2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/slab.h>
27 #include <linux/profile.h>
28 #include <linux/interrupt.h>
29 #include <linux/mempolicy.h>
30 #include <linux/migrate.h>
31 #include <linux/task_work.h>
33 #include <trace/events/sched.h>
38 * Targeted preemption latency for CPU-bound tasks:
39 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
41 * NOTE: this latency value is not the same as the concept of
42 * 'timeslice length' - timeslices in CFS are of variable length
43 * and have no persistent notion like in traditional, time-slice
44 * based scheduling concepts.
46 * (to see the precise effective timeslice length of your workload,
47 * run vmstat and monitor the context-switches (cs) field)
49 unsigned int sysctl_sched_latency = 6000000ULL;
50 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
53 * The initial- and re-scaling of tunables is configurable
54 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
57 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
58 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
59 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
61 enum sched_tunable_scaling sysctl_sched_tunable_scaling
62 = SCHED_TUNABLESCALING_LOG;
65 * Minimal preemption granularity for CPU-bound tasks:
66 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 unsigned int sysctl_sched_min_granularity = 750000ULL;
69 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
72 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
74 static unsigned int sched_nr_latency = 8;
77 * After fork, child runs first. If set to 0 (default) then
78 * parent will (try to) run first.
80 unsigned int sysctl_sched_child_runs_first __read_mostly;
83 * SCHED_OTHER wake-up granularity.
84 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
86 * This option delays the preemption effects of decoupled workloads
87 * and reduces their over-scheduling. Synchronous workloads will still
88 * have immediate wakeup/sleep latencies.
90 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
91 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
93 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
96 * The exponential sliding window over which load is averaged for shares
100 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
102 #ifdef CONFIG_CFS_BANDWIDTH
104 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
105 * each time a cfs_rq requests quota.
107 * Note: in the case that the slice exceeds the runtime remaining (either due
108 * to consumption or the quota being specified to be smaller than the slice)
109 * we will always only issue the remaining available time.
111 * default: 5 msec, units: microseconds
113 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
116 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
122 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
128 static inline void update_load_set(struct load_weight *lw, unsigned long w)
135 * Increase the granularity value when there are more CPUs,
136 * because with more CPUs the 'effective latency' as visible
137 * to users decreases. But the relationship is not linear,
138 * so pick a second-best guess by going with the log2 of the
141 * This idea comes from the SD scheduler of Con Kolivas:
143 static int get_update_sysctl_factor(void)
145 unsigned int cpus = min_t(int, num_online_cpus(), 8);
148 switch (sysctl_sched_tunable_scaling) {
149 case SCHED_TUNABLESCALING_NONE:
152 case SCHED_TUNABLESCALING_LINEAR:
155 case SCHED_TUNABLESCALING_LOG:
157 factor = 1 + ilog2(cpus);
164 static void update_sysctl(void)
166 unsigned int factor = get_update_sysctl_factor();
168 #define SET_SYSCTL(name) \
169 (sysctl_##name = (factor) * normalized_sysctl_##name)
170 SET_SYSCTL(sched_min_granularity);
171 SET_SYSCTL(sched_latency);
172 SET_SYSCTL(sched_wakeup_granularity);
176 void sched_init_granularity(void)
181 #if BITS_PER_LONG == 32
182 # define WMULT_CONST (~0UL)
184 # define WMULT_CONST (1UL << 32)
187 #define WMULT_SHIFT 32
190 * Shift right and round:
192 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
195 * delta *= weight / lw
198 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
199 struct load_weight *lw)
204 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
205 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
206 * 2^SCHED_LOAD_RESOLUTION.
208 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
209 tmp = (u64)delta_exec * scale_load_down(weight);
211 tmp = (u64)delta_exec;
213 if (!lw->inv_weight) {
214 unsigned long w = scale_load_down(lw->weight);
216 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
218 else if (unlikely(!w))
219 lw->inv_weight = WMULT_CONST;
221 lw->inv_weight = WMULT_CONST / w;
225 * Check whether we'd overflow the 64-bit multiplication:
227 if (unlikely(tmp > WMULT_CONST))
228 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
231 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
233 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
237 const struct sched_class fair_sched_class;
239 /**************************************************************
240 * CFS operations on generic schedulable entities:
243 #ifdef CONFIG_FAIR_GROUP_SCHED
245 /* cpu runqueue to which this cfs_rq is attached */
246 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
251 /* An entity is a task if it doesn't "own" a runqueue */
252 #define entity_is_task(se) (!se->my_q)
254 static inline struct task_struct *task_of(struct sched_entity *se)
256 #ifdef CONFIG_SCHED_DEBUG
257 WARN_ON_ONCE(!entity_is_task(se));
259 return container_of(se, struct task_struct, se);
262 /* Walk up scheduling entities hierarchy */
263 #define for_each_sched_entity(se) \
264 for (; se; se = se->parent)
266 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
271 /* runqueue on which this entity is (to be) queued */
272 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
277 /* runqueue "owned" by this group */
278 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
283 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
286 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
288 if (!cfs_rq->on_list) {
290 * Ensure we either appear before our parent (if already
291 * enqueued) or force our parent to appear after us when it is
292 * enqueued. The fact that we always enqueue bottom-up
293 * reduces this to two cases.
295 if (cfs_rq->tg->parent &&
296 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
297 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
298 &rq_of(cfs_rq)->leaf_cfs_rq_list);
300 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
301 &rq_of(cfs_rq)->leaf_cfs_rq_list);
305 /* We should have no load, but we need to update last_decay. */
306 update_cfs_rq_blocked_load(cfs_rq, 0);
310 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
312 if (cfs_rq->on_list) {
313 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
318 /* Iterate thr' all leaf cfs_rq's on a runqueue */
319 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
320 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
322 /* Do the two (enqueued) entities belong to the same group ? */
324 is_same_group(struct sched_entity *se, struct sched_entity *pse)
326 if (se->cfs_rq == pse->cfs_rq)
332 static inline struct sched_entity *parent_entity(struct sched_entity *se)
337 /* return depth at which a sched entity is present in the hierarchy */
338 static inline int depth_se(struct sched_entity *se)
342 for_each_sched_entity(se)
349 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
351 int se_depth, pse_depth;
354 * preemption test can be made between sibling entities who are in the
355 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
356 * both tasks until we find their ancestors who are siblings of common
360 /* First walk up until both entities are at same depth */
361 se_depth = depth_se(*se);
362 pse_depth = depth_se(*pse);
364 while (se_depth > pse_depth) {
366 *se = parent_entity(*se);
369 while (pse_depth > se_depth) {
371 *pse = parent_entity(*pse);
374 while (!is_same_group(*se, *pse)) {
375 *se = parent_entity(*se);
376 *pse = parent_entity(*pse);
380 #else /* !CONFIG_FAIR_GROUP_SCHED */
382 static inline struct task_struct *task_of(struct sched_entity *se)
384 return container_of(se, struct task_struct, se);
387 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
389 return container_of(cfs_rq, struct rq, cfs);
392 #define entity_is_task(se) 1
394 #define for_each_sched_entity(se) \
395 for (; se; se = NULL)
397 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
399 return &task_rq(p)->cfs;
402 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
404 struct task_struct *p = task_of(se);
405 struct rq *rq = task_rq(p);
410 /* runqueue "owned" by this group */
411 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
416 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
420 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
424 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
425 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
428 is_same_group(struct sched_entity *se, struct sched_entity *pse)
433 static inline struct sched_entity *parent_entity(struct sched_entity *se)
439 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
443 #endif /* CONFIG_FAIR_GROUP_SCHED */
445 static __always_inline
446 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec);
448 /**************************************************************
449 * Scheduling class tree data structure manipulation methods:
452 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
454 s64 delta = (s64)(vruntime - max_vruntime);
456 max_vruntime = vruntime;
461 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
463 s64 delta = (s64)(vruntime - min_vruntime);
465 min_vruntime = vruntime;
470 static inline int entity_before(struct sched_entity *a,
471 struct sched_entity *b)
473 return (s64)(a->vruntime - b->vruntime) < 0;
476 static void update_min_vruntime(struct cfs_rq *cfs_rq)
478 u64 vruntime = cfs_rq->min_vruntime;
481 vruntime = cfs_rq->curr->vruntime;
483 if (cfs_rq->rb_leftmost) {
484 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
489 vruntime = se->vruntime;
491 vruntime = min_vruntime(vruntime, se->vruntime);
494 /* ensure we never gain time by being placed backwards. */
495 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
498 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
503 * Enqueue an entity into the rb-tree:
505 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
507 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
508 struct rb_node *parent = NULL;
509 struct sched_entity *entry;
513 * Find the right place in the rbtree:
517 entry = rb_entry(parent, struct sched_entity, run_node);
519 * We dont care about collisions. Nodes with
520 * the same key stay together.
522 if (entity_before(se, entry)) {
523 link = &parent->rb_left;
525 link = &parent->rb_right;
531 * Maintain a cache of leftmost tree entries (it is frequently
535 cfs_rq->rb_leftmost = &se->run_node;
537 rb_link_node(&se->run_node, parent, link);
538 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
541 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
543 if (cfs_rq->rb_leftmost == &se->run_node) {
544 struct rb_node *next_node;
546 next_node = rb_next(&se->run_node);
547 cfs_rq->rb_leftmost = next_node;
550 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
553 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
555 struct rb_node *left = cfs_rq->rb_leftmost;
560 return rb_entry(left, struct sched_entity, run_node);
563 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
565 struct rb_node *next = rb_next(&se->run_node);
570 return rb_entry(next, struct sched_entity, run_node);
573 #ifdef CONFIG_SCHED_DEBUG
574 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
576 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
581 return rb_entry(last, struct sched_entity, run_node);
584 /**************************************************************
585 * Scheduling class statistics methods:
588 int sched_proc_update_handler(struct ctl_table *table, int write,
589 void __user *buffer, size_t *lenp,
592 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
593 int factor = get_update_sysctl_factor();
598 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
599 sysctl_sched_min_granularity);
601 #define WRT_SYSCTL(name) \
602 (normalized_sysctl_##name = sysctl_##name / (factor))
603 WRT_SYSCTL(sched_min_granularity);
604 WRT_SYSCTL(sched_latency);
605 WRT_SYSCTL(sched_wakeup_granularity);
615 static inline unsigned long
616 calc_delta_fair(unsigned long delta, struct sched_entity *se)
618 if (unlikely(se->load.weight != NICE_0_LOAD))
619 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
625 * The idea is to set a period in which each task runs once.
627 * When there are too many tasks (sched_nr_latency) we have to stretch
628 * this period because otherwise the slices get too small.
630 * p = (nr <= nl) ? l : l*nr/nl
632 static u64 __sched_period(unsigned long nr_running)
634 u64 period = sysctl_sched_latency;
635 unsigned long nr_latency = sched_nr_latency;
637 if (unlikely(nr_running > nr_latency)) {
638 period = sysctl_sched_min_granularity;
639 period *= nr_running;
646 * We calculate the wall-time slice from the period by taking a part
647 * proportional to the weight.
651 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
653 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
655 for_each_sched_entity(se) {
656 struct load_weight *load;
657 struct load_weight lw;
659 cfs_rq = cfs_rq_of(se);
660 load = &cfs_rq->load;
662 if (unlikely(!se->on_rq)) {
665 update_load_add(&lw, se->load.weight);
668 slice = calc_delta_mine(slice, se->load.weight, load);
674 * We calculate the vruntime slice of a to-be-inserted task.
678 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
680 return calc_delta_fair(sched_slice(cfs_rq, se), se);
684 static inline void __update_task_entity_contrib(struct sched_entity *se);
686 /* Give new task start runnable values to heavy its load in infant time */
687 void init_task_runnable_average(struct task_struct *p)
691 p->se.avg.decay_count = 0;
692 slice = sched_slice(task_cfs_rq(p), &p->se) >> 10;
693 p->se.avg.runnable_avg_sum = slice;
694 p->se.avg.runnable_avg_period = slice;
695 __update_task_entity_contrib(&p->se);
698 void init_task_runnable_average(struct task_struct *p)
704 * Update the current task's runtime statistics. Skip current tasks that
705 * are not in our scheduling class.
708 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
709 unsigned long delta_exec)
711 unsigned long delta_exec_weighted;
713 schedstat_set(curr->statistics.exec_max,
714 max((u64)delta_exec, curr->statistics.exec_max));
716 curr->sum_exec_runtime += delta_exec;
717 schedstat_add(cfs_rq, exec_clock, delta_exec);
718 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
720 curr->vruntime += delta_exec_weighted;
721 update_min_vruntime(cfs_rq);
724 static void update_curr(struct cfs_rq *cfs_rq)
726 struct sched_entity *curr = cfs_rq->curr;
727 u64 now = rq_clock_task(rq_of(cfs_rq));
728 unsigned long delta_exec;
734 * Get the amount of time the current task was running
735 * since the last time we changed load (this cannot
736 * overflow on 32 bits):
738 delta_exec = (unsigned long)(now - curr->exec_start);
742 __update_curr(cfs_rq, curr, delta_exec);
743 curr->exec_start = now;
745 if (entity_is_task(curr)) {
746 struct task_struct *curtask = task_of(curr);
748 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
749 cpuacct_charge(curtask, delta_exec);
750 account_group_exec_runtime(curtask, delta_exec);
753 account_cfs_rq_runtime(cfs_rq, delta_exec);
757 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
759 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
763 * Task is being enqueued - update stats:
765 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
768 * Are we enqueueing a waiting task? (for current tasks
769 * a dequeue/enqueue event is a NOP)
771 if (se != cfs_rq->curr)
772 update_stats_wait_start(cfs_rq, se);
776 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
778 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
779 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
780 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
781 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
782 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
783 #ifdef CONFIG_SCHEDSTATS
784 if (entity_is_task(se)) {
785 trace_sched_stat_wait(task_of(se),
786 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
789 schedstat_set(se->statistics.wait_start, 0);
793 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
796 * Mark the end of the wait period if dequeueing a
799 if (se != cfs_rq->curr)
800 update_stats_wait_end(cfs_rq, se);
804 * We are picking a new current task - update its stats:
807 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
810 * We are starting a new run period:
812 se->exec_start = rq_clock_task(rq_of(cfs_rq));
815 /**************************************************
816 * Scheduling class queueing methods:
819 #ifdef CONFIG_NUMA_BALANCING
821 * numa task sample period in ms
823 unsigned int sysctl_numa_balancing_scan_period_min = 100;
824 unsigned int sysctl_numa_balancing_scan_period_max = 100*50;
825 unsigned int sysctl_numa_balancing_scan_period_reset = 100*600;
827 /* Portion of address space to scan in MB */
828 unsigned int sysctl_numa_balancing_scan_size = 256;
830 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
831 unsigned int sysctl_numa_balancing_scan_delay = 1000;
833 static void task_numa_placement(struct task_struct *p)
837 if (!p->mm) /* for example, ksmd faulting in a user's mm */
839 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
840 if (p->numa_scan_seq == seq)
842 p->numa_scan_seq = seq;
844 /* FIXME: Scheduling placement policy hints go here */
848 * Got a PROT_NONE fault for a page on @node.
850 void task_numa_fault(int node, int pages, bool migrated)
852 struct task_struct *p = current;
854 if (!numabalancing_enabled)
857 /* FIXME: Allocate task-specific structure for placement policy here */
860 * If pages are properly placed (did not migrate) then scan slower.
861 * This is reset periodically in case of phase changes
864 p->numa_scan_period = min(sysctl_numa_balancing_scan_period_max,
865 p->numa_scan_period + jiffies_to_msecs(10));
867 task_numa_placement(p);
870 static void reset_ptenuma_scan(struct task_struct *p)
872 ACCESS_ONCE(p->mm->numa_scan_seq)++;
873 p->mm->numa_scan_offset = 0;
877 * The expensive part of numa migration is done from task_work context.
878 * Triggered from task_tick_numa().
880 void task_numa_work(struct callback_head *work)
882 unsigned long migrate, next_scan, now = jiffies;
883 struct task_struct *p = current;
884 struct mm_struct *mm = p->mm;
885 struct vm_area_struct *vma;
886 unsigned long start, end;
889 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
891 work->next = work; /* protect against double add */
893 * Who cares about NUMA placement when they're dying.
895 * NOTE: make sure not to dereference p->mm before this check,
896 * exit_task_work() happens _after_ exit_mm() so we could be called
897 * without p->mm even though we still had it when we enqueued this
900 if (p->flags & PF_EXITING)
904 * Reset the scan period if enough time has gone by. Objective is that
905 * scanning will be reduced if pages are properly placed. As tasks
906 * can enter different phases this needs to be re-examined. Lacking
907 * proper tracking of reference behaviour, this blunt hammer is used.
909 migrate = mm->numa_next_reset;
910 if (time_after(now, migrate)) {
911 p->numa_scan_period = sysctl_numa_balancing_scan_period_min;
912 next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
913 xchg(&mm->numa_next_reset, next_scan);
917 * Enforce maximal scan/migration frequency..
919 migrate = mm->numa_next_scan;
920 if (time_before(now, migrate))
923 if (p->numa_scan_period == 0)
924 p->numa_scan_period = sysctl_numa_balancing_scan_period_min;
926 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
927 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
931 * Delay this task enough that another task of this mm will likely win
932 * the next time around.
