline has advanced.
1689 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
1690 /* extend local deadline, drift is bounded above by 2 ticks */
1691 cfs_rq->runtime_expires += TICK_NSEC;
1693 /* global deadline is ahead, expiration has passed */
1694 cfs_rq->runtime_remaining = 0;
1698 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1699 unsigned long delta_exec)
1701 /* dock delta_exec before expiring quota (as it could span periods) */
1702 cfs_rq->runtime_remaining -= delta_exec;
1703 expire_cfs_rq_runtime(cfs_rq);
1705 if (likely(cfs_rq->runtime_remaining > 0))
1709 * if we're unable to extend our runtime we resched so that the active
1710 * hierarchy can be throttled
1712 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
1713 resched_task(rq_of(cfs_rq)->curr);
1716 static __always_inline
1717 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
1719 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
1722 __account_cfs_rq_runtime(cfs_rq, delta_exec);
1725 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
1727 return cfs_bandwidth_used() && cfs_rq->throttled;
1730 /* check whether cfs_rq, or any parent, is throttled */
1731 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
1733 return cfs_bandwidth_used() && cfs_rq->throttle_count;
1737 * Ensure that neither of the group entities corresponding to src_cpu or
1738 * dest_cpu are members of a throttled hierarchy when performing group
1739 * load-balance operations.
1741 static inline int throttled_lb_pair(struct task_group *tg,
1742 int src_cpu, int dest_cpu)
1744 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
1746 src_cfs_rq = tg->cfs_rq[src_cpu];
1747 dest_cfs_rq = tg->cfs_rq[dest_cpu];
1749 return throttled_hierarchy(src_cfs_rq) ||
1750 throttled_hierarchy(dest_cfs_rq);
1753 /* updated child weight may affect parent so we have to do this bottom up */
1754 static int tg_unthrottle_up(struct task_group *tg, void *data)
1756 struct rq *rq = data;
1757 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1759 cfs_rq->throttle_count--;
1761 if (!cfs_rq->throttle_count) {
1762 u64 delta = rq->clock_task - cfs_rq->load_stamp;
1764 /* leaving throttled state, advance shares averaging windows */
1765 cfs_rq->load_stamp += delta;
1766 cfs_rq->load_last += delta;
1768 /* update entity weight now that we are on_rq again */
1769 update_cfs_shares(cfs_rq);
1776 static int tg_throttle_down(struct task_group *tg, void *data)
1778 struct rq *rq = data;
1779 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1781 /* group is entering throttled state, record last load */
1782 if (!cfs_rq->throttle_count)
1783 update_cfs_load(cfs_rq, 0);
1784 cfs_rq->throttle_count++;
1789 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
1791 struct rq *rq = rq_of(cfs_rq);
1792 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1793 struct sched_entity *se;
1794 long task_delta, dequeue = 1;
1796 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1798 /* account load preceding throttle */
1800 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
1803 task_delta = cfs_rq->h_nr_running;
1804 for_each_sched_entity(se) {
1805 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
1806 /* throttled entity or throttle-on-deactivate */
1811 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
1812 qcfs_rq->h_nr_running -= task_delta;
1814 if (qcfs_rq->load.weight)
1819 rq->nr_running -= task_delta;
1821 cfs_rq->throttled = 1;
1822 cfs_rq->throttled_timestamp = rq->clock;
1823 raw_spin_lock(&cfs_b->lock);
1824 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
1825 raw_spin_unlock(&cfs_b->lock);
1828 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
1830 struct rq *rq = rq_of(cfs_rq);
1831 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1832 struct sched_entity *se;
1836 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1838 cfs_rq->throttled = 0;
1839 raw_spin_lock(&cfs_b->lock);
1840 cfs_b->throttled_time += rq->clock - cfs_rq->throttled_timestamp;
1841 list_del_rcu(&cfs_rq->throttled_list);
1842 raw_spin_unlock(&cfs_b->lock);
1843 cfs_rq->throttled_timestamp = 0;
1845 update_rq_clock(rq);
1846 /* update hierarchical throttle state */
1847 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
1849 if (!cfs_rq->load.weight)
1852 task_delta = cfs_rq->h_nr_running;
1853 for_each_sched_entity(se) {
1857 cfs_rq = cfs_rq_of(se);
1859 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
1860 cfs_rq->h_nr_running += task_delta;
1862 if (cfs_rq_throttled(cfs_rq))
1867 rq->nr_running += task_delta;
1869 /* determine whether we need to wake up potentially idle cpu */
1870 if (rq->curr == rq->idle && rq->cfs.nr_running)
1871 resched_task(rq->curr);
1874 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
1875 u64 remaining, u64 expires)
1877 struct cfs_rq *cfs_rq;
1878 u64 runtime = remaining;
1881 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
1883 struct rq *rq = rq_of(cfs_rq);
1885 raw_spin_lock(&rq->lock);
1886 if (!cfs_rq_throttled(cfs_rq))
1889 runtime = -cfs_rq->runtime_remaining + 1;
1890 if (runtime > remaining)
1891 runtime = remaining;
1892 remaining -= runtime;
1894 cfs_rq->runtime_remaining += runtime;
1895 cfs_rq->runtime_expires = expires;
1897 /* we check whether we're throttled above */
1898 if (cfs_rq->runtime_remaining > 0)
1899 unthrottle_cfs_rq(cfs_rq);
1902 raw_spin_unlock(&rq->lock);
1913 * Responsible for refilling a task_group's bandwidth and unthrottling its
1914 * cfs_rqs as appropriate. If there has been no activity within the last
1915 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
1916 * used to track this state.
1918 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
1920 u64 runtime, runtime_expires;
1921 int idle = 1, throttled;
1923 raw_spin_lock(&cfs_b->lock);
1924 /* no need to continue the timer with no bandwidth constraint */
1925 if (cfs_b->quota == RUNTIME_INF)
1928 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
1929 /* idle depends on !throttled (for the case of a large deficit) */
1930 idle = cfs_b->idle && !throttled;
1931 cfs_b->nr_periods += overrun;
1933 /* if we're going inactive then everything else can be deferred */
1937 __refill_cfs_bandwidth_runtime(cfs_b);
1940 /* mark as potentially idle for the upcoming period */
1945 /* account preceding periods in which throttling occurred */
1946 cfs_b->nr_throttled += overrun;
1949 * There are throttled entities so we must first use the new bandwidth
1950 * to unthrottle them before making it generally available. This
1951 * ensures that all existing debts will be paid before a new cfs_rq is
1954 runtime = cfs_b->runtime;
1955 runtime_expires = cfs_b->runtime_expires;
1959 * This check is repeated as we are holding onto the new bandwidth
1960 * while we unthrottle. This can potentially race with an unthrottled
1961 * group trying to acquire new bandwidth from the global pool.
1963 while (throttled && runtime > 0) {
1964 raw_spin_unlock(&cfs_b->lock);
1965 /* we can't nest cfs_b->lock while distributing bandwidth */
1966 runtime = distribute_cfs_runtime(cfs_b, runtime,
1968 raw_spin_lock(&cfs_b->lock);
1970 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
1973 /* return (any) remaining runtime */
1974 cfs_b->runtime = runtime;
1976 * While we are ensured activity in the period following an
1977 * unthrottle, this also covers the case in which the new bandwidth is
1978 * insufficient to cover the existing bandwidth deficit. (Forcing the
1979 * timer to remain active while there are any throttled entities.)
1984 cfs_b->timer_active = 0;
1985 raw_spin_unlock(&cfs_b->lock);
1990 /* a cfs_rq won't donate quota below this amount */
1991 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
1992 /* minimum remaining period time to redistribute slack quota */
1993 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
1994 /* how long we wait to gather additional slack before distributing */
1995 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
1997 /* are we near the end of the current quota period? */
1998 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2000 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2003 /* if the call-back is running a quota refresh is already occurring */
2004 if (hrtimer_callback_running(refresh_timer))
2007 /* is a quota refresh about to occur? */
2008 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2009 if (remaining < min_expire)
2015 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2017 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2019 /* if there's a quota refresh soon don't bother with slack */
2020 if (runtime_refresh_within(cfs_b, min_left))
2023 start_bandwidth_timer(&cfs_b->slack_timer,
2024 ns_to_ktime(cfs_bandwidth_slack_period));
2027 /* we know any runtime found here is valid as update_curr() precedes return */
2028 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2030 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2031 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2033 if (slack_runtime <= 0)
2036 raw_spin_lock(&cfs_b->lock);
2037 if (cfs_b->quota != RUNTIME_INF &&
2038 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2039 cfs_b->runtime += slack_runtime;
2041 /* we are under rq->lock, defer unthrottling using a timer */
2042 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2043 !list_empty(&cfs_b->throttled_cfs_rq))
2044 start_cfs_slack_bandwidth(cfs_b);
2046 raw_spin_unlock(&cfs_b->lock);
2048 /* even if it's not valid for return we don't want to try again */
2049 cfs_rq->runtime_remaining -= slack_runtime;
2052 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2054 if (!cfs_bandwidth_used())
2057 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2060 __return_cfs_rq_runtime(cfs_rq);
2064 * This is done with a timer (instead of inline with bandwidth return) since
2065 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2067 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2069 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2072 /* confirm we're still not at a refresh boundary */
2073 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2076 raw_spin_lock(&cfs_b->lock);
2077 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2078 runtime = cfs_b->runtime;
2081 expires = cfs_b->runtime_expires;
2082 raw_spin_unlock(&cfs_b->lock);
2087 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2089 raw_spin_lock(&cfs_b->lock);
2090 if (expires == cfs_b->runtime_expires)
2091 cfs_b->runtime = runtime;
2092 raw_spin_unlock(&cfs_b->lock);
2096 * When a group wakes up we want to make sure that its quota is not already
2097 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2098 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2100 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2102 if (!cfs_bandwidth_used())
2105 /* an active group must be handled by the update_curr()->put() path */
2106 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2109 /* ensure the group is not already throttled */
2110 if (cfs_rq_throttled(cfs_rq))
2113 /* update runtime allocation */
2114 account_cfs_rq_runtime(cfs_rq, 0);
2115 if (cfs_rq->runtime_remaining <= 0)
2116 throttle_cfs_rq(cfs_rq);
2119 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2120 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2122 if (!cfs_bandwidth_used())
2125 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2129 * it's possible for a throttled entity to be forced into a running
2130 * state (e.g. set_curr_task), in this case we're finished.
2132 if (cfs_rq_throttled(cfs_rq))
2135 throttle_cfs_rq(cfs_rq);
2138 static inline u64 default_cfs_period(void);
2139 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun);
2140 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b);
2142 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2144 struct cfs_bandwidth *cfs_b =
2145 container_of(timer, struct cfs_bandwidth, slack_timer);
2146 do_sched_cfs_slack_timer(cfs_b);
2148 return HRTIMER_NORESTART;
2151 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2153 struct cfs_bandwidth *cfs_b =
2154 container_of(timer, struct cfs_bandwidth, period_timer);
2160 now = hrtimer_cb_get_time(timer);
2161 overrun = hrtimer_forward(timer, now, cfs_b->period);
2166 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2169 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2172 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2174 raw_spin_lock_init(&cfs_b->lock);
2176 cfs_b->quota = RUNTIME_INF;
2177 cfs_b->period = ns_to_ktime(default_cfs_period());
2179 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2180 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2181 cfs_b->period_timer.function = sched_cfs_period_timer;
2182 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2183 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2186 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2188 cfs_rq->runtime_enabled = 0;
2189 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2192 /* requires cfs_b->lock, may release to reprogram timer */
2193 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2196 * The timer may be active because we're trying to set a new bandwidth
2197 * period or because we're racing with the tear-down path
2198 * (timer_active==0 becomes visible before the hrtimer call-back
2199 * terminates). In either case we ensure that it's re-programmed
2201 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2202 raw_spin_unlock(&cfs_b->lock);
2203 /* ensure cfs_b->lock is available while we wait */
2204 hrtimer_cancel(&cfs_b->period_timer);
2206 raw_spin_lock(&cfs_b->lock);
2207 /* if someone else restarted the timer then we're done */
2208 if (cfs_b->timer_active)
2212 cfs_b->timer_active = 1;
2213 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2216 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2218 hrtimer_cancel(&cfs_b->period_timer);
2219 hrtimer_cancel(&cfs_b->slack_timer);
2222 static void unthrottle_offline_cfs_rqs(struct rq *rq)
2224 struct cfs_rq *cfs_rq;
2226 for_each_leaf_cfs_rq(rq, cfs_rq) {
2227 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2229 if (!cfs_rq->runtime_enabled)
2233 * clock_task is not advancing so we just need to make sure
2234 * there's some valid quota amount
2236 cfs_rq->runtime_remaining = cfs_b->quota;
2237 if (cfs_rq_throttled(cfs_rq))
2238 unthrottle_cfs_rq(cfs_rq);
2242 #else /* CONFIG_CFS_BANDWIDTH */
2243 static __always_inline
2244 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec) {}
2245 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2246 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2247 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2249 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2254 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2259 static inline int throttled_lb_pair(struct task_group *tg,
2260 int src_cpu, int dest_cpu)
2265 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2267 #ifdef CONFIG_FAIR_GROUP_SCHED
2268 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2271 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2275 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2276 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2278 #endif /* CONFIG_CFS_BANDWIDTH */
2280 /**************************************************
2281 * CFS operations on tasks:
2284 #ifdef CONFIG_SCHED_HRTICK
2285 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2287 struct sched_entity *se = &p->se;
2288 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2290 WARN_ON(task_rq(p) != rq);
2292 if (cfs_rq->nr_running > 1) {
2293 u64 slice = sched_slice(cfs_rq, se);
2294 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2295 s64 delta = slice - ran;
2304 * Don't schedule slices shorter than 10000ns, that just
2305 * doesn't make sense. Rely on vruntime for fairness.
2308 delta = max_t(s64, 10000LL, delta);
2310 hrtick_start(rq, delta);
2315 * called from enqueue/dequeue and updates the hrtick when the
2316 * current task is from our class and nr_running is low enough
2319 static void hrtick_update(struct rq *rq)
2321 struct task_struct *curr = rq->curr;
2323 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2326 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2327 hrtick_start_fair(rq, curr);
2329 #else /* !CONFIG_SCHED_HRTICK */
2331 hrtick_start_fair(struct rq *rq, struct task_struct *p)
2335 static inline void hrtick_update(struct rq *rq)
2341 * The enqueue_task method is called before nr_running is
2342 * increased. Here we update the fair scheduling stats and
2343 * then put the task into the rbtree:
2346 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2348 struct cfs_rq *cfs_rq;
2349 struct sched_entity *se = &p->se;
2351 for_each_sched_entity(se) {
2354 cfs_rq = cfs_rq_of(se);
2355 enqueue_entity(cfs_rq, se, flags);
2358 * end evaluation on encountering a throttled cfs_rq
2360 * note: in the case of encountering a throttled cfs_rq we will
2361 * post the final h_nr_running increment below.
2363 if (cfs_rq_throttled(cfs_rq))
2365 cfs_rq->h_nr_running++;
2367 flags = ENQUEUE_WAKEUP;
2370 for_each_sched_entity(se) {
2371 cfs_rq = cfs_rq_of(se);
2372 cfs_rq->h_nr_running++;
2374 if (cfs_rq_throttled(cfs_rq))
2377 update_cfs_load(cfs_rq, 0);
2378 update_cfs_shares(cfs_rq);
2386 static void set_next_buddy(struct sched_entity *se);
2389 * The dequeue_task method is called before nr_running is
2390 * decreased. We remove the task from the rbtree and
2391 * update the fair scheduling stats:
2393 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2395 struct cfs_rq *cfs_rq;
2396 struct sched_entity *se = &p->se;
2397 int task_sleep = flags & DEQUEUE_SLEEP;
2399 for_each_sched_entity(se) {
2400 cfs_rq = cfs_rq_of(se);
2401 dequeue_entity(cfs_rq, se, flags);
2404 * end evaluation on encountering a throttled cfs_rq
2406 * note: in the case of encountering a throttled cfs_rq we will
2407 * post the final h_nr_running decrement below.
2409 if (cfs_rq_throttled(cfs_rq))
2411 cfs_rq->h_nr_running--;
2413 /* Don't dequeue parent if it has other entities besides us */
2414 if (cfs_rq->load.weight) {
2416 * Bias pick_next to pick a task from this cfs_rq, as
2417 * p is sleeping when it is within its sched_slice.
2419 if (task_sleep && parent_entity(se))
2420 set_next_buddy(parent_entity(se));
2422 /* avoid re-evaluating load for this entity */
2423 se = parent_entity(se);
2426 flags |= DEQUEUE_SLEEP;
2429 for_each_sched_entity(se) {
2430 cfs_rq = cfs_rq_of(se);
2431 cfs_rq->h_nr_running--;
2433 if (cfs_rq_throttled(cfs_rq))
2436 update_cfs_load(cfs_rq, 0);
2437 update_cfs_shares(cfs_rq);
2446 /* Used instead of source_load when we know the type == 0 */
2447 static unsigned long weighted_cpuload(const int cpu)
2449 return cpu_rq(cpu)->load.weight;
2453 * Return a low guess at the load of a migration-source cpu weighted
2454 * according to the scheduling class and "nice" value.
2456 * We want to under-estimate the load of migration sources, to
2457 * balance conservatively.
2459 static unsigned long source_load(int cpu, int type)
2461 struct rq *rq = cpu_rq(cpu);
2462 unsigned long total = weighted_cpuload(cpu);
2464 if (type == 0 || !sched_feat(LB_BIAS))
2467 return min(rq->cpu_load[type-1], total);
2471 * Return a high guess at the load of a migration-target cpu weighted
2472 * according to the scheduling class and "nice" value.
2474 static unsigned long target_load(int cpu, int type)
2476 struct rq *rq = cpu_rq(cpu);
2477 unsigned long total = weighted_cpuload(cpu);
2479 if (type == 0 || !sched_feat(LB_BIAS))
2482 return max(rq->cpu_load[type-1], total);
2485 static unsigned long power_of(int cpu)
2487 return cpu_rq(cpu)->cpu_power;
2490 static unsigned long cpu_avg_load_per_task(int cpu)
2492 struct rq *rq = cpu_rq(cpu);
2493 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
2496 return rq->load.weight / nr_running;
2502 static void task_waking_fair(struct task_struct *p)
2504 struct sched_entity *se = &p->se;
2505 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2508 #ifndef CONFIG_64BIT
2509 u64 min_vruntime_copy;
2512 min_vruntime_copy = cfs_rq->min_vruntime_copy;
2514 min_vruntime = cfs_rq->min_vruntime;
2515 } while (min_vruntime != min_vruntime_copy);
2517 min_vruntime = cfs_rq->min_vruntime;
2520 se->vruntime -= min_vruntime;
2523 #ifdef CONFIG_FAIR_GROUP_SCHED
2525 * effective_load() calculates the load change as seen from the root_task_group
2527 * Adding load to a group doesn't make a group heavier, but can cause movement
2528 * of group shares between cpus. Assuming the shares were perfectly aligned one
2529 * can calculate the shift in shares.
2531 * Calculate the effective load difference if @wl is added (subtracted) to @tg
2532 * on this @cpu and results in a total addition (subtraction) of @wg to the
2533 * total group weight.
2535 * Given a runqueue weight distribution (rw_i) we can compute a shares
2536 * distribution (s_i) using:
2538 * s_i = rw_i / \Sum rw_j (1)
2540 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
2541 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
2542 * shares distribution (s_i):
2544 * rw_i = { 2, 4, 1, 0 }
2545 * s_i = { 2/7, 4/7, 1/7, 0 }
2547 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
2548 * task used to run on and the CPU the waker is running on), we need to
2549 * compute the effect of waking a task on either CPU and, in case of a sync
2550 * wakeup, compute the effect of the current task going to sleep.
2552 * So for a change of @wl to the local @cpu with an overall group weight change
2553 * of @wl we can compute the new shares distribution (s'_i) using:
2555 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
2557 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
2558 * differences in waking a task to CPU 0. The additional task changes the
2559 * weight and shares distributions like:
2561 * rw'_i = { 3, 4, 1, 0 }
2562 * s'_i = { 3/8, 4/8, 1/8, 0 }
2564 * We can then compute the difference in effective weight by using:
2566 * dw_i = S * (s'_i - s_i) (3)
2568 * Where 'S' is the group weight as seen by its parent.
2570 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
2571 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
2572 * 4/7) times the weight of the group.
2574 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
2576 struct sched_entity *se = tg->se[cpu];
2578 if (!tg->parent) /* the trivial, non-cgroup case */
2581 for_each_sched_entity(se) {
2587 * W = @wg + \Sum rw_j
2589 W = wg + calc_tg_weight(tg, se->my_q);
2594 w = se->my_q->load.weight + wl;
2597 * wl = S * s'_i; see (2)
2600 wl = (w * tg->shares) / W;
2605 * Per the above, wl is the new se->load.weight value; since
2606 * those are clipped to [MIN_SHARES, ...) do so now. See
2607 * calc_cfs_shares().
2609 if (wl < MIN_SHARES)
2613 * wl = dw_i = S * (s'_i - s_i); see (3)
2615 wl -= se->load.weight;
2618 * Recursively apply this logic to all parent groups to compute
2619 * the final effective load change on the root group. Since
2620 * only the @tg group gets extra weight, all parent groups can
2621 * only redistribute existing shares. @wl is the shift in shares
2622 * resulting from this level per the above.
2631 static inline unsigned long effective_load(struct task_group *tg, int cpu,
2632 unsigned long wl, unsigned long wg)
2639 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
2641 s64 this_load, load;
2642 int idx, this_cpu, prev_cpu;
2643 unsigned long tl_per_task;
2644 struct task_group *tg;
2645 unsigned long weight;
2649 this_cpu = smp_processor_id();
2650 prev_cpu = task_cpu(p);
2651 load = source_load(prev_cpu, idx);
2652 this_load = target_load(this_cpu, idx);
2655 * If sync wakeup then subtract the (maximum possible)
2656 * effect of the currently running task from the load
2657 * of the current CPU:
2660 tg = task_group(current);
2661 weight = current->se.load.weight;
2663 this_load += effective_load(tg, this_cpu, -weight, -weight);
2664 load += effective_load(tg, prev_cpu, 0, -weight);
2668 weight = p->se.load.weight;
2671 * In low-load situations, where prev_cpu is idle and this_cpu is idle
2672 * due to the sync cause above having dropped this_load to 0, we'll
2673 * always have an imbalance, but there's really nothing you can do
2674 * about that, so that's good too.