934 p->node_stamp += 2 * TICK_NSEC;
936 start = mm->numa_scan_offset;
937 pages = sysctl_numa_balancing_scan_size;
938 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
942 down_read(&mm->mmap_sem);
943 vma = find_vma(mm, start);
945 reset_ptenuma_scan(p);
949 for (; vma; vma = vma->vm_next) {
950 if (!vma_migratable(vma))
953 /* Skip small VMAs. They are not likely to be of relevance */
954 if (vma->vm_end - vma->vm_start < HPAGE_SIZE)
958 start = max(start, vma->vm_start);
959 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
960 end = min(end, vma->vm_end);
961 pages -= change_prot_numa(vma, start, end);
966 } while (end != vma->vm_end);
971 * It is possible to reach the end of the VMA list but the last few
972 * VMAs are not guaranteed to the vma_migratable. If they are not, we
973 * would find the !migratable VMA on the next scan but not reset the
974 * scanner to the start so check it now.
977 mm->numa_scan_offset = start;
979 reset_ptenuma_scan(p);
980 up_read(&mm->mmap_sem);
984 * Drive the periodic memory faults..
986 void task_tick_numa(struct rq *rq, struct task_struct *curr)
988 struct callback_head *work = &curr->numa_work;
992 * We don't care about NUMA placement if we don't have memory.
994 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
998 * Using runtime rather than walltime has the dual advantage that
999 * we (mostly) drive the selection from busy threads and that the
1000 * task needs to have done some actual work before we bother with
1003 now = curr->se.sum_exec_runtime;
1004 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1006 if (now - curr->node_stamp > period) {
1007 if (!curr->node_stamp)
1008 curr->numa_scan_period = sysctl_numa_balancing_scan_period_min;
1009 curr->node_stamp += period;
1011 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1012 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1013 task_work_add(curr, work, true);
1018 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1021 #endif /* CONFIG_NUMA_BALANCING */
1024 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1026 update_load_add(&cfs_rq->load, se->load.weight);
1027 if (!parent_entity(se))
1028 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1030 if (entity_is_task(se))
1031 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1033 cfs_rq->nr_running++;
1037 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1039 update_load_sub(&cfs_rq->load, se->load.weight);
1040 if (!parent_entity(se))
1041 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1042 if (entity_is_task(se))
1043 list_del_init(&se->group_node);
1044 cfs_rq->nr_running--;
1047 #ifdef CONFIG_FAIR_GROUP_SCHED
1049 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1054 * Use this CPU's actual weight instead of the last load_contribution
1055 * to gain a more accurate current total weight. See
1056 * update_cfs_rq_load_contribution().
1058 tg_weight = atomic_long_read(&tg->load_avg);
1059 tg_weight -= cfs_rq->tg_load_contrib;
1060 tg_weight += cfs_rq->load.weight;
1065 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1067 long tg_weight, load, shares;
1069 tg_weight = calc_tg_weight(tg, cfs_rq);
1070 load = cfs_rq->load.weight;
1072 shares = (tg->shares * load);
1074 shares /= tg_weight;
1076 if (shares < MIN_SHARES)
1077 shares = MIN_SHARES;
1078 if (shares > tg->shares)
1079 shares = tg->shares;
1083 # else /* CONFIG_SMP */
1084 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1088 # endif /* CONFIG_SMP */
1089 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1090 unsigned long weight)
1093 /* commit outstanding execution time */
1094 if (cfs_rq->curr == se)
1095 update_curr(cfs_rq);
1096 account_entity_dequeue(cfs_rq, se);
1099 update_load_set(&se->load, weight);
1102 account_entity_enqueue(cfs_rq, se);
1105 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1107 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1109 struct task_group *tg;
1110 struct sched_entity *se;
1114 se = tg->se[cpu_of(rq_of(cfs_rq))];
1115 if (!se || throttled_hierarchy(cfs_rq))
1118 if (likely(se->load.weight == tg->shares))
1121 shares = calc_cfs_shares(cfs_rq, tg);
1123 reweight_entity(cfs_rq_of(se), se, shares);
1125 #else /* CONFIG_FAIR_GROUP_SCHED */
1126 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1129 #endif /* CONFIG_FAIR_GROUP_SCHED */
1133 * We choose a half-life close to 1 scheduling period.
1134 * Note: The tables below are dependent on this value.
1136 #define LOAD_AVG_PERIOD 32
1137 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1138 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1140 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1141 static const u32 runnable_avg_yN_inv[] = {
1142 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1143 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1144 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1145 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1146 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1147 0x85aac367, 0x82cd8698,
1151 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1152 * over-estimates when re-combining.
1154 static const u32 runnable_avg_yN_sum[] = {
1155 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1156 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1157 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1162 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1164 static __always_inline u64 decay_load(u64 val, u64 n)
1166 unsigned int local_n;
1170 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1173 /* after bounds checking we can collapse to 32-bit */
1177 * As y^PERIOD = 1/2, we can combine
1178 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1179 * With a look-up table which covers k^n (n<PERIOD)
1181 * To achieve constant time decay_load.
1183 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1184 val >>= local_n / LOAD_AVG_PERIOD;
1185 local_n %= LOAD_AVG_PERIOD;
1188 val *= runnable_avg_yN_inv[local_n];
1189 /* We don't use SRR here since we always want to round down. */
1194 * For updates fully spanning n periods, the contribution to runnable
1195 * average will be: \Sum 1024*y^n
1197 * We can compute this reasonably efficiently by combining:
1198 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1200 static u32 __compute_runnable_contrib(u64 n)
1204 if (likely(n <= LOAD_AVG_PERIOD))
1205 return runnable_avg_yN_sum[n];
1206 else if (unlikely(n >= LOAD_AVG_MAX_N))
1207 return LOAD_AVG_MAX;
1209 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1211 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1212 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1214 n -= LOAD_AVG_PERIOD;
1215 } while (n > LOAD_AVG_PERIOD);
1217 contrib = decay_load(contrib, n);
1218 return contrib + runnable_avg_yN_sum[n];
1222 * We can represent the historical contribution to runnable average as the
1223 * coefficients of a geometric series. To do this we sub-divide our runnable
1224 * history into segments of approximately 1ms (1024us); label the segment that
1225 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1227 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1229 * (now) (~1ms ago) (~2ms ago)
1231 * Let u_i denote the fraction of p_i that the entity was runnable.
1233 * We then designate the fractions u_i as our co-efficients, yielding the
1234 * following representation of historical load:
1235 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1237 * We choose y based on the with of a reasonably scheduling period, fixing:
1240 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1241 * approximately half as much as the contribution to load within the last ms
1244 * When a period "rolls over" and we have new u_0`, multiplying the previous
1245 * sum again by y is sufficient to update:
1246 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1247 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1249 static __always_inline int __update_entity_runnable_avg(u64 now,
1250 struct sched_avg *sa,
1254 u32 runnable_contrib;
1255 int delta_w, decayed = 0;
1257 delta = now - sa->last_runnable_update;
1259 * This should only happen when time goes backwards, which it
1260 * unfortunately does during sched clock init when we swap over to TSC.
1262 if ((s64)delta < 0) {
1263 sa->last_runnable_update = now;
1268 * Use 1024ns as the unit of measurement since it's a reasonable
1269 * approximation of 1us and fast to compute.
1274 sa->last_runnable_update = now;
1276 /* delta_w is the amount already accumulated against our next period */
1277 delta_w = sa->runnable_avg_period % 1024;
1278 if (delta + delta_w >= 1024) {
1279 /* period roll-over */
1283 * Now that we know we're crossing a period boundary, figure
1284 * out how much from delta we need to complete the current
1285 * period and accrue it.
1287 delta_w = 1024 - delta_w;
1289 sa->runnable_avg_sum += delta_w;
1290 sa->runnable_avg_period += delta_w;
1294 /* Figure out how many additional periods this update spans */
1295 periods = delta / 1024;
1298 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1300 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1303 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1304 runnable_contrib = __compute_runnable_contrib(periods);
1306 sa->runnable_avg_sum += runnable_contrib;
1307 sa->runnable_avg_period += runnable_contrib;
1310 /* Remainder of delta accrued against u_0` */
1312 sa->runnable_avg_sum += delta;
1313 sa->runnable_avg_period += delta;
1318 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1319 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1321 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1322 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1324 decays -= se->avg.decay_count;
1328 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1329 se->avg.decay_count = 0;
1334 #ifdef CONFIG_FAIR_GROUP_SCHED
1335 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1338 struct task_group *tg = cfs_rq->tg;
1341 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1342 tg_contrib -= cfs_rq->tg_load_contrib;
1344 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1345 atomic_long_add(tg_contrib, &tg->load_avg);
1346 cfs_rq->tg_load_contrib += tg_contrib;
1351 * Aggregate cfs_rq runnable averages into an equivalent task_group
1352 * representation for computing load contributions.
1354 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1355 struct cfs_rq *cfs_rq)
1357 struct task_group *tg = cfs_rq->tg;
1360 /* The fraction of a cpu used by this cfs_rq */
1361 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1362 sa->runnable_avg_period + 1);
1363 contrib -= cfs_rq->tg_runnable_contrib;
1365 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
1366 atomic_add(contrib, &tg->runnable_avg);
1367 cfs_rq->tg_runnable_contrib += contrib;
1371 static inline void __update_group_entity_contrib(struct sched_entity *se)
1373 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1374 struct task_group *tg = cfs_rq->tg;
1379 contrib = cfs_rq->tg_load_contrib * tg->shares;
1380 se->avg.load_avg_contrib = div_u64(contrib,
1381 atomic_long_read(&tg->load_avg) + 1);
1384 * For group entities we need to compute a correction term in the case
1385 * that they are consuming <1 cpu so that we would contribute the same
1386 * load as a task of equal weight.
1388 * Explicitly co-ordinating this measurement would be expensive, but
1389 * fortunately the sum of each cpus contribution forms a usable
1390 * lower-bound on the true value.
1392 * Consider the aggregate of 2 contributions. Either they are disjoint
1393 * (and the sum represents true value) or they are disjoint and we are
1394 * understating by the aggregate of their overlap.
1396 * Extending this to N cpus, for a given overlap, the maximum amount we
1397 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1398 * cpus that overlap for this interval and w_i is the interval width.
1400 * On a small machine; the first term is well-bounded which bounds the
1401 * total error since w_i is a subset of the period. Whereas on a
1402 * larger machine, while this first term can be larger, if w_i is the
1403 * of consequential size guaranteed to see n_i*w_i quickly converge to
1404 * our upper bound of 1-cpu.
1406 runnable_avg = atomic_read(&tg->runnable_avg);
1407 if (runnable_avg < NICE_0_LOAD) {
1408 se->avg.load_avg_contrib *= runnable_avg;
1409 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1413 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1414 int force_update) {}
1415 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1416 struct cfs_rq *cfs_rq) {}
1417 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1420 static inline void __update_task_entity_contrib(struct sched_entity *se)
1424 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1425 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1426 contrib /= (se->avg.runnable_avg_period + 1);
1427 se->avg.load_avg_contrib = scale_load(contrib);
1430 /* Compute the current contribution to load_avg by se, return any delta */
1431 static long __update_entity_load_avg_contrib(struct sched_entity *se)
1433 long old_contrib = se->avg.load_avg_contrib;
1435 if (entity_is_task(se)) {
1436 __update_task_entity_contrib(se);
1438 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1439 __update_group_entity_contrib(se);
1442 return se->avg.load_avg_contrib - old_contrib;
1445 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1448 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1449 cfs_rq->blocked_load_avg -= load_contrib;
1451 cfs_rq->blocked_load_avg = 0;
1454 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1456 /* Update a sched_entity's runnable average */
1457 static inline void update_entity_load_avg(struct sched_entity *se,
1460 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1465 * For a group entity we need to use their owned cfs_rq_clock_task() in
1466 * case they are the parent of a throttled hierarchy.
1468 if (entity_is_task(se))
1469 now = cfs_rq_clock_task(cfs_rq);
1471 now = cfs_rq_clock_task(group_cfs_rq(se));
1473 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
1476 contrib_delta = __update_entity_load_avg_contrib(se);
1482 cfs_rq->runnable_load_avg += contrib_delta;
1484 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1488 * Decay the load contributed by all blocked children and account this so that
1489 * their contribution may appropriately discounted when they wake up.
1491 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1493 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1496 decays = now - cfs_rq->last_decay;
1497 if (!decays && !force_update)
1500 if (atomic_long_read(&cfs_rq->removed_load)) {
1501 unsigned long removed_load;
1502 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
1503 subtract_blocked_load_contrib(cfs_rq, removed_load);
1507 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1509 atomic64_add(decays, &cfs_rq->decay_counter);
1510 cfs_rq->last_decay = now;
1513 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1516 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1518 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
1519 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1522 /* Add the load generated by se into cfs_rq's child load-average */
1523 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1524 struct sched_entity *se,
1528 * We track migrations using entity decay_count <= 0, on a wake-up
1529 * migration we use a negative decay count to track the remote decays
1530 * accumulated while sleeping.
1532 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
1533 * are seen by enqueue_entity_load_avg() as a migration with an already
1534 * constructed load_avg_contrib.
1536 if (unlikely(se->avg.decay_count <= 0)) {
1537 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
1538 if (se->avg.decay_count) {
1540 * In a wake-up migration we have to approximate the
1541 * time sleeping. This is because we can't synchronize
1542 * clock_task between the two cpus, and it is not
1543 * guaranteed to be read-safe. Instead, we can
1544 * approximate this using our carried decays, which are
1545 * explicitly atomically readable.
1547 se->avg.last_runnable_update -= (-se->avg.decay_count)
1549 update_entity_load_avg(se, 0);
1550 /* Indicate that we're now synchronized and on-rq */
1551 se->avg.decay_count = 0;
1556 * Task re-woke on same cpu (or else migrate_task_rq_fair()
1557 * would have made count negative); we must be careful to avoid
1558 * double-accounting blocked time after synchronizing decays.
1560 se->avg.last_runnable_update += __synchronize_entity_decay(se)
1564 /* migrated tasks did not contribute to our blocked load */
1566 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1567 update_entity_load_avg(se, 0);
1570 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1571 /* we force update consideration on load-balancer moves */
1572 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1576 * Remove se's load from this cfs_rq child load-average, if the entity is
1577 * transitioning to a blocked state we track its projected decay using
1580 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1581 struct sched_entity *se,
1584 update_entity_load_avg(se, 1);
1585 /* we force update consideration on load-balancer moves */
1586 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1588 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1590 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1591 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1592 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1596 * Update the rq's load with the elapsed running time before entering
1597 * idle. if the last scheduled task is not a CFS task, idle_enter will
1598 * be the only way to update the runnable statistic.
1600 void idle_enter_fair(struct rq *this_rq)
1602 update_rq_runnable_avg(this_rq, 1);
1606 * Update the rq's load with the elapsed idle time before a task is
1607 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
1608 * be the only way to update the runnable statistic.
1610 void idle_exit_fair(struct rq *this_rq)
1612 update_rq_runnable_avg(this_rq, 0);
1616 static inline void update_entity_load_avg(struct sched_entity *se,
1617 int update_cfs_rq) {}
1618 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1619 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1620 struct sched_entity *se,
1622 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1623 struct sched_entity *se,
1625 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1626 int force_update) {}
1629 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1631 #ifdef CONFIG_SCHEDSTATS
1632 struct task_struct *tsk = NULL;
1634 if (entity_is_task(se))
1637 if (se->statistics.sleep_start) {
1638 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
1643 if (unlikely(delta > se->statistics.sleep_max))
1644 se->statistics.sleep_max = delta;
1646 se->statistics.sleep_start = 0;
1647 se->statistics.sum_sleep_runtime += delta;
1650 account_scheduler_latency(tsk, delta >> 10, 1);
1651 trace_sched_stat_sleep(tsk, delta);
1654 if (se->statistics.block_start) {
1655 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
1660 if (unlikely(delta > se->statistics.block_max))
1661 se->statistics.block_max = delta;
1663 se->statistics.block_start = 0;
1664 se->statistics.sum_sleep_runtime += delta;
1667 if (tsk->in_iowait) {
1668 se->statistics.iowait_sum += delta;
1669 se->statistics.iowait_count++;
1670 trace_sched_stat_iowait(tsk, delta);
1673 trace_sched_stat_blocked(tsk, delta);
1676 * Blocking time is in units of nanosecs, so shift by
1677 * 20 to get a milliseconds-range estimation of the
1678 * amount of time that the task spent sleeping:
1680 if (unlikely(prof_on == SLEEP_PROFILING)) {
1681 profile_hits(SLEEP_PROFILING,
1682 (void *)get_wchan(tsk),
1685 account_scheduler_latency(tsk, delta >> 10, 0);
1691 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1693 #ifdef CONFIG_SCHED_DEBUG
1694 s64 d = se->vruntime - cfs_rq->min_vruntime;
1699 if (d > 3*sysctl_sched_latency)
1700 schedstat_inc(cfs_rq, nr_spread_over);
1705 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1707 u64 vruntime = cfs_rq->min_vruntime;
1710 * The 'current' period is already promised to the current tasks,
1711 * however the extra weight of the new task will slow them down a
1712 * little, place the new task so that it fits in the slot that
1713 * stays open at the end.
1715 if (initial && sched_feat(START_DEBIT))
1716 vruntime += sched_vslice(cfs_rq, se);
1718 /* sleeps up to a single latency don't count. */
1720 unsigned long thresh = sysctl_sched_latency;
1723 * Halve their sleep time's effect, to allow
1724 * for a gentler effect of sleepers:
1726 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1732 /* ensure we never gain time by being placed backwards. */
1733 se->vruntime = max_vruntime(se->vruntime, vruntime);
1736 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1739 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1742 * Update the normalized vruntime before updating min_vruntime
1743 * through calling update_curr().
1745 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1746 se->vruntime += cfs_rq->min_vruntime;
1749 * Update run-time statistics of the 'current'.