2676 * Otherwise check if either cpus are near enough in load to allow this
2677 * task to be woken on this_cpu.
2679 if (this_load > 0) {
2680 s64 this_eff_load, prev_eff_load;
2682 this_eff_load = 100;
2683 this_eff_load *= power_of(prev_cpu);
2684 this_eff_load *= this_load +
2685 effective_load(tg, this_cpu, weight, weight);
2687 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
2688 prev_eff_load *= power_of(this_cpu);
2689 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
2691 balanced = this_eff_load <= prev_eff_load;
2696 * If the currently running task will sleep within
2697 * a reasonable amount of time then attract this newly
2700 if (sync && balanced)
2703 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
2704 tl_per_task = cpu_avg_load_per_task(this_cpu);
2707 (this_load <= load &&
2708 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
2710 * This domain has SD_WAKE_AFFINE and
2711 * p is cache cold in this domain, and
2712 * there is no bad imbalance.
2714 schedstat_inc(sd, ttwu_move_affine);
2715 schedstat_inc(p, se.statistics.nr_wakeups_affine);
2723 * find_idlest_group finds and returns the least busy CPU group within the
2726 static struct sched_group *
2727 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
2728 int this_cpu, int load_idx)
2730 struct sched_group *idlest = NULL, *group = sd->groups;
2731 unsigned long min_load = ULONG_MAX, this_load = 0;
2732 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2735 unsigned long load, avg_load;
2739 /* Skip over this group if it has no CPUs allowed */
2740 if (!cpumask_intersects(sched_group_cpus(group),
2741 tsk_cpus_allowed(p)))
2744 local_group = cpumask_test_cpu(this_cpu,
2745 sched_group_cpus(group));
2747 /* Tally up the load of all CPUs in the group */
2750 for_each_cpu(i, sched_group_cpus(group)) {
2751 /* Bias balancing toward cpus of our domain */
2753 load = source_load(i, load_idx);
2755 load = target_load(i, load_idx);
2760 /* Adjust by relative CPU power of the group */
2761 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
2764 this_load = avg_load;
2765 } else if (avg_load < min_load) {
2766 min_load = avg_load;
2769 } while (group = group->next, group != sd->groups);
2771 if (!idlest || 100*this_load < imbalance*min_load)
2777 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2780 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2782 unsigned long load, min_load = ULONG_MAX;
2786 /* Traverse only the allowed CPUs */
2787 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
2788 load = weighted_cpuload(i);
2790 if (load < min_load || (load == min_load && i == this_cpu)) {
2800 * Try and locate an idle CPU in the sched_domain.
2802 static int select_idle_sibling(struct task_struct *p, int target)
2804 int cpu = smp_processor_id();
2805 int prev_cpu = task_cpu(p);
2806 struct sched_domain *sd;
2807 struct sched_group *sg;
2811 * If the task is going to be woken-up on this cpu and if it is
2812 * already idle, then it is the right target.
2814 if (target == cpu && idle_cpu(cpu))
2818 * If the task is going to be woken-up on the cpu where it previously
2819 * ran and if it is currently idle, then it the right target.
2821 if (target == prev_cpu && idle_cpu(prev_cpu))
2825 * Otherwise, iterate the domains and find an elegible idle cpu.
2827 sd = rcu_dereference(per_cpu(sd_llc, target));
2828 for_each_lower_domain(sd) {
2831 if (!cpumask_intersects(sched_group_cpus(sg),
2832 tsk_cpus_allowed(p)))
2835 for_each_cpu(i, sched_group_cpus(sg)) {
2840 target = cpumask_first_and(sched_group_cpus(sg),
2841 tsk_cpus_allowed(p));
2845 } while (sg != sd->groups);
2852 * sched_balance_self: balance the current task (running on cpu) in domains
2853 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2856 * Balance, ie. select the least loaded group.
2858 * Returns the target CPU number, or the same CPU if no balancing is needed.
2860 * preempt must be disabled.
2863 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
2865 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
2866 int cpu = smp_processor_id();
2867 int prev_cpu = task_cpu(p);
2869 int want_affine = 0;
2870 int sync = wake_flags & WF_SYNC;
2872 if (p->nr_cpus_allowed == 1)
2875 if (sd_flag & SD_BALANCE_WAKE) {
2876 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
2882 for_each_domain(cpu, tmp) {
2883 if (!(tmp->flags & SD_LOAD_BALANCE))
2887 * If both cpu and prev_cpu are part of this domain,
2888 * cpu is a valid SD_WAKE_AFFINE target.
2890 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
2891 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
2896 if (tmp->flags & sd_flag)
2901 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
2904 new_cpu = select_idle_sibling(p, prev_cpu);
2909 int load_idx = sd->forkexec_idx;
2910 struct sched_group *group;
2913 if (!(sd->flags & sd_flag)) {
2918 if (sd_flag & SD_BALANCE_WAKE)
2919 load_idx = sd->wake_idx;
2921 group = find_idlest_group(sd, p, cpu, load_idx);
2927 new_cpu = find_idlest_cpu(group, p, cpu);
2928 if (new_cpu == -1 || new_cpu == cpu) {
2929 /* Now try balancing at a lower domain level of cpu */
2934 /* Now try balancing at a lower domain level of new_cpu */
2936 weight = sd->span_weight;
2938 for_each_domain(cpu, tmp) {
2939 if (weight <= tmp->span_weight)
2941 if (tmp->flags & sd_flag)
2944 /* while loop will break here if sd == NULL */
2951 #endif /* CONFIG_SMP */
2953 static unsigned long
2954 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
2956 unsigned long gran = sysctl_sched_wakeup_granularity;
2959 * Since its curr running now, convert the gran from real-time
2960 * to virtual-time in his units.
2962 * By using 'se' instead of 'curr' we penalize light tasks, so
2963 * they get preempted easier. That is, if 'se' < 'curr' then
2964 * the resulting gran will be larger, therefore penalizing the
2965 * lighter, if otoh 'se' > 'curr' then the resulting gran will
2966 * be smaller, again penalizing the lighter task.
2968 * This is especially important for buddies when the leftmost
2969 * task is higher priority than the buddy.
2971 return calc_delta_fair(gran, se);
2975 * Should 'se' preempt 'curr'.
2989 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
2991 s64 gran, vdiff = curr->vruntime - se->vruntime;
2996 gran = wakeup_gran(curr, se);
3003 static void set_last_buddy(struct sched_entity *se)
3005 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3008 for_each_sched_entity(se)
3009 cfs_rq_of(se)->last = se;
3012 static void set_next_buddy(struct sched_entity *se)
3014 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3017 for_each_sched_entity(se)
3018 cfs_rq_of(se)->next = se;
3021 static void set_skip_buddy(struct sched_entity *se)
3023 for_each_sched_entity(se)
3024 cfs_rq_of(se)->skip = se;
3028 * Preempt the current task with a newly woken task if needed:
3030 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3032 struct task_struct *curr = rq->curr;
3033 struct sched_entity *se = &curr->se, *pse = &p->se;
3034 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3035 int scale = cfs_rq->nr_running >= sched_nr_latency;
3036 int next_buddy_marked = 0;
3038 if (unlikely(se == pse))
3042 * This is possible from callers such as move_task(), in which we
3043 * unconditionally check_prempt_curr() after an enqueue (which may have
3044 * lead to a throttle). This both saves work and prevents false
3045 * next-buddy nomination below.
3047 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3050 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3051 set_next_buddy(pse);
3052 next_buddy_marked = 1;
3056 * We can come here with TIF_NEED_RESCHED already set from new task
3059 * Note: this also catches the edge-case of curr being in a throttled
3060 * group (e.g. via set_curr_task), since update_curr() (in the
3061 * enqueue of curr) will have resulted in resched being set. This
3062 * prevents us from potentially nominating it as a false LAST_BUDDY
3065 if (test_tsk_need_resched(curr))
3068 /* Idle tasks are by definition preempted by non-idle tasks. */
3069 if (unlikely(curr->policy == SCHED_IDLE) &&
3070 likely(p->policy != SCHED_IDLE))
3074 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3075 * is driven by the tick):
3077 if (unlikely(p->policy != SCHED_NORMAL))
3080 find_matching_se(&se, &pse);
3081 update_curr(cfs_rq_of(se));
3083 if (wakeup_preempt_entity(se, pse) == 1) {
3085 * Bias pick_next to pick the sched entity that is
3086 * triggering this preemption.
3088 if (!next_buddy_marked)
3089 set_next_buddy(pse);
3098 * Only set the backward buddy when the current task is still
3099 * on the rq. This can happen when a wakeup gets interleaved
3100 * with schedule on the ->pre_schedule() or idle_balance()
3101 * point, either of which can * drop the rq lock.
3103 * Also, during early boot the idle thread is in the fair class,
3104 * for obvious reasons its a bad idea to schedule back to it.
3106 if (unlikely(!se->on_rq || curr == rq->idle))
3109 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3113 static struct task_struct *pick_next_task_fair(struct rq *rq)
3115 struct task_struct *p;
3116 struct cfs_rq *cfs_rq = &rq->cfs;
3117 struct sched_entity *se;
3119 if (!cfs_rq->nr_running)
3123 se = pick_next_entity(cfs_rq);
3124 set_next_entity(cfs_rq, se);
3125 cfs_rq = group_cfs_rq(se);
3129 if (hrtick_enabled(rq))
3130 hrtick_start_fair(rq, p);
3136 * Account for a descheduled task:
3138 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3140 struct sched_entity *se = &prev->se;
3141 struct cfs_rq *cfs_rq;
3143 for_each_sched_entity(se) {
3144 cfs_rq = cfs_rq_of(se);
3145 put_prev_entity(cfs_rq, se);
3150 * sched_yield() is very simple
3152 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3154 static void yield_task_fair(struct rq *rq)
3156 struct task_struct *curr = rq->curr;
3157 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3158 struct sched_entity *se = &curr->se;
3161 * Are we the only task in the tree?
3163 if (unlikely(rq->nr_running == 1))
3166 clear_buddies(cfs_rq, se);
3168 if (curr->policy != SCHED_BATCH) {
3169 update_rq_clock(rq);
3171 * Update run-time statistics of the 'current'.
3173 update_curr(cfs_rq);
3175 * Tell update_rq_clock() that we've just updated,
3176 * so we don't do microscopic update in schedule()
3177 * and double the fastpath cost.
3179 rq->skip_clock_update = 1;
3185 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3187 struct sched_entity *se = &p->se;
3189 /* throttled hierarchies are not runnable */
3190 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3193 /* Tell the scheduler that we'd really like pse to run next. */
3196 yield_task_fair(rq);
3202 /**************************************************
3203 * Fair scheduling class load-balancing methods:
3206 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
3208 #define LBF_ALL_PINNED 0x01
3209 #define LBF_NEED_BREAK 0x02
3210 #define LBF_SOME_PINNED 0x04
3213 struct sched_domain *sd;
3221 struct cpumask *dst_grpmask;
3223 enum cpu_idle_type idle;
3225 /* The set of CPUs under consideration for load-balancing */
3226 struct cpumask *cpus;
3231 unsigned int loop_break;
3232 unsigned int loop_max;
3236 * move_task - move a task from one runqueue to another runqueue.
3237 * Both runqueues must be locked.
3239 static void move_task(struct task_struct *p, struct lb_env *env)
3241 deactivate_task(env->src_rq, p, 0);
3242 set_task_cpu(p, env->dst_cpu);
3243 activate_task(env->dst_rq, p, 0);
3244 check_preempt_curr(env->dst_rq, p, 0);
3248 * Is this task likely cache-hot:
3251 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
3255 if (p->sched_class != &fair_sched_class)
3258 if (unlikely(p->policy == SCHED_IDLE))
3262 * Buddy candidates are cache hot:
3264 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
3265 (&p->se == cfs_rq_of(&p->se)->next ||
3266 &p->se == cfs_rq_of(&p->se)->last))
3269 if (sysctl_sched_migration_cost == -1)
3271 if (sysctl_sched_migration_cost == 0)
3274 delta = now - p->se.exec_start;
3276 return delta < (s64)sysctl_sched_migration_cost;
3280 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3283 int can_migrate_task(struct task_struct *p, struct lb_env *env)
3285 int tsk_cache_hot = 0;
3287 * We do not migrate tasks that are:
3288 * 1) running (obviously), or
3289 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3290 * 3) are cache-hot on their current CPU.
3292 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
3295 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
3298 * Remember if this task can be migrated to any other cpu in
3299 * our sched_group. We may want to revisit it if we couldn't
3300 * meet load balance goals by pulling other tasks on src_cpu.
3302 * Also avoid computing new_dst_cpu if we have already computed
3303 * one in current iteration.
3305 if (!env->dst_grpmask || (env->flags & LBF_SOME_PINNED))
3308 new_dst_cpu = cpumask_first_and(env->dst_grpmask,
3309 tsk_cpus_allowed(p));
3310 if (new_dst_cpu < nr_cpu_ids) {
3311 env->flags |= LBF_SOME_PINNED;
3312 env->new_dst_cpu = new_dst_cpu;
3317 /* Record that we found atleast one task that could run on dst_cpu */
3318 env->flags &= ~LBF_ALL_PINNED;
3320 if (task_running(env->src_rq, p)) {
3321 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
3326 * Aggressive migration if:
3327 * 1) task is cache cold, or
3328 * 2) too many balance attempts have failed.
3331 tsk_cache_hot = task_hot(p, env->src_rq->clock_task, env->sd);
3332 if (!tsk_cache_hot ||
3333 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
3334 #ifdef CONFIG_SCHEDSTATS
3335 if (tsk_cache_hot) {
3336 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
3337 schedstat_inc(p, se.statistics.nr_forced_migrations);
3343 if (tsk_cache_hot) {
3344 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
3351 * move_one_task tries to move exactly one task from busiest to this_rq, as
3352 * part of active balancing operations within "domain".
3353 * Returns 1 if successful and 0 otherwise.
3355 * Called with both runqueues locked.
3357 static int move_one_task(struct lb_env *env)
3359 struct task_struct *p, *n;
3361 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
3362 if (throttled_lb_pair(task_group(p), env->src_rq->cpu, env->dst_cpu))
3365 if (!can_migrate_task(p, env))
3370 * Right now, this is only the second place move_task()
3371 * is called, so we can safely collect move_task()
3372 * stats here rather than inside move_task().
3374 schedstat_inc(env->sd, lb_gained[env->idle]);
3380 static unsigned long task_h_load(struct task_struct *p);
3382 static const unsigned int sched_nr_migrate_break = 32;
3385 * move_tasks tries to move up to imbalance weighted load from busiest to
3386 * this_rq, as part of a balancing operation within domain "sd".
3387 * Returns 1 if successful and 0 otherwise.
3389 * Called with both runqueues locked.
3391 static int move_tasks(struct lb_env *env)
3393 struct list_head *tasks = &env->src_rq->cfs_tasks;
3394 struct task_struct *p;
3398 if (env->imbalance <= 0)
3401 while (!list_empty(tasks)) {
3402 p = list_first_entry(tasks, struct task_struct, se.group_node);
3405 /* We've more or less seen every task there is, call it quits */
3406 if (env->loop > env->loop_max)
3409 /* take a breather every nr_migrate tasks */
3410 if (env->loop > env->loop_break) {
3411 env->loop_break += sched_nr_migrate_break;
3412 env->flags |= LBF_NEED_BREAK;
3416 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
3419 load = task_h_load(p);
3421 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
3424 if ((load / 2) > env->imbalance)
3427 if (!can_migrate_task(p, env))
3432 env->imbalance -= load;
3434 #ifdef CONFIG_PREEMPT
3436 * NEWIDLE balancing is a source of latency, so preemptible
3437 * kernels will stop after the first task is pulled to minimize
3438 * the critical section.
3440 if (env->idle == CPU_NEWLY_IDLE)
3445 * We only want to steal up to the prescribed amount of
3448 if (env->imbalance <= 0)
3453 list_move_tail(&p->se.group_node, tasks);
3457 * Right now, this is one of only two places move_task() is called,
3458 * so we can safely collect move_task() stats here rather than
3459 * inside move_task().
3461 schedstat_add(env->sd, lb_gained[env->idle], pulled);
3466 #ifdef CONFIG_FAIR_GROUP_SCHED
3468 * update tg->load_weight by folding this cpu's load_avg
3470 static int update_shares_cpu(struct task_group *tg, int cpu)
3472 struct cfs_rq *cfs_rq;
3473 unsigned long flags;
3480 cfs_rq = tg->cfs_rq[cpu];
3482 raw_spin_lock_irqsave(&rq->lock, flags);
3484 update_rq_clock(rq);
3485 update_cfs_load(cfs_rq, 1);
3488 * We need to update shares after updating tg->load_weight in
3489 * order to adjust the weight of groups with long running tasks.
3491 update_cfs_shares(cfs_rq);
3493 raw_spin_unlock_irqrestore(&rq->lock, flags);
3498 static void update_shares(int cpu)
3500 struct cfs_rq *cfs_rq;
3501 struct rq *rq = cpu_rq(cpu);
3505 * Iterates the task_group tree in a bottom up fashion, see
3506 * list_add_leaf_cfs_rq() for details.
3508 for_each_leaf_cfs_rq(rq, cfs_rq) {
3509 /* throttled entities do not contribute to load */
3510 if (throttled_hierarchy(cfs_rq))
3513 update_shares_cpu(cfs_rq->tg, cpu);
3519 * Compute the cpu's hierarchical load factor for each task group.
3520 * This needs to be done in a top-down fashion because the load of a child
3521 * group is a fraction of its parents load.
3523 static int tg_load_down(struct task_group *tg, void *data)
3526 long cpu = (long)data;
3529 load = cpu_rq(cpu)->load.weight;
3531 load = tg->parent->cfs_rq[cpu]->h_load;
3532 load *= tg->se[cpu]->load.weight;
3533 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
3536 tg->cfs_rq[cpu]->h_load = load;
3541 static void update_h_load(long cpu)
3543 struct rq *rq = cpu_rq(cpu);
3544 unsigned long now = jiffies;
3546 if (rq->h_load_throttle == now)
3549 rq->h_load_throttle = now;
3552 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
3556 static unsigned long task_h_load(struct task_struct *p)
3558 struct cfs_rq *cfs_rq = task_cfs_rq(p);
3561 load = p->se.load.weight;
3562 load = div_u64(load * cfs_rq->h_load, cfs_rq->load.weight + 1);
3567 static inline void update_shares(int cpu)
3571 static inline void update_h_load(long cpu)
3575 static unsigned long task_h_load(struct task_struct *p)
3577 return p->se.load.weight;
3581 /********** Helpers for find_busiest_group ************************/
3583 * sd_lb_stats - Structure to store the statistics of a sched_domain
3584 * during load balancing.
3586 struct sd_lb_stats {
3587 struct sched_group *busiest; /* Busiest group in this sd */
3588 struct sched_group *this; /* Local group in this sd */
3589 unsigned long total_load; /* Total load of all groups in sd */
3590 unsigned long total_pwr; /* Total power of all groups in sd */
3591 unsigned long avg_load; /* Average load across all groups in sd */
3593 /** Statistics of this group */
3594 unsigned long this_load;
3595 unsigned long this_load_per_task;
3596 unsigned long this_nr_running;
3597 unsigned long this_has_capacity;
3598 unsigned int this_idle_cpus;
3600 /* Statistics of the busiest group */
3601 unsigned int busiest_idle_cpus;
3602 unsigned long max_load;
3603 unsigned long busiest_load_per_task;
3604 unsigned long busiest_nr_running;
3605 unsigned long busiest_group_capacity;
3606 unsigned long busiest_has_capacity;
3607 unsigned int busiest_group_weight;
3609 int group_imb; /* Is there imbalance in this sd */
3613 * sg_lb_stats - stats of a sched_group required for load_balancing
3615 struct sg_lb_stats {
3616 unsigned long avg_load; /*Avg load across the CPUs of the group */
3617 unsigned long group_load; /* Total load over the CPUs of the group */
3618 unsigned long sum_nr_running; /* Nr tasks running in the group */
3619 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3620 unsigned long group_capacity;
3621 unsigned long idle_cpus;
3622 unsigned long group_weight;
3623 int group_imb; /* Is there an imbalance in the group ? */
3624 int group_has_capacity; /* Is there extra capacity in the group? */
3628 * get_sd_load_idx - Obtain the load index for a given sched domain.
3629 * @sd: The sched_domain whose load_idx is to be obtained.
3630 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3632 static inline int get_sd_load_idx(struct sched_domain *sd,
3633 enum cpu_idle_type idle)
3639 load_idx = sd->busy_idx;
3642 case CPU_NEWLY_IDLE:
3643 load_idx = sd->newidle_idx;
3646 load_idx = sd->idle_idx;
3653 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3655 return SCHED_POWER_SCALE;
3658 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3660 return default_scale_freq_power(sd, cpu);
3663 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3665 unsigned long weight = sd->span_weight;
3666 unsigned long smt_gain = sd->smt_gain;
3673 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3675 return default_scale_smt_power(sd, cpu);
3678 unsigned long scale_rt_power(int cpu)
3680 struct rq *rq = cpu_rq(cpu);
3681 u64 total, available, age_stamp, avg;
3684 * Since we're reading these variables without serialization make sure
3685 * we read them once before doing sanity checks on them.