1751 update_curr(cfs_rq);
1752 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
1753 account_entity_enqueue(cfs_rq, se);
1754 update_cfs_shares(cfs_rq);
1756 if (flags & ENQUEUE_WAKEUP) {
1757 place_entity(cfs_rq, se, 0);
1758 enqueue_sleeper(cfs_rq, se);
1761 update_stats_enqueue(cfs_rq, se);
1762 check_spread(cfs_rq, se);
1763 if (se != cfs_rq->curr)
1764 __enqueue_entity(cfs_rq, se);
1767 if (cfs_rq->nr_running == 1) {
1768 list_add_leaf_cfs_rq(cfs_rq);
1769 check_enqueue_throttle(cfs_rq);
1773 static void __clear_buddies_last(struct sched_entity *se)
1775 for_each_sched_entity(se) {
1776 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1777 if (cfs_rq->last == se)
1778 cfs_rq->last = NULL;
1784 static void __clear_buddies_next(struct sched_entity *se)
1786 for_each_sched_entity(se) {
1787 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1788 if (cfs_rq->next == se)
1789 cfs_rq->next = NULL;
1795 static void __clear_buddies_skip(struct sched_entity *se)
1797 for_each_sched_entity(se) {
1798 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1799 if (cfs_rq->skip == se)
1800 cfs_rq->skip = NULL;
1806 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1808 if (cfs_rq->last == se)
1809 __clear_buddies_last(se);
1811 if (cfs_rq->next == se)
1812 __clear_buddies_next(se);
1814 if (cfs_rq->skip == se)
1815 __clear_buddies_skip(se);
1818 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1821 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1824 * Update run-time statistics of the 'current'.
1826 update_curr(cfs_rq);
1827 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
1829 update_stats_dequeue(cfs_rq, se);
1830 if (flags & DEQUEUE_SLEEP) {
1831 #ifdef CONFIG_SCHEDSTATS
1832 if (entity_is_task(se)) {
1833 struct task_struct *tsk = task_of(se);
1835 if (tsk->state & TASK_INTERRUPTIBLE)
1836 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
1837 if (tsk->state & TASK_UNINTERRUPTIBLE)
1838 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
1843 clear_buddies(cfs_rq, se);
1845 if (se != cfs_rq->curr)
1846 __dequeue_entity(cfs_rq, se);
1848 account_entity_dequeue(cfs_rq, se);
1851 * Normalize the entity after updating the min_vruntime because the
1852 * update can refer to the ->curr item and we need to reflect this
1853 * movement in our normalized position.
1855 if (!(flags & DEQUEUE_SLEEP))
1856 se->vruntime -= cfs_rq->min_vruntime;
1858 /* return excess runtime on last dequeue */
1859 return_cfs_rq_runtime(cfs_rq);
1861 update_min_vruntime(cfs_rq);
1862 update_cfs_shares(cfs_rq);
1866 * Preempt the current task with a newly woken task if needed:
1869 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1871 unsigned long ideal_runtime, delta_exec;
1872 struct sched_entity *se;
1875 ideal_runtime = sched_slice(cfs_rq, curr);
1876 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1877 if (delta_exec > ideal_runtime) {
1878 resched_task(rq_of(cfs_rq)->curr);
1880 * The current task ran long enough, ensure it doesn't get
1881 * re-elected due to buddy favours.
1883 clear_buddies(cfs_rq, curr);
1888 * Ensure that a task that missed wakeup preemption by a
1889 * narrow margin doesn't have to wait for a full slice.
1890 * This also mitigates buddy induced latencies under load.
1892 if (delta_exec < sysctl_sched_min_granularity)
1895 se = __pick_first_entity(cfs_rq);
1896 delta = curr->vruntime - se->vruntime;
1901 if (delta > ideal_runtime)
1902 resched_task(rq_of(cfs_rq)->curr);
1906 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1908 /* 'current' is not kept within the tree. */
1911 * Any task has to be enqueued before it get to execute on
1912 * a CPU. So account for the time it spent waiting on the
1915 update_stats_wait_end(cfs_rq, se);
1916 __dequeue_entity(cfs_rq, se);
1919 update_stats_curr_start(cfs_rq, se);
1921 #ifdef CONFIG_SCHEDSTATS
1923 * Track our maximum slice length, if the CPU's load is at
1924 * least twice that of our own weight (i.e. dont track it
1925 * when there are only lesser-weight tasks around):
1927 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1928 se->statistics.slice_max = max(se->statistics.slice_max,
1929 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1932 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1936 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1939 * Pick the next process, keeping these things in mind, in this order:
1940 * 1) keep things fair between processes/task groups
1941 * 2) pick the "next" process, since someone really wants that to run
1942 * 3) pick the "last" process, for cache locality
1943 * 4) do not run the "skip" process, if something else is available
1945 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1947 struct sched_entity *se = __pick_first_entity(cfs_rq);
1948 struct sched_entity *left = se;
1951 * Avoid running the skip buddy, if running something else can
1952 * be done without getting too unfair.
1954 if (cfs_rq->skip == se) {
1955 struct sched_entity *second = __pick_next_entity(se);
1956 if (second && wakeup_preempt_entity(second, left) < 1)
1961 * Prefer last buddy, try to return the CPU to a preempted task.
1963 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1967 * Someone really wants this to run. If it's not unfair, run it.
1969 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1972 clear_buddies(cfs_rq, se);
1977 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1979 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1982 * If still on the runqueue then deactivate_task()
1983 * was not called and update_curr() has to be done:
1986 update_curr(cfs_rq);
1988 /* throttle cfs_rqs exceeding runtime */
1989 check_cfs_rq_runtime(cfs_rq);
1991 check_spread(cfs_rq, prev);
1993 update_stats_wait_start(cfs_rq, prev);
1994 /* Put 'current' back into the tree. */
1995 __enqueue_entity(cfs_rq, prev);
1996 /* in !on_rq case, update occurred at dequeue */
1997 update_entity_load_avg(prev, 1);
1999 cfs_rq->curr = NULL;
2003 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2006 * Update run-time statistics of the 'current'.
2008 update_curr(cfs_rq);
2011 * Ensure that runnable average is periodically updated.
2013 update_entity_load_avg(curr, 1);
2014 update_cfs_rq_blocked_load(cfs_rq, 1);
2015 update_cfs_shares(cfs_rq);
2017 #ifdef CONFIG_SCHED_HRTICK
2019 * queued ticks are scheduled to match the slice, so don't bother
2020 * validating it and just reschedule.
2023 resched_task(rq_of(cfs_rq)->curr);
2027 * don't let the period tick interfere with the hrtick preemption
2029 if (!sched_feat(DOUBLE_TICK) &&
2030 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2034 if (cfs_rq->nr_running > 1)
2035 check_preempt_tick(cfs_rq, curr);
2039 /**************************************************
2040 * CFS bandwidth control machinery
2043 #ifdef CONFIG_CFS_BANDWIDTH
2045 #ifdef HAVE_JUMP_LABEL
2046 static struct static_key __cfs_bandwidth_used;
2048 static inline bool cfs_bandwidth_used(void)
2050 return static_key_false(&__cfs_bandwidth_used);
2053 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2055 /* only need to count groups transitioning between enabled/!enabled */
2056 if (enabled && !was_enabled)
2057 static_key_slow_inc(&__cfs_bandwidth_used);
2058 else if (!enabled && was_enabled)
2059 static_key_slow_dec(&__cfs_bandwidth_used);
2061 #else /* HAVE_JUMP_LABEL */
2062 static bool cfs_bandwidth_used(void)
2067 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2068 #endif /* HAVE_JUMP_LABEL */
2071 * default period for cfs group bandwidth.
2072 * default: 0.1s, units: nanoseconds
2074 static inline u64 default_cfs_period(void)
2076 return 100000000ULL;
2079 static inline u64 sched_cfs_bandwidth_slice(void)
2081 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2085 * Replenish runtime according to assigned quota and update expiration time.
2086 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2087 * additional synchronization around rq->lock.
2089 * requires cfs_b->lock
2091 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2095 if (cfs_b->quota == RUNTIME_INF)
2098 now = sched_clock_cpu(smp_processor_id());
2099 cfs_b->runtime = cfs_b->quota;
2100 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2103 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2105 return &tg->cfs_bandwidth;
2108 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2109 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2111 if (unlikely(cfs_rq->throttle_count))
2112 return cfs_rq->throttled_clock_task;
2114 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2117 /* returns 0 on failure to allocate runtime */
2118 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2120 struct task_group *tg = cfs_rq->tg;
2121 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2122 u64 amount = 0, min_amount, expires;
2124 /* note: this is a positive sum as runtime_remaining <= 0 */
2125 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2127 raw_spin_lock(&cfs_b->lock);
2128 if (cfs_b->quota == RUNTIME_INF)
2129 amount = min_amount;
2132 * If the bandwidth pool has become inactive, then at least one
2133 * period must have elapsed since the last consumption.
2134 * Refresh the global state and ensure bandwidth timer becomes
2137 if (!cfs_b->timer_active) {
2138 __refill_cfs_bandwidth_runtime(cfs_b);
2139 __start_cfs_bandwidth(cfs_b);
2142 if (cfs_b->runtime > 0) {
2143 amount = min(cfs_b->runtime, min_amount);
2144 cfs_b->runtime -= amount;
2148 expires = cfs_b->runtime_expires;
2149 raw_spin_unlock(&cfs_b->lock);
2151 cfs_rq->runtime_remaining += amount;
2153 * we may have advanced our local expiration to account for allowed
2154 * spread between our sched_clock and the one on which runtime was
2157 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2158 cfs_rq->runtime_expires = expires;
2160 return cfs_rq->runtime_remaining > 0;
2164 * Note: This depends on the synchronization provided by sched_clock and the
2165 * fact that rq->clock snapshots this value.
2167 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2169 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2171 /* if the deadline is ahead of our clock, nothing to do */
2172 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2175 if (cfs_rq->runtime_remaining < 0)
2179 * If the local deadline has passed we have to consider the
2180 * possibility that our sched_clock is 'fast' and the global deadline
2181 * has not truly expired.
2183 * Fortunately we can check determine whether this the case by checking
2184 * whether the global deadline has advanced.
2187 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2188 /* extend local deadline, drift is bounded above by 2 ticks */
2189 cfs_rq->runtime_expires += TICK_NSEC;
2191 /* global deadline is ahead, expiration has passed */
2192 cfs_rq->runtime_remaining = 0;
2196 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2197 unsigned long delta_exec)
2199 /* dock delta_exec before expiring quota (as it could span periods) */
2200 cfs_rq->runtime_remaining -= delta_exec;
2201 expire_cfs_rq_runtime(cfs_rq);
2203 if (likely(cfs_rq->runtime_remaining > 0))
2207 * if we're unable to extend our runtime we resched so that the active
2208 * hierarchy can be throttled
2210 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2211 resched_task(rq_of(cfs_rq)->curr);
2214 static __always_inline
2215 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2217 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2220 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2223 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2225 return cfs_bandwidth_used() && cfs_rq->throttled;
2228 /* check whether cfs_rq, or any parent, is throttled */
2229 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2231 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2235 * Ensure that neither of the group entities corresponding to src_cpu or
2236 * dest_cpu are members of a throttled hierarchy when performing group
2237 * load-balance operations.
2239 static inline int throttled_lb_pair(struct task_group *tg,
2240 int src_cpu, int dest_cpu)
2242 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2244 src_cfs_rq = tg->cfs_rq[src_cpu];
2245 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2247 return throttled_hierarchy(src_cfs_rq) ||
2248 throttled_hierarchy(dest_cfs_rq);
2251 /* updated child weight may affect parent so we have to do this bottom up */
2252 static int tg_unthrottle_up(struct task_group *tg, void *data)
2254 struct rq *rq = data;
2255 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2257 cfs_rq->throttle_count--;
2259 if (!cfs_rq->throttle_count) {
2260 /* adjust cfs_rq_clock_task() */
2261 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
2262 cfs_rq->throttled_clock_task;
2269 static int tg_throttle_down(struct task_group *tg, void *data)
2271 struct rq *rq = data;
2272 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2274 /* group is entering throttled state, stop time */
2275 if (!cfs_rq->throttle_count)
2276 cfs_rq->throttled_clock_task = rq_clock_task(rq);
2277 cfs_rq->throttle_count++;
2282 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2284 struct rq *rq = rq_of(cfs_rq);
2285 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2286 struct sched_entity *se;
2287 long task_delta, dequeue = 1;
2289 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2291 /* freeze hierarchy runnable averages while throttled */
2293 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2296 task_delta = cfs_rq->h_nr_running;
2297 for_each_sched_entity(se) {
2298 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2299 /* throttled entity or throttle-on-deactivate */
2304 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2305 qcfs_rq->h_nr_running -= task_delta;
2307 if (qcfs_rq->load.weight)
2312 rq->nr_running -= task_delta;
2314 cfs_rq->throttled = 1;
2315 cfs_rq->throttled_clock = rq_clock(rq);
2316 raw_spin_lock(&cfs_b->lock);
2317 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2318 raw_spin_unlock(&cfs_b->lock);
2321 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2323 struct rq *rq = rq_of(cfs_rq);
2324 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2325 struct sched_entity *se;
2329 se = cfs_rq->tg->se[cpu_of(rq)];
2331 cfs_rq->throttled = 0;
2333 update_rq_clock(rq);
2335 raw_spin_lock(&cfs_b->lock);
2336 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
2337 list_del_rcu(&cfs_rq->throttled_list);
2338 raw_spin_unlock(&cfs_b->lock);
2340 /* update hierarchical throttle state */
2341 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2343 if (!cfs_rq->load.weight)
2346 task_delta = cfs_rq->h_nr_running;
2347 for_each_sched_entity(se) {
2351 cfs_rq = cfs_rq_of(se);
2353 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2354 cfs_rq->h_nr_running += task_delta;
2356 if (cfs_rq_throttled(cfs_rq))
2361 rq->nr_running += task_delta;
2363 /* determine whether we need to wake up potentially idle cpu */
2364 if (rq->curr == rq->idle && rq->cfs.nr_running)
2365 resched_task(rq->curr);
2368 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2369 u64 remaining, u64 expires)
2371 struct cfs_rq *cfs_rq;
2372 u64 runtime = remaining;
2375 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2377 struct rq *rq = rq_of(cfs_rq);
2379 raw_spin_lock(&rq->lock);
2380 if (!cfs_rq_throttled(cfs_rq))
2383 runtime = -cfs_rq->runtime_remaining + 1;
2384 if (runtime > remaining)
2385 runtime = remaining;
2386 remaining -= runtime;
2388 cfs_rq->runtime_remaining += runtime;
2389 cfs_rq->runtime_expires = expires;
2391 /* we check whether we're throttled above */
2392 if (cfs_rq->runtime_remaining > 0)
2393 unthrottle_cfs_rq(cfs_rq);
2396 raw_spin_unlock(&rq->lock);
2407 * Responsible for refilling a task_group's bandwidth and unthrottling its
2408 * cfs_rqs as appropriate. If there has been no activity within the last
2409 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2410 * used to track this state.
2412 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2414 u64 runtime, runtime_expires;
2415 int idle = 1, throttled;
2417 raw_spin_lock(&cfs_b->lock);
2418 /* no need to continue the timer with no bandwidth constraint */
2419 if (cfs_b->quota == RUNTIME_INF)
2422 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2423 /* idle depends on !throttled (for the case of a large deficit) */
2424 idle = cfs_b->idle && !throttled;
2425 cfs_b->nr_periods += overrun;
2427 /* if we're going inactive then everything else can be deferred */
2431 __refill_cfs_bandwidth_runtime(cfs_b);
2434 /* mark as potentially idle for the upcoming period */
2439 /* account preceding periods in which throttling occurred */
2440 cfs_b->nr_throttled += overrun;
2443 * There are throttled entities so we must first use the new bandwidth
2444 * to unthrottle them before making it generally available. This
2445 * ensures that all existing debts will be paid before a new cfs_rq is
2448 runtime = cfs_b->runtime;
2449 runtime_expires = cfs_b->runtime_expires;
2453 * This check is repeated as we are holding onto the new bandwidth
2454 * while we unthrottle. This can potentially race with an unthrottled
2455 * group trying to acquire new bandwidth from the global pool.
2457 while (throttled && runtime > 0) {
2458 raw_spin_unlock(&cfs_b->lock);
2459 /* we can't nest cfs_b->lock while distributing bandwidth */
2460 runtime = distribute_cfs_runtime(cfs_b, runtime,
2462 raw_spin_lock(&cfs_b->lock);
2464 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2467 /* return (any) remaining runtime */
2468 cfs_b->runtime = runtime;
2470 * While we are ensured activity in the period following an
2471 * unthrottle, this also covers the case in which the new bandwidth is
2472 * insufficient to cover the existing bandwidth deficit. (Forcing the
2473 * timer to remain active while there are any throttled entities.)