3687 age_stamp = ACCESS_ONCE(rq->age_stamp);
3688 avg = ACCESS_ONCE(rq->rt_avg);
3690 total = sched_avg_period() + (rq->clock - age_stamp);
3692 if (unlikely(total < avg)) {
3693 /* Ensures that power won't end up being negative */
3696 available = total - avg;
3699 if (unlikely((s64)total < SCHED_POWER_SCALE))
3700 total = SCHED_POWER_SCALE;
3702 total >>= SCHED_POWER_SHIFT;
3704 return div_u64(available, total);
3707 static void update_cpu_power(struct sched_domain *sd, int cpu)
3709 unsigned long weight = sd->span_weight;
3710 unsigned long power = SCHED_POWER_SCALE;
3711 struct sched_group *sdg = sd->groups;
3713 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3714 if (sched_feat(ARCH_POWER))
3715 power *= arch_scale_smt_power(sd, cpu);
3717 power *= default_scale_smt_power(sd, cpu);
3719 power >>= SCHED_POWER_SHIFT;
3722 sdg->sgp->power_orig = power;
3724 if (sched_feat(ARCH_POWER))
3725 power *= arch_scale_freq_power(sd, cpu);
3727 power *= default_scale_freq_power(sd, cpu);
3729 power >>= SCHED_POWER_SHIFT;
3731 power *= scale_rt_power(cpu);
3732 power >>= SCHED_POWER_SHIFT;
3737 cpu_rq(cpu)->cpu_power = power;
3738 sdg->sgp->power = power;
3741 void update_group_power(struct sched_domain *sd, int cpu)
3743 struct sched_domain *child = sd->child;
3744 struct sched_group *group, *sdg = sd->groups;
3745 unsigned long power;
3746 unsigned long interval;
3748 interval = msecs_to_jiffies(sd->balance_interval);
3749 interval = clamp(interval, 1UL, max_load_balance_interval);
3750 sdg->sgp->next_update = jiffies + interval;
3753 update_cpu_power(sd, cpu);
3759 if (child->flags & SD_OVERLAP) {
3761 * SD_OVERLAP domains cannot assume that child groups
3762 * span the current group.
3765 for_each_cpu(cpu, sched_group_cpus(sdg))
3766 power += power_of(cpu);
3769 * !SD_OVERLAP domains can assume that child groups
3770 * span the current group.
3773 group = child->groups;
3775 power += group->sgp->power;
3776 group = group->next;
3777 } while (group != child->groups);
3780 sdg->sgp->power_orig = sdg->sgp->power = power;
3784 * Try and fix up capacity for tiny siblings, this is needed when
3785 * things like SD_ASYM_PACKING need f_b_g to select another sibling
3786 * which on its own isn't powerful enough.
3788 * See update_sd_pick_busiest() and check_asym_packing().
3791 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
3794 * Only siblings can have significantly less than SCHED_POWER_SCALE
3796 if (!(sd->flags & SD_SHARE_CPUPOWER))
3800 * If ~90% of the cpu_power is still there, we're good.
3802 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
3809 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3810 * @env: The load balancing environment.
3811 * @group: sched_group whose statistics are to be updated.
3812 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3813 * @local_group: Does group contain this_cpu.
3814 * @balance: Should we balance.
3815 * @sgs: variable to hold the statistics for this group.
3817 static inline void update_sg_lb_stats(struct lb_env *env,
3818 struct sched_group *group, int load_idx,
3819 int local_group, int *balance, struct sg_lb_stats *sgs)
3821 unsigned long nr_running, max_nr_running, min_nr_running;
3822 unsigned long load, max_cpu_load, min_cpu_load;
3823 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3824 unsigned long avg_load_per_task = 0;
3828 balance_cpu = group_balance_cpu(group);
3830 /* Tally up the load of all CPUs in the group */
3832 min_cpu_load = ~0UL;
3834 min_nr_running = ~0UL;
3836 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
3837 struct rq *rq = cpu_rq(i);
3839 nr_running = rq->nr_running;
3841 /* Bias balancing toward cpus of our domain */
3843 if (idle_cpu(i) && !first_idle_cpu &&
3844 cpumask_test_cpu(i, sched_group_mask(group))) {
3849 load = target_load(i, load_idx);
3851 load = source_load(i, load_idx);
3852 if (load > max_cpu_load)
3853 max_cpu_load = load;
3854 if (min_cpu_load > load)
3855 min_cpu_load = load;
3857 if (nr_running > max_nr_running)
3858 max_nr_running = nr_running;
3859 if (min_nr_running > nr_running)
3860 min_nr_running = nr_running;
3863 sgs->group_load += load;
3864 sgs->sum_nr_running += nr_running;
3865 sgs->sum_weighted_load += weighted_cpuload(i);
3871 * First idle cpu or the first cpu(busiest) in this sched group
3872 * is eligible for doing load balancing at this and above
3873 * domains. In the newly idle case, we will allow all the cpu's
3874 * to do the newly idle load balance.
3877 if (env->idle != CPU_NEWLY_IDLE) {
3878 if (balance_cpu != env->dst_cpu) {
3882 update_group_power(env->sd, env->dst_cpu);
3883 } else if (time_after_eq(jiffies, group->sgp->next_update))
3884 update_group_power(env->sd, env->dst_cpu);
3887 /* Adjust by relative CPU power of the group */
3888 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
3891 * Consider the group unbalanced when the imbalance is larger
3892 * than the average weight of a task.
3894 * APZ: with cgroup the avg task weight can vary wildly and
3895 * might not be a suitable number - should we keep a
3896 * normalized nr_running number somewhere that negates
3899 if (sgs->sum_nr_running)
3900 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
3902 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task &&
3903 (max_nr_running - min_nr_running) > 1)
3906 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
3908 if (!sgs->group_capacity)
3909 sgs->group_capacity = fix_small_capacity(env->sd, group);
3910 sgs->group_weight = group->group_weight;
3912 if (sgs->group_capacity > sgs->sum_nr_running)
3913 sgs->group_has_capacity = 1;
3917 * update_sd_pick_busiest - return 1 on busiest group
3918 * @env: The load balancing environment.
3919 * @sds: sched_domain statistics
3920 * @sg: sched_group candidate to be checked for being the busiest
3921 * @sgs: sched_group statistics
3923 * Determine if @sg is a busier group than the previously selected
3926 static bool update_sd_pick_busiest(struct lb_env *env,
3927 struct sd_lb_stats *sds,
3928 struct sched_group *sg,
3929 struct sg_lb_stats *sgs)
3931 if (sgs->avg_load <= sds->max_load)
3934 if (sgs->sum_nr_running > sgs->group_capacity)
3941 * ASYM_PACKING needs to move all the work to the lowest
3942 * numbered CPUs in the group, therefore mark all groups
3943 * higher than ourself as busy.
3945 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
3946 env->dst_cpu < group_first_cpu(sg)) {
3950 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
3958 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
3959 * @env: The load balancing environment.
3960 * @balance: Should we balance.
3961 * @sds: variable to hold the statistics for this sched_domain.
3963 static inline void update_sd_lb_stats(struct lb_env *env,
3964 int *balance, struct sd_lb_stats *sds)
3966 struct sched_domain *child = env->sd->child;
3967 struct sched_group *sg = env->sd->groups;
3968 struct sg_lb_stats sgs;
3969 int load_idx, prefer_sibling = 0;
3971 if (child && child->flags & SD_PREFER_SIBLING)
3974 load_idx = get_sd_load_idx(env->sd, env->idle);
3979 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
3980 memset(&sgs, 0, sizeof(sgs));
3981 update_sg_lb_stats(env, sg, load_idx, local_group, balance, &sgs);
3983 if (local_group && !(*balance))
3986 sds->total_load += sgs.group_load;
3987 sds->total_pwr += sg->sgp->power;
3990 * In case the child domain prefers tasks go to siblings
3991 * first, lower the sg capacity to one so that we'll try
3992 * and move all the excess tasks away. We lower the capacity
3993 * of a group only if the local group has the capacity to fit
3994 * these excess tasks, i.e. nr_running < group_capacity. The
3995 * extra check prevents the case where you always pull from the
3996 * heaviest group when it is already under-utilized (possible
3997 * with a large weight task outweighs the tasks on the system).
3999 if (prefer_sibling && !local_group && sds->this_has_capacity)
4000 sgs.group_capacity = min(sgs.group_capacity, 1UL);
4003 sds->this_load = sgs.avg_load;
4005 sds->this_nr_running = sgs.sum_nr_running;
4006 sds->this_load_per_task = sgs.sum_weighted_load;
4007 sds->this_has_capacity = sgs.group_has_capacity;
4008 sds->this_idle_cpus = sgs.idle_cpus;
4009 } else if (update_sd_pick_busiest(env, sds, sg, &sgs)) {
4010 sds->max_load = sgs.avg_load;
4012 sds->busiest_nr_running = sgs.sum_nr_running;
4013 sds->busiest_idle_cpus = sgs.idle_cpus;
4014 sds->busiest_group_capacity = sgs.group_capacity;
4015 sds->busiest_load_per_task = sgs.sum_weighted_load;
4016 sds->busiest_has_capacity = sgs.group_has_capacity;
4017 sds->busiest_group_weight = sgs.group_weight;
4018 sds->group_imb = sgs.group_imb;
4022 } while (sg != env->sd->groups);
4026 * check_asym_packing - Check to see if the group is packed into the
4029 * This is primarily intended to used at the sibling level. Some
4030 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4031 * case of POWER7, it can move to lower SMT modes only when higher
4032 * threads are idle. When in lower SMT modes, the threads will
4033 * perform better since they share less core resources. Hence when we
4034 * have idle threads, we want them to be the higher ones.
4036 * This packing function is run on idle threads. It checks to see if
4037 * the busiest CPU in this domain (core in the P7 case) has a higher
4038 * CPU number than the packing function is being run on. Here we are
4039 * assuming lower CPU number will be equivalent to lower a SMT thread
4042 * Returns 1 when packing is required and a task should be moved to
4043 * this CPU. The amount of the imbalance is returned in *imbalance.
4045 * @env: The load balancing environment.
4046 * @sds: Statistics of the sched_domain which is to be packed
4048 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
4052 if (!(env->sd->flags & SD_ASYM_PACKING))
4058 busiest_cpu = group_first_cpu(sds->busiest);
4059 if (env->dst_cpu > busiest_cpu)
4062 env->imbalance = DIV_ROUND_CLOSEST(
4063 sds->max_load * sds->busiest->sgp->power, SCHED_POWER_SCALE);
4069 * fix_small_imbalance - Calculate the minor imbalance that exists
4070 * amongst the groups of a sched_domain, during
4072 * @env: The load balancing environment.
4073 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4076 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4078 unsigned long tmp, pwr_now = 0, pwr_move = 0;
4079 unsigned int imbn = 2;
4080 unsigned long scaled_busy_load_per_task;
4082 if (sds->this_nr_running) {
4083 sds->this_load_per_task /= sds->this_nr_running;
4084 if (sds->busiest_load_per_task >
4085 sds->this_load_per_task)
4088 sds->this_load_per_task =
4089 cpu_avg_load_per_task(env->dst_cpu);
4092 scaled_busy_load_per_task = sds->busiest_load_per_task
4093 * SCHED_POWER_SCALE;
4094 scaled_busy_load_per_task /= sds->busiest->sgp->power;
4096 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
4097 (scaled_busy_load_per_task * imbn)) {
4098 env->imbalance = sds->busiest_load_per_task;
4103 * OK, we don't have enough imbalance to justify moving tasks,
4104 * however we may be able to increase total CPU power used by
4108 pwr_now += sds->busiest->sgp->power *
4109 min(sds->busiest_load_per_task, sds->max_load);
4110 pwr_now += sds->this->sgp->power *
4111 min(sds->this_load_per_task, sds->this_load);
4112 pwr_now /= SCHED_POWER_SCALE;
4114 /* Amount of load we'd subtract */
4115 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4116 sds->busiest->sgp->power;
4117 if (sds->max_load > tmp)
4118 pwr_move += sds->busiest->sgp->power *
4119 min(sds->busiest_load_per_task, sds->max_load - tmp);
4121 /* Amount of load we'd add */
4122 if (sds->max_load * sds->busiest->sgp->power <
4123 sds->busiest_load_per_task * SCHED_POWER_SCALE)
4124 tmp = (sds->max_load * sds->busiest->sgp->power) /
4125 sds->this->sgp->power;
4127 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4128 sds->this->sgp->power;
4129 pwr_move += sds->this->sgp->power *
4130 min(sds->this_load_per_task, sds->this_load + tmp);
4131 pwr_move /= SCHED_POWER_SCALE;
4133 /* Move if we gain throughput */
4134 if (pwr_move > pwr_now)
4135 env->imbalance = sds->busiest_load_per_task;
4139 * calculate_imbalance - Calculate the amount of imbalance present within the
4140 * groups of a given sched_domain during load balance.