2478 cfs_b->timer_active = 0;
2479 raw_spin_unlock(&cfs_b->lock);
2484 /* a cfs_rq won't donate quota below this amount */
2485 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2486 /* minimum remaining period time to redistribute slack quota */
2487 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2488 /* how long we wait to gather additional slack before distributing */
2489 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2491 /* are we near the end of the current quota period? */
2492 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2494 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2497 /* if the call-back is running a quota refresh is already occurring */
2498 if (hrtimer_callback_running(refresh_timer))
2501 /* is a quota refresh about to occur? */
2502 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2503 if (remaining < min_expire)
2509 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2511 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2513 /* if there's a quota refresh soon don't bother with slack */
2514 if (runtime_refresh_within(cfs_b, min_left))
2517 start_bandwidth_timer(&cfs_b->slack_timer,
2518 ns_to_ktime(cfs_bandwidth_slack_period));
2521 /* we know any runtime found here is valid as update_curr() precedes return */
2522 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2524 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2525 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2527 if (slack_runtime <= 0)
2530 raw_spin_lock(&cfs_b->lock);
2531 if (cfs_b->quota != RUNTIME_INF &&
2532 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2533 cfs_b->runtime += slack_runtime;
2535 /* we are under rq->lock, defer unthrottling using a timer */
2536 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2537 !list_empty(&cfs_b->throttled_cfs_rq))
2538 start_cfs_slack_bandwidth(cfs_b);
2540 raw_spin_unlock(&cfs_b->lock);
2542 /* even if it's not valid for return we don't want to try again */
2543 cfs_rq->runtime_remaining -= slack_runtime;
2546 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2548 if (!cfs_bandwidth_used())
2551 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2554 __return_cfs_rq_runtime(cfs_rq);
2558 * This is done with a timer (instead of inline with bandwidth return) since
2559 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2561 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2563 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2566 /* confirm we're still not at a refresh boundary */
2567 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2570 raw_spin_lock(&cfs_b->lock);
2571 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2572 runtime = cfs_b->runtime;
2575 expires = cfs_b->runtime_expires;
2576 raw_spin_unlock(&cfs_b->lock);
2581 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2583 raw_spin_lock(&cfs_b->lock);
2584 if (expires == cfs_b->runtime_expires)
2585 cfs_b->runtime = runtime;
2586 raw_spin_unlock(&cfs_b->lock);
2590 * When a group wakes up we want to make sure that its quota is not already
2591 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2592 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2594 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2596 if (!cfs_bandwidth_used())
2599 /* an active group must be handled by the update_curr()->put() path */
2600 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2603 /* ensure the group is not already throttled */
2604 if (cfs_rq_throttled(cfs_rq))
2607 /* update runtime allocation */
2608 account_cfs_rq_runtime(cfs_rq, 0);
2609 if (cfs_rq->runtime_remaining <= 0)
2610 throttle_cfs_rq(cfs_rq);
2613 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2614 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2616 if (!cfs_bandwidth_used())
2619 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2623 * it's possible for a throttled entity to be forced into a running
2624 * state (e.g. set_curr_task), in this case we're finished.
2626 if (cfs_rq_throttled(cfs_rq))
2629 throttle_cfs_rq(cfs_rq);
2632 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2634 struct cfs_bandwidth *cfs_b =
2635 container_of(timer, struct cfs_bandwidth, slack_timer);
2636 do_sched_cfs_slack_timer(cfs_b);
2638 return HRTIMER_NORESTART;
2641 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2643 struct cfs_bandwidth *cfs_b =
2644 container_of(timer, struct cfs_bandwidth, period_timer);
2650 now = hrtimer_cb_get_time(timer);
2651 overrun = hrtimer_forward(timer, now, cfs_b->period);
2656 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2659 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2662 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2664 raw_spin_lock_init(&cfs_b->lock);
2666 cfs_b->quota = RUNTIME_INF;
2667 cfs_b->period = ns_to_ktime(default_cfs_period());
2669 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2670 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2671 cfs_b->period_timer.function = sched_cfs_period_timer;
2672 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2673 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2676 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2678 cfs_rq->runtime_enabled = 0;
2679 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2682 /* requires cfs_b->lock, may release to reprogram timer */
2683 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2686 * The timer may be active because we're trying to set a new bandwidth
2687 * period or because we're racing with the tear-down path
2688 * (timer_active==0 becomes visible before the hrtimer call-back
2689 * terminates). In either case we ensure that it's re-programmed
2691 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2692 raw_spin_unlock(&cfs_b->lock);
2693 /* ensure cfs_b->lock is available while we wait */
2694 hrtimer_cancel(&cfs_b->period_timer);
2696 raw_spin_lock(&cfs_b->lock);
2697 /* if someone else restarted the timer then we're done */
2698 if (cfs_b->timer_active)
2702 cfs_b->timer_active = 1;
2703 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2706 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2708 hrtimer_cancel(&cfs_b->period_timer);
2709 hrtimer_cancel(&cfs_b->slack_timer);
2712 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
2714 struct cfs_rq *cfs_rq;
2716 for_each_leaf_cfs_rq(rq, cfs_rq) {
2717 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2719 if (!cfs_rq->runtime_enabled)
2723 * clock_task is not advancing so we just need to make sure
2724 * there's some valid quota amount
2726 cfs_rq->runtime_remaining = cfs_b->quota;
2727 if (cfs_rq_throttled(cfs_rq))
2728 unthrottle_cfs_rq(cfs_rq);
2732 #else /* CONFIG_CFS_BANDWIDTH */
2733 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2735 return rq_clock_task(rq_of(cfs_rq));
2738 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2739 unsigned long delta_exec) {}
2740 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2741 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2742 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2744 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2749 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2754 static inline int throttled_lb_pair(struct task_group *tg,
2755 int src_cpu, int dest_cpu)
2760 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2762 #ifdef CONFIG_FAIR_GROUP_SCHED
2763 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2766 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2770 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2771 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2773 #endif /* CONFIG_CFS_BANDWIDTH */
2775 /**************************************************
2776 * CFS operations on tasks:
2779 #ifdef CONFIG_SCHED_HRTICK
2780 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2782 struct sched_entity *se = &p->se;
2783 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2785 WARN_ON(task_rq(p) != rq);
2787 if (cfs_rq->nr_running > 1) {
2788 u64 slice = sched_slice(cfs_rq, se);
2789 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2790 s64 delta = slice - ran;
2799 * Don't schedule slices shorter than 10000ns, that just
2800 * doesn't make sense. Rely on vruntime for fairness.
2803 delta = max_t(s64, 10000LL, delta);
2805 hrtick_start(rq, delta);
2810 * called from enqueue/dequeue and updates the hrtick when the
2811 * current task is from our class and nr_running is low enough
2814 static void hrtick_update(struct rq *rq)
2816 struct task_struct *curr = rq->curr;
2818 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2821 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2822 hrtick_start_fair(rq, curr);
2824 #else /* !CONFIG_SCHED_HRTICK */
2826 hrtick_start_fair(struct rq *rq, struct task_struct *p)
2830 static inline void hrtick_update(struct rq *rq)
2836 * The enqueue_task method is called before nr_running is
2837 * increased. Here we update the fair scheduling stats and
2838 * then put the task into the rbtree:
2841 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2843 struct cfs_rq *cfs_rq;
2844 struct sched_entity *se = &p->se;
2846 for_each_sched_entity(se) {
2849 cfs_rq = cfs_rq_of(se);
2850 enqueue_entity(cfs_rq, se, flags);
2853 * end evaluation on encountering a throttled cfs_rq
2855 * note: in the case of encountering a throttled cfs_rq we will
2856 * post the final h_nr_running increment below.
2858 if (cfs_rq_throttled(cfs_rq))
2860 cfs_rq->h_nr_running++;
2862 flags = ENQUEUE_WAKEUP;
2865 for_each_sched_entity(se) {
2866 cfs_rq = cfs_rq_of(se);
2867 cfs_rq->h_nr_running++;
2869 if (cfs_rq_throttled(cfs_rq))
2872 update_cfs_shares(cfs_rq);
2873 update_entity_load_avg(se, 1);
2877 update_rq_runnable_avg(rq, rq->nr_running);
2883 static void set_next_buddy(struct sched_entity *se);
2886 * The dequeue_task method is called before nr_running is
2887 * decreased. We remove the task from the rbtree and
2888 * update the fair scheduling stats:
2890 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2892 struct cfs_rq *cfs_rq;
2893 struct sched_entity *se = &p->se;
2894 int task_sleep = flags & DEQUEUE_SLEEP;
2896 for_each_sched_entity(se) {
2897 cfs_rq = cfs_rq_of(se);
2898 dequeue_entity(cfs_rq, se, flags);
2901 * end evaluation on encountering a throttled cfs_rq
2903 * note: in the case of encountering a throttled cfs_rq we will
2904 * post the final h_nr_running decrement below.
2906 if (cfs_rq_throttled(cfs_rq))
2908 cfs_rq->h_nr_running--;
2910 /* Don't dequeue parent if it has other entities besides us */
2911 if (cfs_rq->load.weight) {
2913 * Bias pick_next to pick a task from this cfs_rq, as
2914 * p is sleeping when it is within its sched_slice.
2916 if (task_sleep && parent_entity(se))
2917 set_next_buddy(parent_entity(se));
2919 /* avoid re-evaluating load for this entity */
2920 se = parent_entity(se);
2923 flags |= DEQUEUE_SLEEP;
2926 for_each_sched_entity(se) {
2927 cfs_rq = cfs_rq_of(se);
2928 cfs_rq->h_nr_running--;
2930 if (cfs_rq_throttled(cfs_rq))
2933 update_cfs_shares(cfs_rq);
2934 update_entity_load_avg(se, 1);
2939 update_rq_runnable_avg(rq, 1);
2945 /* Used instead of source_load when we know the type == 0 */
2946 static unsigned long weighted_cpuload(const int cpu)
2948 return cpu_rq(cpu)->cfs.runnable_load_avg;
2952 * Return a low guess at the load of a migration-source cpu weighted
2953 * according to the scheduling class and "nice" value.
2955 * We want to under-estimate the load of migration sources, to
2956 * balance conservatively.
2958 static unsigned long source_load(int cpu, int type)
2960 struct rq *rq = cpu_rq(cpu);
2961 unsigned long total = weighted_cpuload(cpu);
2963 if (type == 0 || !sched_feat(LB_BIAS))
2966 return min(rq->cpu_load[type-1], total);
2970 * Return a high guess at the load of a migration-target cpu weighted
2971 * according to the scheduling class and "nice" value.
2973 static unsigned long target_load(int cpu, int type)
2975 struct rq *rq = cpu_rq(cpu);
2976 unsigned long total = weighted_cpuload(cpu);
2978 if (type == 0 || !sched_feat(LB_BIAS))
2981 return max(rq->cpu_load[type-1], total);
2984 static unsigned long power_of(int cpu)
2986 return cpu_rq(cpu)->cpu_power;
2989 static unsigned long cpu_avg_load_per_task(int cpu)
2991 struct rq *rq = cpu_rq(cpu);
2992 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
2993 unsigned long load_avg = rq->cfs.runnable_load_avg;
2996 return load_avg / nr_running;
3001 static void record_wakee(struct task_struct *p)
3004 * Rough decay (wiping) for cost saving, don't worry
3005 * about the boundary, really active task won't care
3008 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3009 current->wakee_flips = 0;
3010 current->wakee_flip_decay_ts = jiffies;
3013 if (current->last_wakee != p) {
3014 current->last_wakee = p;
3015 current->wakee_flips++;
3019 static void task_waking_fair(struct task_struct *p)
3021 struct sched_entity *se = &p->se;
3022 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3025 #ifndef CONFIG_64BIT
3026 u64 min_vruntime_copy;
3029 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3031 min_vruntime = cfs_rq->min_vruntime;
3032 } while (min_vruntime != min_vruntime_copy);
3034 min_vruntime = cfs_rq->min_vruntime;
3037 se->vruntime -= min_vruntime;
3041 #ifdef CONFIG_FAIR_GROUP_SCHED
3043 * effective_load() calculates the load change as seen from the root_task_group
3045 * Adding load to a group doesn't make a group heavier, but can cause movement
3046 * of group shares between cpus. Assuming the shares were perfectly aligned one
3047 * can calculate the shift in shares.
3049 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3050 * on this @cpu and results in a total addition (subtraction) of @wg to the
3051 * total group weight.
3053 * Given a runqueue weight distribution (rw_i) we can compute a shares
3054 * distribution (s_i) using:
3056 * s_i = rw_i / \Sum rw_j (1)
3058 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3059 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3060 * shares distribution (s_i):
3062 * rw_i = { 2, 4, 1, 0 }
3063 * s_i = { 2/7, 4/7, 1/7, 0 }
3065 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3066 * task used to run on and the CPU the waker is running on), we need to
3067 * compute the effect of waking a task on either CPU and, in case of a sync
3068 * wakeup, compute the effect of the current task going to sleep.
3070 * So for a change of @wl to the local @cpu with an overall group weight change
3071 * of @wl we can compute the new shares distribution (s'_i) using:
3073 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3075 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3076 * differences in waking a task to CPU 0. The additional task changes the
3077 * weight and shares distributions like:
3079 * rw'_i = { 3, 4, 1, 0 }
3080 * s'_i = { 3/8, 4/8, 1/8, 0 }
3082 * We can then compute the difference in effective weight by using:
3084 * dw_i = S * (s'_i - s_i) (3)
3086 * Where 'S' is the group weight as seen by its parent.
3088 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3089 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3090 * 4/7) times the weight of the group.
3092 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3094 struct sched_entity *se = tg->se[cpu];
3096 if (!tg->parent) /* the trivial, non-cgroup case */
3099 for_each_sched_entity(se) {
3105 * W = @wg + \Sum rw_j
3107 W = wg + calc_tg_weight(tg, se->my_q);
3112 w = se->my_q->load.weight + wl;
3115 * wl = S * s'_i; see (2)
3118 wl = (w * tg->shares) / W;
3123 * Per the above, wl is the new se->load.weight value; since
3124 * those are clipped to [MIN_SHARES, ...) do so now. See
3125 * calc_cfs_shares().
3127 if (wl < MIN_SHARES)
3131 * wl = dw_i = S * (s'_i - s_i); see (3)
3133 wl -= se->load.weight;
3136 * Recursively apply this logic to all parent groups to compute
3137 * the final effective load change on the root group. Since
3138 * only the @tg group gets extra weight, all parent groups can
3139 * only redistribute existing shares. @wl is the shift in shares
3140 * resulting from this level per the above.
3149 static inline unsigned long effective_load(struct task_group *tg, int cpu,
3150 unsigned long wl, unsigned long wg)
3157 static int wake_wide(struct task_struct *p)
3159 int factor = this_cpu_read(sd_llc_size);
3162 * Yeah, it's the switching-frequency, could means many wakee or
3163 * rapidly switch, use factor here will just help to automatically
3164 * adjust the loose-degree, so bigger node will lead to more pull.
3166 if (p->wakee_flips > factor) {
3168 * wakee is somewhat hot, it needs certain amount of cpu
3169 * resource, so if waker is far more hot, prefer to leave
3172 if (current->wakee_flips > (factor * p->wakee_flips))
3179 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3181 s64 this_load, load;
3182 int idx, this_cpu, prev_cpu;
3183 unsigned long tl_per_task;
3184 struct task_group *tg;
3185 unsigned long weight;
3189 * If we wake multiple tasks be careful to not bounce
3190 * ourselves around too much.
3196 this_cpu = smp_processor_id();
3197 prev_cpu = task_cpu(p);
3198 load = source_load(prev_cpu, idx);
3199 this_load = target_load(this_cpu, idx);
3202 * If sync wakeup then subtract the (maximum possible)
3203 * effect of the currently running task from the load
3204 * of the current CPU:
3207 tg = task_group(current);
3208 weight = current->se.load.weight;
3210 this_load += effective_load(tg, this_cpu, -weight, -weight);
3211 load += effective_load(tg, prev_cpu, 0, -weight);
3215 weight = p->se.load.weight;
3218 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3219 * due to the sync cause above having dropped this_load to 0, we'll
3220 * always have an imbalance, but there's really nothing you can do
3221 * about that, so that's good too.
3223 * Otherwise check if either cpus are near enough in load to allow this
3224 * task to be woken on this_cpu.
3226 if (this_load > 0) {
3227 s64 this_eff_load, prev_eff_load;
3229 this_eff_load = 100;
3230 this_eff_load *= power_of(prev_cpu);
3231 this_eff_load *= this_load +
3232 effective_load(tg, this_cpu, weight, weight);
3234 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3235 prev_eff_load *= power_of(this_cpu);
3236 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3238 balanced = this_eff_load <= prev_eff_load;
3243 * If the currently running task will sleep within
3244 * a reasonable amount of time then attract this newly
3247 if (sync && balanced)
3250 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3251 tl_per_task = cpu_avg_load_per_task(this_cpu);
3254 (this_load <= load &&
3255 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3257 * This domain has SD_WAKE_AFFINE and
3258 * p is cache cold in this domain, and
3259 * there is no bad imbalance.
3261 schedstat_inc(sd, ttwu_move_affine);
3262 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3270 * find_idlest_group finds and returns the least busy CPU group within the
3273 static struct sched_group *
3274 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3275 int this_cpu, int load_idx)
3277 struct sched_group *idlest = NULL, *group = sd->groups;
3278 unsigned long min_load = ULONG_MAX, this_load = 0;
3279 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3282 unsigned long load, avg_load;
3286 /* Skip over this group if it has no CPUs allowed */
3287 if (!cpumask_intersects(sched_group_cpus(group),
3288 tsk_cpus_allowed(p)))
3291 local_group = cpumask_test_cpu(this_cpu,
3292 sched_group_cpus(group));
3294 /* Tally up the load of all CPUs in the group */
3297 for_each_cpu(i, sched_group_cpus(group)) {
3298 /* Bias balancing toward cpus of our domain */
3300 load = source_load(i, load_idx);
3302 load = target_load(i, load_idx);
3307 /* Adjust by relative CPU power of the group */
3308 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3311 this_load = avg_load;
3312 } else if (avg_load < min_load) {
3313 min_load = avg_load;
3316 } while (group = group->next, group != sd->groups);
3318 if (!idlest || 100*this_load < imbalance*min_load)
3324 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3327 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3329 unsigned long load, min_load = ULONG_MAX;
3333 /* Traverse only the allowed CPUs */
3334 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3335 load = weighted_cpuload(i);
3337 if (load < min_load || (load == min_load && i == this_cpu)) {
3347 * Try and locate an idle CPU in the sched_domain.
3349 static int select_idle_sibling(struct task_struct *p, int target)
3351 struct sched_domain *sd;
3352 struct sched_group *sg;
3353 int i = task_cpu(p);
3355 if (idle_cpu(target))
3359 * If the prevous cpu is cache affine and idle, don't be stupid.
3361 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3365 * Otherwise, iterate the domains and find an elegible idle cpu.
3367 sd = rcu_dereference(per_cpu(sd_llc, target));
3368 for_each_lower_domain(sd) {
3371 if (!cpumask_intersects(sched_group_cpus(sg),
3372 tsk_cpus_allowed(p)))
3375 for_each_cpu(i, sched_group_cpus(sg)) {
3376 if (i == target || !idle_cpu(i))
3380 target = cpumask_first_and(sched_group_cpus(sg),
3381 tsk_cpus_allowed(p));
3385 } while (sg != sd->groups);
3392 * sched_balance_self: balance the current task (running on cpu) in domains
3393 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3396 * Balance, ie. select the least loaded group.
3398 * Returns the target CPU number, or the same CPU if no balancing is needed.
3400 * preempt must be disabled.
3403 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
3405 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3406 int cpu = smp_processor_id();
3407 int prev_cpu = task_cpu(p);
3409 int want_affine = 0;
3410 int sync = wake_flags & WF_SYNC;
3412 if (p->nr_cpus_allowed == 1)
3415 if (sd_flag & SD_BALANCE_WAKE) {
3416 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
3422 for_each_domain(cpu, tmp) {
3423 if (!(tmp->flags & SD_LOAD_BALANCE))
3427 * If both cpu and prev_cpu are part of this domain,
3428 * cpu is a valid SD_WAKE_AFFINE target.
3430 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3431 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3436 if (tmp->flags & sd_flag)
3441 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3444 new_cpu = select_idle_sibling(p, prev_cpu);
3449 int load_idx = sd->forkexec_idx;
3450 struct sched_group *group;
3453 if (!(sd->flags & sd_flag)) {
3458 if (sd_flag & SD_BALANCE_WAKE)
3459 load_idx = sd->wake_idx;
3461 group = find_idlest_group(sd, p, cpu, load_idx);
3467 new_cpu = find_idlest_cpu(group, p, cpu);
3468 if (new_cpu == -1 || new_cpu == cpu) {
3469 /* Now try balancing at a lower domain level of cpu */
3474 /* Now try balancing at a lower domain level of new_cpu */
3476 weight = sd->span_weight;
3478 for_each_domain(cpu, tmp) {
3479 if (weight <= tmp->span_weight)
3481 if (tmp->flags & sd_flag)
3484 /* while loop will break here if sd == NULL */
3493 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3494 * cfs_rq_of(p) references at time of call are still valid and identify the
3495 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3496 * other assumptions, including the state of rq->lock, should be made.
3499 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
3501 struct sched_entity *se = &p->se;
3502 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3505 * Load tracking: accumulate removed load so that it can be processed
3506 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3507 * to blocked load iff they have a positive decay-count. It can never
3508 * be negative here since on-rq tasks have decay-count == 0.
3510 if (se->avg.decay_count) {
3511 se->avg.decay_count = -__synchronize_entity_decay(se);
3512 atomic_long_add(se->avg.load_avg_contrib,
3513 &cfs_rq->removed_load);
3516 #endif /* CONFIG_SMP */
3518 static unsigned long
3519 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3521 unsigned long gran = sysctl_sched_wakeup_granularity;
3524 * Since its curr running now, convert the gran from real-time
3525 * to virtual-time in his units.
3527 * By using 'se' instead of 'curr' we penalize light tasks, so
3528 * they get preempted easier. That is, if 'se' < 'curr' then
3529 * the resulting gran will be larger, therefore penalizing the
3530 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3531 * be smaller, again penalizing the lighter task.
3533 * This is especially important for buddies when the leftmost
3534 * task is higher priority than the buddy.
3536 return calc_delta_fair(gran, se);
3540 * Should 'se' preempt 'curr'.
3554 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3556 s64 gran, vdiff = curr->vruntime - se->vruntime;
3561 gran = wakeup_gran(curr, se);
3568 static void set_last_buddy(struct sched_entity *se)
3570 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3573 for_each_sched_entity(se)
3574 cfs_rq_of(se)->last = se;
3577 static void set_next_buddy(struct sched_entity *se)
3579 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3582 for_each_sched_entity(se)
3583 cfs_rq_of(se)->next = se;
3586 static void set_skip_buddy(struct sched_entity *se)
3588 for_each_sched_entity(se)
3589 cfs_rq_of(se)->skip = se;
3593 * Preempt the current task with a newly woken task if needed:
3595 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3597 struct task_struct *curr = rq->curr;
3598 struct sched_entity *se = &curr->se, *pse = &p->se;
3599 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3600 int scale = cfs_rq->nr_running >= sched_nr_latency;
3601 int next_buddy_marked = 0;
3603 if (unlikely(se == pse))
3607 * This is possible from callers such as move_task(), in which we
3608 * unconditionally check_prempt_curr() after an enqueue (which may have
3609 * lead to a throttle). This both saves work and prevents false
3610 * next-buddy nomination below.
3612 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3615 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3616 set_next_buddy(pse);
3617 next_buddy_marked = 1;
3621 * We can come here with TIF_NEED_RESCHED already set from new task
3624 * Note: this also catches the edge-case of curr being in a throttled
3625 * group (e.g. via set_curr_task), since update_curr() (in the
3626 * enqueue of curr) will have resulted in resched being set. This
3627 * prevents us from potentially nominating it as a false LAST_BUDDY
3630 if (test_tsk_need_resched(curr))
3633 /* Idle tasks are by definition preempted by non-idle tasks. */
3634 if (unlikely(curr->policy == SCHED_IDLE) &&
3635 likely(p->policy != SCHED_IDLE))
3639 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3640 * is driven by the tick):
3642 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
3645 find_matching_se(&se, &pse);
3646 update_curr(cfs_rq_of(se));
3648 if (wakeup_preempt_entity(se, pse) == 1) {
3650 * Bias pick_next to pick the sched entity that is
3651 * triggering this preemption.
3653 if (!next_buddy_marked)
3654 set_next_buddy(pse);
3663 * Only set the backward buddy when the current task is still
3664 * on the rq. This can happen when a wakeup gets interleaved
3665 * with schedule on the ->pre_schedule() or idle_balance()
3666 * point, either of which can * drop the rq lock.
3668 * Also, during early boot the idle thread is in the fair class,
3669 * for obvious reasons its a bad idea to schedule back to it.
3671 if (unlikely(!se->on_rq || curr == rq->idle))
3674 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3678 static struct task_struct *pick_next_task_fair(struct rq *rq)
3680 struct task_struct *p;
3681 struct cfs_rq *cfs_rq = &rq->cfs;
3682 struct sched_entity *se;
3684 if (!cfs_rq->nr_running)
3688 se = pick_next_entity(cfs_rq);
3689 set_next_entity(cfs_rq, se);
3690 cfs_rq = group_cfs_rq(se);
3694 if (hrtick_enabled(rq))
3695 hrtick_start_fair(rq, p);
3701 * Account for a descheduled task:
3703 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3705 struct sched_entity *se = &prev->se;
3706 struct cfs_rq *cfs_rq;
3708 for_each_sched_entity(se) {
3709 cfs_rq = cfs_rq_of(se);
3710 put_prev_entity(cfs_rq, se);
3715 * sched_yield() is very simple
3717 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3719 static void yield_task_fair(struct rq *rq)
3721 struct task_struct *curr = rq->curr;
3722 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3723 struct sched_entity *se = &curr->se;
3726 * Are we the only task in the tree?
3728 if (unlikely(rq->nr_running == 1))
3731 clear_buddies(cfs_rq, se);
3733 if (curr->policy != SCHED_BATCH) {
3734 update_rq_clock(rq);
3736 * Update run-time statistics of the 'current'.
3738 update_curr(cfs_rq);
3740 * Tell update_rq_clock() that we've just updated,
3741 * so we don't do microscopic update in schedule()
3742 * and double the fastpath cost.
3744 rq->skip_clock_update = 1;
3750 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3752 struct sched_entity *se = &p->se;
3754 /* throttled hierarchies are not runnable */
3755 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3758 /* Tell the scheduler that we'd really like pse to run next. */
3761 yield_task_fair(rq);
3767 /**************************************************
3768 * Fair scheduling class load-balancing methods.
3772 * The purpose of load-balancing is to achieve the same basic fairness the
3773 * per-cpu scheduler provides, namely provide a proportional amount of compute
3774 * time to each task. This is expressed in the following equation:
3776 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
3778 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
3779 * W_i,0 is defined as:
3781 * W_i,0 = \Sum_j w_i,j (2)
3783 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
3784 * is derived from the nice value as per prio_to_weight[].
3786 * The weight average is an exponential decay average of the instantaneous
3789 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
3791 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
3792 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
3793 * can also include other factors [XXX].
3795 * To achieve this balance we define a measure of imbalance which follows
3796 * directly from (1):
3798 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
3800 * We them move tasks around to minimize the imbalance. In the continuous
3801 * function space it is obvious this converges, in the discrete case we get
3802 * a few fun cases generally called infeasible weight scenarios.
3805 * - infeasible weights;
3806 * - local vs global optima in the discrete case. ]
3811 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
3812 * for all i,j solution, we create a tree of cpus that follows the hardware
3813 * topology where each level pairs two lower groups (or better). This results
3814 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
3815 * tree to only the first of the previous level and we decrease the frequency
3816 * of load-balance at each level inv. proportional to the number of cpus in
3822 * \Sum { --- * --- * 2^i } = O(n) (5)
3824 * `- size of each group
3825 * | | `- number of cpus doing load-balance
3827 * `- sum over all levels
3829 * Coupled with a limit on how many tasks we can migrate every balance pass,
3830 * this makes (5) the runtime complexity of the balancer.
3832 * An important property here is that each CPU is still (indirectly) connected
3833 * to every other cpu in at most O(log n) steps:
3835 * The adjacency matrix of the resulting graph is given by:
3838 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
3841 * And you'll find that:
3843 * A^(log_2 n)_i,j != 0 for all i,j (7)
3845 * Showing there's indeed a path between every cpu in at most O(log n) steps.
3846 * The task movement gives a factor of O(m), giving a convergence complexity
3849 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
3854 * In order to avoid CPUs going idle while there's still work to do, new idle
3855 * balancing is more aggressive and has the newly idle cpu iterate up the domain
3856 * tree itself instead of relying on other CPUs to bring it work.
3858 * This adds some complexity to both (5) and (8) but it reduces the total idle
3866 * Cgroups make a horror show out of (2), instead of a simple sum we get:
3869 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
3874 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
3876 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
3878 * The big problem is S_k, its a global sum needed to compute a local (W_i)
3881 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
3882 * rewrite all of this once again.]
3885 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
3887 #define LBF_ALL_PINNED 0x01
3888 #define LBF_NEED_BREAK 0x02
3889 #define LBF_DST_PINNED 0x04
3890 #define LBF_SOME_PINNED 0x08
3893 struct sched_domain *sd;
3901 struct cpumask *dst_grpmask;
3903 enum cpu_idle_type idle;
3905 /* The set of CPUs under consideration for load-balancing */
3906 struct cpumask *cpus;
3911 unsigned int loop_break;
3912 unsigned int loop_max;
3916 * move_task - move a task from one runqueue to another runqueue.
3917 * Both runqueues must be locked.
3919 static void move_task(struct task_struct *p, struct lb_env *env)
3921 deactivate_task(env->src_rq, p, 0);
3922 set_task_cpu(p, env->dst_cpu);
3923 activate_task(env->dst_rq, p, 0);
3924 check_preempt_curr(env->dst_rq, p, 0);
3928 * Is this task likely cache-hot:
3931 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
3935 if (p->sched_class != &fair_sched_class)
3938 if (unlikely(p->policy == SCHED_IDLE))
3942 * Buddy candidates are cache hot:
3944 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
3945 (&p->se == cfs_rq_of(&p->se)->next ||
3946 &p->se == cfs_rq_of(&p->se)->last))
3949 if (sysctl_sched_migration_cost == -1)
3951 if (sysctl_sched_migration_cost == 0)
3954 delta = now - p->se.exec_start;
3956 return delta < (s64)sysctl_sched_migration_cost;
3960 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3963 int can_migrate_task(struct task_struct *p, struct lb_env *env)
3965 int tsk_cache_hot = 0;
3967 * We do not migrate tasks that are:
3968 * 1) throttled_lb_pair, or
3969 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3970 * 3) running (obviously), or
3971 * 4) are cache-hot on their current CPU.
3973 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
3976 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
3979 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
3981 env->flags |= LBF_SOME_PINNED;
3984 * Remember if this task can be migrated to any other cpu in
3985 * our sched_group. We may want to revisit it if we couldn't
3986 * meet load balance goals by pulling other tasks on src_cpu.
3988 * Also avoid computing new_dst_cpu if we have already computed
3989 * one in current iteration.
3991 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
3994 /* Prevent to re-select dst_cpu via env's cpus */
3995 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
3996 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
3997 env->flags |= LBF_DST_PINNED;
3998 env->new_dst_cpu = cpu;
4006 /* Record that we found atleast one task that could run on dst_cpu */
4007 env->flags &= ~LBF_ALL_PINNED;
4009 if (task_running(env->src_rq, p)) {
4010 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4015 * Aggressive migration if:
4016 * 1) task is cache cold, or
4017 * 2) too many balance attempts have failed.
4020 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4021 if (!tsk_cache_hot ||
4022 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4024 if (tsk_cache_hot) {
4025 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4026 schedstat_inc(p, se.statistics.nr_forced_migrations);
4032 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4037 * move_one_task tries to move exactly one task from busiest to this_rq, as
4038 * part of active balancing operations within "domain".
4039 * Returns 1 if successful and 0 otherwise.
4041 * Called with both runqueues locked.
4043 static int move_one_task(struct lb_env *env)
4045 struct task_struct *p, *n;
4047 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4048 if (!can_migrate_task(p, env))
4053 * Right now, this is only the second place move_task()
4054 * is called, so we can safely collect move_task()
4055 * stats here rather than inside move_task().
4057 schedstat_inc(env->sd, lb_gained[env->idle]);
4063 static unsigned long task_h_load(struct task_struct *p);
4065 static const unsigned int sched_nr_migrate_break = 32;
4068 * move_tasks tries to move up to imbalance weighted load from busiest to
4069 * this_rq, as part of a balancing operation within domain "sd".
4070 * Returns 1 if successful and 0 otherwise.
4072 * Called with both runqueues locked.
4074 static int move_tasks(struct lb_env *env)
4076 struct list_head *tasks = &env->src_rq->cfs_tasks;
4077 struct task_struct *p;
4081 if (env->imbalance <= 0)
4084 while (!list_empty(tasks)) {
4085 p = list_first_entry(tasks, struct task_struct, se.group_node);
4088 /* We've more or less seen every task there is, call it quits */
4089 if (env->loop > env->loop_max)
4092 /* take a breather every nr_migrate tasks */
4093 if (env->loop > env->loop_break) {
4094 env->loop_break += sched_nr_migrate_break;
4095 env->flags |= LBF_NEED_BREAK;
4099 if (!can_migrate_task(p, env))
4102 load = task_h_load(p);
4104 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4107 if ((load / 2) > env->imbalance)
4112 env->imbalance -= load;
4114 #ifdef CONFIG_PREEMPT
4116 * NEWIDLE balancing is a source of latency, so preemptible
4117 * kernels will stop after the first task is pulled to minimize
4118 * the critical section.
4120 if (env->idle == CPU_NEWLY_IDLE)
4125 * We only want to steal up to the prescribed amount of
4128 if (env->imbalance <= 0)
4133 list_move_tail(&p->se.group_node, tasks);
4137 * Right now, this is one of only two places move_task() is called,
4138 * so we can safely collect move_task() stats here rather than
4139 * inside move_task().
4141 schedstat_add(env->sd, lb_gained[env->idle], pulled);
4146 #ifdef CONFIG_FAIR_GROUP_SCHED
4148 * update tg->load_weight by folding this cpu's load_avg
4150 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
4152 struct sched_entity *se = tg->se[cpu];
4153 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
4155 /* throttled entities do not contribute to load */
4156 if (throttled_hierarchy(cfs_rq))
4159 update_cfs_rq_blocked_load(cfs_rq, 1);
4162 update_entity_load_avg(se, 1);
4164 * We pivot on our runnable average having decayed to zero for
4165 * list removal. This generally implies that all our children
4166 * have also been removed (modulo rounding error or bandwidth
4167 * control); however, such cases are rare and we can fix these
4170 * TODO: fix up out-of-order children on enqueue.
4172 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
4173 list_del_leaf_cfs_rq(cfs_rq);
4175 struct rq *rq = rq_of(cfs_rq);
4176 update_rq_runnable_avg(rq, rq->nr_running);
4180 static void update_blocked_averages(int cpu)
4182 struct rq *rq = cpu_rq(cpu);
4183 struct cfs_rq *cfs_rq;
4184 unsigned long flags;
4186 raw_spin_lock_irqsave(&rq->lock, flags);
4187 update_rq_clock(rq);
4189 * Iterates the task_group tree in a bottom up fashion, see
4190 * list_add_leaf_cfs_rq() for details.
4192 for_each_leaf_cfs_rq(rq, cfs_rq) {
4194 * Note: We may want to consider periodically releasing
4195 * rq->lock about these updates so that creating many task
4196 * groups does not result in continually extending hold time.