4141 * @env: load balance environment
4142 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4144 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4146 unsigned long max_pull, load_above_capacity = ~0UL;
4148 sds->busiest_load_per_task /= sds->busiest_nr_running;
4149 if (sds->group_imb) {
4150 sds->busiest_load_per_task =
4151 min(sds->busiest_load_per_task, sds->avg_load);
4155 * In the presence of smp nice balancing, certain scenarios can have
4156 * max load less than avg load(as we skip the groups at or below
4157 * its cpu_power, while calculating max_load..)
4159 if (sds->max_load < sds->avg_load) {
4161 return fix_small_imbalance(env, sds);
4164 if (!sds->group_imb) {
4166 * Don't want to pull so many tasks that a group would go idle.
4168 load_above_capacity = (sds->busiest_nr_running -
4169 sds->busiest_group_capacity);
4171 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
4173 load_above_capacity /= sds->busiest->sgp->power;
4177 * We're trying to get all the cpus to the average_load, so we don't
4178 * want to push ourselves above the average load, nor do we wish to
4179 * reduce the max loaded cpu below the average load. At the same time,
4180 * we also don't want to reduce the group load below the group capacity
4181 * (so that we can implement power-savings policies etc). Thus we look
4182 * for the minimum possible imbalance.
4183 * Be careful of negative numbers as they'll appear as very large values
4184 * with unsigned longs.
4186 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
4188 /* How much load to actually move to equalise the imbalance */
4189 env->imbalance = min(max_pull * sds->busiest->sgp->power,
4190 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
4191 / SCHED_POWER_SCALE;
4194 * if *imbalance is less than the average load per runnable task
4195 * there is no guarantee that any tasks will be moved so we'll have
4196 * a think about bumping its value to force at least one task to be
4199 if (env->imbalance < sds->busiest_load_per_task)
4200 return fix_small_imbalance(env, sds);
4204 /******* find_busiest_group() helpers end here *********************/
4207 * find_busiest_group - Returns the busiest group within the sched_domain
4208 * if there is an imbalance. If there isn't an imbalance, and
4209 * the user has opted for power-savings, it returns a group whose
4210 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4211 * such a group exists.
4213 * Also calculates the amount of weighted load which should be moved
4214 * to restore balance.
4216 * @env: The load balancing environment.
4217 * @balance: Pointer to a variable indicating if this_cpu
4218 * is the appropriate cpu to perform load balancing at this_level.
4220 * Returns: - the busiest group if imbalance exists.
4221 * - If no imbalance and user has opted for power-savings balance,
4222 * return the least loaded group whose CPUs can be
4223 * put to idle by rebalancing its tasks onto our group.
4225 static struct sched_group *
4226 find_busiest_group(struct lb_env *env, int *balance)
4228 struct sd_lb_stats sds;
4230 memset(&sds, 0, sizeof(sds));
4233 * Compute the various statistics relavent for load balancing at
4236 update_sd_lb_stats(env, balance, &sds);
4239 * this_cpu is not the appropriate cpu to perform load balancing at
4245 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
4246 check_asym_packing(env, &sds))
4249 /* There is no busy sibling group to pull tasks from */
4250 if (!sds.busiest || sds.busiest_nr_running == 0)
4253 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
4256 * If the busiest group is imbalanced the below checks don't
4257 * work because they assumes all things are equal, which typically
4258 * isn't true due to cpus_allowed constraints and the like.
4263 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
4264 if (env->idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
4265 !sds.busiest_has_capacity)
4269 * If the local group is more busy than the selected busiest group
4270 * don't try and pull any tasks.
4272 if (sds.this_load >= sds.max_load)
4276 * Don't pull any tasks if this group is already above the domain
4279 if (sds.this_load >= sds.avg_load)
4282 if (env->idle == CPU_IDLE) {
4284 * This cpu is idle. If the busiest group load doesn't
4285 * have more tasks than the number of available cpu's and
4286 * there is no imbalance between this and busiest group
4287 * wrt to idle cpu's, it is balanced.
4289 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
4290 sds.busiest_nr_running <= sds.busiest_group_weight)
4294 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
4295 * imbalance_pct to be conservative.
4297 if (100 * sds.max_load <= env->sd->imbalance_pct * sds.this_load)
4302 /* Looks like there is an imbalance. Compute it */
4303 calculate_imbalance(env, &sds);
4313 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4315 static struct rq *find_busiest_queue(struct lb_env *env,
4316 struct sched_group *group)
4318 struct rq *busiest = NULL, *rq;
4319 unsigned long max_load = 0;
4322 for_each_cpu(i, sched_group_cpus(group)) {
4323 unsigned long power = power_of(i);
4324 unsigned long capacity = DIV_ROUND_CLOSEST(power,
4329 capacity = fix_small_capacity(env->sd, group);
4331 if (!cpumask_test_cpu(i, env->cpus))
4335 wl = weighted_cpuload(i);
4338 * When comparing with imbalance, use weighted_cpuload()
4339 * which is not scaled with the cpu power.
4341 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
4345 * For the load comparisons with the other cpu's, consider
4346 * the weighted_cpuload() scaled with the cpu power, so that
4347 * the load can be moved away from the cpu that is potentially
4348 * running at a lower capacity.
4350 wl = (wl * SCHED_POWER_SCALE) / power;
4352 if (wl > max_load) {
4362 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4363 * so long as it is large enough.
4365 #define MAX_PINNED_INTERVAL 512
4367 /* Working cpumask for load_balance and load_balance_newidle. */
4368 DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4370 static int need_active_balance(struct lb_env *env)
4372 struct sched_domain *sd = env->sd;
4374 if (env->idle == CPU_NEWLY_IDLE) {
4377 * ASYM_PACKING needs to force migrate tasks from busy but
4378 * higher numbered CPUs in order to pack all tasks in the
4379 * lowest numbered CPUs.
4381 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
4385 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
4388 static int active_load_balance_cpu_stop(void *data);
4391 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4392 * tasks if there is an imbalance.
4394 static int load_balance(int this_cpu, struct rq *this_rq,
4395 struct sched_domain *sd, enum cpu_idle_type idle,
4398 int ld_moved, cur_ld_moved, active_balance = 0;
4399 int lb_iterations, max_lb_iterations;
4400 struct sched_group *group;
4402 unsigned long flags;
4403 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4405 struct lb_env env = {
4407 .dst_cpu = this_cpu,
4409 .dst_grpmask = sched_group_cpus(sd->groups),
4411 .loop_break = sched_nr_migrate_break,
4415 cpumask_copy(cpus, cpu_active_mask);
4416 max_lb_iterations = cpumask_weight(env.dst_grpmask);
4418 schedstat_inc(sd, lb_count[idle]);
4421 group = find_busiest_group(&env, balance);
4427 schedstat_inc(sd, lb_nobusyg[idle]);
4431 busiest = find_busiest_queue(&env, group);
4433 schedstat_inc(sd, lb_nobusyq[idle]);
4437 BUG_ON(busiest == env.dst_rq);
4439 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
4443 if (busiest->nr_running > 1) {
4445 * Attempt to move tasks. If find_busiest_group has found
4446 * an imbalance but busiest->nr_running <= 1, the group is
4447 * still unbalanced. ld_moved simply stays zero, so it is
4448 * correctly treated as an imbalance.
4450 env.flags |= LBF_ALL_PINNED;
4451 env.src_cpu = busiest->cpu;
4452 env.src_rq = busiest;
4453 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
4455 update_h_load(env.src_cpu);
4457 local_irq_save(flags);
4458 double_rq_lock(env.dst_rq, busiest);
4461 * cur_ld_moved - load moved in current iteration
4462 * ld_moved - cumulative load moved across iterations
4464 cur_ld_moved = move_tasks(&env);
4465 ld_moved += cur_ld_moved;
4466 double_rq_unlock(env.dst_rq, busiest);
4467 local_irq_restore(flags);
4469 if (env.flags & LBF_NEED_BREAK) {
4470 env.flags &= ~LBF_NEED_BREAK;
4475 * some other cpu did the load balance for us.
4477 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
4478 resched_cpu(env.dst_cpu);
4481 * Revisit (affine) tasks on src_cpu that couldn't be moved to
4482 * us and move them to an alternate dst_cpu in our sched_group
4483 * where they can run. The upper limit on how many times we
4484 * iterate on same src_cpu is dependent on number of cpus in our
4487 * This changes load balance semantics a bit on who can move
4488 * load to a given_cpu. In addition to the given_cpu itself
4489 * (or a ilb_cpu acting on its behalf where given_cpu is
4490 * nohz-idle), we now have balance_cpu in a position to move
4491 * load to given_cpu. In rare situations, this may cause
4492 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
4493 * _independently_ and at _same_ time to move some load to
4494 * given_cpu) causing exceess load to be moved to given_cpu.
4495 * This however should not happen so much in practice and
4496 * moreover subsequent load balance cycles should correct the
4497 * excess load moved.
4499 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0 &&
4500 lb_iterations++ < max_lb_iterations) {
4502 env.dst_rq = cpu_rq(env.new_dst_cpu);
4503 env.dst_cpu = env.new_dst_cpu;
4504 env.flags &= ~LBF_SOME_PINNED;
4506 env.loop_break = sched_nr_migrate_break;
4508 * Go back to "more_balance" rather than "redo" since we
4509 * need to continue with same src_cpu.
4514 /* All tasks on this runqueue were pinned by CPU affinity */
4515 if (unlikely(env.flags & LBF_ALL_PINNED)) {
4516 cpumask_clear_cpu(cpu_of(busiest), cpus);
4517 if (!cpumask_empty(cpus)) {
4519 env.loop_break = sched_nr_migrate_break;
4527 schedstat_inc(sd, lb_failed[idle]);
4529 * Increment the failure counter only on periodic balance.
4530 * We do not want newidle balance, which can be very
4531 * frequent, pollute the failure counter causing
4532 * excessive cache_hot migrations and active balances.
4534 if (idle != CPU_NEWLY_IDLE)
4535 sd->nr_balance_failed++;
4537 if (need_active_balance(&env)) {
4538 raw_spin_lock_irqsave(&busiest->lock, flags);
4540 /* don't kick the active_load_balance_cpu_stop,
4541 * if the curr task on busiest cpu can't be
4544 if (!cpumask_test_cpu(this_cpu,
4545 tsk_cpus_allowed(busiest->curr))) {
4546 raw_spin_unlock_irqrestore(&busiest->lock,
4548 env.flags |= LBF_ALL_PINNED;
4549 goto out_one_pinned;
4553 * ->active_balance synchronizes accesses to
4554 * ->active_balance_work. Once set, it's cleared
4555 * only after active load balance is finished.