4198 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
4201 raw_spin_unlock_irqrestore(&rq->lock, flags);
4205 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
4206 * This needs to be done in a top-down fashion because the load of a child
4207 * group is a fraction of its parents load.
4209 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
4211 struct rq *rq = rq_of(cfs_rq);
4212 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
4213 unsigned long now = jiffies;
4216 if (cfs_rq->last_h_load_update == now)
4219 cfs_rq->h_load_next = NULL;
4220 for_each_sched_entity(se) {
4221 cfs_rq = cfs_rq_of(se);
4222 cfs_rq->h_load_next = se;
4223 if (cfs_rq->last_h_load_update == now)
4228 cfs_rq->h_load = cfs_rq->runnable_load_avg;
4229 cfs_rq->last_h_load_update = now;
4232 while ((se = cfs_rq->h_load_next) != NULL) {
4233 load = cfs_rq->h_load;
4234 load = div64_ul(load * se->avg.load_avg_contrib,
4235 cfs_rq->runnable_load_avg + 1);
4236 cfs_rq = group_cfs_rq(se);
4237 cfs_rq->h_load = load;
4238 cfs_rq->last_h_load_update = now;
4242 static unsigned long task_h_load(struct task_struct *p)
4244 struct cfs_rq *cfs_rq = task_cfs_rq(p);
4246 update_cfs_rq_h_load(cfs_rq);
4247 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
4248 cfs_rq->runnable_load_avg + 1);
4251 static inline void update_blocked_averages(int cpu)
4255 static unsigned long task_h_load(struct task_struct *p)
4257 return p->se.avg.load_avg_contrib;
4261 /********** Helpers for find_busiest_group ************************/
4263 * sg_lb_stats - stats of a sched_group required for load_balancing
4265 struct sg_lb_stats {
4266 unsigned long avg_load; /*Avg load across the CPUs of the group */
4267 unsigned long group_load; /* Total load over the CPUs of the group */
4268 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
4269 unsigned long load_per_task;
4270 unsigned long group_power;
4271 unsigned int sum_nr_running; /* Nr tasks running in the group */
4272 unsigned int group_capacity;
4273 unsigned int idle_cpus;
4274 unsigned int group_weight;
4275 int group_imb; /* Is there an imbalance in the group ? */
4276 int group_has_capacity; /* Is there extra capacity in the group? */
4280 * sd_lb_stats - Structure to store the statistics of a sched_domain
4281 * during load balancing.
4283 struct sd_lb_stats {
4284 struct sched_group *busiest; /* Busiest group in this sd */
4285 struct sched_group *local; /* Local group in this sd */
4286 unsigned long total_load; /* Total load of all groups in sd */
4287 unsigned long total_pwr; /* Total power of all groups in sd */
4288 unsigned long avg_load; /* Average load across all groups in sd */
4290 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
4291 struct sg_lb_stats local_stat; /* Statistics of the local group */
4294 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
4297 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
4298 * local_stat because update_sg_lb_stats() does a full clear/assignment.
4299 * We must however clear busiest_stat::avg_load because
4300 * update_sd_pick_busiest() reads this before assignment.
4302 *sds = (struct sd_lb_stats){
4314 * get_sd_load_idx - Obtain the load index for a given sched domain.
4315 * @sd: The sched_domain whose load_idx is to be obtained.
4316 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4318 * Return: The load index.
4320 static inline int get_sd_load_idx(struct sched_domain *sd,
4321 enum cpu_idle_type idle)
4327 load_idx = sd->busy_idx;
4330 case CPU_NEWLY_IDLE:
4331 load_idx = sd->newidle_idx;
4334 load_idx = sd->idle_idx;
4341 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
4343 return SCHED_POWER_SCALE;
4346 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
4348 return default_scale_freq_power(sd, cpu);
4351 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
4353 unsigned long weight = sd->span_weight;
4354 unsigned long smt_gain = sd->smt_gain;
4361 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
4363 return default_scale_smt_power(sd, cpu);
4366 static unsigned long scale_rt_power(int cpu)
4368 struct rq *rq = cpu_rq(cpu);
4369 u64 total, available, age_stamp, avg;
4372 * Since we're reading these variables without serialization make sure
4373 * we read them once before doing sanity checks on them.
4375 age_stamp = ACCESS_ONCE(rq->age_stamp);
4376 avg = ACCESS_ONCE(rq->rt_avg);
4378 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
4380 if (unlikely(total < avg)) {
4381 /* Ensures that power won't end up being negative */
4384 available = total - avg;
4387 if (unlikely((s64)total < SCHED_POWER_SCALE))
4388 total = SCHED_POWER_SCALE;
4390 total >>= SCHED_POWER_SHIFT;
4392 return div_u64(available, total);
4395 static void update_cpu_power(struct sched_domain *sd, int cpu)
4397 unsigned long weight = sd->span_weight;
4398 unsigned long power = SCHED_POWER_SCALE;
4399 struct sched_group *sdg = sd->groups;
4401 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
4402 if (sched_feat(ARCH_POWER))
4403 power *= arch_scale_smt_power(sd, cpu);
4405 power *= default_scale_smt_power(sd, cpu);
4407 power >>= SCHED_POWER_SHIFT;
4410 sdg->sgp->power_orig = power;
4412 if (sched_feat(ARCH_POWER))
4413 power *= arch_scale_freq_power(sd, cpu);
4415 power *= default_scale_freq_power(sd, cpu);
4417 power >>= SCHED_POWER_SHIFT;
4419 power *= scale_rt_power(cpu);
4420 power >>= SCHED_POWER_SHIFT;
4425 cpu_rq(cpu)->cpu_power = power;
4426 sdg->sgp->power = power;
4429 void update_group_power(struct sched_domain *sd, int cpu)
4431 struct sched_domain *child = sd->child;
4432 struct sched_group *group, *sdg = sd->groups;
4433 unsigned long power, power_orig;
4434 unsigned long interval;
4436 interval = msecs_to_jiffies(sd->balance_interval);
4437 interval = clamp(interval, 1UL, max_load_balance_interval);
4438 sdg->sgp->next_update = jiffies + interval;
4441 update_cpu_power(sd, cpu);
4445 power_orig = power = 0;
4447 if (child->flags & SD_OVERLAP) {
4449 * SD_OVERLAP domains cannot assume that child groups
4450 * span the current group.
4453 for_each_cpu(cpu, sched_group_cpus(sdg)) {
4454 struct sched_group *sg = cpu_rq(cpu)->sd->groups;
4456 power_orig += sg->sgp->power_orig;
4457 power += sg->sgp->power;
4461 * !SD_OVERLAP domains can assume that child groups
4462 * span the current group.
4465 group = child->groups;
4467 power_orig += group->sgp->power_orig;
4468 power += group->sgp->power;
4469 group = group->next;
4470 } while (group != child->groups);
4473 sdg->sgp->power_orig = power_orig;
4474 sdg->sgp->power = power;
4478 * Try and fix up capacity for tiny siblings, this is needed when
4479 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4480 * which on its own isn't powerful enough.
4482 * See update_sd_pick_busiest() and check_asym_packing().
4485 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
4488 * Only siblings can have significantly less than SCHED_POWER_SCALE
4490 if (!(sd->flags & SD_SHARE_CPUPOWER))
4494 * If ~90% of the cpu_power is still there, we're good.
4496 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
4503 * Group imbalance indicates (and tries to solve) the problem where balancing
4504 * groups is inadequate due to tsk_cpus_allowed() constraints.
4506 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
4507 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
4510 * { 0 1 2 3 } { 4 5 6 7 }
4513 * If we were to balance group-wise we'd place two tasks in the first group and
4514 * two tasks in the second group. Clearly this is undesired as it will overload
4515 * cpu 3 and leave one of the cpus in the second group unused.
4517 * The current solution to this issue is detecting the skew in the first group
4518 * by noticing the lower domain failed to reach balance and had difficulty
4519 * moving tasks due to affinity constraints.
4521 * When this is so detected; this group becomes a candidate for busiest; see
4522 * update_sd_pick_busiest(). And calculcate_imbalance() and
4523 * find_busiest_group() avoid some of the usual balance conditions to allow it
4524 * to create an effective group imbalance.
4526 * This is a somewhat tricky proposition since the next run might not find the
4527 * group imbalance and decide the groups need to be balanced again. A most
4528 * subtle and fragile situation.
4531 static inline int sg_imbalanced(struct sched_group *group)
4533 return group->sgp->imbalance;
4537 * Compute the group capacity.
4539 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
4540 * first dividing out the smt factor and computing the actual number of cores
4541 * and limit power unit capacity with that.
4543 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
4545 unsigned int capacity, smt, cpus;
4546 unsigned int power, power_orig;
4548 power = group->sgp->power;
4549 power_orig = group->sgp->power_orig;
4550 cpus = group->group_weight;
4552 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
4553 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
4554 capacity = cpus / smt; /* cores */
4556 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
4558 capacity = fix_small_capacity(env->sd, group);
4564 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4565 * @env: The load balancing environment.
4566 * @group: sched_group whose statistics are to be updated.
4567 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4568 * @local_group: Does group contain this_cpu.
4569 * @sgs: variable to hold the statistics for this group.
4571 static inline void update_sg_lb_stats(struct lb_env *env,
4572 struct sched_group *group, int load_idx,
4573 int local_group, struct sg_lb_stats *sgs)
4575 unsigned long nr_running;
4579 memset(sgs, 0, sizeof(*sgs));
4581 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
4582 struct rq *rq = cpu_rq(i);
4584 nr_running = rq->nr_running;
4586 /* Bias balancing toward cpus of our domain */
4588 load = target_load(i, load_idx);
4590 load = source_load(i, load_idx);
4592 sgs->group_load += load;
4593 sgs->sum_nr_running += nr_running;
4594 sgs->sum_weighted_load += weighted_cpuload(i);
4599 /* Adjust by relative CPU power of the group */
4600 sgs->group_power = group->sgp->power;
4601 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
4603 if (sgs->sum_nr_running)
4604 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
4606 sgs->group_weight = group->group_weight;
4608 sgs->group_imb = sg_imbalanced(group);
4609 sgs->group_capacity = sg_capacity(env, group);
4611 if (sgs->group_capacity > sgs->sum_nr_running)
4612 sgs->group_has_capacity = 1;
4616 * update_sd_pick_busiest - return 1 on busiest group
4617 * @env: The load balancing environment.
4618 * @sds: sched_domain statistics
4619 * @sg: sched_group candidate to be checked for being the busiest
4620 * @sgs: sched_group statistics
4622 * Determine if @sg is a busier group than the previously selected
4625 * Return: %true if @sg is a busier group than the previously selected
4626 * busiest group. %false otherwise.
4628 static bool update_sd_pick_busiest(struct lb_env *env,
4629 struct sd_lb_stats *sds,
4630 struct sched_group *sg,
4631 struct sg_lb_stats *sgs)
4633 if (sgs->avg_load <= sds->busiest_stat.avg_load)
4636 if (sgs->sum_nr_running > sgs->group_capacity)
4643 * ASYM_PACKING needs to move all the work to the lowest
4644 * numbered CPUs in the group, therefore mark all groups
4645 * higher than ourself as busy.
4647 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
4648 env->dst_cpu < group_first_cpu(sg)) {
4652 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
4660 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
4661 * @env: The load balancing environment.
4662 * @balance: Should we balance.
4663 * @sds: variable to hold the statistics for this sched_domain.
4665 static inline void update_sd_lb_stats(struct lb_env *env,
4666 struct sd_lb_stats *sds)
4668 struct sched_domain *child = env->sd->child;
4669 struct sched_group *sg = env->sd->groups;
4670 struct sg_lb_stats tmp_sgs;
4671 int load_idx, prefer_sibling = 0;
4673 if (child && child->flags & SD_PREFER_SIBLING)
4676 load_idx = get_sd_load_idx(env->sd, env->idle);
4679 struct sg_lb_stats *sgs = &tmp_sgs;
4682 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
4685 sgs = &sds->local_stat;
4687 if (env->idle != CPU_NEWLY_IDLE ||
4688 time_after_eq(jiffies, sg->sgp->next_update))
4689 update_group_power(env->sd, env->dst_cpu);
4692 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
4698 * In case the child domain prefers tasks go to siblings
4699 * first, lower the sg capacity to one so that we'll try
4700 * and move all the excess tasks away. We lower the capacity
4701 * of a group only if the local group has the capacity to fit
4702 * these excess tasks, i.e. nr_running < group_capacity. The
4703 * extra check prevents the case where you always pull from the
4704 * heaviest group when it is already under-utilized (possible
4705 * with a large weight task outweighs the tasks on the system).
4707 if (prefer_sibling && sds->local &&
4708 sds->local_stat.group_has_capacity)
4709 sgs->group_capacity = min(sgs->group_capacity, 1U);
4711 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
4713 sds->busiest_stat = *sgs;
4717 /* Now, start updating sd_lb_stats */
4718 sds->total_load += sgs->group_load;
4719 sds->total_pwr += sgs->group_power;
4722 } while (sg != env->sd->groups);
4726 * check_asym_packing - Check to see if the group is packed into the
4729 * This is primarily intended to used at the sibling level. Some
4730 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4731 * case of POWER7, it can move to lower SMT modes only when higher
4732 * threads are idle. When in lower SMT modes, the threads will
4733 * perform better since they share less core resources. Hence when we
4734 * have idle threads, we want them to be the higher ones.
4736 * This packing function is run on idle threads. It checks to see if
4737 * the busiest CPU in this domain (core in the P7 case) has a higher
4738 * CPU number than the packing function is being run on. Here we are
4739 * assuming lower CPU number will be equivalent to lower a SMT thread
4742 * Return: 1 when packing is required and a task should be moved to
4743 * this CPU. The amount of the imbalance is returned in *imbalance.
4745 * @env: The load balancing environment.
4746 * @sds: Statistics of the sched_domain which is to be packed
4748 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
4752 if (!(env->sd->flags & SD_ASYM_PACKING))
4758 busiest_cpu = group_first_cpu(sds->busiest);
4759 if (env->dst_cpu > busiest_cpu)
4762 env->imbalance = DIV_ROUND_CLOSEST(
4763 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
4770 * fix_small_imbalance - Calculate the minor imbalance that exists
4771 * amongst the groups of a sched_domain, during
4773 * @env: The load balancing environment.
4774 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4777 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4779 unsigned long tmp, pwr_now = 0, pwr_move = 0;
4780 unsigned int imbn = 2;
4781 unsigned long scaled_busy_load_per_task;
4782 struct sg_lb_stats *local, *busiest;
4784 local = &sds->local_stat;
4785 busiest = &sds->busiest_stat;
4787 if (!local->sum_nr_running)
4788 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
4789 else if (busiest->load_per_task > local->load_per_task)
4792 scaled_busy_load_per_task =
4793 (busiest->load_per_task * SCHED_POWER_SCALE) /
4794 busiest->group_power;
4796 if (busiest->avg_load + scaled_busy_load_per_task >=
4797 local->avg_load + (scaled_busy_load_per_task * imbn)) {
4798 env->imbalance = busiest->load_per_task;
4803 * OK, we don't have enough imbalance to justify moving tasks,
4804 * however we may be able to increase total CPU power used by
4808 pwr_now += busiest->group_power *
4809 min(busiest->load_per_task, busiest->avg_load);
4810 pwr_now += local->group_power *
4811 min(local->load_per_task, local->avg_load);
4812 pwr_now /= SCHED_POWER_SCALE;
4814 /* Amount of load we'd subtract */
4815 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
4816 busiest->group_power;
4817 if (busiest->avg_load > tmp) {
4818 pwr_move += busiest->group_power *
4819 min(busiest->load_per_task,
4820 busiest->avg_load - tmp);
4823 /* Amount of load we'd add */
4824 if (busiest->avg_load * busiest->group_power <
4825 busiest->load_per_task * SCHED_POWER_SCALE) {
4826 tmp = (busiest->avg_load * busiest->group_power) /
4829 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
4832 pwr_move += local->group_power *
4833 min(local->load_per_task, local->avg_load + tmp);
4834 pwr_move /= SCHED_POWER_SCALE;
4836 /* Move if we gain throughput */
4837 if (pwr_move > pwr_now)
4838 env->imbalance = busiest->load_per_task;
4842 * calculate_imbalance - Calculate the amount of imbalance present within the
4843 * groups of a given sched_domain during load balance.
4844 * @env: load balance environment
4845 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4847 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4849 unsigned long max_pull, load_above_capacity = ~0UL;
4850 struct sg_lb_stats *local, *busiest;
4852 local = &sds->local_stat;
4853 busiest = &sds->busiest_stat;
4855 if (busiest->group_imb) {
4857 * In the group_imb case we cannot rely on group-wide averages
4858 * to ensure cpu-load equilibrium, look at wider averages. XXX
4860 busiest->load_per_task =
4861 min(busiest->load_per_task, sds->avg_load);
4865 * In the presence of smp nice balancing, certain scenarios can have
4866 * max load less than avg load(as we skip the groups at or below
4867 * its cpu_power, while calculating max_load..)
4869 if (busiest->avg_load <= sds->avg_load ||
4870 local->avg_load >= sds->avg_load) {
4872 return fix_small_imbalance(env, sds);
4875 if (!busiest->group_imb) {
4877 * Don't want to pull so many tasks that a group would go idle.
4878 * Except of course for the group_imb case, since then we might
4879 * have to drop below capacity to reach cpu-load equilibrium.