4557 if (!busiest->active_balance) {
4558 busiest->active_balance = 1;
4559 busiest->push_cpu = this_cpu;
4562 raw_spin_unlock_irqrestore(&busiest->lock, flags);
4564 if (active_balance) {
4565 stop_one_cpu_nowait(cpu_of(busiest),
4566 active_load_balance_cpu_stop, busiest,
4567 &busiest->active_balance_work);
4571 * We've kicked active balancing, reset the failure
4574 sd->nr_balance_failed = sd->cache_nice_tries+1;
4577 sd->nr_balance_failed = 0;
4579 if (likely(!active_balance)) {
4580 /* We were unbalanced, so reset the balancing interval */
4581 sd->balance_interval = sd->min_interval;
4584 * If we've begun active balancing, start to back off. This
4585 * case may not be covered by the all_pinned logic if there
4586 * is only 1 task on the busy runqueue (because we don't call
4589 if (sd->balance_interval < sd->max_interval)
4590 sd->balance_interval *= 2;
4596 schedstat_inc(sd, lb_balanced[idle]);
4598 sd->nr_balance_failed = 0;
4601 /* tune up the balancing interval */
4602 if (((env.flags & LBF_ALL_PINNED) &&
4603 sd->balance_interval < MAX_PINNED_INTERVAL) ||
4604 (sd->balance_interval < sd->max_interval))
4605 sd->balance_interval *= 2;
4613 * idle_balance is called by schedule() if this_cpu is about to become
4614 * idle. Attempts to pull tasks from other CPUs.
4616 void idle_balance(int this_cpu, struct rq *this_rq)
4618 struct sched_domain *sd;
4619 int pulled_task = 0;
4620 unsigned long next_balance = jiffies + HZ;
4622 this_rq->idle_stamp = this_rq->clock;
4624 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4628 * Drop the rq->lock, but keep IRQ/preempt disabled.
4630 raw_spin_unlock(&this_rq->lock);
4632 update_shares(this_cpu);
4634 for_each_domain(this_cpu, sd) {
4635 unsigned long interval;
4638 if (!(sd->flags & SD_LOAD_BALANCE))
4641 if (sd->flags & SD_BALANCE_NEWIDLE) {
4642 /* If we've pulled tasks over stop searching: */
4643 pulled_task = load_balance(this_cpu, this_rq,
4644 sd, CPU_NEWLY_IDLE, &balance);
4647 interval = msecs_to_jiffies(sd->balance_interval);
4648 if (time_after(next_balance, sd->last_balance + interval))
4649 next_balance = sd->last_balance + interval;
4651 this_rq->idle_stamp = 0;
4657 raw_spin_lock(&this_rq->lock);
4659 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4661 * We are going idle. next_balance may be set based on
4662 * a busy processor. So reset next_balance.
4664 this_rq->next_balance = next_balance;
4669 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
4670 * running tasks off the busiest CPU onto idle CPUs. It requires at
4671 * least 1 task to be running on each physical CPU where possible, and
4672 * avoids physical / logical imbalances.
4674 static int active_load_balance_cpu_stop(void *data)
4676 struct rq *busiest_rq = data;
4677 int busiest_cpu = cpu_of(busiest_rq);
4678 int target_cpu = busiest_rq->push_cpu;
4679 struct rq *target_rq = cpu_rq(target_cpu);
4680 struct sched_domain *sd;
4682 raw_spin_lock_irq(&busiest_rq->lock);
4684 /* make sure the requested cpu hasn't gone down in the meantime */
4685 if (unlikely(busiest_cpu != smp_processor_id() ||
4686 !busiest_rq->active_balance))
4689 /* Is there any task to move? */
4690 if (busiest_rq->nr_running <= 1)
4694 * This condition is "impossible", if it occurs
4695 * we need to fix it. Originally reported by
4696 * Bjorn Helgaas on a 128-cpu setup.
4698 BUG_ON(busiest_rq == target_rq);
4700 /* move a task from busiest_rq to target_rq */
4701 double_lock_balance(busiest_rq, target_rq);
4703 /* Search for an sd spanning us and the target CPU. */
4705 for_each_domain(target_cpu, sd) {
4706 if ((sd->flags & SD_LOAD_BALANCE) &&
4707 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4712 struct lb_env env = {
4714 .dst_cpu = target_cpu,
4715 .dst_rq = target_rq,
4716 .src_cpu = busiest_rq->cpu,
4717 .src_rq = busiest_rq,
4721 schedstat_inc(sd, alb_count);
4723 if (move_one_task(&env))
4724 schedstat_inc(sd, alb_pushed);
4726 schedstat_inc(sd, alb_failed);
4729 double_unlock_balance(busiest_rq, target_rq);
4731 busiest_rq->active_balance = 0;
4732 raw_spin_unlock_irq(&busiest_rq->lock);
4738 * idle load balancing details
4739 * - When one of the busy CPUs notice that there may be an idle rebalancing
4740 * needed, they will kick the idle load balancer, which then does idle
4741 * load balancing for all the idle CPUs.
4744 cpumask_var_t idle_cpus_mask;
4746 unsigned long next_balance; /* in jiffy units */
4747 } nohz ____cacheline_aligned;
4749 static inline int find_new_ilb(int call_cpu)
4751 int ilb = cpumask_first(nohz.idle_cpus_mask);
4753 if (ilb < nr_cpu_ids && idle_cpu(ilb))
4760 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
4761 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
4762 * CPU (if there is one).
4764 static void nohz_balancer_kick(int cpu)
4768 nohz.next_balance++;
4770 ilb_cpu = find_new_ilb(cpu);
4772 if (ilb_cpu >= nr_cpu_ids)
4775 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
4778 * Use smp_send_reschedule() instead of resched_cpu().
4779 * This way we generate a sched IPI on the target cpu which
4780 * is idle. And the softirq performing nohz idle load balance
4781 * will be run before returning from the IPI.
4783 smp_send_reschedule(ilb_cpu);
4787 static inline void nohz_balance_exit_idle(int cpu)
4789 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
4790 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
4791 atomic_dec(&nohz.nr_cpus);
4792 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
4796 static inline void set_cpu_sd_state_busy(void)
4798 struct sched_domain *sd;
4799 int cpu = smp_processor_id();
4801 if (!test_bit(NOHZ_IDLE, nohz_flags(cpu)))
4803 clear_bit(NOHZ_IDLE, nohz_flags(cpu));
4806 for_each_domain(cpu, sd)
4807 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
4811 void set_cpu_sd_state_idle(void)
4813 struct sched_domain *sd;
4814 int cpu = smp_processor_id();
4816 if (test_bit(NOHZ_IDLE, nohz_flags(cpu)))
4818 set_bit(NOHZ_IDLE, nohz_flags(cpu));
4821 for_each_domain(cpu, sd)
4822 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
4827 * This routine will record that the cpu is going idle with tick stopped.
4828 * This info will be used in performing idle load balancing in the future.
4830 void nohz_balance_enter_idle(int cpu)
4833 * If this cpu is going down, then nothing needs to be done.
4835 if (!cpu_active(cpu))
4838 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
4841 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
4842 atomic_inc(&nohz.nr_cpus);
4843 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
4846 static int __cpuinit sched_ilb_notifier(struct notifier_block *nfb,
4847 unsigned long action, void *hcpu)
4849 switch (action & ~CPU_TASKS_FROZEN) {
4851 nohz_balance_exit_idle(smp_processor_id());
4859 static DEFINE_SPINLOCK(balancing);
4862 * Scale the max load_balance interval with the number of CPUs in the system.
4863 * This trades load-balance latency on larger machines for less cross talk.
4865 void update_max_interval(void)
4867 max_load_balance_interval = HZ*num_online_cpus()/10;
4871 * It checks each scheduling domain to see if it is due to be balanced,
4872 * and initiates a balancing operation if so.
4874 * Balancing parameters are set up in arch_init_sched_domains.
4876 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4879 struct rq *rq = cpu_rq(cpu);
4880 unsigned long interval;
4881 struct sched_domain *sd;
4882 /* Earliest time when we have to do rebalance again */
4883 unsigned long next_balance = jiffies + 60*HZ;
4884 int update_next_balance = 0;
4890 for_each_domain(cpu, sd) {
4891 if (!(sd->flags & SD_LOAD_BALANCE))
4894 interval = sd->balance_interval;
4895 if (idle != CPU_IDLE)
4896 interval *= sd->busy_factor;
4898 /* scale ms to jiffies */
4899 interval = msecs_to_jiffies(interval);
4900 interval = clamp(interval, 1UL, max_load_balance_interval);
4902 need_serialize = sd->flags & SD_SERIALIZE;
4904 if (need_serialize) {
4905 if (!spin_trylock(&balancing))
4909 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4910 if (load_balance(cpu, rq, sd, idle, &balance)) {
4912 * We've pulled tasks over so either we're no
4915 idle = CPU_NOT_IDLE;
4917 sd->last_balance = jiffies;
4920 spin_unlock(&balancing);
4922 if (time_after(next_balance, sd->last_balance + interval)) {
4923 next_balance = sd->last_balance + interval;
4924 update_next_balance = 1;
4928 * Stop the load balance at this level. There is another
4929 * CPU in our sched group which is doing load balancing more
4938 * next_balance will be updated only when there is a need.
4939 * When the cpu is attached to null domain for ex, it will not be
4942 if (likely(update_next_balance))
4943 rq->next_balance = next_balance;
4948 * In CONFIG_NO_HZ case, the idle balance kickee will do the
4949 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4951 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
4953 struct rq *this_rq = cpu_rq(this_cpu);
4957 if (idle != CPU_IDLE ||
4958 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
4961 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
4962 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
4966 * If this cpu gets work to do, stop the load balancing
4967 * work being done for other cpus. Next load
4968 * balancing owner will pick it up.
4973 rq = cpu_rq(balance_cpu);
4975 raw_spin_lock_irq(&rq->lock);
4976 update_rq_clock(rq);
4977 update_idle_cpu_load(rq);
4978 raw_spin_unlock_irq(&rq->lock);
4980 rebalance_domains(balance_cpu, CPU_IDLE);
4982 if (time_after(this_rq->next_balance, rq->next_balance))
4983 this_rq->next_balance = rq->next_balance;
4985 nohz.next_balance = this_rq->next_balance;
4987 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
4991 * Current heuristic for kicking the idle load balancer in the presence
4992 * of an idle cpu is the system.
4993 * - This rq has more than one task.
4994 * - At any scheduler domain level, this cpu's scheduler group has multiple
4995 * busy cpu's exceeding the group's power.
4996 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
4997 * domain span are idle.
4999 static inline int nohz_kick_needed(struct rq *rq, int cpu)
5001 unsigned long now = jiffies;
5002 struct sched_domain *sd;
5004 if (unlikely(idle_cpu(cpu)))
5008 * We may be recently in ticked or tickless idle mode. At the first
5009 * busy tick after returning from idle, we will update the busy stats.