4881 load_above_capacity =
4882 (busiest->sum_nr_running - busiest->group_capacity);
4884 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
4885 load_above_capacity /= busiest->group_power;
4889 * We're trying to get all the cpus to the average_load, so we don't
4890 * want to push ourselves above the average load, nor do we wish to
4891 * reduce the max loaded cpu below the average load. At the same time,
4892 * we also don't want to reduce the group load below the group capacity
4893 * (so that we can implement power-savings policies etc). Thus we look
4894 * for the minimum possible imbalance.
4896 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
4898 /* How much load to actually move to equalise the imbalance */
4899 env->imbalance = min(
4900 max_pull * busiest->group_power,
4901 (sds->avg_load - local->avg_load) * local->group_power
4902 ) / SCHED_POWER_SCALE;
4905 * if *imbalance is less than the average load per runnable task
4906 * there is no guarantee that any tasks will be moved so we'll have
4907 * a think about bumping its value to force at least one task to be
4910 if (env->imbalance < busiest->load_per_task)
4911 return fix_small_imbalance(env, sds);
4914 /******* find_busiest_group() helpers end here *********************/
4917 * find_busiest_group - Returns the busiest group within the sched_domain
4918 * if there is an imbalance. If there isn't an imbalance, and
4919 * the user has opted for power-savings, it returns a group whose
4920 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4921 * such a group exists.
4923 * Also calculates the amount of weighted load which should be moved
4924 * to restore balance.
4926 * @env: The load balancing environment.
4928 * Return: - The busiest group if imbalance exists.
4929 * - If no imbalance and user has opted for power-savings balance,
4930 * return the least loaded group whose CPUs can be
4931 * put to idle by rebalancing its tasks onto our group.
4933 static struct sched_group *find_busiest_group(struct lb_env *env)
4935 struct sg_lb_stats *local, *busiest;
4936 struct sd_lb_stats sds;
4938 init_sd_lb_stats(&sds);
4941 * Compute the various statistics relavent for load balancing at
4944 update_sd_lb_stats(env, &sds);
4945 local = &sds.local_stat;
4946 busiest = &sds.busiest_stat;
4948 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
4949 check_asym_packing(env, &sds))
4952 /* There is no busy sibling group to pull tasks from */
4953 if (!sds.busiest || busiest->sum_nr_running == 0)
4956 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
4959 * If the busiest group is imbalanced the below checks don't
4960 * work because they assume all things are equal, which typically
4961 * isn't true due to cpus_allowed constraints and the like.
4963 if (busiest->group_imb)
4966 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
4967 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
4968 !busiest->group_has_capacity)
4972 * If the local group is more busy than the selected busiest group
4973 * don't try and pull any tasks.
4975 if (local->avg_load >= busiest->avg_load)
4979 * Don't pull any tasks if this group is already above the domain
4982 if (local->avg_load >= sds.avg_load)
4985 if (env->idle == CPU_IDLE) {
4987 * This cpu is idle. If the busiest group load doesn't
4988 * have more tasks than the number of available cpu's and
4989 * there is no imbalance between this and busiest group
4990 * wrt to idle cpu's, it is balanced.
4992 if ((local->idle_cpus < busiest->idle_cpus) &&
4993 busiest->sum_nr_running <= busiest->group_weight)
4997 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
4998 * imbalance_pct to be conservative.
5000 if (100 * busiest->avg_load <=
5001 env->sd->imbalance_pct * local->avg_load)
5006 /* Looks like there is an imbalance. Compute it */
5007 calculate_imbalance(env, &sds);
5016 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5018 static struct rq *find_busiest_queue(struct lb_env *env,
5019 struct sched_group *group)
5021 struct rq *busiest = NULL, *rq;
5022 unsigned long busiest_load = 0, busiest_power = 1;
5025 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5026 unsigned long power = power_of(i);
5027 unsigned long capacity = DIV_ROUND_CLOSEST(power,
5032 capacity = fix_small_capacity(env->sd, group);
5035 wl = weighted_cpuload(i);
5038 * When comparing with imbalance, use weighted_cpuload()
5039 * which is not scaled with the cpu power.
5041 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5045 * For the load comparisons with the other cpu's, consider
5046 * the weighted_cpuload() scaled with the cpu power, so that
5047 * the load can be moved away from the cpu that is potentially
5048 * running at a lower capacity.
5050 * Thus we're looking for max(wl_i / power_i), crosswise
5051 * multiplication to rid ourselves of the division works out
5052 * to: wl_i * power_j > wl_j * power_i; where j is our
5055 if (wl * busiest_power > busiest_load * power) {
5057 busiest_power = power;
5066 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5067 * so long as it is large enough.
5069 #define MAX_PINNED_INTERVAL 512
5071 /* Working cpumask for load_balance and load_balance_newidle. */
5072 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5074 static int need_active_balance(struct lb_env *env)
5076 struct sched_domain *sd = env->sd;
5078 if (env->idle == CPU_NEWLY_IDLE) {
5081 * ASYM_PACKING needs to force migrate tasks from busy but
5082 * higher numbered CPUs in order to pack all tasks in the
5083 * lowest numbered CPUs.
5085 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
5089 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
5092 static int active_load_balance_cpu_stop(void *data);
5094 static int should_we_balance(struct lb_env *env)
5096 struct sched_group *sg = env->sd->groups;
5097 struct cpumask *sg_cpus, *sg_mask;
5098 int cpu, balance_cpu = -1;
5101 * In the newly idle case, we will allow all the cpu's
5102 * to do the newly idle load balance.
5104 if (env->idle == CPU_NEWLY_IDLE)
5107 sg_cpus = sched_group_cpus(sg);
5108 sg_mask = sched_group_mask(sg);
5109 /* Try to find first idle cpu */
5110 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
5111 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
5118 if (balance_cpu == -1)
5119 balance_cpu = group_balance_cpu(sg);
5122 * First idle cpu or the first cpu(busiest) in this sched group
5123 * is eligible for doing load balancing at this and above domains.
5125 return balance_cpu == env->dst_cpu;
5129 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5130 * tasks if there is an imbalance.
5132 static int load_balance(int this_cpu, struct rq *this_rq,
5133 struct sched_domain *sd, enum cpu_idle_type idle,
5134 int *continue_balancing)
5136 int ld_moved, cur_ld_moved, active_balance = 0;
5137 struct sched_domain *sd_parent = sd->parent;
5138 struct sched_group *group;
5140 unsigned long flags;
5141 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
5143 struct lb_env env = {
5145 .dst_cpu = this_cpu,
5147 .dst_grpmask = sched_group_cpus(sd->groups),
5149 .loop_break = sched_nr_migrate_break,
5154 * For NEWLY_IDLE load_balancing, we don't need to consider
5155 * other cpus in our group
5157 if (idle == CPU_NEWLY_IDLE)
5158 env.dst_grpmask = NULL;
5160 cpumask_copy(cpus, cpu_active_mask);
5162 schedstat_inc(sd, lb_count[idle]);
5165 if (!should_we_balance(&env)) {
5166 *continue_balancing = 0;
5170 group = find_busiest_group(&env);
5172 schedstat_inc(sd, lb_nobusyg[idle]);
5176 busiest = find_busiest_queue(&env, group);
5178 schedstat_inc(sd, lb_nobusyq[idle]);
5182 BUG_ON(busiest == env.dst_rq);
5184 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
5187 if (busiest->nr_running > 1) {
5189 * Attempt to move tasks. If find_busiest_group has found
5190 * an imbalance but busiest->nr_running <= 1, the group is
5191 * still unbalanced. ld_moved simply stays zero, so it is
5192 * correctly treated as an imbalance.
5194 env.flags |= LBF_ALL_PINNED;
5195 env.src_cpu = busiest->cpu;
5196 env.src_rq = busiest;
5197 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
5200 local_irq_save(flags);
5201 double_rq_lock(env.dst_rq, busiest);
5204 * cur_ld_moved - load moved in current iteration
5205 * ld_moved - cumulative load moved across iterations
5207 cur_ld_moved = move_tasks(&env);
5208 ld_moved += cur_ld_moved;
5209 double_rq_unlock(env.dst_rq, busiest);
5210 local_irq_restore(flags);
5213 * some other cpu did the load balance for us.
5215 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
5216 resched_cpu(env.dst_cpu);
5218 if (env.flags & LBF_NEED_BREAK) {
5219 env.flags &= ~LBF_NEED_BREAK;
5224 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5225 * us and move them to an alternate dst_cpu in our sched_group
5226 * where they can run. The upper limit on how many times we
5227 * iterate on same src_cpu is dependent on number of cpus in our
5230 * This changes load balance semantics a bit on who can move
5231 * load to a given_cpu. In addition to the given_cpu itself
5232 * (or a ilb_cpu acting on its behalf where given_cpu is
5233 * nohz-idle), we now have balance_cpu in a position to move
5234 * load to given_cpu. In rare situations, this may cause
5235 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5236 * _independently_ and at _same_ time to move some load to
5237 * given_cpu) causing exceess load to be moved to given_cpu.
5238 * This however should not happen so much in practice and
5239 * moreover subsequent load balance cycles should correct the
5240 * excess load moved.
5242 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
5244 /* Prevent to re-select dst_cpu via env's cpus */
5245 cpumask_clear_cpu(env.dst_cpu, env.cpus);
5247 env.dst_rq = cpu_rq(env.new_dst_cpu);
5248 env.dst_cpu = env.new_dst_cpu;
5249 env.flags &= ~LBF_DST_PINNED;
5251 env.loop_break = sched_nr_migrate_break;
5254 * Go back to "more_balance" rather than "redo" since we
5255 * need to continue with same src_cpu.
5261 * We failed to reach balance because of affinity.
5264 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
5266 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
5267 *group_imbalance = 1;
5268 } else if (*group_imbalance)
5269 *group_imbalance = 0;
5272 /* All tasks on this runqueue were pinned by CPU affinity */
5273 if (unlikely(env.flags & LBF_ALL_PINNED)) {
5274 cpumask_clear_cpu(cpu_of(busiest), cpus);
5275 if (!cpumask_empty(cpus)) {
5277 env.loop_break = sched_nr_migrate_break;
5285 schedstat_inc(sd, lb_failed[idle]);
5287 * Increment the failure counter only on periodic balance.
5288 * We do not want newidle balance, which can be very
5289 * frequent, pollute the failure counter causing
5290 * excessive cache_hot migrations and active balances.
5292 if (idle != CPU_NEWLY_IDLE)
5293 sd->nr_balance_failed++;
5295 if (need_active_balance(&env)) {
5296 raw_spin_lock_irqsave(&busiest->lock, flags);
5298 /* don't kick the active_load_balance_cpu_stop,
5299 * if the curr task on busiest cpu can't be
5302 if (!cpumask_test_cpu(this_cpu,
5303 tsk_cpus_allowed(busiest->curr))) {
5304 raw_spin_unlock_irqrestore(&busiest->lock,
5306 env.flags |= LBF_ALL_PINNED;
5307 goto out_one_pinned;
5311 * ->active_balance synchronizes accesses to
5312 * ->active_balance_work. Once set, it's cleared
5313 * only after active load balance is finished.
5315 if (!busiest->active_balance) {
5316 busiest->active_balance = 1;
5317 busiest->push_cpu = this_cpu;
5320 raw_spin_unlock_irqrestore(&busiest->lock, flags);
5322 if (active_balance) {
5323 stop_one_cpu_nowait(cpu_of(busiest),
5324 active_load_balance_cpu_stop, busiest,
5325 &busiest->active_balance_work);
5329 * We've kicked active balancing, reset the failure
5332 sd->nr_balance_failed = sd->cache_nice_tries+1;
5335 sd->nr_balance_failed = 0;
5337 if (likely(!active_balance)) {
5338 /* We were unbalanced, so reset the balancing interval */
5339 sd->balance_interval = sd->min_interval;
5342 * If we've begun active balancing, start to back off. This
5343 * case may not be covered by the all_pinned logic if there
5344 * is only 1 task on the busy runqueue (because we don't call
5347 if (sd->balance_interval < sd->max_interval)
5348 sd->balance_interval *= 2;
5354 schedstat_inc(sd, lb_balanced[idle]);
5356 sd->nr_balance_failed = 0;
5359 /* tune up the balancing interval */
5360 if (((env.flags & LBF_ALL_PINNED) &&
5361 sd->balance_interval < MAX_PINNED_INTERVAL) ||
5362 (sd->balance_interval < sd->max_interval))
5363 sd->balance_interval *= 2;
5371 * idle_balance is called by schedule() if this_cpu is about to become
5372 * idle. Attempts to pull tasks from other CPUs.
5374 void idle_balance(int this_cpu, struct rq *this_rq)
5376 struct sched_domain *sd;
5377 int pulled_task = 0;
5378 unsigned long next_balance = jiffies + HZ;
5381 this_rq->idle_stamp = rq_clock(this_rq);
5383 if (this_rq->avg_idle < sysctl_sched_migration_cost)
5387 * Drop the rq->lock, but keep IRQ/preempt disabled.
5389 raw_spin_unlock(&this_rq->lock);
5391 update_blocked_averages(this_cpu);
5393 for_each_domain(this_cpu, sd) {
5394 unsigned long interval;
5395 int continue_balancing = 1;
5396 u64 t0, domain_cost;
5398 if (!(sd->flags & SD_LOAD_BALANCE))
5401 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
5404 if (sd->flags & SD_BALANCE_NEWIDLE) {
5405 t0 = sched_clock_cpu(this_cpu);
5407 /* If we've pulled tasks over stop searching: */
5408 pulled_task = load_balance(this_cpu, this_rq,
5410 &continue_balancing);
5412 domain_cost = sched_clock_cpu(this_cpu) - t0;
5413 if (domain_cost > sd->max_newidle_lb_cost)
5414 sd->max_newidle_lb_cost = domain_cost;
5416 curr_cost += domain_cost;
5419 interval = msecs_to_jiffies(sd->balance_interval);
5420 if (time_after(next_balance, sd->last_balance + interval))
5421 next_balance = sd->last_balance + interval;
5423 this_rq->idle_stamp = 0;
5429 raw_spin_lock(&this_rq->lock);
5431 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
5433 * We are going idle. next_balance may be set based on
5434 * a busy processor. So reset next_balance.
5436 this_rq->next_balance = next_balance;
5439 if (curr_cost > this_rq->max_idle_balance_cost)
5440 this_rq->max_idle_balance_cost = curr_cost;
5444 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5445 * running tasks off the busiest CPU onto idle CPUs. It requires at
5446 * least 1 task to be running on each physical CPU where possible, and
5447 * avoids physical / logical imbalances.
5449 static int active_load_balance_cpu_stop(void *data)
5451 struct rq *busiest_rq = data;
5452 int busiest_cpu = cpu_of(busiest_rq);
5453 int target_cpu = busiest_rq->push_cpu;
5454 struct rq *target_rq = cpu_rq(target_cpu);
5455 struct sched_domain *sd;
5457 raw_spin_lock_irq(&busiest_rq->lock);
5459 /* make sure the requested cpu hasn't gone down in the meantime */
5460 if (unlikely(busiest_cpu != smp_processor_id() ||
5461 !busiest_rq->active_balance))
5464 /* Is there any task to move? */
5465 if (busiest_rq->nr_running <= 1)
5469 * This condition is "impossible", if it occurs
5470 * we need to fix it. Originally reported by
5471 * Bjorn Helgaas on a 128-cpu setup.
5473 BUG_ON(busiest_rq == target_rq);
5475 /* move a task from busiest_rq to target_rq */
5476 double_lock_balance(busiest_rq, target_rq);
5478 /* Search for an sd spanning us and the target CPU. */
5480 for_each_domain(target_cpu, sd) {
5481 if ((sd->flags & SD_LOAD_BALANCE) &&
5482 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
5487 struct lb_env env = {
5489 .dst_cpu = target_cpu,
5490 .dst_rq = target_rq,
5491 .src_cpu = busiest_rq->cpu,
5492 .src_rq = busiest_rq,
5496 schedstat_inc(sd, alb_count);
5498 if (move_one_task(&env))
5499 schedstat_inc(sd, alb_pushed);
5501 schedstat_inc(sd, alb_failed);
5504 double_unlock_balance(busiest_rq, target_rq);
5506 busiest_rq->active_balance = 0;
5507 raw_spin_unlock_irq(&busiest_rq->lock);
5511 #ifdef CONFIG_NO_HZ_COMMON
5513 * idle load balancing details
5514 * - When one of the busy CPUs notice that there may be an idle rebalancing
5515 * needed, they will kick the idle load balancer, which then does idle
5516 * load balancing for all the idle CPUs.
5519 cpumask_var_t idle_cpus_mask;
5521 unsigned long next_balance; /* in jiffy units */
5522 } nohz ____cacheline_aligned;
5524 static inline int find_new_ilb(int call_cpu)
5526 int ilb = cpumask_first(nohz.idle_cpus_mask);
5528 if (ilb < nr_cpu_ids && idle_cpu(ilb))
5535 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5536 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5537 * CPU (if there is one).
5539 static void nohz_balancer_kick(int cpu)
5543 nohz.next_balance++;
5545 ilb_cpu = find_new_ilb(cpu);
5547 if (ilb_cpu >= nr_cpu_ids)
5550 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
5553 * Use smp_send_reschedule() instead of resched_cpu().
5554 * This way we generate a sched IPI on the target cpu which
5555 * is idle. And the softirq performing nohz idle load balance
5556 * will be run before returning from the IPI.
5558 smp_send_reschedule(ilb_cpu);
5562 static inline void nohz_balance_exit_idle(int cpu)
5564 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5565 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5566 atomic_dec(&nohz.nr_cpus);
5567 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5571 static inline void set_cpu_sd_state_busy(void)
5573 struct sched_domain *sd;
5576 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5578 if (!sd || !sd->nohz_idle)
5582 for (; sd; sd = sd->parent)
5583 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
5588 void set_cpu_sd_state_idle(void)
5590 struct sched_domain *sd;
5593 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5595 if (!sd || sd->nohz_idle)
5599 for (; sd; sd = sd->parent)
5600 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
5606 * This routine will record that the cpu is going idle with tick stopped.