5011 set_cpu_sd_state_busy();
5012 nohz_balance_exit_idle(cpu);
5015 * None are in tickless mode and hence no need for NOHZ idle load
5018 if (likely(!atomic_read(&nohz.nr_cpus)))
5021 if (time_before(now, nohz.next_balance))
5024 if (rq->nr_running >= 2)
5028 for_each_domain(cpu, sd) {
5029 struct sched_group *sg = sd->groups;
5030 struct sched_group_power *sgp = sg->sgp;
5031 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
5033 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
5034 goto need_kick_unlock;
5036 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
5037 && (cpumask_first_and(nohz.idle_cpus_mask,
5038 sched_domain_span(sd)) < cpu))
5039 goto need_kick_unlock;
5041 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
5053 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
5057 * run_rebalance_domains is triggered when needed from the scheduler tick.
5058 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
5060 static void run_rebalance_domains(struct softirq_action *h)
5062 int this_cpu = smp_processor_id();
5063 struct rq *this_rq = cpu_rq(this_cpu);
5064 enum cpu_idle_type idle = this_rq->idle_balance ?
5065 CPU_IDLE : CPU_NOT_IDLE;
5067 rebalance_domains(this_cpu, idle);
5070 * If this cpu has a pending nohz_balance_kick, then do the
5071 * balancing on behalf of the other idle cpus whose ticks are
5074 nohz_idle_balance(this_cpu, idle);
5077 static inline int on_null_domain(int cpu)
5079 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
5083 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
5085 void trigger_load_balance(struct rq *rq, int cpu)
5087 /* Don't need to rebalance while attached to NULL domain */
5088 if (time_after_eq(jiffies, rq->next_balance) &&
5089 likely(!on_null_domain(cpu)))
5090 raise_softirq(SCHED_SOFTIRQ);
5092 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
5093 nohz_balancer_kick(cpu);
5097 static void rq_online_fair(struct rq *rq)
5102 static void rq_offline_fair(struct rq *rq)
5106 /* Ensure any throttled groups are reachable by pick_next_task */
5107 unthrottle_offline_cfs_rqs(rq);
5110 #endif /* CONFIG_SMP */
5113 * scheduler tick hitting a task of our scheduling class:
5115 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
5117 struct cfs_rq *cfs_rq;
5118 struct sched_entity *se = &curr->se;
5120 for_each_sched_entity(se) {
5121 cfs_rq = cfs_rq_of(se);
5122 entity_tick(cfs_rq, se, queued);
5125 if (sched_feat_numa(NUMA))
5126 task_tick_numa(rq, curr);
5130 * called on fork with the child task as argument from the parent's context
5131 * - child not yet on the tasklist
5132 * - preemption disabled
5134 static void task_fork_fair(struct task_struct *p)
5136 struct cfs_rq *cfs_rq;
5137 struct sched_entity *se = &p->se, *curr;
5138 int this_cpu = smp_processor_id();
5139 struct rq *rq = this_rq();
5140 unsigned long flags;
5142 raw_spin_lock_irqsave(&rq->lock, flags);
5144 update_rq_clock(rq);
5146 cfs_rq = task_cfs_rq(current);
5147 curr = cfs_rq->curr;
5149 if (unlikely(task_cpu(p) != this_cpu)) {
5151 __set_task_cpu(p, this_cpu);
5155 update_curr(cfs_rq);
5158 se->vruntime = curr->vruntime;
5159 place_entity(cfs_rq, se, 1);
5161 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
5163 * Upon rescheduling, sched_class::put_prev_task() will place
5164 * 'current' within the tree based on its new key value.
5166 swap(curr->vruntime, se->vruntime);
5167 resched_task(rq->curr);
5170 se->vruntime -= cfs_rq->min_vruntime;
5172 raw_spin_unlock_irqrestore(&rq->lock, flags);
5176 * Priority of the task has changed. Check to see if we preempt
5180 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
5186 * Reschedule if we are currently running on this runqueue and
5187 * our priority decreased, or if we are not currently running on
5188 * this runqueue and our priority is higher than the current's
5190 if (rq->curr == p) {
5191 if (p->prio > oldprio)
5192 resched_task(rq->curr);
5194 check_preempt_curr(rq, p, 0);
5197 static void switched_from_fair(struct rq *rq, struct task_struct *p)
5199 struct sched_entity *se = &p->se;
5200 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5203 * Ensure the task's vruntime is normalized, so that when its
5204 * switched back to the fair class the enqueue_entity(.flags=0) will
5205 * do the right thing.
5207 * If it was on_rq, then the dequeue_entity(.flags=0) will already
5208 * have normalized the vruntime, if it was !on_rq, then only when
5209 * the task is sleeping will it still have non-normalized vruntime.
5211 if (!se->on_rq && p->state != TASK_RUNNING) {
5213 * Fix up our vruntime so that the current sleep doesn't
5214 * cause 'unlimited' sleep bonus.
5216 place_entity(cfs_rq, se, 0);
5217 se->vruntime -= cfs_rq->min_vruntime;
5222 * We switched to the sched_fair class.
5224 static void switched_to_fair(struct rq *rq, struct task_struct *p)
5230 * We were most likely switched from sched_rt, so
5231 * kick off the schedule if running, otherwise just see
5232 * if we can still preempt the current task.
5235 resched_task(rq->curr);
5237 check_preempt_curr(rq, p, 0);
5240 /* Account for a task changing its policy or group.
5242 * This routine is mostly called to set cfs_rq->curr field when a task
5243 * migrates between groups/classes.
5245 static void set_curr_task_fair(struct rq *rq)
5247 struct sched_entity *se = &rq->curr->se;
5249 for_each_sched_entity(se) {
5250 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5252 set_next_entity(cfs_rq, se);
5253 /* ensure bandwidth has been allocated on our new cfs_rq */
5254 account_cfs_rq_runtime(cfs_rq, 0);
5258 void init_cfs_rq(struct cfs_rq *cfs_rq)
5260 cfs_rq->tasks_timeline = RB_ROOT;
5261 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
5262 #ifndef CONFIG_64BIT
5263 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
5267 #ifdef CONFIG_FAIR_GROUP_SCHED
5268 static void task_move_group_fair(struct task_struct *p, int on_rq)
5271 * If the task was not on the rq at the time of this cgroup movement
5272 * it must have been asleep, sleeping tasks keep their ->vruntime
5273 * absolute on their old rq until wakeup (needed for the fair sleeper
5274 * bonus in place_entity()).
5276 * If it was on the rq, we've just 'preempted' it, which does convert
5277 * ->vruntime to a relative base.
5279 * Make sure both cases convert their relative position when migrating
5280 * to another cgroup's rq. This does somewhat interfere with the
5281 * fair sleeper stuff for the first placement, but who cares.
5284 * When !on_rq, vruntime of the task has usually NOT been normalized.
5285 * But there are some cases where it has already been normalized:
5287 * - Moving a forked child which is waiting for being woken up by
5288 * wake_up_new_task().
5289 * - Moving a task which has been woken up by try_to_wake_up() and
5290 * waiting for actually being woken up by sched_ttwu_pending().
5292 * To prevent boost or penalty in the new cfs_rq caused by delta
5293 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
5295 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
5299 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
5300 set_task_rq(p, task_cpu(p));
5302 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
5305 void free_fair_sched_group(struct task_group *tg)
5309 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
5311 for_each_possible_cpu(i) {
5313 kfree(tg->cfs_rq[i]);
5322 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
5324 struct cfs_rq *cfs_rq;
5325 struct sched_entity *se;
5328 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
5331 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
5335 tg->shares = NICE_0_LOAD;
5337 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
5339 for_each_possible_cpu(i) {
5340 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
5341 GFP_KERNEL, cpu_to_node(i));
5345 se = kzalloc_node(sizeof(struct sched_entity),
5346 GFP_KERNEL, cpu_to_node(i));
5350 init_cfs_rq(cfs_rq);
5351 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
5362 void unregister_fair_sched_group(struct task_group *tg, int cpu)
5364 struct rq *rq = cpu_rq(cpu);
5365 unsigned long flags;
5368 * Only empty task groups can be destroyed; so we can speculatively
5369 * check on_list without danger of it being re-added.
5371 if (!tg->cfs_rq[cpu]->on_list)
5374 raw_spin_lock_irqsave(&rq->lock, flags);
5375 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
5376 raw_spin_unlock_irqrestore(&rq->lock, flags);
5379 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
5380 struct sched_entity *se, int cpu,
5381 struct sched_entity *parent)
5383 struct rq *rq = cpu_rq(cpu);
5388 /* allow initial update_cfs_load() to truncate */
5389 cfs_rq->load_stamp = 1;
5391 init_cfs_rq_runtime(cfs_rq);
5393 tg->cfs_rq[cpu] = cfs_rq;
5396 /* se could be NULL for root_task_group */
5401 se->cfs_rq = &rq->cfs;
5403 se->cfs_rq = parent->my_q;
5406 update_load_set(&se->load, 0);
5407 se->parent = parent;
5410 static DEFINE_MUTEX(shares_mutex);
5412 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
5415 unsigned long flags;
5418 * We can't change the weight of the root cgroup.
5423 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
5425 mutex_lock(&shares_mutex);
5426 if (tg->shares == shares)
5429 tg->shares = shares;
5430 for_each_possible_cpu(i) {
5431 struct rq *rq = cpu_rq(i);
5432 struct sched_entity *se;
5435 /* Propagate contribution to hierarchy */
5436 raw_spin_lock_irqsave(&rq->lock, flags);
5437 for_each_sched_entity(se)
5438 update_cfs_shares(group_cfs_rq(se));
5439 raw_spin_unlock_irqrestore(&rq->lock, flags);
5443 mutex_unlock(&shares_mutex);
5446 #else /* CONFIG_FAIR_GROUP_SCHED */
5448 void free_fair_sched_group(struct task_group *tg) { }
5450 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
5455 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
5457 #endif /* CONFIG_FAIR_GROUP_SCHED */
5460 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
5462 struct sched_entity *se = &task->se;
5463 unsigned int rr_interval = 0;
5466 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
5469 if (rq->cfs.load.weight)
5470 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5476 * All the scheduling class methods:
5478 const struct sched_class fair_sched_class = {
5479 .next = &idle_sched_class,
5480 .enqueue_task = enqueue_task_fair,
5481 .dequeue_task = dequeue_task_fair,
5482 .yield_task = yield_task_fair,
5483 .yield_to_task = yield_to_task_fair,
5485 .check_preempt_curr = check_preempt_wakeup,
5487 .pick_next_task = pick_next_task_fair,
5488 .put_prev_task = put_prev_task_fair,
5491 .select_task_rq = select_task_rq_fair,
5493 .rq_online = rq_online_fair,
5494 .rq_offline = rq_offline_fair,
5496 .task_waking = task_waking_fair,
5499 .set_curr_task = set_curr_task_fair,
5500 .task_tick = task_tick_fair,
5501 .task_fork = task_fork_fair,
5503 .prio_changed = prio_changed_fair,
5504 .switched_from = switched_from_fair,
5505 .switched_to = switched_to_fair,
5507 .get_rr_interval = get_rr_interval_fair,
5509 #ifdef CONFIG_FAIR_GROUP_SCHED
5510 .task_move_group = task_move_group_fair,
5514 #ifdef CONFIG_SCHED_DEBUG
5515 void print_cfs_stats(struct seq_file *m, int cpu)
5517 struct cfs_rq *cfs_rq;
5520 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
5521 print_cfs_rq(m, cpu, cfs_rq);
5526 __init void init_sched_fair_class(void)
5529 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
5532 nohz.next_balance = jiffies;
5533 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
5534 cpu_notifier(sched_ilb_notifier, 0);