5607 * This info will be used in performing idle load balancing in the future.
5609 void nohz_balance_enter_idle(int cpu)
5612 * If this cpu is going down, then nothing needs to be done.
5614 if (!cpu_active(cpu))
5617 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
5620 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
5621 atomic_inc(&nohz.nr_cpus);
5622 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5625 static int sched_ilb_notifier(struct notifier_block *nfb,
5626 unsigned long action, void *hcpu)
5628 switch (action & ~CPU_TASKS_FROZEN) {
5630 nohz_balance_exit_idle(smp_processor_id());
5638 static DEFINE_SPINLOCK(balancing);
5641 * Scale the max load_balance interval with the number of CPUs in the system.
5642 * This trades load-balance latency on larger machines for less cross talk.
5644 void update_max_interval(void)
5646 max_load_balance_interval = HZ*num_online_cpus()/10;
5650 * It checks each scheduling domain to see if it is due to be balanced,
5651 * and initiates a balancing operation if so.
5653 * Balancing parameters are set up in init_sched_domains.
5655 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
5657 int continue_balancing = 1;
5658 struct rq *rq = cpu_rq(cpu);
5659 unsigned long interval;
5660 struct sched_domain *sd;
5661 /* Earliest time when we have to do rebalance again */
5662 unsigned long next_balance = jiffies + 60*HZ;
5663 int update_next_balance = 0;
5664 int need_serialize, need_decay = 0;
5667 update_blocked_averages(cpu);
5670 for_each_domain(cpu, sd) {
5672 * Decay the newidle max times here because this is a regular
5673 * visit to all the domains. Decay ~1% per second.
5675 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
5676 sd->max_newidle_lb_cost =
5677 (sd->max_newidle_lb_cost * 253) / 256;
5678 sd->next_decay_max_lb_cost = jiffies + HZ;
5681 max_cost += sd->max_newidle_lb_cost;
5683 if (!(sd->flags & SD_LOAD_BALANCE))
5687 * Stop the load balance at this level. There is another
5688 * CPU in our sched group which is doing load balancing more
5691 if (!continue_balancing) {
5697 interval = sd->balance_interval;
5698 if (idle != CPU_IDLE)
5699 interval *= sd->busy_factor;
5701 /* scale ms to jiffies */
5702 interval = msecs_to_jiffies(interval);
5703 interval = clamp(interval, 1UL, max_load_balance_interval);
5705 need_serialize = sd->flags & SD_SERIALIZE;
5707 if (need_serialize) {
5708 if (!spin_trylock(&balancing))
5712 if (time_after_eq(jiffies, sd->last_balance + interval)) {
5713 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
5715 * The LBF_DST_PINNED logic could have changed
5716 * env->dst_cpu, so we can't know our idle
5717 * state even if we migrated tasks. Update it.
5719 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
5721 sd->last_balance = jiffies;
5724 spin_unlock(&balancing);
5726 if (time_after(next_balance, sd->last_balance + interval)) {
5727 next_balance = sd->last_balance + interval;
5728 update_next_balance = 1;
5733 * Ensure the rq-wide value also decays but keep it at a
5734 * reasonable floor to avoid funnies with rq->avg_idle.
5736 rq->max_idle_balance_cost =
5737 max((u64)sysctl_sched_migration_cost, max_cost);
5742 * next_balance will be updated only when there is a need.
5743 * When the cpu is attached to null domain for ex, it will not be
5746 if (likely(update_next_balance))
5747 rq->next_balance = next_balance;
5750 #ifdef CONFIG_NO_HZ_COMMON
5752 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
5753 * rebalancing for all the cpus for whom scheduler ticks are stopped.
5755 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
5757 struct rq *this_rq = cpu_rq(this_cpu);
5761 if (idle != CPU_IDLE ||
5762 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
5765 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
5766 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
5770 * If this cpu gets work to do, stop the load balancing
5771 * work being done for other cpus. Next load
5772 * balancing owner will pick it up.
5777 rq = cpu_rq(balance_cpu);
5779 raw_spin_lock_irq(&rq->lock);
5780 update_rq_clock(rq);
5781 update_idle_cpu_load(rq);
5782 raw_spin_unlock_irq(&rq->lock);
5784 rebalance_domains(balance_cpu, CPU_IDLE);
5786 if (time_after(this_rq->next_balance, rq->next_balance))
5787 this_rq->next_balance = rq->next_balance;
5789 nohz.next_balance = this_rq->next_balance;
5791 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
5795 * Current heuristic for kicking the idle load balancer in the presence
5796 * of an idle cpu is the system.
5797 * - This rq has more than one task.
5798 * - At any scheduler domain level, this cpu's scheduler group has multiple
5799 * busy cpu's exceeding the group's power.
5800 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
5801 * domain span are idle.
5803 static inline int nohz_kick_needed(struct rq *rq, int cpu)
5805 unsigned long now = jiffies;
5806 struct sched_domain *sd;
5808 if (unlikely(idle_cpu(cpu)))
5812 * We may be recently in ticked or tickless idle mode. At the first
5813 * busy tick after returning from idle, we will update the busy stats.
5815 set_cpu_sd_state_busy();
5816 nohz_balance_exit_idle(cpu);
5819 * None are in tickless mode and hence no need for NOHZ idle load
5822 if (likely(!atomic_read(&nohz.nr_cpus)))
5825 if (time_before(now, nohz.next_balance))
5828 if (rq->nr_running >= 2)
5832 for_each_domain(cpu, sd) {
5833 struct sched_group *sg = sd->groups;
5834 struct sched_group_power *sgp = sg->sgp;
5835 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
5837 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
5838 goto need_kick_unlock;
5840 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
5841 && (cpumask_first_and(nohz.idle_cpus_mask,
5842 sched_domain_span(sd)) < cpu))
5843 goto need_kick_unlock;
5845 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
5857 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
5861 * run_rebalance_domains is triggered when needed from the scheduler tick.
5862 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
5864 static void run_rebalance_domains(struct softirq_action *h)
5866 int this_cpu = smp_processor_id();
5867 struct rq *this_rq = cpu_rq(this_cpu);
5868 enum cpu_idle_type idle = this_rq->idle_balance ?
5869 CPU_IDLE : CPU_NOT_IDLE;
5871 rebalance_domains(this_cpu, idle);
5874 * If this cpu has a pending nohz_balance_kick, then do the
5875 * balancing on behalf of the other idle cpus whose ticks are
5878 nohz_idle_balance(this_cpu, idle);
5881 static inline int on_null_domain(int cpu)
5883 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
5887 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
5889 void trigger_load_balance(struct rq *rq, int cpu)
5891 /* Don't need to rebalance while attached to NULL domain */
5892 if (time_after_eq(jiffies, rq->next_balance) &&
5893 likely(!on_null_domain(cpu)))
5894 raise_softirq(SCHED_SOFTIRQ);
5895 #ifdef CONFIG_NO_HZ_COMMON
5896 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
5897 nohz_balancer_kick(cpu);
5901 static void rq_online_fair(struct rq *rq)
5906 static void rq_offline_fair(struct rq *rq)
5910 /* Ensure any throttled groups are reachable by pick_next_task */
5911 unthrottle_offline_cfs_rqs(rq);
5914 #endif /* CONFIG_SMP */
5917 * scheduler tick hitting a task of our scheduling class:
5919 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
5921 struct cfs_rq *cfs_rq;
5922 struct sched_entity *se = &curr->se;
5924 for_each_sched_entity(se) {
5925 cfs_rq = cfs_rq_of(se);
5926 entity_tick(cfs_rq, se, queued);
5929 if (numabalancing_enabled)
5930 task_tick_numa(rq, curr);
5932 update_rq_runnable_avg(rq, 1);
5936 * called on fork with the child task as argument from the parent's context
5937 * - child not yet on the tasklist
5938 * - preemption disabled
5940 static void task_fork_fair(struct task_struct *p)
5942 struct cfs_rq *cfs_rq;
5943 struct sched_entity *se = &p->se, *curr;
5944 int this_cpu = smp_processor_id();
5945 struct rq *rq = this_rq();
5946 unsigned long flags;
5948 raw_spin_lock_irqsave(&rq->lock, flags);
5950 update_rq_clock(rq);
5952 cfs_rq = task_cfs_rq(current);
5953 curr = cfs_rq->curr;
5956 * Not only the cpu but also the task_group of the parent might have
5957 * been changed after parent->se.parent,cfs_rq were copied to
5958 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
5959 * of child point to valid ones.
5962 __set_task_cpu(p, this_cpu);
5965 update_curr(cfs_rq);
5968 se->vruntime = curr->vruntime;
5969 place_entity(cfs_rq, se, 1);
5971 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
5973 * Upon rescheduling, sched_class::put_prev_task() will place
5974 * 'current' within the tree based on its new key value.
5976 swap(curr->vruntime, se->vruntime);
5977 resched_task(rq->curr);
5980 se->vruntime -= cfs_rq->min_vruntime;
5982 raw_spin_unlock_irqrestore(&rq->lock, flags);
5986 * Priority of the task has changed. Check to see if we preempt
5990 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
5996 * Reschedule if we are currently running on this runqueue and
5997 * our priority decreased, or if we are not currently running on
5998 * this runqueue and our priority is higher than the current's
6000 if (rq->curr == p) {
6001 if (p->prio > oldprio)
6002 resched_task(rq->curr);
6004 check_preempt_curr(rq, p, 0);
6007 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6009 struct sched_entity *se = &p->se;
6010 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6013 * Ensure the task's vruntime is normalized, so that when its
6014 * switched back to the fair class the enqueue_entity(.flags=0) will
6015 * do the right thing.
6017 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6018 * have normalized the vruntime, if it was !on_rq, then only when
6019 * the task is sleeping will it still have non-normalized vruntime.
6021 if (!se->on_rq && p->state != TASK_RUNNING) {
6023 * Fix up our vruntime so that the current sleep doesn't
6024 * cause 'unlimited' sleep bonus.
6026 place_entity(cfs_rq, se, 0);
6027 se->vruntime -= cfs_rq->min_vruntime;
6032 * Remove our load from contribution when we leave sched_fair
6033 * and ensure we don't carry in an old decay_count if we
6036 if (se->avg.decay_count) {
6037 __synchronize_entity_decay(se);
6038 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
6044 * We switched to the sched_fair class.
6046 static void switched_to_fair(struct rq *rq, struct task_struct *p)
6052 * We were most likely switched from sched_rt, so
6053 * kick off the schedule if running, otherwise just see
6054 * if we can still preempt the current task.
6057 resched_task(rq->curr);
6059 check_preempt_curr(rq, p, 0);
6062 /* Account for a task changing its policy or group.
6064 * This routine is mostly called to set cfs_rq->curr field when a task
6065 * migrates between groups/classes.
6067 static void set_curr_task_fair(struct rq *rq)
6069 struct sched_entity *se = &rq->curr->se;
6071 for_each_sched_entity(se) {
6072 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6074 set_next_entity(cfs_rq, se);
6075 /* ensure bandwidth has been allocated on our new cfs_rq */
6076 account_cfs_rq_runtime(cfs_rq, 0);
6080 void init_cfs_rq(struct cfs_rq *cfs_rq)
6082 cfs_rq->tasks_timeline = RB_ROOT;
6083 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6084 #ifndef CONFIG_64BIT
6085 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
6088 atomic64_set(&cfs_rq->decay_counter, 1);
6089 atomic_long_set(&cfs_rq->removed_load, 0);
6093 #ifdef CONFIG_FAIR_GROUP_SCHED
6094 static void task_move_group_fair(struct task_struct *p, int on_rq)
6096 struct cfs_rq *cfs_rq;
6098 * If the task was not on the rq at the time of this cgroup movement
6099 * it must have been asleep, sleeping tasks keep their ->vruntime
6100 * absolute on their old rq until wakeup (needed for the fair sleeper
6101 * bonus in place_entity()).
6103 * If it was on the rq, we've just 'preempted' it, which does convert
6104 * ->vruntime to a relative base.
6106 * Make sure both cases convert their relative position when migrating
6107 * to another cgroup's rq. This does somewhat interfere with the
6108 * fair sleeper stuff for the first placement, but who cares.
6111 * When !on_rq, vruntime of the task has usually NOT been normalized.
6112 * But there are some cases where it has already been normalized:
6114 * - Moving a forked child which is waiting for being woken up by
6115 * wake_up_new_task().
6116 * - Moving a task which has been woken up by try_to_wake_up() and
6117 * waiting for actually being woken up by sched_ttwu_pending().
6119 * To prevent boost or penalty in the new cfs_rq caused by delta
6120 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
6122 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
6126 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
6127 set_task_rq(p, task_cpu(p));
6129 cfs_rq = cfs_rq_of(&p->se);
6130 p->se.vruntime += cfs_rq->min_vruntime;
6133 * migrate_task_rq_fair() will have removed our previous
6134 * contribution, but we must synchronize for ongoing future
6137 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
6138 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
6143 void free_fair_sched_group(struct task_group *tg)
6147 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
6149 for_each_possible_cpu(i) {
6151 kfree(tg->cfs_rq[i]);
6160 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6162 struct cfs_rq *cfs_rq;
6163 struct sched_entity *se;
6166 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
6169 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
6173 tg->shares = NICE_0_LOAD;
6175 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
6177 for_each_possible_cpu(i) {
6178 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
6179 GFP_KERNEL, cpu_to_node(i));
6183 se = kzalloc_node(sizeof(struct sched_entity),
6184 GFP_KERNEL, cpu_to_node(i));
6188 init_cfs_rq(cfs_rq);
6189 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
6200 void unregister_fair_sched_group(struct task_group *tg, int cpu)
6202 struct rq *rq = cpu_rq(cpu);
6203 unsigned long flags;
6206 * Only empty task groups can be destroyed; so we can speculatively
6207 * check on_list without danger of it being re-added.
6209 if (!tg->cfs_rq[cpu]->on_list)
6212 raw_spin_lock_irqsave(&rq->lock, flags);
6213 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
6214 raw_spin_unlock_irqrestore(&rq->lock, flags);
6217 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
6218 struct sched_entity *se, int cpu,
6219 struct sched_entity *parent)
6221 struct rq *rq = cpu_rq(cpu);
6225 init_cfs_rq_runtime(cfs_rq);
6227 tg->cfs_rq[cpu] = cfs_rq;
6230 /* se could be NULL for root_task_group */
6235 se->cfs_rq = &rq->cfs;
6237 se->cfs_rq = parent->my_q;
6240 update_load_set(&se->load, 0);
6241 se->parent = parent;
6244 static DEFINE_MUTEX(shares_mutex);
6246 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6249 unsigned long flags;
6252 * We can't change the weight of the root cgroup.
6257 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
6259 mutex_lock(&shares_mutex);
6260 if (tg->shares == shares)
6263 tg->shares = shares;
6264 for_each_possible_cpu(i) {
6265 struct rq *rq = cpu_rq(i);
6266 struct sched_entity *se;
6269 /* Propagate contribution to hierarchy */
6270 raw_spin_lock_irqsave(&rq->lock, flags);
6272 /* Possible calls to update_curr() need rq clock */
6273 update_rq_clock(rq);
6274 for_each_sched_entity(se)
6275 update_cfs_shares(group_cfs_rq(se));
6276 raw_spin_unlock_irqrestore(&rq->lock, flags);
6280 mutex_unlock(&shares_mutex);
6283 #else /* CONFIG_FAIR_GROUP_SCHED */
6285 void free_fair_sched_group(struct task_group *tg) { }
6287 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6292 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
6294 #endif /* CONFIG_FAIR_GROUP_SCHED */
6297 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
6299 struct sched_entity *se = &task->se;
6300 unsigned int rr_interval = 0;
6303 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6306 if (rq->cfs.load.weight)
6307 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
6313 * All the scheduling class methods:
6315 const struct sched_class fair_sched_class = {
6316 .next = &idle_sched_class,
6317 .enqueue_task = enqueue_task_fair,
6318 .dequeue_task = dequeue_task_fair,
6319 .yield_task = yield_task_fair,
6320 .yield_to_task = yield_to_task_fair,
6322 .check_preempt_curr = check_preempt_wakeup,
6324 .pick_next_task = pick_next_task_fair,
6325 .put_prev_task = put_prev_task_fair,
6328 .select_task_rq = select_task_rq_fair,
6329 .migrate_task_rq = migrate_task_rq_fair,
6331 .rq_online = rq_online_fair,
6332 .rq_offline = rq_offline_fair,
6334 .task_waking = task_waking_fair,
6337 .set_curr_task = set_curr_task_fair,
6338 .task_tick = task_tick_fair,
6339 .task_fork = task_fork_fair,
6341 .prio_changed = prio_changed_fair,
6342 .switched_from = switched_from_fair,
6343 .switched_to = switched_to_fair,
6345 .get_rr_interval = get_rr_interval_fair,
6347 #ifdef CONFIG_FAIR_GROUP_SCHED
6348 .task_move_group = task_move_group_fair,
6352 #ifdef CONFIG_SCHED_DEBUG
6353 void print_cfs_stats(struct seq_file *m, int cpu)
6355 struct cfs_rq *cfs_rq;
6358 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
6359 print_cfs_rq(m, cpu, cfs_rq);
6364 __init void init_sched_fair_class(void)
6367 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
6369 #ifdef CONFIG_NO_HZ_COMMON
6370 nohz.next_balance = jiffies;
6371 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
6372 cpu_notifier(sched_ilb_notifier, 0);